Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
A lifetime perspective
Douglas A. Beatton
BBus (Finance and Management)
MBus (Research)
School of Economics and Finance
Queensland University of Technology
A
Dissertation
in fulfilment of the requirements
for the
Degree
of
Doctor of Philosophy
[i]
Chapter 7
Verse 16
Confucius said,
Give me a few more years to learn at the age of fifty and I will be
unlikely to have major faults
(Cheung, 2010)
[ii]
[iii]
Abstract
The three studies in this thesis focus on happiness and age and seek to contribute to
our understanding of happiness change over the lifetime. The first study contributes
by offering an explanation for what was evolving to a stylised fact in the economics
literature, the U-shape of happiness in age. No U-shape is evident if one makes a
visual inspection of the age happiness relationship in the German socio-economic
panel data, and, it seems counter-intuitive that we just have to wait until we get old to
be happy. Eliminating the very young, the very old, and the first timers from the
analysis did not explain away regression results supporting the U-shape of happiness
in age, but fixed effect analysis did. Analysis revealed found that reverse causality
arising from time-invariant individual traits explained the U-shape of happiness in
age in the German population, and the results were robust across six econometric
methods. Robustness was added to the German fixed effect finding by replicating it
with the Australian and the British socio-economic panel data sets.
During analysis of the German data an unexpected finding emerged, an exceedingly
large negative linear effect of age on happiness in fixed-effect regressions. There is a
large self-reported happiness decline by those who remain in the German panel. A
similar decline over time was not evident in the Australian or the British data. After
testing away age, time and cohort effects, a time-in-panel effect was found. Germans
who remain in the panel for longer progressively report lower levels of happiness.
Because time-in-panel effects have not been included in happiness regression
specifications, our estimates may be biased; perhaps some economics of the
happiness studies, that used German panel data, need revisiting.
The second study builds upon the fixed-effect finding of the first study and extends
our view of lifetime happiness to a cohort little visited by economists, children.
Initial analysis extends our view of lifetime happiness beyond adulthood and
revealed a happiness decline in adolescent (15 to 23 year-old) Australians that is
twice the size of the happiness decline we see in older Australians (75 to 86 yearolds), who we expect to be unhappy due to declining income, failing health and the
onset of death. To resolve a difference of opinion in the literature as to whether
[iv]
However, the
[v]
Table of Contents
Abstract ................................................................................................................................... iv
Table of Contents .................................................................................................................... vi
List of Tables ........................................................................................................................... x
Table of Figures .................................................................................................................... xiv
Glossary of Terms and Abbreviations ................................................................................ xviii
Statement of Original Authorship .......................................................................................... xx
Acknowledgements .............................................................................................................. xxii
Chapter One ............................................................................................................................. 1
Introduction .............................................................................................................................. 1
1.1
Chapter Two............................................................................................................................. 3
Preliminary review of the happiness literature......................................................................... 3
2.1
Happiness ..................................................................................................................... 3
2.2
2.3
2.4
2.5
Chapter Three......................................................................................................................... 15
The puzzle of the U-shaped relationship between happiness and age ................................... 15
3.1
Introduction ................................................................................................................ 15
3.2
3.3
3.4
3.5
[vi]
3.6
3.7
3.8
3.9
3.10
3.11
4.2
4.3
4.4
4.5
4.6
4.7
5.2
5.3
5.4
5.5
[vii]
[viii]
[ix]
List of Tables
Table 3.1a&b: Life Satisfaction regression results (t-values) from recent studies ................ 19
Table 3.2: Sample averages from the entire GSOEP, HILDA and BHPS samples ............... 25
Table 3.3: Summary of changes in the GSOEP Age and Age2 coefficients as controls are
progressively added ............................................................................................................... 31
Table 3.4: Comparison of changes in the GSOEP & HILDA Age and Age2 coefficients as
controls are progressively added ............................................................................................ 35
Table 3.5: Summary of changes in the GSOEP, HILDA & BHPS Age and Age2 coefficients
as controls are progressively added ....................................................................................... 38
Table 3.6: Summary of changes in the GSOEP, HILDA & BHPS Age and Age2 coefficients
as controls are progressively added; for ages 22 to 80 years ................................................. 43
Table 3.7: Summary of changes in the GSOEP Age and Age2 coefficients as controls are
progressively added; with fixed effects ................................................................................. 46
Table 3.8: Summary of changes in the GSOEP and HILDA Age and Age2 coefficients as
controls are progressively added; with fixed effects .............................................................. 49
Table 3.9: Summary of changes in the GSOEP, HILDA & BHPS Age and Age2 coefficients
as controls are progressively added; with fixed effects ......................................................... 52
Table 3.10: Coefficients for the key 5 variables (pooled & fixed effects) for the three data
sets. ........................................................................................................................................ 56
Table 3.11: Descriptive statistics for the entire and first-timer GSOEP samples .................. 81
Table 3.12: Descriptive statistics for the entire and first-timer HILDA samples .................. 82
Table 3.13: Descriptive statistics for the entire and first-timer BHPS samples ..................... 83
Table 3.14: The determinants of Life Satisfaction for West-Germans in the GSOEP; Pooled
OLS Regression entire sample, N = 176,770...................................................................... 86
Table 3.15: Determinants of Life Satisfaction for West-Germans in the GSOEP; Pooled OLS
Regressions ages 22 to 80, N = 160,332 ............................................................................. 87
Table 3.16: The determinants of Life Satisfaction for West-Germans in the GSOEP; Fixedeffect Regressions entire sample, N = 176,770 .................................................................. 88
[x]
Table 3.17: The determinants of Life Satisfaction for West-Germans in the GSOEP; Pooled
Regression first time respondents, N = 18,821 ................................................................... 89
Table 3.18: The determinants of Life Satisfaction; Pooled OLS regression results for all
individuals in the HILDA; N = 75,529 .................................................................................. 90
Table 3.19: Determinants of Life Satisfaction for Australians in the HILDA; Pooled
Regressions, ages 22 to 80, N = 65.679 ................................................................................. 91
Table 3.20: The determinants of Life Satisfaction for Australians in the HILDA; Fixed-effect
Regressions entire sample, N = 75,529............................................................................... 92
Table 3.21: The determinants of Life Satisfaction for Australians in the HILDA; Pooled
Regression first time respondents, N = 14,857 ................................................................... 93
Table 3.22: The determinants of Life Satisfaction for Britons in the BHPS; Pooled
Regression entire sample, N = 153,886 .............................................................................. 94
Table 3.23: The determinants of Life Satisfaction for Britons in the BHPS; Pooled
Regressions ages 22 to 80, N = 138,481 ............................................................................. 95
Table 3.24: The determinants of Life Satisfaction for Britons in the BHPS; Fixed-effect
Regressions entire sample, N = 153,886............................................................................. 96
Table 3.25: The determinants of Life Satisfaction for Britons in the BHPS; Pooled
Regression first time respondents, N = 22,922 ................................................................... 97
Table 3.26: Summary of changes in the GSOEP Age and Age2 coefficients as controls are
progressively added ............................................................................................................... 98
Table 3.27: Summary of changes in the HILDA Age and Age2 coefficients as controls are
progressively added ............................................................................................................... 99
Table 3.28: Summary of changes in BHPS Age and Age2 coefficients as controls are
progressively added ............................................................................................................. 100
Table 3.29: The determinants of Life Satisfaction for West-Germans in the GSOEP; Usual
suspects plus health specification, OLS, OLS with categorical health, Ordered Logit, BUC
estimator & OLS with fixed effect entire sample, N = 176,770 ....................................... 102
Table 3.30: The determinants of Life Satisfaction for West-Germans in the GSOEP, Agebands: (1) without control; (2) with controls per the Usual suspects plus health
specification; (3) plus fixed effects (3), N = 176,770 .......................................................... 104
[xi]
Table 3.31: The determinants of Life Satisfaction for British in the BHPS; Usual suspects
plus health specification, OLS, OLS with categorical health, Ordered Logit, BUC estimator
& OLS with fixed effect entire sample, N = 153,886 ...................................................... 105
Table 3.32: The determinants of Life Satisfaction for British in the BHPS; Age-bands: (1)
without controls; (2) with controls per the Usual suspects plus health specification; (3) plus
fixed effects, N = 153,886 ................................................................................................... 107
Table 3.33: The determinants of Life Satisfaction for Australians in the HILDA; Usual
suspects plus health specification, OLS, OLS with categorical health, Ordered Logit, BUC
estimator & OLS with fixed effect entire sample, N = 72,529 ......................................... 108
Table 3.34: The determinants of Life Satisfaction for Australians in the HILDA, Age-bands:
(1) without controls; (2) with controls per the Usual suspects plus health specification; (3)
plus fixed effects, N = 75,729 .............................................................................................. 110
Table 4.35a & b: Summary of happiness studies of the young from the economics literature
(EconLit) identifying the study population as adolescents or children ................................ 114
Table 4.36: Observations by age & year for the 15 to 23 year-old sample from HILDA waves
2-8 ........................................................................................................................................ 117
Table 4.37: Sample averages for the 15 to 23 year-old cohort and the entire HILDA sample
............................................................................................................................................. 118
Table 4.38: Cross-country OLS happiness results ordered (1) to (6) by the size of the
standardised beta coefficient ................................................................................................ 129
Table 4.39: Personality factors, related behaviours, & Happiness Survey question numbers
............................................................................................................................................. 132
Table 4.40: Happiness Survey question numbers & behaviours for the school environment
and interaction with friends life satisfaction domains ......................................................... 134
Table 4.41: Life Satisfaction for the 9 to 14 year old children in the Smart Train data .... 136
Table 4.42: Descriptive statistics for selected questions from the Smart Train crosssectional data; N = 389 ........................................................................................................ 138
Table 4.43: The determinants of Life Satisfaction for children aged 9 to 14 years in the Smart
Train dataset; OLS regression, N = 389 .............................................................................. 167
Table 4.44: The determinants of Life Satisfaction for children aged 9 to 14 years in the Smart
Train dataset; OLS regression, N = 389 .............................................................................. 168
[xii]
Table 4.45: Other determinant variables of Life Satisfaction for the children aged 9 to 14
years in the cross-sectional Smart Train dataset; OLS regression, N = 389 ........................ 169
Table 4.46: The determinants of Life Satisfaction; Pooled OLS regression results for 15 to 23
year-old cohort and All (15 to 92 year-olds) in the HILDA; N = 12,330 ............................ 170
Table 5.47: Sample averages for individuals in the HILDA; N = 55,177 ........................... 176
Table 5.48: Sample averages for life events affecting individuals in the HILDA; N = 55,177
............................................................................................................................................. 177
Table 5.49: Stress levels defined by the Social Readjustment Rating Scale........................ 183
Table 5.50: Stress levels defined by the Social Readjustment Rating Scale (continued) .... 184
Table 5.51: The HILDA personality questionnaire (HILDA, 2008a) .................................. 191
Table 5.52: Descriptive statistics for aggregate variables used in models (1) to (6); N = 70
............................................................................................................................................. 205
Table 5.53: OLS regressions results for nested Aggregate Models of Happiness (5) for
Australians aged 15 to 84; N = 70 ....................................................................................... 207
Table 5.54: OLS regressions for the Aggregate Model of Happiness (5.2) with the eleven
most important life events; N = 70 ...................................................................................... 208
Table 5.55: The determinants of Life Satisfaction for Australians; Pooled OLS regression
results for individuals in the HILDA; N = 55,177 ............................................................ 213
Table 5.56: The determinants of Life Satisfaction for Australians; Fixed-effect regression
results for individuals in the balanced HILDA panel; N = 55,177 ..................................... 215
[xiii]
Table of Figures
Figure 3.1: Life satisfaction question from page 35 of wave 18 of the GSOEP Living in
Germany Survey questionnaire on the social situation of households ................................... 27
Figure 3.2: Life satisfaction question from page 74 of wave 8 of the HILDA Continuing
Person Questionnaire ............................................................................................................. 28
Figure 3.3: Life satisfaction question from the British Household Panel Survey: wave 18
questionnaire, p. 64 ................................................................................................................ 28
Figure 3.4: Average life satisfaction by age in the GSOEP for the pooled sample .............. 29
Figure 3.5: Life satisfaction in the GSOEP for the pooled sample with added controls........ 30
Figure 3.6: Life satisfaction in the HILDA for the pooled sample ........................................ 32
Figure 3.7: Life satisfaction in the HILDA for the pooled sample with added controls ........ 34
Figure 3.8: Life satisfaction in the BHPS for the pooled sample........................................... 36
Figure 3.9: Life satisfaction in the BHPS for the pooled sample with added controls .......... 37
Figure 3.10: Life satisfaction in the GSOEP for the pooled sample for the mid-age range... 40
Figure 3.11: Life satisfaction in the HILDA for the pooled sample for the mid-age range ... 41
Figure 3.12: Life satisfaction in the BHPS for the pooled sample for the mid-age range ..... 42
Figure 3.13: Life satisfaction in the GSOEP for the balanced panel ..................................... 45
Figure 3.14: Age and observed correlates in the GSOEP ...................................................... 47
Figure 3.15: Life satisfaction in the HILDA for the balanced panel...................................... 48
Figure 3.16: Age and observed correlates in the HILDA ...................................................... 50
Figure 3.17: Life satisfaction in the BHPS for the balanced panel ........................................ 51
Figure 3.18: Age and observed correlates in the BHPS ......................................................... 53
Figure 3.19: Comparison of Age and individual observed correlates across data sets .......... 54
Figure 3.20: Predicted happiness effects of the non-age variables in the GSOEP................. 57
Figure 3.21: Predicted happiness effects of the non-age variables in the HILDA ................. 58
Figure 3.22: Predicted happiness effects of the non-age variables in the BHPS ................... 58
[xiv]
Figure 3.23: Year and life satisfaction in the GSOEP for the pooled sample ........................ 63
Figure 3.24 (top) and 3.24 (bottom): life satisfaction in the GSOEP for first-time respondents
............................................................................................................................................... 66
Figure 3.25: The degree of selection in the GSOEP for stayers in the panel ......................... 68
Figure 3.26: Year and life satisfaction in the HILDA for the pooled sample ........................ 71
Figure 3.27 (top) and 3.25 (bottom): life satisfaction in the HILDA for first-time respondents
............................................................................................................................................... 72
Figure 3.28: The degree of selection in the HILDA for stayers in the panel ......................... 73
Figure 3.29: Year and life satisfaction in the HILDA for the pooled sample ........................ 74
Figure 3.30 (top) and 3.30 (bottom): life satisfaction in the BHPS for first-time respondents
............................................................................................................................................... 75
Figure 3.31: The degree of selection in the BHPS for stayers in the panel ........................... 76
Figure 4.32: Log of average Annual Household Income and proportion of self-reported
pregnancy (self or partner) for Australians aged 15 to 93 years, 2002-2008 HILDA, N =
77,132 .................................................................................................................................. 119
Figure 4.33: Life Satisfaction of 15 to 93 year-old Australians (2002-2008 HILDA panel
data) ..................................................................................................................................... 120
Figure 4.34a: Front side of the Happiness Postcard themed with the same graphics as the
Happiness promotional poster & the icon clicked to initiate the web-based Happiness
Survey ................................................................................................................................. 123
Figures 4.35(a) & (b): Average Life Satisfaction for 9 to 14 year old Australian children in
the Smart Train data and 15 to 90 year-old Australians in the 2002-2008 HILDA panel data
............................................................................................................................................. 140
Figure 4.36: The predicted changes in childhood happiness from each domain factor as the
children move up in school grade ........................................................................................ 145
Figure 4.37a: The online Happiness Survey: initial screen and question q1 ..................... 155
Figure 5.38: Average happiness for Australians aged 15 to 84 ........................................... 186
Figure 5.39: Average stress level for Australians aged 15 to 84 ......................................... 186
Figure 5.40: Average stress from positive and negative life events; Australians aged 15 to 84
............................................................................................................................................. 188
[xv]
Figures 5.41a to e: Change in personality factors over time for Australians aged 15 to 84;
scale is 1 to 7 ........................................................................................................................ 194
Figures 5.42: The role of the direct and indirect effects from personality on life events and
the stress arising from those life events ............................................................................... 198
Figure 5.43a to k: Graphics of the stress at each age arising from the eleven most important
life events ............................................................................................................................. 209
[xvi]
[xvii]
BHPS1
GDP
GSOEP2
Happiness
Life satisfaction
HILDA3
PanelWhiz4 Software used to extract variables from the panel data sets
QUT
SRRS
UK
USA
I thank the UK Data Archive, the Economic and Social Research Council and the University of
Essex for the use of the BHPS data - Study Number 5151 - British Household Panel Survey: Waves 118, 1991-2009.
2
This thesis uses unit record data from the Household, Income and Labour Dynamics in Australia
(HILDA) Survey. The HILDA Project was initiated and is funded by the Australian Government
Department of Families, Housing, Community Services and Indigenous Affairs (FaHCSIA) and is
managed by the Melbourne Institute of Applied Economic and Social Research (MIAESR). The
findings and views reported in this paper are those of the author and should not be attributed to either
FaHCSIA or the MIAESR. I thank FaHCSIA & the Melbourne Institute director, Professor Deborah
Cobb-Clark, and her staff for making the data available.
4
The HILDA and BHPS data was extracted using the Add-On package PanelWhiz v3.0 (Nov 2010)
for Stata. PanelWhiz was written by Dr. John P. Haisken-DeNew (john@panelwhiz.eu). The
PanelWhiz generated DO file to retrieve the HILDA data used here and any Panelwhiz Plugins are
available upon request. Any data or computational errors in this paper are my own. Haisken-DeNew
and Hahn (2006) describes PanelWhiz in detail.
xviii
xix
The work contained in this thesis has not been previously submitted for a Degree or
Diploma at any other higher education institution. To the best of my knowledge and
belief the thesis contains no material previously published or written by another
person except where due reference is made in the thesis itself.
Signed :
______________________________
Douglas A. Beatton
xx
xxi
Acknowledgements
Many people contributed to the completion of this thesis. To all those involved I offer my
sincere gratitude and thanks. There are those, whose contribution was exceptional, I
acknowledge their considerable assistance. First, thank you to my wife, Debbie, and our sons
Stuart and Anthony for their understanding and patience. Your love and support have been
the greatest driving force. To my mentor and supervisor during my PhD studies, Professor
Paul Frijters, thank you for your guidance and advice, for believing that I could learn
something about economics, redeem my lost math skills, develop my empirical abilities, and,
actually write something that others might choose to read. My sincere thanks also go to Dr
David Johnston and Professor Uwe Dulleck who willingly substituted as my Principal
Supervisor after Paul moved to another university. I particularly thank my co-supervisor,
Professor Lisa Bradley, for helping me traverse the inter-disciplinary precipice between
psychology and economics. Thanks for the roller coaster ride; it has been an exhilarating
four years.
To those who helped develop my economics knowledge, thank you. To Professor Stan Hurn,
Dr David Johnston and Dr Vlad Pavlov, thank you for my new found econometric skills. To
Professor Uwe Dulleck, thank you for the helpful discussions and the generous loan of your
personal library of economics texts. To the research students in the School of Economics and
Finance, thanks for the helpful advice and debate that contributed to my learning. My
particular thanks go to Redzo Mujcic, Markus Schaffner, Bin Dong, Jonas Fooken, David
Savage and Wasantha Athukorala who were always available to offer help and
encouragement when things became difficult.
An empirical thesis like this is reliant upon good data. I therefore thank the providers of the
data sets used in this thesis: the German Socio-Economic Panel, the British Household Panel
Survey, and, the Household Income Dynamics in Australia Survey. I also thank Leesa
Watkin, QUT Smart Train Project Manager, and Annie Harris, Senior Project Officer from
the Queensland Government, Department of Tourism, Regional Development and Industry
for their assistance in collecting data from the Queensland children who visited the QUT
Smart Train. Finally, the assistance of the School of Economics & Finance, the QUT
Business School, the Research Students Centre and the Library staff at QUT is gratefully
acknowledged.
xxii
xxiii
Chapter One
Introduction
1.1
This thesis is about happiness and age. The thesis begins with a preliminary review
of the economics of happiness literature (Chapter 2). The review reveals gaps in the
happiness literature that are the basis for the research questions pursued in the
Chapter 3, 4, and 5 studies. The Chapter 3 study addresses the first research gap, the
U-shape of happiness in age. After offering a methodological explanation for the Ushape of happiness in age in Chapter 3, Chapter 4 extends our view of lifetime
happiness to an age-cohort infrequently visited in the economics of happiness
literature, children. The Chapter 4 study uses unique cross-sectional data to reveal
life satisfaction domain factors that contribute to explaining happiness change in 9 to
14 year-old children. Chapter 5 builds on the findings from the Chapter 3 and 4
studies by offering a life event based explanation of changes in happiness over the
lifetime. Chapter 6 summarises the Chapter 3, 4 and 5 findings and they are
discussed in Chapter 7.
Consistent with the example of eminent happiness researchers such as Andrew Clark, Ed Diener,
David Blanchflower, Richard Easterlin, Bruno Frey, Paul Frijters, Andrew Oswald, Bernard van
Praag, and others, I interchangeably refer to life satisfaction as happiness or wellbeing.
Lifetime happiness is topical. The December 2010 issue of The Economist magazine
devoted three pages to The U-bend of life .... Why, beyond middle age, people get
happier as they get older (Economist, 2011, pp. 33-36). This article began by
identifying what The Economist called the dismal way ... the economics discipline
talks about happiness ... using money as a proxy for utility. Like The Economist, this
behavioural economics researcher is unconvinced that material wealth is all that
makes us happy. Some policy makers agree. For example, the French President,
Nicolas Sarkozy (2009) has recently commissioned two Nobel-prize-winning
economists (Amartya Sen and Joseph Stiglitz) to develop a broader happiness
measure. In addition, The Economist revealed that, David Cameron, the British Prime
Minister, has shown similar leadership by initiating the collection of wellbeing data.
After revealing that policy makers are beginning to believe that money and GDP
growth are not the panacea to measuring the overall wellbeing of society, The
Economist illustrates that happiness over a lifetime is shaped like the U-bend in our
bathroom washbasin. To explain this U-shape of happiness in age, The Economist
draws from the current economics literature. Unlike The Economist, this thesis is
unbounded from a single happiness perspective. This thesis offers explanations for
the U-shape of happiness in age that are based upon theory from other behavioural
disciplines. The explanations offered are supported by empirical evidence from oftused socio-economic datasets6. But first, I begin with a preliminary review of the
happiness literature7 to reveal the research gaps that are the basis for the three studies
documented in this thesis.
As well as my own data set, I use panel data from the Household Income Dynamics in Australia
Survey (HILDA, 2009), the German Socio-Economic Panel (GSOEP, 2008), and the British
Household Panel Survey: Institute for Social & Economic Research (BHPS, 2010).
7
Chapters 3, 4 & 5 also contain a detailed literature review relevant to the research questions
addressed in those chapters.
Chapter Two
Preliminary review of the happiness literature8
2.1
Happiness
The notion of happiness transcends culture, geography & time. The above quote
from John Stuart Mill (2009) is reflective of the importance nineteenth century
British society placed upon happiness. The founding fathers of the United States of
America considered it such an important human ideal that they included the pursuit
of Happiness in the preamble of the 1776 Declaration of Independence. More
recently the Nobel Prize winner Joseph Stiglitz and the French President (Sarkozy,
2009) urged world leaders to revolutionise their thinking on economic success by
extending economic measures of growth beyond gross domestic product to include
factors such as healthcare, social situation and happiness.
Happiness is not just the reserve of western society. Confucius, the fifth century BC
Chinese philosopher and teacher motivated the practice of a life of virtue by stating
that the souls of departed relatives were largely dependent for their happiness on the
conduct of their living descendants (Shinn, 2009). The Hindu philosophy of
prarabdha karma amplifies the inter-temporal importance of happiness by stating that
happiness is a consequence both of actions performed in the present and past lives
(Srivastava & Misra, 2003). In addition, the Buddha taught that happiness (sukkha) is
not derived from an escalating spiral of consumption but from a pervasive state of
mind that emerges when the trials, tribulations, and events that confront us in our
lives can be borne with ease. The notion of happiness extends across time and
Note: There is a detailed literature review for each study in Chapters 3, 4 and 5.
10
See Kashdana, Biswas-Diener, & Laura A. King (2008) for more on the objectivist versus the
subjectivist or mentalist happiness positions.
11
Extended by John Stuart Mill to Utilitarianism (Mill, 2010; UCL, 2009).
In a society where individuals (i) have limited resources and opportunities, we should
consider the overall wellbeing of society when selecting alternatives that efficiently
maximise the utility value (U) we derive from our choice behaviours, and thereby
maximise our and the overall well-being of society:
max
.
U
i =1
Economics of happiness researchers view life satisfaction scales as measures of subjective utility;
they propose using them as an alternative to the orthodox revealed preferences approach to utility
(Varian, 1992, p.132)
being that arises from the environment in which we live at a particular point in time.
However, there emerged a difference of opinion across the scientific disciplines on
what constitutes a quality or happy life and how its determinants should be
measured.
The objective approach to measuring the determinants of wellbeing uses selected
objective variables drawn from the regular census and other socio-demographic
surveys.
level, health, age, gender, housing, leisure activities, and income. The subjective
approach to measuring the determinants of wellbeing also uses data collected with
polls and surveys, but with this subjective data, people are asked about the quality of
their life, their experiences and what is going on in their lives. Examples of selfreported variables include asking people about their mental health, their overall
health, self-rated stress, financial well-being and overall life satisfaction (Andrews &
Withey, 1976; Campbell, Converse, & Rodgers, 1976). This research takes a multidiscipline approach by using combinations of objective and subjective variables to
explain happiness change over the lifetime.
As to measuring happiness itself, economists could choose from numerous happiness
or wellbeing measures. Measures include: the Satisfaction With Life Scale (SWLS)
(Diener, et al., 1985); Life Satisfaction Index (LSI- A) (Leugarten, 1987 #645);
Temporal Satisfaction with Life Scale (TSWLS) (Pavot, Diener, & Suh, 1998;
(Watson,
1988)
Positive
Affect/Negative
Affect
Schedule
(PANAS);
All things considered, how satisfied are you with your life?
In Chapter 4, I argue and provide evidence for why the global happiness question is
suitable for measuring happiness; the aggregate utility arising from all the good and
bad things that occur in our lives (Fordyce 1988). Using data collected with the
global happiness question, economics of happiness research has typically focused on
finding the causes and correlates of happiness. The model of individual happiness
(2.1) is:
LSit = C + 1 X it + 2 Leit + it
2.1
where,
LSit
Constant
Xit
Leit
it
error term
2007); good health (Gerdtham & Johannesson, 2001; Veenhoven, 2008) and, social
relationships (Powdthavee, 2008). Socio-demographic variables found to have a
negative effect on happiness include: unemployment (Clark & Oswald, 1994; Di
Tella, MacCulloch, & Oswald, 2001); victim of crime (Powdthavee, 2003); obesity
(Oswald & Powdthavee, 2007); children (Phelps, 2001; Tsang, 2003), divorce
(Gardner & Oswald, 2006), ill-health (Gerdtham & Johannesson, 2001); disability
(Oswald & Powdthavee, 2008), and; the death of a loved one or the imminent onset
of ones own demise (Frijters, Johnston, & Shields, 2008).
In addition to time-variant socio-demographic variables, the inclusion of life event
questions in socio-economic surveys has provided the opportunity to analyse the
effect(s) of self-reported life events (Leit) on individual happiness. Life events
occurring in the previous period and found to have a negative effect on individual
happiness include: finances just worsened (Frijters, Johnston, & Shields, 2009); just
separated or divorced (Blanchflower & Oswald, 2004); just fired from a job (Clark &
Oswald, 2002); was just a victim of crime (Frijters, et al., 2008), and; spouse, friend
or close relative just died (Frijters, et al., 2008). Life events occurring in the previous
period and found to have a positive effect on happiness include: finances just
improved; just got married; just had a baby (Frijters, et al., 2009), and; just retired
(Beatton & Frijters, 2009). Some of these life event shocks have been priced. Clark
& Oswald (2002) concluded that widowhood (death of spouse) brings a degree of
unhappiness that requires, on average, an extra 170 000 per annum to offset, and,
Frijters et al. (2008a) concluded that the average criminal event can be offset by a
windfall income gain of about 14,000 US. The inclusion of life events in happiness
regressions has contributed to the level of happiness explanation.
2.2
The inclusion of age effects in happiness regressions has also contributed to the level
of happiness explanation. Until the early 2000s, economic opinion about the effect of
age on happiness was divided. Clark (2006) found a U-shaped pattern for the UK,
whilst Winkelmann and Winkelmann (1998) found no U-shape in happiness but
simply a very strong negative effect of age. Easterlin and Schaefer & Macunovich
(1993), using 20 years of the US General Social Survey even concluded that life
8
satisfaction is almost flat in age, with neither a U-shape nor a negative slope.
Alesina, Di Tella, & MacCulloch, (2004) and van Praag, Frijters, & Ferrer-iCarbonell (2000) even found an inverted U-shape. Despite economic opinion on the
effect of age on happiness being divided, age effects began to regularly appear in
models of individual happiness (2.2).
2.2
In 2006, Clark claimed robustness with respect to methodology for the finding that
happiness is U-shaped in age when he concluded that Panel analysis controlling for
fixed effects continues to produce a U-shaped relationship between well-being and
age.
The weight of economic opinion appeared to have tilted towards acceptance of the
U-shape of happiness in age as a stylised fact. However, such a stylised fact conflicts
with an old psychology literature that finds no happiness-age relationship (Cantril,
1965). Palmore and Luikart (1972) comment in their review; Several variables
thought to be related to life satisfaction had little or no relationship: age, sex, total
social contacts. Studying the age-effect relationship on the happiness of 2,272
individuals aged 25 to 74 years, Mroczek & Kolarz (1998) found that that
personality, contextual, and socio-demographic variables, as well as their
interactions, are all that is needed to fully understand the age-happiness relationship.
In Chapter 3, I examine this difference of opinion between the age-happiness
literatures and provide an empirically-supported explanation for the U-shaped
relationship between age and happiness over a lifetime.
2.3
13
Oft-used socio-economic panel data sets s usually include observations on individuals aged 15 years
and over.
10
Lee & Oguzoglu (2007) examined how the receipt of income support payments
affected the well-being of youths with a median age of 14 years. Flouri (2004)
examined the role of parenting in later-life subjective well-being and found
correlations between closeness to parents at age 7 and the happiness of the same
individuals at age 42, and, Csikszentmihalyi & Hunter (2003) found correlations
between proximal environmental factors and the happiness of US grade school
children. There are a small number of studies on adolescent happiness but even less
on childhood happiness, and, these studies do not widen our view of lifetime
happiness to include the young. In Chapter 4, I address this gap in the literature by
extending our view of lifetime happiness to nine-year-old children and ask, is there is
a decrease or an increase in the happiness of young individuals, and, what explains
changes in childhood happiness?
2.4
With our view of happiness extended to young children, Chapter 5 builds on the
findings of the first study by offering explanations to changes in happiness over a
lifetime. The aggregate model of happiness over a lifetime that is offered
incorporates life event shocks, peer effects and fixed traits. As we will see in Chapter
3, fixed traits are very important to happiness and the question arises what these
fixed traits might be. Obvious fixed-effect candidates are the traits that psychologists
now consider stable over a lifetime, personality (McCrae, Costa, Mroczek, & Little,
2006). Previously, the prevailing consensus was that life events like marriage,
parenting, retirement, and chronic illnesses would profoundly affect personality.
After many decades of research, psychologists generally agree that personality traits
are invariant across age (McCrae et al., 2002), life events have been found to have no
effect on personality (Costa, Herbst, McCrae, & Siegler, 2000).
However,
personality can affect how an individual behaves when confronted with a life event
shock. In their four-year longitudinal study of young adults, Magnus, Diener, Fujita,
& Pavot (1993) examined the causal pathways between personality and life events.
They found the personality trait of extraversion predisposed individuals to
experience life events more positively, whereas neuroticism predisposed individuals
to experience life events more negatively. These findings from the psychology
literature provide some evidence that, while personality has been found to have direct
11
effects on happiness, the effects from life events on happiness may be mediated by
personality.
The study of Howell (2006) goes beyond the above-mentioned studies that
considered interactions between personality and life event shocks; they found that
daily social interactions mediated the relationship between personality trait of
extraversion and life satisfaction. Individual happiness is not only affected by
personality, life event shocks, and their interactions, but also by those individuals
with whom we socially interact on a daily basis, our peers. In clarification, I offer an
example from Rayo & Beckers seminal peer effect paper (Rayo & Becker, 2007).
Individuals confronted with an improvement in finances life event shock are not
just concerned with their absolute level of income but also with the difference
between their income and the income of their peers. Rayo & Becker (2007) theorise
that the basis for this comparison by peers is a social (income) norm, or set point,
that changes over time. This peer comparison example fits well with the Easterlin
Paradox, a phenomena seen in a number of western countries over the past few
decades (Easterlin, 1995). In France, Germany, Japan, the United Kingdom and the
United States of America higher incomes have not translated to increased happiness.
Individual happiness has not changed as real per capita income grew (Lewer,
Gerlich, & Gretz, 2009). Individuals appear to quickly adapt to changes in social
norms and as a result are no (un)happier.
We can see from the previous paragraphs that happiness researchers have considered
peer effects, personality and life events shocks in isolation, but, not together. In
Chapter 5, I take a unique approach to the study of lifetime happiness by combining
all three (personality, life events shocks and peer effects) to offer an aggregate model
that significantly explains changes in happiness over a lifetime.
12
2.5
The preliminary review of literature revealed the research gaps that are the basis for
the studies in Chapters 3, 4 and 5. The Chapter 3 study seeks to contribute to the
ongoing debate on whether lifetime happiness is U-shaped in age. The Chapter 3
study extends our view of lifetime happiness by examining an age cohort seldom
visited in the economics of happiness literature, childhood happiness. With our view
of lifetime happiness extended to children, Chapter 5 takes a unique approach to the
study of happiness by combining personality and life event shock interactions with
peer effects in an aggregate model of lifetime happiness that significantly contributes
to the explanation of changes in happiness over a lifetime.
13
14
Chapter Three
The puzzle of the U-shaped relationship between
happiness and age
In this chapter, the puzzle of the relationship between age and happiness is
considered. Whilst the majority of psychologists have concluded there is not much of
a relationship at all, the economic literature has unearthed a possible U-shape of
happiness in age. In this chapter, I look for the U-shape in three panel data sets, the
German Socioeconomic Panel (GSOEP), the British Household Panel Survey
(BHPS) and the Household Income Labour Dynamics Australia (HILDA) and
investigate several possible explanations for it14.
3.1
Introduction
What is the relationship between happiness and age? Do we become more miserable
as we age, or is our happiness relatively constant throughout our lives with only the
occasional special event (marriage, birth, promotion, illness) temporarily raising or
reducing our happiness, or, do we actually get happier as life gets on and we learn to
be content with what we have?
14
This chapter has been submitted as the joint paper with my supervisor, Professor Paul Frijters. The
paper is titled The mystery of the U-shaped relationship between happiness and age and is at the
review and resubmit stage with the Journal of Economic Behavior and Organisation. We would like to
thank conference attendees, anonymous referees, and, seminar participants for useful comments and
suggestions.
15
The weight of evidence from the recent economics of happiness literature supports a
belief that the age-happiness relationship is U-shaped15. This finding holds for the
United States, Germany, Britain, Australia, Europe, and South Africa. The stylised
finding is that individuals gradually become unhappier after their 18th birthday, with
a minimum around 50, followed by a gradual upturn in old age. The predicted effect
of age can be quite large. For example, the difference in average happiness between
an 18 year old and a 50 year old can be as much as 1.5 points on a 10-point-scale.
This recent economics literature, however, conflicts with an old psychology literature
that finds no happiness-age relationship (Cantril, 1965). Palmore and Luikart (1972)
comment in their review, Several variables thought to be related to life satisfaction
had little or no relationship: age, sex, total social contacts, .....Diener, Sapyta, &
Suh (1998) provide a common-held psychological explanation for changes in
happiness over our lifetime. The changes are a reflection of the positive and negative
effects that emerge from changes in our life situations. More recently, Dear,
Henderson, & Korten (2002) conclude that the prevalence of high life satisfaction
simply becomes less common at higher ages. From this reading, it is clear that either
the psychologists have overlooked something important for a long time or that the
economists have somehow gotten it wrong recently; this chapter intends to clarify
and find out, which it is.
I begin by re-examining the age-happiness relationship and then delve into the
methodological aspects of the problem. Essentially, we want to know if the U-shape
that economic scholars find is an artefact or real and what the actual relationship
between age and life satisfaction is. The age-happiness relationship is examined
using an often-used dataset, the German Socio Economic Panel that has an extensive
set of variables on the individual level. This data-richness allows us to not only
15
Recent papers on this in the economic literature include: (Bell & Blanchflower, 2007;
Blanchflower, 2008; Blanchflower & Oswald, 2001; 2004; 2007; 2008; 2009; Clark, 2006; Dear,
Henderson, & Korten, 2002; Di Tella, MacCulloch, & Oswald, 2001; Ferrer-i-Carbonell & Frijters,
2004; Ferrer-i-Carbonell, 2005; Gerdtham & Johannesson, 2001; Hayo & Seifert, 2003; Helliwell,
2003; Oswald, 1997; Oswald & Powdthavee, 2008; Powdthavee, 2003; Seifert, 2003; Senik, 2004;
Theodossiou, 1998; Van Landeghem, 2008; Winkelmann & Winkelmann, 1998; Wolpert, 2010). An
introduction to the found effects of correlates of happiness can be found in Frey & Stutzer (2002). For
a recent general introduction to the economic literature on happiness, see Clark et al. (2008). For a full
list of earlier papers in the field of happiness, see Veenhovens Database of Happiness (introduced in
Veenhoven et al. 1994).
16
replicate the findings of other studies based on cross-sectional data, but also allows
us to explore the dynamic interplay between age, covariates, unobserved
heterogeneity and happiness.
The progression in this chapter is to let the puzzle of the age-happiness relationship
unfold. I first briefly review the recent literature where I summarise the methodology
and reveal the main findings from other happiness studies. Then I present the data we
have and show that we can also generate a U-shape of happiness in age when we run
similar regressions to those in the literature. I then go through successive
explanations for the U-shape, including: the possibility that it depends upon
including the happiness decline found in early adulthood (age 18 to 22); that it is an
artefact of not allowing for fixed effects; or that it is a truly robust finding. In the
conclusion, I summarise the findings and explain what they mean for future research
into the economics of happiness.
3.2
Literature review
Despite this early difference of opinion, nearly all recent papers come down on the
side of a U-shape relationship between happiness and age. Blanchflower and Oswald
(2001; 2004) simply state that Wellbeing is U-shaped in age. Gerdtham and
Johannesson (2001) also report a U-shape in age with a minimum around the age of
55. Hayo and Seifert ( 2003) and Seifert (2003) also report a U-shape and call the U17
18
Table 3.1a&b: Life Satisfaction regression results (t-values) from recent studies
Coefficients - Pooled
Dependent
(t-value)
(t-value)
variable (DV)
Sample
Author, date
Age
Age Squared
Age
Age Squared
and controls
DV:
(Blanchflower
data from 8
& Oswald,
European
2009)
nations
(Blanchflower
& Oswald,
2008)
data from 16
countries
OLS
16
-0.00800
OLS
15
0.0000815
& Oswald,
2001)
Ordered
Ordered
Logit
Logit
-0.0576
0.0006
(8.85)
(9.95)
USA: General
Logit
Ordered Logit
Social Survey
(men+women
(men+women
1972-2006
averaged)
averaged)
N = 45,474
USA -0.0211
USA: 0.0003
(4.39)
(5.92)
Ordered
(Blanchflower
& Oswald,
2001)
Europe:
Logit
Ordered Logit
Eurobarometer
(men+women
(men+women
1976-2002
averaged)
averaged)
N = 589,446
Eur: -0.045
Eur: 0.00052
(31.31)
(10.1)
Ordered
(Blanchflower
& Oswald,
2001)
World Value
Survey
1981- 2004
N = 163,852
Satisfaction
without
controls
Ordered
(Blanchflower
Life
Logit
(men+women
averaged)
WVS:
-0.0505
(10.1)
DV:
Life
Satisfaction
with personal
controls
DV:
Happiness
Controls: yes
(specification
without
cohort)
DV: Life
Satisfaction
Controls: yes
(specification
without
cohort)
DV: Life
Ordered Logit
Satisfaction
(men+women
Controls: yes
averaged)
(specification
WVS: 0.0003
without
(5.92)
cohort)
16
19
Table 3.1a&b (continued): Continuation of Life Satisfaction regression results (t-values) from
recent studies
Sample
Author, date
(t-value)
(t-value)
Age
Age Squared
UK:
(Blanchflower
Eurobarometer
& Oswald,
Survey
2004)
1975-1998
(Clark, 2006)
Coefficients - Pooled
Age
Ordered
Logits - All
UK: -0.0424
UK: 0.0005
Satisfaction
(2.84)
(15.38)
Controls: yes
N = 54,549
N = 54,549
DV: Life
Applied age
cohorts to
waves 1 to 14
and controls
Ordered
British
(BHPS)
Squared
variable (DV)
Logits - All
Household
Panel Survey
Age
Dependent
-0.075
0.00091
(-25)
(30.33)
N = 82,096
N = 82,096
OLS
OLS
DV: Life
Satisfaction
derive fixed
Controls: yes
effect
coefficients
DV: Life
(Di Tella, et
Eurobarometer
-0.02
0.0002
Satisfaction
al., 2001)
Survey Series
(20.0)
(33.33)
Controls: yes
1975-1991
N = 264,710
N = 264,710
-0.011
0.0001
(z-stat: -2.38)
(z stat: 2.03)
N = 20,634
N = 20, 634
(Powdthavee
, 2005)
Statistics
South Africa
OHS study of
1997
Russian
longitudinal
(Senik, 2004)
monitoring
survey
(RLMS).
(Winkelmann
German Socio-
&
Economic
Winkelmann,
Panel
1998)
1984-89 waves
of the GSOEP
Ordered
Ordered
Probit (2)
Probit (2)
-0.050
.001
(8.33)
(p < .01)
N = 17,897
N = 17,897
DV: Life
Satisfaction
Controls: yes
DV: Life
Satisfaction
Controls: yes
fixed effects
logit model 2
-0.098
0.0012
-0.118
(-9.8)
(12)
(-3.19)
N = 20,944
N = 20,944
N = 20,944
20
fixed effects
logit model
DV: Binary
Life
-0.0001
Satisfaction
(0.25)
Controls: yes
N = 20,944
Tables 3.1a & 3.1b confirm the very strong effect that age is found to have upon life
satisfaction in recent studies and that the effect of linear age is always negative,
whilst that of age-squared is positive, indicating a U-shape. Bearing in mind that the
age at which the minimum occurs is given by the coefficient of linear age divided by
twice the coefficient of age-squared; it appears that the majority of the studies find
that 55 is the age at which minimum happiness occurs. Tables 3.1a & 3.1b also
underscore that the effects are mainly found in cross-sections when controls are
added for individual socio-economic variables.
Despite the reliance in the literature on using age and age-squared in order to unearth
a U-shape, other approaches have been used. Wunder et al. (2009) include a fourthorder polynomial of age in their happiness regressions, where they find that the
higher order terms are also significant and hence that the U-shape is not a perfect
description of the actual relationships (they find a clear negative slope at the very
high age ranges). Yet, since this chapter is interested in seeing where a particular
finding in the literature comes from, I follow the convention of focussing on just a
second-order polynomial (age and age-squared).
With the literature review coming down heavily in support of the presence of a Ushape of happiness in age, this chapter proceeds by seeing if we can replicate the Ushape of happiness in age using the oft-used German Socio-Economic Panel data
(GSOEP) panel data set. While the GSOEP is the base data source for findings and
explanations, robustness added by replicating the GSOEP findings with two
additional socio-economic panel data sets: the Household, Income and Labour
Dynamics in Australia Survey panel data (HILDA), and; the British Household Panel
Survey (BHPS) panel data. These three panel data sets are now described.
21
3.3
3.3.1.1
The average self-assessed happiness in the GSOEP (7.16) is on the high side of the 0
to 10 scale. Forty nine per cent of respondents are male and the average age across
the entire unbalanced panel is 44.26 years, with the average age of first-time
respondents 5.79 years lower. Average years of education are 10.93 years. The
average employment level is 47%, 4% are unemployed, and, 33% are nonparticipants in the labour force, with the remaining retired. Average household
annual income is 49,800 Euros. Average health (2.59) is on the poorer health side of
the health question response scale (1 to 5), but only 4% self-identify with a disability
(invalid). The majority of households are married (65%), with less than one child on
average (0.65). Just 42% own or are paying off the home in which they live. On
average, the life event that occurs most often is the birth of a child (4%) with the next
highest event occurrence being job loss (2%), a life event occurrence the same as
marriage. Let us now review the HILDA panel data.
22
3.3.2.1
The average self-assessed happiness in the HILDA (7.91) is on the high side of the 0
to 10 scale. Forty seven per cent of respondents are male and the average age across
the entire unbalanced panel is 45.50 years; with the average age of first-time
respondents 1.44 years lower. Surprisingly, average years of education for
Australians (12.82) is 1.89 years higher than in the German sample, and, the
Australian standard deviation is lower (1.80 versus 2.46). The average Australian
employment is 66%, 3% unemployed, with the remaining non-participating or
retired. Average annual household income is $AUD 63,357, but the distribution is
positively skewed (range 0 to 583,260; 50th percentile 55,535; 75th percentile 81,966,
and; 95th percentile 138,624). Average health (2.64) is towards the poorer health side
of the health question response scale (1 to 5), and, relative to the German sample
(4%), many more Australians (24%) self-identify with a disability (invalid)18.
17
The questionnaire for wave 1 of the HILDA panel survey did not include several important
variables often used in happiness regressions (life events).
18
The difference between high disability levels in the Australian population and populations from
other western nations is worthy of further study.
23
Compared to the German sample (65%), fewer Australian households are married
(55%), but, on average, Australian families have more children (0.65 for the
Germans versus 0.77 for Australians). Household ownership in the Australian sample
(75%) is much higher than in the German sample (42%); home-ownership is a longheld social norm in Australia. The life event shock that most affects Australians is
family death (11%), closely followed by divorce (9%). The occurrence of marriage
breakdown in the Australian sample (3%) is half as big again as in the German
sample. This is also the case with job loss; Australians are more likely to be
dismissed (fired) from their job (0.02 in the German sample versus 0.03 in the
Australian sample). This all appears to indicate that Australian families may be less
stable than German families. Compared to Germans, Australians have a higher
incidence of marriage breakdown, more children per household, and, are more likely
to lose their jobs then find new employment (they change jobs more often). Let us
now review the BHPS panel data.
19
The BHPS Waves 1 to 5 and 11 did not include the happiness question.
24
Table 3.2: Sample averages from the entire GSOEP, HILDA and BHPS samples
GSOEP
HILDA
BHPS
Mean
s. d.
Mean
s. d.
Mean
s. d.
7.16
1.85
7.91
1.47
5.23
1.29
44.26
16.91
45.50
16.92
46.45
17.80
10.68
0.53
10.77
1.07
9.84
1.61
0.49
0.50
0.47
0.50
0.45
0.50
10.93
2.46
12.82
1.80
13.26
2.40
0.65
0.99
0.77
1.12
0.53
0.94
married (1=yes)
0.65
0.48
0.55
0.50
0.55
0.50
employed (1=yes)
0.47
0.50
0.66
0.47
0.58
0.49
0.04
0.20
0.03
0.16
0.03
0.18
10.04
0.09
male (1=yes)
level of education (years)
unemployed (1=yes)
average regional income
20
4149.99
20
Self-reported health
21
0.42
0.49
0.75
0.43
0.73
0.44
1484.61 2910.48
4.94
39.82
40.70
179.67
2.59
0.95
2.64
0.95
2.20
0.95
0.04
0.20
0.24
0.43
0.02
0.14
0.01
0.08
0.11
0.31
divorced (1=yes)
0.05
0.22
0.09
0.29
0.06
0.23
0.01
0.12
0.04
0.19
0.02
0.13
0.06
0.24
0.05
0.22
0.07
0.26
0.02
0.15
0.03
0.16
0.00
0.07
0.01
0.08
0.01
0.11
0.04
0.20
0.00
0.06
0.01
0.09
0.04
0.19
0.04
0.19
0.00
0.07
0.01
0.11
0.05
0.22
0.00
0.04
0.02
0.12
0.03
0.17
invalid (1=yes)
23
22
22
22
22
N=
22
22
176,770
72,529
153,886
20
Monetary denominations are GSOEP, Euros; HILDA, $AUD, and; BHPS, British pounds.
Health is reverse coded: 1 = excellent to 5 = poor.
22
The self-report life event variables in the GSOEP and the HILDA are not in the BHPS panel data.
21
25
23
The BHPS happiness response scale is different from the GSOEP & the HILDA; the measurement
of happiness in the panel surveys is discussed in the next section (3.3.4)
24
This high level of disability in Australia (relative to Germany and the United Kingdom) may be
related to the Australian benefits system and is worthy of future research.
25
The just had a baby question in the GSOEP and the HILDA asks if you or your partner has just
had a baby. In the BHPS, with just one partner eligible for maternity leave, only one individual would
be identified as just having a baby, thus leading to a birth life event occurrence of approximately
26
preliminary analysis using the German data set, I reveal the cross-panel differences
in the measurement of our dependent variable, happiness.
3.3.4.1
The happiness question (Figure 3.1) in the GSOEP survey questionnaire is based on
the Fordyce Global Happiness Scale. The happiness question seeks to measure the
aggregate utility from all the good and bad things that occur throughout our lives
(Fordyce, 1988). The GSOEP happiness question is:
Figure 3.1: Life satisfaction question from page 35 of wave 18 of the GSOEP Living in Germany
Survey questionnaire on the social situation of households
half of that which would be self-reported by both male & female parents in the GSOEP or the HILDA
(assuming that at least one partner takes maternity leave). Doubling the BHPS just had a baby life
event to account for the mothers/fathers who do not both report to be on maternity leave still results in
a just had a baby occurrence of half that of Germany & Australia. This difference in national birth
rates and its effect on economic growth could be worthy of future research.
27
3.3.4.2
The HILDA life satisfaction question (Figure 3.2) is very similar to the GSOEP life
satisfaction question in Figure 3.1. It asks the respondent to pick a number between
0 and 10 that indicates your level of satisfaction. The more you are satisfied you are,
the higher the number you should pick. The less satisfied you are, the lower the
number.
Figure 3.2: Life satisfaction question from page 74 of wave 8 of the HILDA Continuing Person
Questionnaire
3.3.4.3
The BHPS life satisfaction question is worded and scaled differently to the GSOEP
and the HILDA. The BHPS question asks: On a scale of 1 to 7 where 1 = not
satisfied at all and 7 = completely satisfied, please tell me the number which you feel
best describes how dissatisfied or satisfied you are with the following aspects of
your current situation (plus 0, doesnt apply, and; 8, dont know). The overall life
satisfaction question is:
Figure 3.3: Life satisfaction question from the British Household Panel Survey: wave 18
questionnaire, p. 64
With the happiness question differences revealed, let us begin analysis by seeking to
confirm the presence of the U-shape of happiness in age using the German data.
28
3.4
Raw average
life satisfaction
Life Satisfacition
7.6
7.4
7.2
7
6.8
6.6
6.4
6.2
Age+Age2
6
5.8
AGE
Figure 3.4: Average life satisfaction by age in the GSOEP for the pooled sample
26
I follow the example of many psychologists, the arguments and empirical evidence from Ferrer-iCarbonell & Frijters (2004), and, the evidence from the robustness results in Appendix C of this
chapter, to interpret the meaning of the happiness question answers as interpersonally cardinally
comparable.
27
Thus, the thin curved line depicts {Life Satisfaction (age 20) + (age * (age-20)) + (age2 * (age2- 202)}
where age runs from 18 to 92 in the GSOEP, 18 to 92 in the HILDA and 18 to 90 in the BHPS.
29
The GSOEP findings are quite typical of those from other scholars in the economics
literature. There is sharp decline in raw average happiness from 18 years to 22 years
and in those close to death (where there are not many individuals left), but for the age
range 22 to 60 there is no strong relationship between age and happiness to be seen.
If we overlay the regression results, when we include linear age as the only variable,
we get a significantly negative coefficient. If we overlay a regression line with both
age and age-squared, we find a significant nonlinear pattern. The regression results
are:
(15.7)
(11.6)
R = 0.0037, N = 176770
2
Age-squared is highly significant, but the age at which the minimum occurs is about
70 with this simple specification28, implying that for the vast majority of the sample,
there is not so much of a U-shape but rather a horizontal j-shape. What if we add
additional regressors to this simple specification? Figure 3.5 shows the predicted
age-happiness profiles when we successively add additional variables. The detailed
specifications for the Pooled OLS regressions are in Table 3.14 in Chapter 3
Appendix B and Table 3.3 on page 31 summarises Age and Age2 coefficient changes
as controls are progressively added.
Age and Life Satisfaction: what if more 'controls' are
added?
Life Satisfacition
Raw average
life satisfaction
7.5
Usual suspects
6.5
Us sus + Health
5.5
Kitchen sink
5
AGE
Figure 3.5: Life satisfaction in the GSOEP for the pooled sample with added controls
28
The simple specification with just age and age2 effects comes from the recent economic of
happiness literature and provides the foundation for a systematic analytical process that seeks to reveal
the presence of the U-shape of happiness in age in the panel data sets.
30
In Figure 3.5, the thin solid line is the Usual suspects specification which includes
the socio-economic variables commonly found in the happiness regressions noted in
Tables 3.1a and 3.1b. The variables are log-income, gender, education in years, the
number of children, a marriage dummy, and three indicators of work-status
(employed,
non-participant
and
unemployed).
With
the
Usual
suspects
coefficient
t-value
Age + Age
age
age*age
-0.0217 ** 15.69
0.0002 ** 11.63
Usual suspects
age
age*age
-0.0541 ** 32.80
0.0005 ** 29.20
-0.0600 ** 36.77
0.0006 ** 34.13
Kitchen sink
age
age*age
N
Level of significance:
-0.0454 ** 25.64
0.0005 ** 25.39
176,770
31
What we see in Figure 3.5 is congruent with the findings of other scholars in the
field: when standard regressors are added, a very strong U-shape effect emerges with
predicted age effects far bigger than anything observable in the raw data. The Ushape of happiness in age is indeed in the German data, which makes it likely that the
U-shape may similarly manifest in the HILDA dataset.
Life Satisfacition
9.50
Raw average
life satisfaction
9.00
8.50
8.00
Linear age
term
7.50
7.00
6.50
Age+Age2
6.00
5.50
5.00
AGE
Figure 3.6: Life satisfaction in the HILDA for the pooled sample
32
With the GSOEP, there was a sharp decline in raw average happiness from 18 years
to 22 years; this is not evident in the HILDA29. The HILDA shows arguably the
cleanest U-shape with a predicted minimum at age 36 and no clear happiness
decrease in old age. Indeed, the linear happiness profile is quite strongly increasing
by age, counter to the general profile in the GSOEP.
Like the GSOEP, when we include linear age as the only variable we get a
significant positive coefficient (Table 3.18). If we next overlay a regression line with
both age and age-squared, we find a significant nonlinear profile (the thin solid line
in Figure 3.6). The regression results are:
(25.79)
The HILDA Age2 coefficient (0.0004, t-value 25.79) is larger and more significant
than the Age2 coefficient from the German data (0.00016, t-value 11.6). This is
probably because the raw happiness from the HILDA is more U-shaped over a
lifetime. Figure 3.7 returns our focus to the HILDA by showing how the predicted
age-happiness profiles change as we progressively add variables. The detailed
specifications for the Pooled OLS regressions are in Table 3.18 (Chapter 3
Appendix B, p.90), and Table 3.4 on page 35 summarises Age and Age2 coefficient
changes for both the GSOEP and the HILDA as we progressively add controls.
29
In Chapter 5 I ask the question; is there a similar dramatic drop in the happiness of young
Australians, when does it occur, and, why.
33
Life Satisfacition
9.50
9.00
8.50
Usual suspects
8.00
7.50
Usual suspects +
Health
7.00
6.50
6.00
"Kitchen sink"
5.50
5.00
AGE
Figure 3.7: Life satisfaction in the HILDA for the pooled sample with added controls
34
Table 3.4: Comparison of changes in the GSOEP & HILDA Age and Age2 coefficients as
controls are progressively added
Specification
Pooled OLS
Pooled OLS
(All)
(All)
GSOEP
HILDA
coefficient
t-value coefficient
t-value
Age + Age2
age
age*age
-0.0217 ** 15.69
0.0002 ** 11.63
-0.0320 **
19.0
0.0004
**
25.8
-0.0554 **
29.6
0.0007
**
34.0
-0.0403 **
22.7
0.0006
30.3
Usual suspects
age
age*age
-0.0541 ** 32.80
0.0005 ** 29.20
-0.0600 ** 36.77
0.0006 ** 34.13
**
Kitchen sink
age
age*age
Level of significance:
-0.0454 ** 25.64
-0.0311 **
16.0
0.0005 ** 25.39
0.0005 **
24.0
176,770
75,529
Next, our specification is expanded to include the life event shocks from the HILDA
data that are also in the GSOEP data. This Kitchen Sink specification adds six life
event shocks: 1) just married; 2) just divorced; 3) just separated; 4) loss of a
spouse/child; 5) the birth of a child, and; 6) being fired from your job. Comparing the
Kitchen sink prediction with the others in Figure 3.7, we see that the U-shape has
decreased slightly. The minimum predicted happiness of 7.80 now occurs earlier, at
age 30. Two of the five life event shocks (just married & just had a baby) have a
significant positive effect on happiness and the other four have a significant and large
negative effect (Appendix B, Table 3.18). So far, we have confirmed the presence of
the U-shape of happiness in age in the GSOEP and HILDA data sets. Adding a third,
I ask, is the U-shape also in the British data.
35
Raw average
life
satisfaction
Life Satisfacition
6.00
5.50
5.00
Linear age
term
4.50
4.00
Age+Age2
3.50
3.00
AGE
Figure 3.8: Life satisfaction in the BHPS for the pooled sample
When we include linear age as the only variable, we get a significant positive
coefficient (Appendix B, Table 3.22). If we next overlay a regression line with both
age and age-squared30; for the second time, we find a significant nonlinear profile
(the thin solid line in Figure 8). The regression results are:
(29.00)
30
Raw average life satisfaction as depicted in Figure 3.8 could be seen to show two turning points that
might be better explained by a specification with an age3 term. Age3 has not been included in this
specification because this study focuses on the age2-based economics of happiness literature.
36
Life Satisfacition
6.00
5.50
Usual suspects
5.00
4.50
Usual suspects +
Health
4.00
"Kitchen sink"
3.50
3.00
AGE
Figure 3.9: Life satisfaction in the BHPS for the pooled sample with added controls
37
Table 3.5: Summary of changes in the GSOEP, HILDA & BHPS Age and Age2 coefficients as
controls are progressively added
Specification
Pooled OLS
Pooled OLS
Pooled OLS
(All)
(All)
(All)
GSOEP
HILDA
BHPS
coefficient
t-value coefficient
t-value
coefficient
t-value
Age + Age2
age
-0.0217 ** 15.69
-0.0320 **
19.0
-0.0221 ** 21.96
age*age
0.00016 ** 11.63
0.0004
**
25.8
0.0003 ** 29.00
-0.0541 ** 32.80
-0.0554 **
29.6
-0.0554 ** 49.28
0.0007
**
34.0
0.0006 ** 55.15
-0.0403 **
22.7
-0.0438 ** 41.01
0.0006
30.3
0.0005 ** 48.21
Usual suspects
age
age*age
0.0005 ** 29.20
-0.0600 ** 36.77
0.0006 ** 34.13
**
Kitchen sink
age
age*age
-0.0454 ** 25.64
-0.0311 **
16.0
-0.0350 ** 31.02
0.0005 ** 25.39
0.0005 **
24.0
0.0005 ** 39.43
176,770
Level of significance:
75, 529
153,886
When we add health and wealth to form the Usual suspects + health specification,
the U-shape is still evident and, like the HILDA, becomes shallower (the long dashed
line in Figure 3.9). The Age2 coefficient is 17% smaller and the Age coefficient is
20% less negative. Minimum predicted happiness increases by 0.05 (just 1%) from
4.95 at age 55 in the Usual suspects specification to 5.00 at age 55 with the Usual
suspects + health specification.
Next, the specification is expanded to add some of the variables from the Kitchen
sink specification: divorced, separated, partner dead, pregnant, and, just had a baby.
Unlike the GSOEP & the HILDA, we cannot add the other five life event shocks
because they are not in the BHPS data. Comparing the Kitchen sink prediction with
the others in Figure 3.9, we see that the U-shape has again decreased. Minimum
predicted happiness has increased 0.07 (1.5%) to 5.07 and now occurs earlier, at age
thirty-nine. We have evidence of a U-shape of happiness in age in all three data sets,
the GSOEP, the HILDA and the BHPS.
38
In summary, the findings from all three data sets are consistent with other scholars in
the field. When standard regressors are added a very strong U-shape effect emerges
with predicted age effects bigger than anything observable in the raw data. The
answer to this puzzle is now to be revealed using the GSOEP data. Then, robustness
is added to the GSOEP findings by replicating them using the HILDA and BHPS
data sets. The main regression with covariates I talk about in the remainder of this
chapter is the preferred specification of the Usual suspects because it has the
strongest U-shape in our chosen reference data set, the GSOEP. I now seek
explanations for the U-shape of happiness in age.
3.5
3.5.1a
A naive first-thought is that there is a particular issue with the early ages, i.e. age 18
to 22, and with high ages, i.e. those above 80. This is because the happiness decline
is particularly steep for the early years and erratic at the later years, which makes one
wonder if the young are being overly optimistic about their actual levels of happiness
and that the happiness of the very old is hard to tell from the few data points in that
range. To examine this possibility, Figure 3.10 shows the GSOEP results of the
regressions when we drop the under-22 year olds and those over 80 from the data
(about 9% of the panel). Table 3.15 (Chapter 3 Appendix B, p.87) shows the exact
specifications and detailed regression results and Table 3.6 (at the end of section
3.5.1c) summarises changes in the age and age2 coefficients as we change
specifications.
39
Raw average
life satisfaction
Life Satisfacition
7.5
Usual suspects
6.5
Us sus + Health
5.5
Kitchen sink
5
AGE
Figure 3.10: Life satisfaction in the GSOEP for the pooled sample for the mid-age range
The U-shape for the full sample and those aged 22 to 80 is almost identical.
Comparing the Usual suspects specification in Figures 3.5 & 3.10, minimum
predicted life satisfaction has increased just 0.07 (1%). There is no clear qualitative
difference between the results. Hence, the U-shape cannot be explained by the
extremities of the age range and must be due to relations in large parts of the age
range. However, is the result the same for the HILDA data set?
3.5.1b
When we drop the under-22 year olds and those over 80 from the HILDA (10% of
the panel data) we see (Figure 3.11) a very slight deepening of the U-shape of the
Usual suspects, Usual suspects + Health and Kitchen sink specifications. Table
3.19 (Chapter 3 Appendix B) shows the exact specifications and regression results
and Table 3.6 (at the end of section 3.5.1c) summarises changes in the age and age2
coefficients for each specification.
40
Life Satisfacition
9.00
8.50
Usual suspects
8.00
7.50
Usual suspects
+ Health
7.00
6.50
6.00
"Kitchen sink"
5.50
5.00
AGE
Figure 3.11: Life satisfaction in the HILDA for the pooled sample for the mid-age range
Comparing Figures 3.7 & 3.11, for our preferred Usual suspect specification,
minimum predicted happiness varied just 0.6% (7.63 for the full sample versus 7.58
for 22 to 80 year olds). There is no clear qualitative difference between the results
when we exclude the very young and the old from the HILDA. Hence, like the
GSOEP, the U-shape in the HILDA cannot be explained by the extremities of the age
range and must be due to relationships in large parts of the age range. Let us see if
the same holds for the third data set, the BHPS.
3.5.1c
When we similarly drop the under-22 year olds and those over 80 from the BHPS
(10% of the panel data) we see (Figure 3.12) a very slight deepening of the U-shape
of the Usual suspects, Usual suspects + Health and Kitchen sink specifications.
Table 3.23 (Chapter 3 Appendix B) shows the exact specifications and regression
results and Table 3.6 ( p.43) summarises changes in the age and age2 coefficients for
each specification.
41
Raw average
life satisfaction
Life Satisfacition
6.00
5.50
Usual suspects
5.00
4.50
Usual suspects
+ Health
4.00
"Kitchen sink"
3.50
3.00
AGE
Figure 3.12: Life satisfaction in the BHPS for the pooled sample for the mid-age range
The minimum still occurs around the same age (44 for the entire sample and 42 when
those below 22 and over 80 are excluded). For our preferred Usual suspect
specification, minimum predicted happiness varied just 2% (4.87 for the full sample
versus 4.77 for 22 to 80 year olds). This is larger than the variation in the GSOEP
and the HILDA but the U-shape is still clearly evident. When we exclude the very
young and the old, there appears to be no clear qualitative difference between the
BHPS results and those from the GSOEP and the HILDA.
Summarising, the age and age2 coefficients remain strongly significant across all
specifications with all three data sets (Table 3.6). Therefore, the U-shape evident in
all three data sets cannot be explained away by the extremities of the age range and
must be due to relationships in large parts of the age range. We now consider
explanation 2, and refocus back on the full sample and consider omitted variables
and reverse causality.
42
Table 3.6: Summary of changes in the GSOEP, HILDA & BHPS Age and Age2 coefficients as
controls are progressively added; for ages 22 to 80 years
Specification
Pooled OLS
Pooled OLS
Pooled OLS
(Ages 22 to 80)
(Ages 22 to 80)
(Ages 22 to 80)
GSOEP
HILDA
BHPS
coefficient
t-value coefficient
t-value
coefficient
t-value
Age + Age2
age
-0.0306** 15.46
-0.0417 ** 17.50
-0.0373 ** 26.21
0.0003** 13.89
0.0006 ** 23.45
0.0005 ** 32.70
age
0.0745** 33.91
-0.0680** 26.69
-0.0772 ** 51.31
age*age
0.0008** 32.30
0.0008** 31.09
0.0009 ** 58.50
-0.0768** 35.27
-0.0496** 20.48
-0.0620 ** 43.38
0.0008** 33.73
0.0007** 26.72
0.0007 ** 50.57
-0.0618** 26.75
-0.0386** 14.91
0.0521 ** 34.94
0.0007** 27.03
0.0006** 21.61
0.0007 ** 42.99
age*age
Usual suspects
Kitchen sink
age
age*age
160,332
Level of significance:
65,679
43
138,481
3.5.2a
heterogeneity
An important finding in the literature is that happiness is strongly affected by stable
personality traits (Argyle, Kahneman, Diener, & Schwarz, 1999; Ferrer-i-Carbonell
& Frijters, 2004; Frey & Stutzer, 2002). These fixed individual traits are usually part
of the error term. A stylised finding from both the economic and the psychological
literature is that accounting for fixed traits has a very strong impact on the
coefficients found for socio-economic variables (Ferrer-i-Carbonell & Frijters, 2004;
Clark, Frijters, & Shields, 2008). A leading explanation for this is the possibility of
reverse causality arising from unobserved heterogeneity. For example, the individual
personality traits that make you happier (Lyubomirsky, et al., 2005) also make it
more likely that you will have a higher income, a job, a partner, better health, greater
wealth, and a higher level of education.
Could the problem of reverse causality caused by unobserved fixed traits explain
something about the U-shape? At first glance, one would think not because fixed
personality traits are by design uncorrelated with age and it seems unrealistic to
suggest that happiness causes age. However, the personality traits that affect
happiness can be correlated with variables that are correlated with age, such as
income, a job, a partner, good health and wealth. How would this work? Consider the
problem in its simplest form. Suppose for the purposes of this subsection the truth is
that the following relationship holds
f i ageit2 , cov( f i , xit ) > 0, cov( ageit , xit ) > 0, E [u it | ageit , xit , f i ] = 0
where we have for simplicity subsumed a linear age term into xit and all variables are
normalised to have expectation 0 implying there is no constant term either; there are
individual fixed traits f i unrelated to age-squared but related to a composite timevarying socio-economic variable called xit . There is an error term uit orthogonal to
everything else. What are now the estimated coefficients if we mistakenly run a
regression without accounting for fixed-effects? The asymptotic values are,
44
cov( f i , xit )
cov(ageit2 , xit )
+ (1 lim 1 )
var( xit )
var( xit )
2
cov( xit , ageit )
lim 1 = 1 + ( lim )
var(ageit )
lim = +
which shows that even though ageit2 is not correlated with the omitted fixed effect,
the coefficient on ageit2 can nevertheless be biased when it is related to included
time-varying variables that are correlated with the omitted fixed-effect. The
equations become rather elaborate if we add a linear age term and a constant but the
basic principle remains that a bias in the age-term can occur if the added variables
are correlated with age and with the omitted fixed-effect.
Intuitively, there are two steps in the possible emergence of the bias. The first is that,
as shown just above, the inclusion of fixed effects will change the coefficients of the
non-age variables xit. The second is that xit itself changes systematically with agesquared, which leads to a bias in the estimated coefficient of age-squared. To explore
this possibility I run fixed-effect analyses on the GSOEP to see how this changes the
U-shape findings. Figure 3.13 depicts the fixed-effect regressions, the specifications
are in Table 3.16 (Chapter 3 Appendix B. p.88) and Table 3.7 provides a summary of
changes in the age and age2 coefficients across specifications.
Can reverse causality explain the U-shape?
8
Raw average
life satisfaction
Life Satisfacition
7.5
Us sus + health
of pooled
6.5
Us sus +health
of fixed effect
5.5
age+age2 of
fixed effect
5
AGE
Figure 3.13: Life satisfaction in the GSOEP for the balanced panel
45
The results for Figure 3.13 are both confirming and surprising. The graph shows the
raw relationship between age and happiness with three overlaid three lines. The Ushaped line is the same one depicted previously and is the pooled regression with the
Usual suspects + health specification. Overlaid are two lines from fixed-effect
regressions. The thick dark dashed line is the result of running the same regression as
for the pooled regression but including fixed effects. As one can see, the U-shape
completely disappears, i.e. the age-squared coefficient becomes tiny and insignificant
(Table 3.7). It however replaces the U-shape by a similarly puzzling effect, which is
a very strongly significant negative linear relationship. The third thin solid line,
which shows the result of just running a fixed effect regression with only age and
age-squared as regressors, confirms this. The U-shape slightly reverses into an
inverted U shape, but a very strong negative age relationship emerges (Table 3.7).
Table 3.7: Summary of changes in the GSOEP Age and Age2 coefficients as controls are
progressively added; with fixed effects
Fixed Effect
(All)
GSOEP
Specification
coefficient
t-value
Age + Age2
age
-0.0166**
6.60
age*age
-0.0003**
9.71
-0.00328**
11.61
-0.0001*
2.22
age
-0.0298**
9.95
age*age
0.00006
0.74
-0.0184**
5.81
Usual suspects
age
age*age
Usual suspects + health
Kitchen sink
age
age*age
N
-0.00002*
2.40
176,770
Level of significance:
46
Before I turn to explain the new puzzle of the strong negative linear relationship in
the Section 3.7, I first want to confirm that the disappearance of the U-shape is
indeed because of reverse causality. To begin, Table 3.14 (OLS) and Table 3.16
(Fixed effects) in the Chapter 3 Appendix B reflect the findings of many other
studies (Ferrer-i-Carbonell, 2005): the coefficients of most socio-economic variables
become much smaller when one adds fixed effects. The income coefficient drops by
45% and the importance of marriage, a job, health and education all reduce.
However, do these variables correlate with age?
Figure 3.14 depicts changes in lifetime happiness for the four most significant socioeconomic variables from the regression of the Usual suspects + health
specification.
0.8
ln Household
income;
mean = 8.20;
sd = 0.53
0.3
Employed;
mean = 0.47;
sd = 0.50
-0.2
Education;
mean =
10.93; sd =
2.46
-0.7
Health
(inverse);
mean = 2.20;
sd = 0.95
-1.2
AGE
Figure 3.14: Age and observed correlates in the GSOEP
Figure 3.14 indeed shows a strong relationship between age and employment (reverse
U-shaped), education (reverse J-shaped), and health (which has a reverse U-shape in
the main age range between 30 and 50). Household income declines in age with an
inverted U-shape in the middle age range. The relationships all go in the direction we
anticipated above: an artificially high coefficient for employment, education, income
and health would all give rise to a false U-shape in age. However, is there evidence
that reverse causality explains the U-shape of happiness in age in our other data sets?
47
3.5.2b
heterogeneity
When we run the same fixed-effect analyses with the HILDA, we get similar
results31. Figure 3.15 shows the raw relationship between age and happiness and has
three overlaid lines.
Raw
average life
satisfaction
9.50
Life Satisfacition
9.00
8.50
Usual
suspects +
health of
pooled
8.00
7.50
Usual
suspects +
health of
fixed effect
7.00
6.50
age+age2 of
fixed effect
6.00
5.50
5.00
AGE
Figure 3.15: Life satisfaction in the HILDA for the balanced panel
The thin-dashed U-shape line is the same as we showed previously, the pooled OLS
regression with the preferred Usual suspects + health specification. The other two
lines are from the fixed-effect regressions. The thick-dashed line is the result of
running a fixed effect regression. Like the GSOEP, the HILDA age2 coefficient goes
non-significant (.00007, t-value 0.94). Just like the GSOEP, the U-shape disappears
(Table 3.8).
31
48
Table 3.8: Summary of changes in the GSOEP and HILDA Age and Age2 coefficients as controls
are progressively added; with fixed effects
Specification
Fixed Effect
Fixed Effect
(All)
(All)
GSOEP
HILDA
coefficient
t-value
coefficient
t-value
Age + Age2
age
-0.0166**
6.60
-0.0076
1.02
age*age
-0.0003**
9.71
-0.0001
1.30
-0.00328**
11.61
-0.0173*
2.12
-0.0001*
2.22
0.00001
0.17
age
-0.0298**
9.95
-0.0202*
2.50
age*age
0.00006
0.74
0.00007
0.94
-0.0184**
5.81
-0.0033
0.40
2.40
-0.00006
0.72
Usual suspects
age
age*age
Usual suspects + health
Kitchen sink
age
age*age
N
-0.00002*
176,770
Level of significance:
75,529
Like the GSOEP regression results, the coefficients of the socio-economic variables
from the HILDA are also much smaller with fixed effects. With the GSOEP data, the
income coefficient dropped by more than 40% and the importance of marriage, a job,
health and education all reduced. With the HILDA data, the income coefficient
reduction is similar to the GSOEP, 42%. In addition, the health coefficient dropped
49% from the 0.53 in the pooled regression to 0.27 with fixed effect32. Like the
GSOEP, the HILDA variables correlate with age. Figure 3.16 shows the simple
averages by age of the four most significant socio-economic variables in the
regressions.
32
Comparing the HILDA OLS results in Table 3.17 with the HILDA fixed-effect results in Table 3.20
of Chapter 3 Appendix B.
49
ln Household
income;
mean = 10.77;
sd = 1.07
0.5
Employed;
mean = 0.66;
sd = 0.47
-0.5
Education;
mean = 12.82;
sd = 1.80
-1
Health
(inverse);
mean = 2.64;
sd = 0.95
-1.5
AGE
Figure 3.16: Age and observed correlates in the HILDA
Like the GSOEP (Figure 3.14), the HILDA Figure 3.16 shows a strong relationship
between age and employment (reverse U-shape), education (reverse J-shape), and
health (which consistently declines slowly over time). Like the GSOEP, household
income declines in age with an inverted U-shape in the middle age range. Overall,
the relationships all go in the same direction as we anticipated and saw in the
GSOEP. An artificially high coefficient for employment, education, income and
health would all give rise to a false U-shape in age. Consistent with the GSOEP
findings in Chapter 3, the U-shape is considered to arise from reverse causality with
the included covariates. However, does this hold for the BHPS?
50
explanation
2:
it
is
also
about
unobserved
heterogeneity
I again run the same fixed-effect analyses on the BHPS as we ran on the GSOEP and
the HILDA. We get similar results33. As before, Figure 3.17 shows the raw
relationship between age and happiness and has three overlaid lines.
Raw
average
life
satisfaction
Life Satisfacition
6.00
Usual
suspects +
health of
pooled
5.50
5.00
Usual
suspects +
health of
fixed effect
4.50
age+age2
of fixed
effect
4.00
3.50
AGE
Figure 3.17: Life satisfaction in the BHPS for the balanced panel
The thin-dashed U-shape line is the same as we showed previously, the pooled OLS
regression with the preferred Usual suspects + health specification. The other two
lines are from the fixed-effect regressions. The thick-dashed line is the result of
running a fixed effect regression on our preferred Usual suspects + health
specification; the U-shape disappears. Like the GSOEP and the HILDA, the BHPS
age2 coefficient is also non-significant (.000008, t-value 0.27). The results with the
BHPS panel data are consistent with those from the GSOEP and the HILDA data; the
U-shape disappears (Table 3.9).
33
51
Table 3.9: Summary of changes in the GSOEP, HILDA & BHPS Age and Age2 coefficients as
controls are progressively added; with fixed effects
Specification
Fixed Effect
Fixed Effect
Fixed Effect
(All)
(All)
(All)
GSOEP
HILDA
BHPS
coefficient
t-value
coefficient
t-value
coefficient
t-value
Age + Age2
age
-0.0166**
6.60
-0.0076
1.02
0.0044
1.42
age*age
-0.0003**
9.71
-0.0001
1.30
-0.0001**
4.61
-0.00328**
11.61
-0.0173*
2.12
-0.0064+
1.88
-0.0001*
2.22
0.00001
0.17
-0.0001+
1.75
age
-0.0298**
9.95
-0.0202*
2.50
-0.0127**
3.86
age*age
0.00006
0.74
0.00007
0.94
0.00008
0.27
-0.0184**
5.81
-0.0033
0.40
-0.0082*
2.49
2.40
-0.00006
0.72
Usual suspects
age
age*age
Usual suspects + health
Kitchen sink
age
age*age
N
-0.00002*
176,770
Level of significance:
0.000008
75,529
0.31
153,886
Like the GSOEP and the HILDA regression results, the coefficients of the socioeconomic variables from the BHPS are also much smaller with fixed effects34. With
the GSOEP data, the income coefficient dropped by 45% (40% with the HILDA) and
the importance of marriage, a job, health and education all fell in both data sets. With
the BHPS data, the income coefficient reduction is 58%. Like the GSOEP and the
HILDA, GSOEP variables correlate with age. Figure 18 shows the simple averages
by age of the four most significant socio-economic variables in the BHPS
regressions. The correlation of the changes in lifetime happiness with the four most
significant socio-economic variables in the GSOEP regressions can also be seen with
the BHPS and the HILDA (Figures 3.18 & 3.19a, b, c & d).
34
Comparing the BHPS OLS results in Table 3.22 with the BHPS Fixed-effect results in Table 3.24 of
Chapter 3 Appendix B.
52
Life
satisfaction;
mean = 5.23;
sd = 1.29
0.10
ln Household
income;
mean = 9.84;
sd = 1.61
Employed;
mean = 0.58;
sd = 0.49
-0.40
-0.90
Education;
mean =
13.26; sd =
2.40
-1.40
Health
(inverse);
mean = 2.20;
sd = 0.95
AGE
Figure 3.18: Age and observed correlates in the BHPS
53
BHPS: life
satisfaction
90
87
84
81
78
75
72
69
63
66
60
57
GSOEP:
life
satisfaction
-0.90
HILDA:
household
income
GSOEP:
household
income
-1.40
HILDA:
education
-0.40
90
87
84
78
75
72
69
66
81
90
87
84
81
78
75
72
69
66
63
60
57
54
48
45
42
39
36
33
30
27
GSOEP:
life
satisfaction
HILDA: life
satisfaction
0.10
24
90
87
84
81
78
75
69
72
66
63
60
57
54
51
48
45
42
39
36
33
30
27
24
HILDA: life
satisfaction
BHPS:
health
0.60
21
0.10
BHPS: life
satisfaction
1.10
18
BHPS:
education
0.30
21
63
0.50
18
60
GSOEP:
employed
AGE
0.70
-0.70
57
HILDA:
employed
-1.45
BHPS: life
satisfaction
-0.50
54
48
-0.95
-0.30
51
42
45
39
36
33
30
27
GSOEP:
life
satisfaction
AGE
-0.10
HILDA: life
satisfaction
-0.45
51
54
51
48
45
42
36
39
33
30
27
24
21
-0.40
0.05
24
HILDA: life
satisfaction
0.10
BHPS:
employed
21
BHPS:
household
income
BHPS: life
satisfaction
0.55
18
0.60
18
GSOEP:
life
satisfaction
HILDA:
health
-0.90
GSOEP:
education
-1.10
GSOEP:
health
-0.90
AGE
AGE
Figure 3.19: Comparison of Age and individual observed correlates across data sets
54
Whilst a solution has been found to the original puzzle, i.e. the U-shape is an artefact
of unobserved heterogeneity and reverse causality with the included covariates, let us
consider more deeply how the unobserved heterogeneity could bias the pooled
results.
3.5.3 How does the unobserved heterogeneity bias the pooled results?
The mechanism hypothesised in the previous sub-section was that fixed traits lead to
a reverse causality between variables and life-satisfaction. For example, individuals
have high incomes and get married partially because they have high levels of
happiness. The biases in the coefficients of these reverse causality variables would
lead to a bias in the age profile because those variables change systematically with
age. Let us see if we can confirm whether those mechanisms are visible in the data.
Two steps lead to the emergence of bias in the age-coefficients. The first is whether
the coefficients of other variables change when fixed-effects are included. Table 3.10
summarises the estimates of particular coefficients with and without fixed-effects.
The variables shown are those often applied in economic research: employment,
unemployment, marriage, income, and education. These are also the most significant
variables in the usual suspects specification.
Looking at Table 3.10, we can see large changes in coefficients for all three datasets
when fixed-effects are included. For income, the coefficient drops 37% in the
GSOEP (0.28 in fixed-effects compared to 0.44 in the pooled regressions), 40% in
the HILDA and 58% in the BHPS. For marriage, the coefficient drops 16% in the
GSOEP (0.25 in fixed-effects compared to 0.29 in the pooled regressions), 32% in
the HILDA and 55% in the BHPS. Interestingly, the absolute coefficients of all these
five variables reduce in all three datasets when including fixed effects. There is a
clear change in the coefficients of variables from pooled to fixed-effects.
55
Table 3.10: Coefficients for the key 5 variables (pooled & fixed effects) for the three data sets.
OLS
GSOEP
Specification
Age + Age
coefficient
HILDA
t-value
coefficient
BHPS
t-value
coefficient
GSOEP
t-value
coefficient
HILDA
BHPS
t-value
coefficient
t-value
coefficient
t-value
age
-0.0217**
15.69
-0.0320**
19.0
-0.0221**
21.96
-0.0166**
6.60
-0.0076
1.02
0.0044
1.42
age*age
0.00016**
11.63
0.0004**
25.8
0.0003**
29.00
-0.0003**
9.71
-0.0001
1.30
-0.0001**
4.61
-0.0541**
32.80
-0.0554**
29.6
-0.0554**
49.28
-0.00328**
11.61
-0.0173*
2.12
-0.0064
1.88
age*age
0.0005**
29.20
0.0007**
34.0
0.0006**
55.15
-0.0001**
2.22
0.00001
0.17
-0.0001
1.75
income
0.4619**
52.2
0.0805**
15.00
0.0943**
19.44
0.2414**
23.13
0.0257**
4.26
0.0164**
employed
0.0650**
4.8
0.1355**
9.27
0.2397**
27.19
0.0991**
6.69
0.0536*
2.33
0.0618**
4.63
married
0.3106**
27.9
0.4429**
36.20
-0.0033*
2.27
0.2385**
14.95
0.2056
**
6.76
0.3567**
46.48
-0.0600**
36.77
-0.0403**
22.7
-0.0438**
41.01
-0.0298**
9.95
-0.0202*
2.50
-0.0127**
3.86
0.00006
0.74
0.00007
0.94
0.00008
0.27
Usual suspects
age
Usual suspects
+ health
age
age*age
0.0006**
34.13
0.0006**
30.3
0.0005**
48.21
income
0.4420**
45.91
0.0329**
4.77
0.0399**
5.47
0.2750**
employed
0.0791**
7.60
-0.1338**
9.20
0.0295*
2.02
0.1001**
6.75
-0.0238
married
0.2915**
26.50
0.3624**
30.32
-0.3063**
41.48
0.2457**
15.42
0.1963
-0.0454**
25.64
-0.0311**
16.0
-0.0350**
31.02
-0.0184**
5.81
-0.0033
-0.00006
23.67
0.0191**
**
2.64
3.22
0.0166**
2.7
1.06
0.0355**
2.77
6.52
0.1390**
8.47
Kitchen sink
age
0.40
-0.0082*
2.49
age*age
0.0005**
25.39
0.0005**
24.0
0.0005**
39.43
-0.00002**
2.40
0.72
0.000008
0.31
income
0.4307**
44.53
0.0309**
5.69
0.0251**
5.25
0.2585**
22.15
0.0151*
2.55
0.0060
0.98
employed
0.0688**
5.11
-0.1166**
8.00
0.0372**
4.38
0.0925**
6.24
0.0394
1.75
0.0365**
2.86
married
0.1180**
7.65
0.2514**
14.85
-0.0195**
14.27
0.0327
1.37
-0.1095**
2.84
0.2014**
22.43
N=
Level of significance: + p < 0 .1
176,770
75,529
153,886
56
176,770
75,529
153,886
The second step is to see if changes in the coefficients of these non-age variables
lead to a difference in the predicted age-profile. The clearest way to depict if this
occurs is to show the predicted effect of all non-age variables in the pooled
regressions versus the fixed-effects regressions.
prediction lines are shown for our three datasets. The first is from the usual
suspects regressions that do not include fixed effects (column 3 of Tables 3.14, 3.18,
and 3.22 in the Appendix B), and the second from the usual suspects regressions
that do include fixed effects (column 3 of Tables 3.16, 3.20, and 3.24 in the Chapter
3 Appendix B). In all cases, we let the prediction lines start at the same point at age
18 to aid the interpretation.35
AGE
OLS
Figure 3.20: Predicted happiness effects of the non-age variables in the GSOEP
35
The model is Life Satisfaction = age * age + age2 * age2 +xitx+eit where eit is the error term that
either includes fixed-effects or not. The prediction lines show the average over i of xitx by age, which
uses the fact, that xit changes by age.
57
AGE
OLS
Figure 3.21: Predicted happiness effects of the non-age variables in the HILDA
AGE
OLS
Figure 3.22: Predicted happiness effects of the non-age variables in the BHPS
Looking at the results for the GSOEP in Figure 3.20 first, the main thing to note is
that the predicted OLS line looks very much like an inverted U-shape: the increase
from age 18 to the top at age 48 is about 0.32 and the subsequent decrease to age 80
is about 0.9. Since the regression coefficients of age-and age-square essentially try to
58
fit the actual age-profiles conditional on this predicted effect from the non-age
variables, this inverted U-shape forces a finding of a U-shape in the age coefficients.
When including fixed-effects one can see that the inverted U-shape is much less
pronounced (though not entirely gone): the increase from age 18 to the top is about
0.22 and the subsequent decrease is 0.54. This reduction in the predicted inverted Ushape from the non-age variables in turn will lead to a reduction in the U-shape
found for age when including fixed-effects.
Qualitatively, the same results appear for the HILDA (Figure 3.21). The reduction in
the predicted happiness contribution of the non-age variables from the top to age 90
is 0.5 with the pooled regression results and only 0.35 with the fixed-effects. Just like
the GSOEP and the HILDA, the results from the BHPS (Figure 3.22) show the
inverted U-shape of the happiness contribution of the non-age variables is much
stronger without fixed-effects than with fixed-effects. Both the upswing and the
downswing are more pronounced. Summarising, we can indeed see that the inclusion
of fixed-effects reduces the coefficients of variables that themselves systematically
vary by age (incomes and marriage peak in middle age) and that this in turn reduces
the predicted inverted U-profile of their effects.
59
The positive peak is at 70 years for the BHPS & GSOEP and age 65 in the HILDA.
Perhaps the results vary because Clark only used age bands up to age 65.
60
who were already happy. In all three data sets, this reverse causality shows up in
cross-sections as inflated coefficients for income, marriage, and getting a job. In
order to fit the actual age profile of happiness, the bias in coefficients for socioeconomic variables forces the predicted age profile to become U-shaped. When one
controls for fixed-effects, the non-linearity all but disappears for all three data sets.
The bottom line is that the supposed happiness decline in middle age is far less of a
real finding than has been proposed and that it is not the most prominent age-related
feature of either the raw happiness data or the results of fixed-effect regressions. The
raw data in Germany and Britain is much more supportive of a wave-pattern in
happiness (a happiness peak around the age of 70), whilst the main finding from
fixed-effects regressions is a large and steady decrease in happiness as people get
old. The reasons for such a happiness decline in panel datasets needs further study.
3.6
3.6.1 Explanation for the negative slope I: time and cohort effects
Before studying why happiness changes over the lifetime in Chapters 4 and 5, I
pursue an additional research question that arose from the more immediate puzzle
that emerged from the analysis of the GSOEP data. This new puzzle is the strong
negative relationship between age and happiness over time in fixed-effects
regressions even though the relationship in the pooled cross-sectional data is much
less pronounced. Where does this strong negative slope come from? We have seen
that it is not due to any other included variable because it remains when we do not
include other variables. We know for the same reason that it is not due to reverse
causality. It also cannot be due to some simple missing variable, like average
income, because that would go in the opposite direction (average income rose). What
candidate explanations remain?
Recall the details, a strong negative relationship emerged between age and happiness
over time in fixed-effects regressions with the GSOEP data (Section 3.5.2a). The Ushape of happiness in age reversed into an inverted U shape, and, a very strong
negative relationship could be seen (Figure 3.13, p.45); While GSOEP age2
61
coefficient went non-significant with fixed effects; the age coefficient became
strongly significant across all specifications (Table 3.7, p.46) 38.
A popular explanation in the recent literature is that there are important time and
cohort effects on happiness (Cribier, 2005). Right at the outset, one should point out
that such explanations are somewhat unsatisfactory because both time and cohort
effects are in a sense aggregate unobservables. For instance, a cohort effect is just a
missing aggregate variable specific to an age group, where we do not know what the
missing variable is. The missing variable could be the mental experience from a
particular traumatic event or the effect of a particular diet in a certain era, or some
cultural trait particular to an era (like expectations), etc. It would be preferable to
measure the supposed elements making up a cohort effect before becoming
convinced cohort effects actually exist. We also need to know what the cohort effects
consist of if we are to make any policy-relevant inferences about how happy cohort
effects might be created.
Another problematic aspect of the notion of cohort-effects and time-effects is that
they are statistically difficult to identify. It has been known for a long time that it is
not possible to simultaneously identify age, year, and cohort effects because one can
be written as a function of the other two. Only by introducing somewhat arbitrary
restrictions on cohort effects can they be separated from age and time-effects.
What we can do is to accept that we cannot tease apart the age and cohort effects and
simply presume that the effects of age pick up cohort effects. One can then divide the
age distribution into different cohorts and label the effect of being born in a
particular interval as due to a cohort effect. Blanchflower & Oswald (2007) for
instance define cohorts by age-intervals, i.e. 10-year intervals, and label the effects of
these age-intervals as cohorts. This practically means that there is a sharp dividing
line in the influence of whatever causes a cohort effect on particular days in the
century. Someone born on January 5, 1910 for instance could be in one cohort, and
38
Looking back at the summary fixed effects regression results (Table 3.9) on page 52 we can see a
very large and significant age coefficient (-0.0166, t-value 6.60) with the German (GSOEP) data. We
do not see this with the Australia and British data, the age coefficients are very small and nonsignificant.
62
someone born the day after in another. If one does not include age as a regressor, this
procedure is akin to using a semi-parametric function of age (Clark, 2006). However,
if one then also adds linear and quadratic age variables to a regression containing
these age-intervals and proceeds with assuming that the age-intervals pick up
something very different from age effects, then these arbitrary age-intervals become
binding: the assumption that cohort effects jump at particular ages is then what
separately identifies age effects from time and cohort effects.
So far, I have ignored the possibility of cohort or time effects. As far as there are
linear time effects, then these would be indistinguishable in a fixed-effect framework
from age effects. This is because we can write ageit = agei 0 + t where the first
terms ( agei 0 ) is fixed and absorbed by the fixed-effect term, whilst, the second term
(t) is a straightforward time effect.
In order to ascertain whether there are likely to be time or cohort effects, I next look
at the evolution over time of aggregate life satisfaction of the complete pooled panel.
If there are strong time or cohort effects capable of explaining the large decline that
we saw under the fixed-effects regressions, then we should see such a decline in the
aggregate data.
7.5
7.4
7.3
7.2
7.1
6.9
6.8
6.7
1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002
Year
Figure 3.23: Year and life satisfaction in the GSOEP for the pooled sample
63
Figure 3.21 shows the evolution over time for the GSOEP. As we can see, there is
indeed a strong decline. Bearing in mind that the standard deviation of the mean life
satisfaction in a year with so many observations (10,000 per year) is less than 0.02,
the year on year changes are highly significant. The overall decline also fits
somewhat, though it is not quite enough: in 19 years, the aggregate drop is only 0.4
whilst the drop predicted by the fixed-effect regression is about 0.6 for 19 years.
There might hence be cohort or time effects39 responsible for the found drop by age,
but the predicted future drop in satisfaction would be enormous, i.e. at current trends
a predicted drop by more than 1 point in the next 30 years. In addition, this predicted
drop does not tally with what we know from other surveys (like the Eurobarometer
Survey), where it has been found that aggregate life satisfaction is quite constant in
Western countries over time, including Germany40. There is hence still something not
quite right about this explanation because whilst it fits the GSOEP as a whole, it
seems to violate what we know to hold at the aggregate for happiness cross-sections.
What other explanation is left?
3.6.2 Explanation for the negative slope II: is something wrong with the
panel data
Forced to reject or seriously doubt all other reasonable explanations, I now turn to
the most uncomfortable potential explanation, which is that there is something wrong
with repeat happiness responses in the GSOEP panel data. What if individuals who
remain in the panel, and, who keep answering the survey, are different from those
who drop out of the panel or do not keep answering the satisfaction question?
The GSOEP has a large numbers of dropouts each year. For instance, of the roughly
10,000 individuals in the original 1984 GSOEP sample, only about 4,000 remain in
our data for the full 19 years. The GSOEP replaces those who no longer answer with
39
Time and cohort effects cannot be meaningfully separated if we also include non-linear age
effects.
40
Clark et al. (2008) show that life-satisfaction profiles have been virtually flat in cross-sections in the
last 30 years for Germany, France the US, and many other Western countries.
64
e
t
it
with eit being i.i.d. shocks, then a high (close to 1) would indicate a strong
persistence in shocks and the selection we would worry about is on eit . If, wave after
wave, it is the case that individuals are more likely to stay another year in the
GSOEP when eit is lower, then we would indeed be able to get the strong negative
age slope observed in the fixed-effects regression.
How can we verify this possibility? We verify it by comparing the answers of those
who stay in the panel with those who enter for the first time. Every year, there are
several thousand new entrants in the GSOEP who have never answered the
questionnaire before. Some of these are new samples, some are partners of regular
respondents, and some are children becoming old enough to be in the sample. What
65
is the relationship between age and life satisfaction for them, and, what does
aggregate satisfaction over the years look like for them?
Raw average
life satisfaction
total sample
Life Satisfacition
7.5
6.5
Raw of firsttimers
5.5
age+age2 of
full sample
5
AGE
Surprisingly: yes! No significant increase for first-time entrants with a very significant decrease for
all!
7.8
7.6
Life sat of first-time in
GSOEP
7.4
7.2
Life sat of all in GSOEP
6.8
6.6
6.4
Year
Figure 3.24 (top) and 3.24 (bottom): life satisfaction in the GSOEP for first-time respondents
66
Figure 3.24 (top) shows the aggregate level of satisfaction by age of those who
answer for the first time, as well as for the entire pool. As we can see, there is still a
decline in happiness at a very young age and at very old ages (which has very few
observations and therefore looks erratic), but there is no decline at all from age 25 to
75. If we then overlay the predicted regression lines, the predicted regression line for
the whole sample shows the horizontal j-shape, but the predicted regression line for
the first-time panel entrants is almost flat. Not quite flat, because there is still a
significantly negative age trend, but the coefficient is about 85% smaller than that for
the full sample. When adding a quadratic term all age effects become insignificant
for the new entrants (the exact specifications are in Table 3.16 (Chapter 3 Appendix
B). Hence, there is a small age effect on happiness, concentrated at the very young
and the very old, but it is not U-shaped. Rather, it is simply a decline.
Figure 3.24 (bottom) confirms the impression of the top graph: when looking at the
average satisfaction over the years for first-time respondents, there is no time profile
of happiness anymore, in line with what is found for cross-sectional studies which by
design only question individuals once. This is consistent with the notion that there
are no age, time, or cohort effects, or that they at least cancel each other out. The
decline in the aggregate-panel is thus most likely due to selection on time-varying
unobservables. How bad is the decline in satisfaction for repeat respondents? Finally,
Figure 3.25 shows aggregate satisfaction depending on the number of years someone
has answered the questionnaire.
67
7.6
7.4
7.2
6.8
6.6
6.4
1
10
11
12
13
14
15
16
17
18
19
Figure 3.25: The degree of selection in the GSOEP for stayers in the panel
Figure 3.25 confirms the running hypothesis: there is a large decline in reported
satisfaction as an individual is in the panel for longer. We now know this is not age
related or related to observables, but is a selection on unobservables. Given that there
is no discernible bounce-back, it furthermore has to be a relatively persistent timevarying unobservables responsible for this decline. The decline is large enough to
explain the fixed-effect profile: in eighteen years, there is a 0.64 reduction in life
satisfaction41, which translates to a decline of 2.2 over 60 years, almost exactly the
predicted amount from the fixed-effect regression.
41
This 2.2 predicted reduction in happiness over a lifetime is an out of (age) range forecast; the 19
years of the GSOEP panel are assumed to be representative of a typical lifetime.
68
related variables in the panel, there is no reliable outside information with which to
check the panels selection. The one negative shock for which there is some
information is divorce rates. According to UN statistics, the yearly divorce rate
(official annulments of registered marriages) per 1000 is 5.2 in Germany42. If we take
into account that the panel does not contain individuals below 18 (the 0-17 year olds
make up some 18% in Germany), then this would imply that divorce rates should be
6.2 per 1000 in the GSOEP. As it is, the yearly self-reported divorce rate in the
GSOEP is about right, i.e. about 6 per 1000 people on average and rising over time in
the panel. Hence, self-reported divorce rates are not higher for the sample than for
the population. This means we cannot claim firm outside evidence of negative
dynamic selection.
42
The UN numbers refer to the number of granted annulments. Since one annulment affects the
marital status of two people, we have doubled this to arrive at the number of individuals getting
divorced.
69
would eventually have to be some bottoming out of the happiness decline for
people who stay in the panel. This is not noticeable in the graph above. Whilst the
interpretation that we are looking at more truthful responses over time is probably
more soothing to the collectors of the GSOEP because it takes away the suspicion
that the representativeness of the panel is in doubt. The implications for the
happiness literature of this second interpretation is worse than the first interpretation.
If it were just the case that panels suffer from non-random attrition based on timevarying unobservable happiness determinants then there is still a lot of useful
information in cross-sectional surveys and one can hope that the selection does not
involve cross-terms between observable time varying variables and unobservable
ones, allowing useful interpretations of panel coefficients.
If it is, alternatively, the case that we cannot trust the responses on happiness
questions of the first 19 years of responses, then we effectively cannot trust over 99%
of the data in this field. In addition, the changes over time are big, certainly big
enough to get seriously worried: an overstatement of happiness by 0.7 could drop a
country for instance from being the happiest country in the world to being one of the
unhappier countries in the world, nullifying the validity of all the rank-tables. Now,
of course, if the overstatement is a universal constant and does not differ by
country, then the implications are less strong but the hypothesis of a universal
constant overstatement cannot be claimed a priori. If a possible tendency to overstate
is furthermore correlated with observables, then cross-sectional and panel analyses
all become highly suspect. Alternatively, is this just an artefact of the GSOEP? To
test this, we look for cohort and time effects in the BHPS and the HILDA data sets.
3.7
With the GSOEP data (Section 3.5.2a), the U-shape of happiness in age reversed into
an inverted U shape and a very strong negative relationship emerged (Figure 3.13);
The GSOEP age2 coefficient went non-significant with fixed effects, the age
coefficient became strongly significant across all specifications (Table 3.7). This
occurred to a much lesser extent with the HILDA, the (thin solid) line of the age +
age2 fixed-effect specification in Figure 3.15 only exhibits a very slight negative
relationship. Like the GSOEP, the HILDA age2 coefficient does go non-significant
70
with fixed effects, but the age coefficient was not as strongly significant across all
specifications (Table 3.8). Therefore, if our GSOEP explanation is to hold, we should
not see strong time or cohort effects with the HILDA data because the large decline
we saw under the fixed-effects regressions with the GSOEP are not evident in the
HILDA data; we should not see a decline in the aggregate HILDA data.
8.1
7.9
Life satisfaction
of all in HILDA
7.8
7.7
7.6
7.5
2002
2003
2004
2005
2006
2007
2008
Year
Figure 3.26: Year and life satisfaction in the HILDA for the pooled sample
With the GSOEP, cohort effects were clearly visible (Figure 3.23); we do not see a
similar decline in the aggregate HILDA data (Figure 3.36). The GSOEP had a 0.4
drop in happiness over the 19-year panel, the HILDA only exhibits a very small
(0.08) happiness increase in 2002-3 but the life satisfaction level across the whole
panel period is almost the same. There appears to be little if any time or cohort
effects in the HILDA. If so, we should not see large differences between the first
timers and those who remain the HILDA panel.
71
Raw average
life satisfaction
total sample
Life Satisfacition
9.50
9.00
8.50
age + age2 of
first-timers
8.00
7.50
7.00
Raw average
life satisfaction
of first-timers
6.50
6.00
age+age2 of
full sample
5.50
5.00
AGE
8.1
Life satisfaction
of first-timers in
HILDA
7.9
7.8
Life satisfaction
of all in HILDA
7.7
7.6
7.5
2002
2003
2004
2005
2006
2007
2008
Year
Figure 3.27 (top) and 3.25 (bottom): life satisfaction in the HILDA for first-time respondents
Figure 3.27 (top) shows the aggregate level of satisfaction by age of those who
answer for the first time, as well as for the entire HILDA pool. The raw average
lifetime satisfaction of first timers (thin dashed line) exhibits lower volatility than the
GSOEP, and, unlike the GSOEP (Figure 3.22), the time effects (age + age2) of the
72
HILDA stayers and first timers overlay one another43. Figure 3.27 (bottom) confirms
the impression of the top graph: when we look at the average satisfaction over the
years for first-time respondents, there is only a very small change in life satisfaction
over the 8-year panel: 0.13, less than 10% of a standard deviation. We do not see
large differences between the first timers and those who remain the HILDA panel.
As a final check that time and cohort effects do not strongly manifest in the HILDA,
we look for the evidence that we found with the GSOEP data; how much does life
satisfaction decline with years in the panel.
8.10
8.00
7.90
7.80
7.70
7.60
7.50
1
Unlike the GSOEP (Figure 3.25), there is no large decline in reported satisfaction as
an individual is in the HILDA panel for longer (Figure 3.28). There are little if any
cohort and time effects evident in the HILDA data. Of course, the lack of time and
cohort effects may be due to the brevity of the HILDA panel. The 12-year BHPS
offers a panel length between that of the GSOEP (18 years) and the HILDA (8
years). I again ask, are the cohort and time effects found in the GSOEP also in the
BHPS.
43
This may be a testament to the high level of attention the Melbourne Institute places on retaining
individuals in the HILDA panel and ensuring dropouts are replaced with individuals of similar
demographic characteristics (Watson & Wooden, 2010)
.
73
3.8
With the GSOEP data (Section 3.6), the U-shape of happiness in age reversed into an
inverted U shape and a very strong negative relationship emerged (Figure 3.13). The
GSOEP age2 coefficient went non-significant with fixed effects, the age coefficient
became strongly significant across all specifications (Table 3.7). Like the HILDA,
this occurred to a much lesser extent with the BHPS, the (thin solid) line of the age
+ age2 fixed-effect specification in Figure 3.29 only shows a very slight negative
relationship. Like the GSOEP, the BHPS age2 coefficient does go non-significant
with fixed effects, but the age coefficient is not as strongly significant across all
specifications (Table 3.9) as it was with the GSOEP. Therefore, if our GSOEP
explanation is to hold, we should not see the strong time or cohort effects with the
BHPS data because the large decline we saw under the fixed-effects regressions with
the GSOEP are not evident in the BHPS data; we should not see a decline in the
aggregate BHPS data.
5.4
5.3
Life satisfaction
of all in BHPS
5.2
5.1
5
1996
1997
1998
1999
2000
2002
2003
2004
2005
2006
2007
2008
Year
Figure 3.29: Year and life satisfaction in the HILDA for the pooled sample
74
With the GSOEP, cohort effects were clearly visible (Figure 3.23); we do not see a
similar decline in the aggregate BHPS data (Figure 3.29). The GSOEP had a 0.4 drop
in happiness over the 19-year panel but, while there are small changes year-on-year,
life satisfaction levels across the 12-year BHPS panel period remains almost the
same. There appears to be little if any time or cohort effects in the BHPS. If so, we
should not see large differences between the first timers and those who remain the
BHPS panel.
Raw average
life
satisfaction
total sample
Life Satisfacition
6.00
5.50
age + age2 of
first-timers
5.00
4.50
Raw average
life
satisfaction of
first-timers
4.00
age+age2 of
full sample
3.50
3.00
AGE
5.4
5.3
Life satisfaction
of first-timers in
BHPS
5.2
Life satisfaction
of all in BHPS
5.1
5
1996
1997
1998
1999
2000
2002
2003
2004
2005
2006
2007
2008
Year
Figure 3.30 (top) and 3.30 (bottom): life satisfaction in the BHPS for first-time respondents
75
Figure 3.30 (top) shows the aggregate level of satisfaction by age of those who
answer for the first time, as well as for the entire BHPS pool. The raw average life
satisfaction of first timers (thin dashed line) does not exhibit the high level of
volatility evident in the GSOEP; the time effects (age + age2) of the BHPS stayers
and first timers overlay one another. Figure 3.30 (bottom) confirms the impression of
the top graph; when we looking at the average satisfaction over the years for firsttime respondents, there is only a very small increase in life satisfaction over the 8year panel: 0.07, less than 5% of a standard deviation. We do not see large
differences between the first timers and those who remain the BHPS panel. As a final
check that time and cohort effects do not strongly manifest in the BHPS, we look for
the evidence that we found with the GSOEP data: how much does life satisfaction
decline with years in the panel.
5.40
5.30
Life satisfaction of
stayers in BHPS
5.20
5.10
5.00
1
10
11
12
Unlike the GSOEP, which had a decline of 0.6 (32% of a standard deviation), there is
no large decline in reported satisfaction as an individual remains in the BHPS panel
for longer (Figure 3.31). Like the HILDA, there appears to be little if any cohort and
time effects in the BHPS data.
76
3.9
In trying to explain the new puzzle of the negative age slope in the GSOEP, I was
able to discount the likelihood of cohort and time effects: there simply is no
sufficiently strong time profile in the aggregate responses to explain the fixed-effect
results and, there is no time trend at all for first-time responses. This left sample
selection as the reason for the anomalously high age slope. I confirmed that those
who answer the GSOEP for the first time showed a smaller age profile in happiness.
There is a small decrease in happiness after age eighteen and after about age eighty,
but no significant change between the ages of twenty-five and seventy-five for firsttime responses. On the other hand, those who stayed in the panel reported lower
levels of life satisfaction with the decline being 0.64 for those in the GSOEP for
nineteen years. This perfectly fit the fixed-effects regressions.
The two possible interpretations of this were either that there was a selection of
stayers on somewhat persistent time-varying happiness unobservables or that those
who keep answering the GSOEP questionnaire become progressively more honest
about their actual, lower, level of satisfaction. If there is a selection on unobservables
then this is problematic for the reliability of long panels and analyses based on them.
If individuals are untruthful for the first nineteen years of responding to
questionnaires then this more or less affects all the analyses in the field, especially
cross-sectional studies that make up the bulk of the literature. The exact time profile
of the happiness of the stayers (a virtually continuous decline) slightly favours the
notion of selection on time-varying unobservables because one would have expected
the effect of becoming open and honest to the interviewer to gradually level off
before nineteen years.
Can we think of a reasonable third alternative explanation that is neither damning for
the collection of long representative panels or for the happiness field as a whole? I
cannot think of one. The found effect of age in fixed-effect regressions is simply too
large and too out of line with everything else we know to be believable. The
difference between first-time respondents and stayers and between the number of
years someone stays in the panel does not allow for explanations based on fixed traits
or observables. There has to be either a problem on the left-hand side (i.e. the
77
3.10
Chapter 3 Limitations
This study has limitations. There are happiness question wording and format
differences across the Australian, British and German panel survey questionnaires.
Format and wording differences bring into question whether the three surveys are
measuring the same phenomena. Wording and format differences alter the
psychometric properties of a survey question and initiate different responses from a
subject (Nunnally, 1978). It is possible to measure the instrument bias arising from
wording and format differences if subjects respond to all the happiness questions at
the same time. Confirmatory factor analysis can quantify the size of any instrument
bias and we could include this as a control variable in our regressions. Unfortunately,
the data to do this analysis is not in the German, British or Australian panel data. The
instrument bias arising from differences in the format and wording of the life
satisfaction questions remains unmeasured and is worthy of future research.
78
response scale (1-7) for the happiness question in the British panel survey
questionnaire is different from the scale (0-10) used for the happiness question in the
German and Australian panel survey questionnaires. The British happiness variable
could be rescaled 0-10 but this could introduce error in cross-panel comparisons.
Using beta coefficients or expressing within panel happiness changes as percentages
or coefficients of variation can assist with cross-panel comparisons. However,
analysis would be simplified if the authors of the socio-economic survey
questionnaires standardised on questions with the same scale, format and wording.
The next limitation concerns the graphics depicting happiness over a lifetime; they
are out of range predictions. In example, the figures depicting happiness over a
lifetime for the German population use nineteen years of data to depict a typical
lifetime. Some individuals may have appeared in the panel once, others many times.
Ideally, we could work with a balanced panel of data where a large number of
individuals stay in a panel for their complete lifetime. However, given the short
duration and the high subject attrition rate in socio-economic panels like the German
GSOEP, the British BHPS and the Australian HILDA, this ideal situation may never
be realised. We will just have to continue to use short unbalanced panels to depict a
typical lifetime.
3.11
Chapter 3 Summary
Chapter 3 questioned the relationship between happiness and age. A review of recent
economic literature revealed the finding of several economists that happiness is Ushaped in age; we get happier as we age. The psychological literature disagrees; their
consensus opinion is that age has little to do with happiness. Using a nested model
approach, we re-examined the age-happiness relationship in the often-used German
(GSOEP) panel data set and found evidence that reverse causality caused by
unobserved fixed traits44 explained the U-shape. Robustness was added to the
GSOEP fixed effect finding by replicating it with the Australian (HILDA), and,
44
The effects of observed fixed traits on the life event shocks affecting happiness are examined in
Chapter 5.
79
British (BHPS) panel data sets. In addition, the years-in-panel effect identified in the
GSOEP did not manifest in the HILDA and BHPS data.
I now pursue a difference between German and Australian lifetime happiness noted
earlier in this chapter. Germans have a steep decline in happiness between the ages of
18 and 22 that is not evident in the Australian HILDA data. In Chapter 4, I ask the
question; is there a similar dramatic drop in the happiness of young Australians,
when does it begin, and, what could lie behind the steep decrease in the happiness of
the young?
80
First-time
Respondents
Mean
s.d.
7.16
1.85
7.45
2.01
44.26
16.91
38.45
17.28
2244.67
1659.88
1769.01
1577.34
10.68
0.53
10.64
0.51
0.49
0.50
0.49
0.50
10.93
2.46
10.51
2.41
0.65
0.99
0.71
1.01
married (1=yes)
0.65
0.48
0.55
0.50
employed (1=yes)
0.47
0.50
0.44
0.50
0.33
0.47
0.37
0.48
unemployed (1=yes)
0.04
0.20
0.04
0.21
4149.99
477.88
3980.37
332.64
0.42
0.49
0.34
0.48
2359.80
10700.08
1258.23
6706.26
1484.61
2910.48
921.85
2134.08
2.59
0.95
2.22
0.94
invalid (1=yes)
0.04
0.20
0.04
0.19
0.01
0.08
0.00
0.06
divorced (1=yes)
0.05
0.22
0.04
0.19
0.01
0.12
0.01
0.11
0.06
0.24
0.05
0.22
0.02
0.15
0.05
0.21
0.00
0.07
0.00
0.06
0.01
0.11
0.01
0.08
0.00
0.06
0.00
0.03
0.04
0.19
0.02
0.14
pregnant (1=yes)
0.01
0.11
0.01
0.01
0.02
0.12
0.00
0.07
45
(stated)
Sample Size:
176,770
45
81
18,824
Table 3.12: Descriptive statistics for the entire and first-timer HILDA samples
Entire
Sample
Mean
s.d.
overall life satisfaction
First-time
Respondents
Mean
s.d.
7.91
1.47
7.88
1.58
45.50
16.92
43.06
16.92
2356.65
1666.50
2140.16
1631.59
10.77
1.07
5.23
3.09
0.47
0.50
0.48
0.50
12.82
1.80
12.70
1.74
0.77
1.12
0.75
1.11
married (1=yes)
0.55
0.50
0.50
0.50
employed (1=yes)
0.66
0.47
0.65
0.48
unemployed (1=yes)
0.03
0.16
0.04
0.20
1108.65
1113.69
1026.75
1009.84
0.75
0.43
0.71
0.45
4.94
39.82
4.35
37.67
2.64
0.95
2.61
0.97
invalid (1=yes)
0.24
0.43
0.21
0.41
0.11
0.31
0.11
0.31
divorced (1=yes)
0.09
0.29
0.09
0.29
0.04
0.19
0.04
0.20
0.05
0.22
0.05
0.21
0.03
0.16
0.04
0.19
just divorced
0.01
0.08
0.001
0.04
0.04
0.20
0.06
0.23
0.01
0.09
0.01
0.10
0.04
0.19
0.04
0.20
pregnant (1=yes)
0.05
0.22
0.06
0.24
0.03
0.17
0.04
0.19
47
(stated)
Sample Size:
72,108
18,821
46
47
Includes wages and salary, income from investments, as well as, Government transfers.
Health is reverse coded; 1 = excellent to 5 = poor.
82
Table 3.13: Descriptive statistics for the entire and first-timer BHPS samples 48
Entire
Sample
Mean
overall life satisfaction
s.d.
First-time
Respondents
Mean
s.d.
5.23
1.29
5.24
1.37
46.45
17.80
42.92
17.99
2474.85
1789.30
2165.49
1746.34
9.84
1.61
9.77
1.14
male (1=yes)
0.45
0.50
0.46
0.50
13.26
2.40
12.89
2.31
0.53
0.94
0.52
0.95
married (1=yes)
0.55
0.50
0.49
0.50
employed (1=yes)
0.58
0.49
0.57
0.49
age
age*age
unemployed (1=yes)
0.03
0.18
0.05
0.21
10.04
0.09
10.04
0.09
0.73
0.44
0.64
0.48
40.70
179.67
65.15
242.16
2.20
0.95
2.17
0.96
invalid (1=yes)
0.02
0.14
0.04
0.21
divorced (1=yes)
0.06
0.23
0.05
0.22
0.02
0.13
0.02
0.14
0.07
0.26
0.07
0.25
0.005
0.07
0.004
0.06
pregnant (1=yes)
0.002
0.04
0.001
0.03
49
Sample Size
48
153,886
22,922
The self-report life event variables in the GSOEP and the HILDA are not in the BHPS panel data.
49
. The health average is for the stated health question is for all BHPS waves except wave 9.
The wave 9 questionnaire asked a different (relative) health question and its response distribution was
significantly different from the health question in the other waves. To maximise sample size in the
regressions and to avoid the loss of a wave of data in the middle of the BHPS panel, a health
missing/dummy replacement routine was used to enable retention of wave 9 observations (STATA
code available upon request).
83
84
85
Table 3.14: The determinants of Life Satisfaction for West-Germans in the GSOEP; Pooled OLS Regression entire sample, N = 176,770
Age
Variable:
age
coefficient
-0.0059
Age + Age
t-value
22.68
age*age
coefficient
Usual Suspects
t-value
coefficient
t-value
coefficient
t-value
Kitchen Sink
coefficient
t-value
-0.0217
15.69
-0.0541
32.8
-0.0600
36.77
-0.0454
25.64
0.0002
11.63
0.0005
29.2
0.0006
34.13
0.0005
25.39
0.4619
52.2
0.4420
45.91
0.4307
44.53
-0.0603
6.3
-0.0719
7.60
-0.0719
7.47
11.17
-0.0249
13.7
-0.0227
12.47
0.0204
number of children
-0.0640
13.1
-0.0498
10.30
-0.0382
7.50
0.3106
27.9
0.2915
26.50
0.1180
7.65
married
employed
0.0650
4.8
0.0791
5.88
0.0688
5.11
non-participant
-0.0033
0.2
-0.0027
0.19
-0.0305
2.14
unemployed
-1.0076
42.3
regional income
-0.9508
40.37
-0.9225
38.37
-0.0001
11.24
-0.0001
11.15
home owner
0.0943
8.45
0.0956
8.58
asset income
0.0000
1.95
0.0000
1.87
15.23
imputed rent
0.0000
15.60
0.0000
health
-0.0228
30.42
-0.0231
30.71
invalid
-1.2427
55.93
-1.2444
56.12
family death
-0.3158
4.69
divorce
-0.2558
10.63
separated
-0.4539
11.26
partner dead
-0.0468
1.84
just married
0.4061
13.70
just divorced
0.0474
0.75
just separated
-0.4112
9.92
-0.9895
10.55
0.1354
5.65
pregnant
0.2118
5.05
-0.2747
7.86
4.7469
56.33
7.4165
0.00
602.43
7.7472
249.95
0.00
4.2642
0.0481
86
53.5
4.8670
0.07
58.03
0.08
Table 3.15: Determinants of Life Satisfaction for West-Germans in the GSOEP; Pooled OLS Regressions ages 22 to 80, N = 160,332
Age
Variable:
age
coefficient
-0.0034
Age + Age
t-value
coefficient
11.21
age*age
Usual Suspects
Kitchen Sink
t-value
coefficient
t-value
coefficient
t-value
coefficient
t-value
-0.0306
15.46
-0.07451
33.91
-0.0768
35.27
-0.0618
26.75
0.0003
13.89
0.0008
32.30
0.0008
33.73
0.0007
27.03
0.4759
50.35
0.4593
44.66
0.4479
43.41
-0.1196
11.66
-0.1208
11.92
-0.1196
11.64
education
-0.0286
15.13
-0.0248
13.07
0.0224
11.78
number of children
-0.0484
9.44
-0.0403
7.93
-0.0315
5.88
0.3373
29.39
0.3096
27.22
0.1531
9.61
male
married
employed
0.1442
9.86
0.1434
9.85
0.1303
8.96
non-participant
-0.0204
1.32
-0.0127
0.83
-0.0458
2.94
unemployed
-0.9302
37.63
regional income
-0.8791
35.91
-0.8539
34.21
-0.0001
12.07
-0.0001
12.01
home owner
0.1152
9.90
0.1173
10.09
asset income
0.0000
2.38
0.0000
2.33
14.18
imputed rent
0.0000
14.43
0.0000
health
-0.0233
30.01
-0.0237
30.40
invalid
-1.1565
47.22
-1.1579
47.36
family death
-0.3037
4.19
divorce
-0.2193
9.01
separated
-0.4069
10.02
partner dead
-0.0029
0.11
just married
0.3976
13.19
just divorced
0.0618
0.98
just separated
-0.3974
9.55
-1.0887
10.87
0.1538
6.38
pregnant
0.2388
5.68
-0.2663
7.45
4.8739
53.19
7.2962
0.00
502.70
7.8762
178.17
0.00
4.4076
0.05
87
50.67
5.0061
0.07
54.98
0.08
Table 3.16: The determinants of Life Satisfaction for West-Germans in the GSOEP; Fixed-effect Regressions entire sample, N = 176,770
Age
Variable:
age
coefficient
-0.0398
Age + Age
t-value
50.43
age*age
coefficient
Usual Suspects
t-value
coefficient
t-value
coefficient
t-value
Kitchen Sink
coefficient
t-value
-0.0166
6.60
-0.0328
-11.61
-0.0298
9.95
-0.0184
5.81
-0.0003
9.71
-0.0001
-2.22
0.0000
0.74
-0.0001
2.40
0.2414
23.13
0.2750
23.67
0.2585
22.15
(omitted)
(omitted)
(omitted)
education
-0.0017
-0.35
-0.0009
0.19
-0.0024
0.51
number of children
-0.0255
-3.9
-0.0231
3.52
-0.0102
1.52
married
0.2385
14.95
0.2457
15.42
0.0327
1.37
employed
0.0991
6.69
0.1001
6.75
0.0925
6.24
non-participant
0.0214
1.46
0.0244
1.67
0.0150
1.00
-0.6623
-29.28
-0.6455
28.61
-0.6282
27.40
0.0000
5.31
0.0000
4.98
unemployed
regional income
home owner
0.0115
0.71
0.0292
1.82
asset income
0.0000
0.64
0.0000
0.78
imputed rent
0.0000
4.26
0.0000
3.80
health
-0.0124
11.07
-0.0122
10.91
invalid
-0.7192
29.20
-0.7296
29.65
family death
-0.2978
5.24
divorce
-0.0008
0.02
separated
-0.3273
7.63
partner dead
-0.1631
3.74
just married
0.3553
13.83
just divorced
-0.0573
1.04
just separated
-0.2858
7.96
-0.9168
11.54
0.1127
5.55
pregnant
0.0541
1.37
-0.1769
5.88
6.1589
49.54
8.9180
0.00
254.02
8.4541
142.62
0.00
6.6296
0.0135
88
58.97
6.1697
0.03
49.82
0.03
Table 3.17: The determinants of Life Satisfaction for West-Germans in the GSOEP; Pooled Regression first time respondents, N = 18,821
Age
Variable:
age
coefficient
-0.0046
Age + Age
t-value
5.43
age*age
coefficient
Usual Suspects
t-value
coefficient
t-value
coefficient
t-value
Kitchen Sink
coefficient
t-value
-0.0030
0.73
-0.0437
8.06
-0.0490
9.05
-0.0367
6.26
0.0000
0.40
0.0004
6.86
0.0005
8.17
0.0004
6.09
0.4720
16.06
0.4472
14.40
0.4414
14.13
-0.0629
2.02
-0.0730
2.36
-0.0811
2.58
3.78
-0.0296
4.66
-0.0262
4.12
0.0242
number of children
-0.0600
3.89
-0.0436
2.85
-0.0255
1.61
married
0.4189
11.28
0.3997
10.83
0.2359
4.74
employed
0.0501
1.07
0.0779
1.67
0.0759
1.62
non-participant
-0.0581
1.26
-0.0497
1.08
-0.0565
1.23
unemployed
-1.2065
15.26
-1.1363
14.46
-1.1409
14.38
0.0000
0.91
0.0000
0.60
regional income
home owner
0.1407
3.66
0.1348
3.50
asset income
0.0000
1.37
0.0000
1.29
imputed rent
0.0000
2.66
0.0000
2.65
health
-0.0065
1.98
-0.0074
2.22
invalid
-1.3171
17.40
-1.3228
17.46
family death
-0.3478
1.35
divorce
-0.3366
3.81
separated
-0.5037
3.67
partner dead
-0.1496
1.66
just married
0.2765
3.56
just divorced
0.0996
0.43
just separated
0.0272
0.14
-0.3384
0.71
-0.1287
1.14
pregnant
0.2552
1.26
0.0198
0.09
4.1449
12.81
constant
2
7.6261
0.00
213.61
7.5963
91.73
0.00
4.1537
0.05
89
15.68
4.2166
0.06
13.06
0.07
Table 3.18: The determinants of Life Satisfaction; Pooled OLS regression results for all individuals in the HILDA; N = 75,529
Age
Variable:
coefficient
age
0.0107
Age + Age
t-value
33.37
coefficient
Usual Suspects
t-value
coefficient
t-value
coefficient
t-value
Kitchen Sink
coefficient
t-value
-0.0320
-18.98
-0.0554
-29.57
-0.0403
-22.72
-0.0311
16.00
0.0004
25.79
0.0007
33.95
0.0006
30.26
0.0005
23.98
0.0805
15.00
0.0329
6.03
0.0309
5.69
male
-0.1471
-13.52
-0.1016
-9.93
-0.0981
-9.53
education
-0.0232
-7.52
-0.0582
-19.74
-0.0606
-20.6
number of children
-0.0590
-10.81
-0.0623
-12.15
-0.0600
-11.25
age*age
income
married
0.4496
36.2
0.3624
30.32
0.2514
14.85
employed
0.1355
9.27
-0.1338
-9.2
-0.1166
-8
-0.3244
-9.36
-0.4208
-12.92
-0.3560
-10.87
0.0000
5.87
0.0000
5.2
home owner
0.1397
10.95
0.1253
9.82
imputed rent
-0.0001
-0.42
-0.0001
-0.43
health
-0.5278
-88.46
-0.5251
-88.37
invalid
-0.1073
-7.92
-0.1046
-7.76
0.0110
0.68
divorced
-0.0771
-3.51
separated
-0.2947
-9.12
0.0237
0.78
unemployed
regional income
family death
partner dead
just married
0.1175
3.7
just divorced
-0.1853
-2.72
just separated
-0.4209
-15.2
-0.2597
-4.59
0.1278
3.81
pregnant
0.1382
4.88
7.4240
0.0151
Adjusted R
477.31
8.3255
217.8
8.0870
110
10.0957
135.44
-0.2875
-9.52
10.0082
132.16
0.0241
0.0531
0.1661
0.1745
0.0240
0.0530
0.1661
0.1742
90
Table 3.19: Determinants of Life Satisfaction for Australians in the HILDA; Pooled Regressions, ages 22 to 80, N = 65.679
Age
Variable:
age
coefficient
0.0135
Age + Age
t-value
35.46
coefficient
Usual Suspects
t-value
coefficient
t-value
coefficient
Kitchen Sink
t-value
coefficient
t-value
-0.0417
-17.5
-0.0680
-26.69
-0.0496
-20.48
-0.0386
-14.91
0.0006
23.45
0.0008
31.09
0.0007
26.72
0.0006
21.61
0.0815
14.02
0.0349
5.92
0.0329
5.61
male
-0.1660
-14.52
-0.1039
-9.64
-0.0992
-9.14
education
-0.0218
-6.9
-0.0569
-18.77
-0.0589
-19.49
number of children
-0.0429
-7.66
-0.0514
-9.72
-0.0489
-8.96
age*age
income
married
0.4494
35.47
0.3598
29.26
0.2514
14.65
employed
0.1758
11.44
-0.1162
-7.55
-0.0995
-6.46
-0.3029
-7.8
-0.4141
-11.32
-0.3443
-9.36
0.0000
5.47
0.0000
4.71
home owner
0.1386
10.12
0.1197
8.75
imputed rent
0.0000
0.0000
0.07
health
-0.5226
-83.07
-0.5206
-83.1
invalid
-0.0922
-6.51
-0.0896
-6.36
0.0142
0.84
divorce
-0.0720
-3.25
separated
-0.2898
-8.93
0.0515
1.6
unemployed
regional income
family death
partner dead
just married
0.1292
3.96
just divorced
-0.1711
-2.48
just separated
-0.4360
-14.51
-0.2332
-3.82
0.1311
3.77
pregnant
0.1492
5.04
-0.2977
-9.12
10.0494
112
Adjusted R2
7.2676
0.0188
389.98
8.4890
153.54
8.2251
94.47
10.1845
115.78
0.0269
0.0587
0.1678
0.1767
0.0269
0.0586
0.1677
0.1764
91
Table 3.20: The determinants of Life Satisfaction for Australians in the HILDA; Fixed-effect Regressions entire sample, N = 75,529
Age
Variable:
age
coefficient
-0.0167
Age + Age
t-value
-6.90
age*age
coefficient
Usual Suspects
t-value
coefficient
t-value
coefficient
t-value
Kitchen Sink
coefficient
t-value
-0.0076
-1.02
-0.0173
-2.12
-0.0202
-2.50
-0.0033
-0.40
-0.0001
-1.30
0.00001
-0.17
0.00007
0.94
-0.00006
-0.72
0.0257
4.26
0.0191
3.22
0.0151
2.55
income
education
-0.0204
-1.72
-0.0198
-1.71
-0.0246
-2.11
number of children
-0.0329
-2.24
-0.0312
-2.18
-0.0436
-2.92
married
0.2056
6.76
0.1963
6.52
-0.1095
-2.84
employed
0.0536
2.33
0.0238
1.06
0.0394
1.75
-0.1344
-3.19
unemployed
-0.1422
-3.43
-0.1208
-2.91
regional income
0.0000
3.30
0.0000
2.65
home owner
0.0687
3.01
0.0640
2.83
imputed rent
-0.0001
-0.79
-0.0001
-0.62
health
-0.2694
-28.16
-0.2679
-28.13
invalid
-0.0492
-3.29
-0.0488
-3.28
family death
-0.0046
-0.33
divorce
-0.1608
-2.71
separated
-0.3479
-4.83
partner dead
-0.3512
-3.85
just married
0.1361
4.56
just divorced
-0.2707
-2.94
just separated
-0.3063
-8.45
-0.2326
-3.02
0.1338
4.79
pregnant
just fired from job
constant
8.6701
2
Overall R
0.0151
78.75
8.4827
47.34
0.0175
8.5967
0.0047
92
41.04
9.2482
0.0344
44.12
0.1258
4.94
-0.0389
-1.22
9.1148
43.45
0.0325
Table 3.21: The determinants of Life Satisfaction for Australians in the HILDA; Pooled Regression first time respondents, N = 14,857
Age
Variable:
coefficient
age
0.0103
Age + Age
t-value
13.55
age*age
coefficient
Usual Suspects
t-value
coefficient
t-value
coefficient
t-value
Kitchen Sink
coefficient
t-value
-0.0303
-7.62
-0.0582
-13.12
-0.0469
-11.08
-0.0325
-6.94
0.0004
10.41
0.0007
14.76
0.0006
14.23
0.0005
10.65
0.0986
7.11
0.0525
3.64
0.0476
3.32
income
male
-0.1251
-4.83
-0.0756
-3.07
-0.0823
-3.33
education
-0.0326
-4.29
-0.0684
-9.34
-0.0730
-10.01
number of children
-0.0549
-4.2
-0.0554
-4.48
-0.0446
-3.44
married
0.5072
16.96
0.4188
14.45
0.2048
5.02
employed
0.1338
3.86
-0.1431
-4.13
-0.1130
-3.25
-0.3529
-5.03
-0.4572
-6.87
-0.3968
-5.94
0.0001
3.31
0.0000
3.03
unemployed
regional income
home owner
0.1249
4.25
0.1069
3.64
imputed rent
-0.0002
-0.74
-0.0003
-0.88
health
-0.5263
-37.18
-0.5218
-37.09
invalid
-0.1025
-3.04
-0.1005
-3
family death
-0.0358
-0.94
divorced
-0.1699
-3.23
separated
-0.3920
-5.61
partner dead
-0.0865
-1.16
just married
0.2846
4.29
just divorced
-0.2612
-0.81
just separated
-0.5640
-10.31
-0.3410
-2.88
0.1571
2.04
pregnant
0.1567
2.5
7.4309
0.0122
Adjusted R
210.77
8.2610
94.81
8.0377
43.93
10.0849
53.56
-0.2236
-3.52
9.9728
52.35
0.0194
0.0505
0.1496
0.1629
0.0192
0.0499
0.1488
0.1615
93
Table 3.22: The determinants of Life Satisfaction for Britons in the BHPS; Pooled Regression entire sample, N = 153,886
Age
Variable:
coefficient
age
0.0066
Age + Age
t-value
35.86
age*age
coefficient
Usual Suspects
t-value
-0.0221
-21.96
0.0003
29.00
coefficient
-0.0554
t-value
-49.28
coefficient
t-value
-0.0438
-41.01
Kitchen Sink
coefficient
t-value
-0.0349
-31.02
39.43
0.0006
55.15
0.0005
48.21
0.0005
income
0.0943
19.44
0.0399
8.39
0.0251
5.25
male
-0.0438
-6.66
-0.0540
-8.69
-0.0625
-9.95
education
-0.0033
-2.27
-0.0199
-14.52
-0.0195
-14.27
number of children
-0.0581
-15.27
-0.0600
-16.63
-0.0510
-13.89
married
0.3567
46.48
0.3063
41.8
0.2014
22.43
employed
0.2397
27.19
0.0295
3.48
0.0372
4.38
unemployed
-0.2823
-14.46
-0.3430
-18.55
-0.3406
-18.46
regional income
-0.1659
-5.05
-0.1606
-4.9
home owner
0.1521
18.47
0.1445
17.56
imputed rent
0.0001
5.13
0.0001
4.4
health
-0.4653
-130.81
-0.4635
-130.57
invalid
-0.3415
-15.82
-0.3373
-15.66
divorced
-0.2966
-19.74
separated
-0.4714
-19.72
partner dead
-0.1139
-7.23
0.3748
8.39
pregnant
0.2263
3.22
constant
4.9199
0.0083
Adjusted R
537.15
5.5346
239.77
5.0476
95.85
8.2857
25.17
8.2286
25.05
0.0137
0.0481
0.1529
0.157
0.0137
0.048
0.1528
0.1569
94
Table 3.23: The determinants of Life Satisfaction for Britons in the BHPS; Pooled Regressions ages 22 to 80, N = 138,481
Age
Variable:
age
coefficient
0.0087
Age + Age
t-value
39.44
age*age
coefficient
Usual Suspects
t-value
coefficient
t-value
coefficient
Kitchen Sink
t-value
coefficient
t-value
-0.0373
-26.21
-0.0772
-51.31
-0.0620
-43.38
-0.0521
-34.94
0.0005
32.7
0.0009
58.5
0.0007
50.57
0.0007
42.99
income
male
0.1134
21.22
0.0561
10.75
0.0384
7.31
-0.0727
-10.56
-0.0726
-11.16
-0.0810
-12.31
education
-0.0042
-2.83
-0.0204
-14.41
-0.0198
-13.97
number of children
-0.0306
-7.76
-0.0419
-11.21
-0.0342
-9.04
married
0.3439
43.95
0.2951
39.41
0.1984
21.9
employed
0.3340
35
0.0825
8.91
0.0898
9.7
-0.1903
-8.86
-0.2907
-14.28
-0.2903
-14.29
-0.1874
-5.45
-0.1798
-5.24
0.1539
17.38
0.1449
16.38
unemployed
regional income
home owner
imputed rent
0.0001
3.86
0.0001
3.23
health
-0.4629
-124.22
-0.4613
-124.05
invalid
-0.3054
-13.52
-0.3019
-13.39
divorced
-0.2824
-18.68
separated
-0.4506
-18.8
partner dead
-0.0957
-5.69
0.3484
7.62
pregnant
0.2327
3.13
8.5603
24.83
constant
2
Adjusted R2
4.8093
0.0111
440.65
5.8274
176.68
5.1761
86.01
8.6378
25
0.0187
0.0599
0.1632
0.1672
0.0187
0.0599
0.1631
0.1671
95
Table 3.24: The determinants of Life Satisfaction for Britons in the BHPS; Fixed-effect Regressions entire sample, N = 153,886
Age
Variable:
age
coefficient
-0.0097
Age + Age
t-value
-9.65
age*age
coefficient
Usual Suspects
t-value
coefficient
t-value
coefficient
Kitchen Sink
t-value
coefficient
t-value
0.0044
1.42
-0.0064
-1.88
-0.0127
-3.86
-0.0082
-2.49
-0.0001
-4.61
-0.0001
-1.75
0.00008
0.27
0.000008
-0.31
0.0164
2.64
0.0166
2.7
0.0060
0.98
income
male
(omitted)
(omitted)
(omitted)
education
-0.0029
-0.76
-0.0016
-0.43
-0.0017
-0.47
number of children
-0.0103
-1.41
-0.0120
-1.67
-0.0099
-1.38
-0.24
married
0.1430
8.58
0.1390
8.47
-0.0041
employed
0.0618
4.63
0.0355
2.77
0.0365
2.86
-0.1754
-7.28
-0.1968
-8.35
-0.1958
-8.33
-0.1389
-1.23
-0.1365
-1.22
0.0340
2.06
0.0340
2.08
unemployed
regional income
home owner
0.0001
2.37
0.0001
2.24
health
imputed rent
-0.1995
-40.87
-0.1994
-40.9
invalid
-0.1337
-3.66
-0.1316
-3.61
divorced
-0.2244
-6.73
separated
-0.4065
-11.11
partner dead
-0.3500
-8.25
0.2348
7.22
pregnant
0.1158
2.28
7.4777
6.64
constant
5.6756
2
Overall R
0.0083
121.91
5.3853
72.17
0.0114
5.7226
0.0011
96
69.53
7.4496
0.0374
6.59
0.0353
Table 3.25: The determinants of Life Satisfaction for Britons in the BHPS; Pooled Regression first time respondents, N = 22,922
Age
Variable:
age
coefficient
0.0073
Age + Age
t-value
14.61
age*age
coefficient
Usual Suspects
t-value
coefficient
t-value
coefficient
t-value
Kitchen Sink
coefficient
t-value
-0.0211
-7.76
-0.0531
-17.19
-0.0424
-14.13
-0.0312
-9.89
0.0003
10.63
0.0006
19.24
0.0005
16.88
0.0004
13.35
0.1274
10.56
0.0860
7.18
0.0661
5.48
-0.0512
-2.84
-0.0630
-3.64
-0.0811
-4.63
income
male
education
-0.0078
-1.93
-0.0239
-6.1
-0.0237
-6.07
number of children
-0.0739
-7.22
-0.0690
-6.99
-0.0549
-5.49
married
0.3461
16.38
0.2930
14.28
0.1498
6.1
employed
0.1969
8.42
0.0043
0.19
0.0134
0.58
-0.3322
-7.45
-0.3968
-9.22
-0.3960
-9.23
-0.3102
-3.14
-0.2944
-2.99
0.1745
8.25
0.1654
7.84
unemployed
regional income
home owner
imputed rent
0.0001
2.33
0.0001
1.74
health
-0.4163
-39.93
-0.4142
-39.86
invalid
-0.4888
-11.29
-0.4769
-11.05
divorced
-0.4021
-9.26
separated
-0.6378
-10.07
partner dead
-0.2113
-4.69
0.5297
3.89
-0.0098
-0.04
9.1019
9.24
constant
2
Adjusted R2
4.9264
0.0092
211.71
5.5010
93.5
4.8111
36.65
9.2465
9.35
0.0141
0.0495
0.1285
0.1353
0.0140
0.0491
0.1279
0.1345
97
Table 3.26: Summary of changes in the GSOEP Age and Age2 coefficients as controls are
progressively added
Specification
Pooled OLS
Pooled OLS
Fixed Effect
(All)
(Ages 22 to 80)
(All)
coefficient
t-value coefficient
t-value
coefficient
t-value
Age + Age2
age
age*age
-0.0217** 15.69
-0.0306** 15.46
-0.0166**
6.60
0.0002** 11.63
0.0003** 13.89
-0.0003**
9.71
-0.0541** 32.80
0.0745** 33.91
-0.00328**
11.61
0.0005** 29.20
0.0008** 32.30
-0.0001*
2.22
-0.0600** 36.77
-0.0768** 35.27
-0.0298**
9.95
0.0006** 34.13
0.0008** 33.73
0.00000
0.74
-0.0454** 25.64
-0.0618** 26.75
-0.0184**
5.81
0.0005** 25.39
0.0007** 27.03
-0.00002**
2.40
Usual suspects
age
age*age
Usual suspects + health
age
age*age
Kitchen sink
age
age*age
N
176,770
Level of significance:
160,332
98
176,770
Table 3.27: Summary of changes in the HILDA Age and Age2 coefficients as controls are
progressively added
Specification
Pooled OLS
Pooled OLS
Fixed Effect
(All)
(Ages 22 to 80)
(All)
coefficient
t-value coefficient
t-value
coefficient
t-value
Age + Age2
age
-0.0320
**
19.0
-0.0417** 17.50
-0.00755
1.02
age*age
0.0004
**
25.8
0.0006** 23.45
-0.00010
1.30
age
-0.0554
**
29.6
-0.0680** 26.69
-0.01728
age*age
0.0007
**
34.0
0.0008** 31.09
-0.00001
age
-0.0403
**
22.7
-0.0496** 20.48
-0.02017
age*age
0.0006
**
30.3
0.0007** 26.72
0.00007
Usual suspects
*
2.12
0.17
2.50
0.94
Kitchen sink
age
-0.0311 **
16.0
-0.0386** 14.91
-0.00328
0.40
age*age
0.0005 **
24.0
0.0006** 21.61
-0.00006
0.72
72529
Level of significance:
65679
99
72529
Table 3.28: Summary of changes in BHPS Age and Age2 coefficients as controls are
progressively added
Specification
Pooled OLS
Pooled OLS
Fixed Effect
(All)
(Ages 22 to 80)
(All)
coefficient
t-value coefficient
t-value
coefficient
t-value
Age + Age2
age
age*age
-0.0221 ** 21.96
-0.0373 ** 26.21
0.0003 ** 29.00
0.0005 ** 32.70
-0.00015 ** 4.61
0.00444
1.42
-0.0554 ** 49.28
-0.0772 ** 51.31
-0.00638 + 1.88
0.0006 ** 55.15
0.0009 ** 58.50
-0.00006 + 1.75
-0.0438 ** 41.01
-0.0620 ** 43.38
-0.01274 ** 3.86
0.0005 ** 48.21
0.0007 ** 50.57
0.000009
-0.0350 ** 31.02
0.0521 ** 34.94
-0.00824 * 2.49
0.0005 ** 39.43
0.0007 ** 42.99
-0.00001
Usual suspects
age
age*age
Usual suspects + health
age
age*age
0.27
Kitchen sink
age
age*age
N
153886
Level of significance:
138481
100
0.31
153886
Appendix C contains robustness analyses on the three datasets. Table 3.29 provide
the robustness calculations for the GSOEP, Table 3.31 for the BHPS, and Table 3.33
for the HILDA; column 1 shows the pooled OLS results (with robust standard errors)
that were included in the main text as the Usual suspects + health results. Column 2
re-runs that regression with the health variable as categorical rather than continuous.
Column 3 re-runs column 1 with an ordered logit specification rather than a least
squares specification. Column 4 implements the recent conditional fixed-effect logit
model of Baetschmann, Staub, & Winkelmann (2011), whilst column five compares
this with the fixed-effect results. Per the method of Clark (2006), Tables 3.30, 3.32,
& 3.34 show OLS results of the effect of age-bands on life satisfaction for the
GSOEP, BHPS and HILDA.
101
Table 3.29: The determinants of Life Satisfaction for West-Germans in the GSOEP; Usual suspects plus health specification, OLS, OLS with categorical health,
Ordered Logit, BUC estimator & OLS with fixed effect entire sample, N = 176,770
OLS
(1)
Usual Suspects +
Health
Variable:
age
coefficient
t-value
OLS
(1)
with
Categorical Health
coefficient
t-value
Fixed effects
BUC estimator2
Ordered Logit
coefficient
t-value
m-effect1
coefficient
t-value
coefficient
t-value
-0.0497
-15.38
-0.0514
-16.02
-0.0646
-18.06
0.0003
-0.0444
-6.94
-0.030
-9.95
age*age
0.0005
16.11
0.0006
16.69
0.0006
16.4
0.0000
0.00004
0.56
0.00002
0.74
income
0.4088
24.67
0.4073
24.69
0.4407
24.66
-0.0022
0.3483
16.34
0.275
23.67
-0.0855
-4.44
-0.0816
-4.27
-0.0728
-3.42
0.0004
(omitted)
male
(omitted)
0.0074
2.07
0.0078
2.19
0.0229
5.8
-0.0001
-0.0036
-0.4
-0.001
-0.19
-0.0601
-6.79
-0.0607
-6.87
-0.0510
-5.34
0.0003
-0.0365
-2.81
-0.023
-3.52
married
0.2997
14.18
0.2951
14.04
0.2837
12.44
-0.0015
0.3141
10.13
0.246
15.42
employed
0.0698
3.42
0.0722
3.54
0.0638
2.85
-0.0003
0.1368
5.21
0.100
6.75
non-participant
0.0385
1.84
0.0468
2.24
0.0183
0.79
-0.0001
0.0385
1.54
0.024
1.67
-0.8543
-23.44
-0.8427
-23.18
-0.8626
-24.2
0.0066
-0.7480
-19.36
-0.646
-28.61
-0.0001
-9.51
-0.0001
-9.24
-0.0001
-10.7
0.0000
0.0000
-3.72
0.000
-5.31
education
number of children
unemployed
regional income
home owner
0.0999
5.15
0.0972
5.02
0.0873
4.11
-0.0004
0.0129
0.42
0.011
0.71
asset income
0.0000
-1.54
0.0000
-1.59
0.0000
-0.98
0.0000
0.0000
0.62
0.000
0.64
0.0000
8.44
0.0000
8.67
health
-0.7395
-71.83
invalid
-1.0499
0.05
imputed rent
-0.9913
-21.37
102
0.0000
9.66
0.0000
0.0000
1.96
0.000
4.26
-0.0274
-25.47
0.0001
-0.0162
-10.25
-0.012
-11.07
-1.1027
-23.36
0.0096
-0.8048
-17.44
-0.719
-29.2
OLS
(1)
Usual Suspects +
Health
Variable:
coefficient
t-value
OLS
(1)
with
Categorical Health
coefficient
t-value
Fixed effects
BUC estimator2
Ordered Logit
coefficient
t-value
m-effect1
coefficient
t-value
t-value
Categorical health
health==1 (very good)
0.8695
39.14
health==2 (good)
0.2545
18.29
health==3 (satisfactory)
-0.3703
-22.68
health==4 (poor)
-1.1258
-44.80
-2.5839
-45.67
5.0755
37.21
constant
6.7867
R2 / Pseudo R2/Overall R2
0.1373
48.48
6.1697
0.1403
0.0182
103
0.0287
0.302
49.82
Table 3.30: The determinants of Life Satisfaction for West-Germans in the GSOEP, Age-bands:
(1) without control; (2) with controls per the Usual suspects plus health specification; (3) plus
fixed effects (3), N = 176,770
Specification
(2)
(3)
(1)
Controls per
(2) plus
No controls
Fixed Effects
coefficient
t-value
Age 18 to 22
0.5815 **
10.86
Age 23 to 27
0.4742 **
Age 28 to 32
coefficient
t-value
coefficient
t-value
0.2228 **
4.26
3.8592E-01 +
1.95
8.97
0.0314
0.61
3.1578E-01 +
1.65
0.4436 *
8.43
-0.0739
-1.44
3.3470E-01 *
1.82
Age 33 to 37
0.4015 **
7.63
-0.1923 **
-3.71
3.4949E-01 *
1.98
Age 38 to 42
0.3902 **
7.39
-0.2602 **
-4.99
3.7309E-01 *
2.2
Age 43 to 47
0.3395 **
6.4
-0.3652 **
-7.01
3.5606E-01 *
2.19
Age 48 to 52
0.2953 **
5.55
-0.4120 **
-7.9
3.4070E-01 *
2.18
Age 53 to 57
0.2058 **
3.84
-0.4479 **
-8.56
3.1915E-01 *
2.13
Age 58 to 62
0.3160 **
5.87
-0.2097 **
-4.02
5.0512E-01 **
3.5
Age 63 to 67
0.4864 **
8.9
0.0049
0.09
6.8212E-01 **
4.92
Age 68 to 72
0.5088 **
9.08
0.0859
1.61
7.1373E-01 **
5.35
Age 73 to 77
0.4581 **
7.88
0.1359 **
2.46
6.5856E-01 **
5.15
Age 78 to 82
0.1885 **
3.01
0.0263
0.44
3.9369E-01 **
3.23
2.8359E-01 **
2.64
Age 83 to 87
(omitted)
Age 88 to 93
-0.0425
Year in Panel
-0.0227 **
constant
2
2
R / R Between
6.9257
(omitted)
-0.45
0.1384 +
-23.79
-0.0169 **
133.75
3.72
0.009
(omitted)
-15.51 -3.0472E-02 **
40.30
4.6327E+00
0.077
153,886
Level of significance:
1.56
153,886
104
0.07
153,886
0
-10.42
21.97
Table 3.31: The determinants of Life Satisfaction for British in the BHPS; Usual suspects plus health specification, OLS, OLS with categorical health, Ordered
Logit, BUC estimator & OLS with fixed effect entire sample, N = 153,886
OLS
(1)
Usual Suspects +
Health
Variable:
coefficient
t-value
OLS
(1)
with
Categorical Health
coefficient
t-value
Fixed effects
BUC estimator2
Ordered Logit
coefficient
t-value
m-effect1
coefficient
t-value
coefficient
t-value
age
-0.0438
-21.17
-0.0431
-20.89
-0.0705
-21.11
0.0008
-0.00887
-0.93
-0.0127
-3.86
age*age
0.00052
24.08
0.00051
23.77
0.0009
23.9
-0.00001
-0.00013
-1.35
0.00008
0.27
income
0.0399
5.47
0.0415
5.70
0.0545
4.82
-0.0006
0.0227
1.21
0.0166
2.7
male
-0.0540
-4.44
-0.0525
-4.33
-0.0961
-5.09
0.0010
(omitted)
education
-0.0199
-7.80
-0.0195
-7.65
-0.0353
-8.99
0.0004
0.0147
1.19
-0.0016
-0.43
number of children
-0.0600
-9.35
-0.0602
-9.38
-0.0916
-9.47
0.0010
-0.0246
-1.07
-0.012
-1.67
0.3063
21.98
0.3036
21.84
0.4685
21.6
-0.0052
0.3387
6.79
0.139
8.47
married
(omitted)
0.0295
2.02
0.0190
1.31
-0.0317
-1.41
0.0003
-0.0109
-0.28
0.0355
2.77
unemployed
-0.3430
-11.85
-0.3542
-12.27
-0.5273
-12.52
0.0073
-0.4871
-7.49
-0.1968
-8.35
regional income
-0.1659
-2.59
-0.1723
-2.69
-0.3035
-3.08
0.0033
-0.3949
-1.16
-0.1389
-1.23
home owner
0.1521
9.52
0.0001
3.74
0.2070
8.29
-0.0023
0.0803
1.61
0.034
2.06
imputed rent
0.0001
3.92
0.0001
3.74
0.0001
3.41
0.0000
0.0001
1.63
0.0001
2.37
health
-0.4653
-75.05
-0.7068
-76.63
0.0076
-0.4507
-32.34
-0.1995
-40.87
invalid
-0.3415
-8.51
-0.0431
-20.89
-0.5088
-8.52
0.0070
-0.1909
-2.3
-0.1337
-3.66
employed
105
OLS
(1)
Usual Suspects +
Health
Variable:
coefficient
t-value
OLS
(1)
with
Categorical Health
coefficient
t-value
Fixed effects
BUC estimator2
Ordered Logit
coefficient
t-value
m-effect1
coefficient
t-value
t-value
Categorical health
health==1 (very good)
1.3937
61.96
health==2 (good)
1.0413
48.52
health==3 (satisfactory)
0.5586
26.46
(omitted)
health==4 (poor)
health==5 (very poor)
constant
8.2857
R2 / Pseudo R2/Overall R2
0.1529
12.9
-0.5990
-13.80
6.3989
9.98
7.4496
0.1547
0.0503
106
0.0252
6.59
0.374
Table 3.32: The determinants of Life Satisfaction for British in the BHPS; Age-bands: (1)
without controls; (2) with controls per the Usual suspects plus health specification; (3) plus
fixed effects, N = 153,886
(1) No controls
Specification
coefficient
Age 18 to 22
-0.0641
Age 23 to 27
t-value
(2)
(3)
(2) plus
Fixed Effects
coefficient
t-value
coefficient
t-value
-1.26
-0.4913 **
-16.29
0.2373 *
2.11
-0.1115**
-2.2
-0.6036 **
-19.83
0.1817 +
1.72
Age 28 to 32
-0.0867*
-1.72
-0.6543 **
-21.54
0.1776 +
1.77
Age 33 to 37
-0.1775**
-3.52
-0.7518 **
-24.53
0.1373
1.45
Age 38 to 42
-0.2463**
-4.87
-0.8222 **
-26.76
0.1072
1.21
Age 43 to 47
-0.2801**
-5.52
-0.8597 **
-28.05
0.0867
1.05
Age 48 to 52
-0.2260**
-4.45
-0.8091 **
-26.41
0.0994
1.3
Age 53 to 57
-0.1336**
-2.62
-0.6647 **
-21.69
0.1558 *
2.2
Age 58 to 62
0.0678
1.32
-0.4291 **
-14.1
0.2951 **
4.55
Age 63 to 67
0.2037**
3.95
-0.2387 **
-7.79
0.3553 **
5.97
Age 68 to 72
0.3429**
6.63
-0.0407
-1.33
0.4109 **
7.67
Age 73 to 77
0.3058**
5.87
0.0163
0.52
0.3271 **
6.77
Age 78 to 82
0.2209**
4.12
0.0021
0.06
0.2124**
5.43
Age 83 to 87
0.1181**
2.06
(omitted)
Year in Panel
-0.0060**
-5.62
constant
2
R / R Between
N
5.3192
106.67
0.021
153,886
Level of significance:
omitted
-0.0177 **
-16.91
7.81
25.25
-0.0139 **
6.7974
0.16
153,886
0.189
153,886
107
-6.34
5.91
Table 3.33: The determinants of Life Satisfaction for Australians in the HILDA; Usual suspects plus health specification, OLS, OLS with categorical health,
Ordered Logit, BUC estimator & OLS with fixed effect entire sample, N = 72,529
OLS
(1)
Usual Suspects +
Health
Variable:
coefficient
t-value
OLS
(1)
with
categorical Health
coefficient
Fixed effects
BUC estimator2
Ordered Logit
t-value
coefficient
t-value
m-effect1
coefficient
t-value
t-value
-0.0403
-13.19
-0.0397
-13.10
-0.0542
-12.69
0.000042
-0.03198
-1.88
-0.0202
-2.5
age*age
0.0006
17.22
0.000546
17.07
0.00077
16.82
-0.000001
0.00006
0.34
0.00007
0.94
income
0.0329
4.77
0.0321
4.65
0.0413
4.29
-0.000032
0.03720
3.19
0.0191
3.22
age
male
-0.1016
-5.78
-0.0986
-5.62
-0.1427
-5.96
0.000111
(omitted)
education
-0.0582
-11.84
-0.0565
-11.50
-0.0893
-13.26
0.000069
-0.04512
-1.64
-0.0198
-1.71
number of children
-0.0623
-7.32
-0.0645
-7.59
-0.0854
-7.3
0.000066
-0.07669
-2.62
-0.0312
-2.18
0.3624
17.79
0.3606
17.77
0.4784
17.13
-0.000381
0.36791
6.23
0.1963
6.52
married
(omitted)
employed
-0.1338
-5.91
-0.1589
-7.08
-0.2888
-9.27
0.000214
0.01390
0.31
0.0238
1.06
unemployed
-0.4208
-8.15
-0.4467
-8.66
-0.5628
-8.68
0.000574
-0.25771
-3.61
-0.1422
-3.43
0.0000
4.49
0.0000
4.74
0.0000
3.51
0.000000
0.00006
3.04
0.0000
3.3
regional income
home owner
0.1397
6.59
0.1361
6.42
0.1826
6.48
-0.000148
0.13279
2.93
0.0687
3.01
imputed rent
-0.0001
-0.29
0.0000
-0.14
0.0000
-0.16
0.000000
-0.00024
-0.82
-0.0001
-0.79
health
-0.5278
-54.71
-0.7339
-56.08
0.000566
-0.53700
-28.33
-0.2694
-28.16
invalid
-0.1073
-5.70
-0.0749
-4.03
-0.1027
-4.09
0.000081
-0.07743
-2.52
-0.0492
-3.29
108
OLS
(1)
Usual Suspects +
Health
Variable:
coefficient
t-value
OLS
(1)
with
categorical Health
coefficient
Fixed effects
BUC estimator2
Ordered Logit
t-value
coefficient
t-value
m-effect1
coefficient
t-value
t-value
Categorical health
health==1 (very good)
2.4188
34.21
health==2 (good)
1.9884
29.08
health==3 (satisfactory)
1.5218
22.36
health==4 (poor)
0.9288
13.68
(omitted)
constant
R2 / Pseudo R2/Overall R2
1
10.0957
0.1661
95.55
7.0492
58.69
0.1694
9.2482
0.0553
BUC estimator implemented using feologit in Stata (Baetschmann, Staub et al. 2011)
Estimates with clustered standard errors.
109
0.0270
0.344
44.12
Table 3.34: The determinants of Life Satisfaction for Australians in the HILDA, Age-bands: (1)
without controls; (2) with controls per the Usual suspects plus health specification; (3) plus
fixed effects, N = 75,729
Specification
(2)
(3)
(1)
Controls per
(2) plus
No controls
Fixed Effects
coefficient
t-value
coefficient
t-value coefficient
t-value
Age 18 to 22
-0.4468**
-7.37
-0.8740 **
-14.94
0.2433
0.96
Age 23 to 27
-0.5804**
-9.54
-0.9962 **
-16.89
0.2021
0.82
Age 28 to 32
-0.5715**
-9.41
-1.0528 **
-17.96
0.1529
0.64
Age 33 to 37
-0.5648**
-9.36
-1.0616 **
-18.15
0.1466
0.62
Age 38 to 42
-0.7049**
-11.66
-1.1689 **
-19.84
0.0979
0.43
Age 43 to 47
-0.6845**
-11.33
-1.1116 **
-18.86
0.1748
0.78
Age 48 to 52
-0.6213**
-10.21
-1.0253 **
-17.4
0.2441
1.12
Age 53 to 57
-0.4268**
-6.96
-0.8304 **
-14.21
0.3548 +
1.67
Age 58 to 62
-0.2739**
-4.43
-0.6389 **
-10.98
0.4728 *
2.29
Age 63 to 67
-0.0929
-1.49
-0.4481 **
-7.72
0.5795**
2.89
Age 68 to 72
0.0825
1.31
-0.2382 **
-4.07
0.5409**
2.77
Age 73 to 77
0.0952
1.48
-0.1568 **
-2.64
0.4889**
2.58
Age 78 to 82
0.0789
1.17
-0.0383
-0.61
0.4259 *
2.35
Age 83 to 87
(omitted)
0.2647
1.58
(omitted)
Age 88 to 93
-0.0239
-0.2
Year in Panel
-0.0044*
-1.54
constant
R2/ R2 Between
N
8.3780
141.57
0.029
0.0616
-0.0128 **
10.4078
0.55
-4.83
118.98
0.169
75,729
Level of significance:
-0.0185** -4.83
8.2548
30.7
0.182
75,729
110
(omitted)
75,729
Chapter Four
Unhappy Young Australians
4.1
Introduction
In Chapter 3, we saw a decline in the happiness of young Germans (Figure 3.4) and
Australians (Figure 3.6) that was much steeper than the decline in older individuals
whose happiness we could expect to decline due to falling income, failing health or
the imminent onset of death. This exploratory study asks why the happiness of young
Australians exhibits such a steep decline.
Such a question is study-worthy because there is evidence from other scientific
disciplines that the happiness of individuals in their childhood can affect the
happiness of those same individuals in adulthood. For example, Cheng & Furnham
(2004) showed that maternal care during early childhood was directly correlated with
adult happiness, and, Flouri's (2004) study of British children found that maternal
care at age seven predicted higher life satisfaction in 42 year-old males, and,
closeness to your mother at age 16 predicted higher life satisfaction at age 42 in both
men and women. Trzcinski & Holst (2007) examined the level of subjective wellbeing among young people in transition to adulthood, and like Flouri (2004), they
found that the quality of parental-adolescent relationships was a predictor of the
happiness of young people in transition to adulthood. There is some evidence that
happiness in childhood should be maximised because it maximises our happiness as
adults. For an economist, it would follow that a policy of keeping families together
and thereby increasing child and adult wellbeing is beneficial to the overall
wellbeing of society. However, childhood happiness is little studied by economics of
happiness researchers.
In this chapter, I seek to extend our view of lifetime happiness to an age cohort little
visited by economists, 9 to 14 year old children. The chapter proceeds with a review
of childhood happiness studies in the economics literature. After revealing four
important factors that affect childhood happiness and conflicting results in the
economics of happiness literature, I explain how a survey was developed to collect
111
data from the 9 to 14 year old Australians. After explaining how the data were
collected, a model of childhood happiness and used to examine the data from the 9 to
14 year olds. The results are interpreted and conclusion offered that explain the
relationship between the life satisfaction domains of children and childhood
happiness. Finally, the 15 to 23 year-old cohort from the HILDA socio-economic
panel is examined with a model of individual happiness. First, let us review
childhood happiness studies from the economics literature
4.2
Reviewing studies listed in Tables 4.35a & b, over the past twelve years there have
been happiness studies of Australian, European, Italian, Scottish, British, and, United
States adolescents.
Community Household Panel (ECHP) data to reveal that adolescents are happier
when they make the decision to leave the family home. A study by Dockery (2005)
used data from the (1997 -2004) Longitudinal Surveys of Australian Youth and wave
1 of the HILDA and found evidence of declining levels of happiness in adolescents
during periods of unemployment and the importance of the quality and type of work
to the happiness of adolescents. Ulker (2008) also used the Australian HILDA data to
reveal the importance of the role of family members in the well-being of Australian
adolescents; e.g. adolescents are unhappy when their parents divorce. Bassi & Delle
Fave's (2004) longitudinal study used an experience sampling method (aka, daily
112
A few studies have focussed on childhood happiness. Fogle, Scott Huebner, &
Laughlin's (2002) cross-sectional study revealed the positive interrelationships
between the personality trait of extraversion, social self-efficacy and social
competence, to the life satisfaction of children aged 10 to 15 years from public
schools in mid-sized South-eastern United States cities. A positive attitude,
confidence in own abilities and the skills to interact with your peer group were found
to be important to childhood happiness. Also focussing on school children, Huebner,
Valois, Paxton, & Drane's (2005) cross-sectional study of public middle school
students (from South Carolina, U.S.A) found that family, friends, school and the
environment in which children live and learn are important to childhood happiness.
Lee & Oguzoglu's (2007) longitudinal study of Australian youths ventured outside
the school environment and found that income support payments contributed to
childhood happiness. Flouris (2004) study focussed on the importance of family to
childhood happiness. Using the GHQ-12 measure of mental and physical wellbeing,
Flouri found that children who were insulated from psychological stress within the
family domain were happier and those (7 year-old) children who were more involved
with their mother were happier (as a 42-year-old) in adulthood. These studies mirror
many of the findings from the studies of adult happiness. Individuals with a job,
sufficient income, and stable relationships tend to be happier. What is somewhat
different to adult happiness is the importance of family, school, interaction with
school friends, and the environment to childhood happiness; these are the factors that
are considered in this chapter.
113
Children
Adolescent
Table 4.35a & b: Summary of happiness studies of the young from the economics literature (EconLit) identifying the study population as adolescents or children
Author
Data source
Sample
(Ebner, 2008)
European Community
Household Panel (ECHP)
(Ulker, 2008)
Individuals aged 15 to
24 years. N = 6,013
Longitudinal Surveys of
Australian Youth.
(1997 -2004)
5,865 Australian
youths with a median
age of 14. N = 26,146
(Dockery, 2005)
Longitudinal Surveys of
Australian Youth.
8,567 individuals
aged 16 to 19 years.
N = 30,406
(Huebner, Valois,
Paxton, & Drane,
2005)
School children
N = 2278
Participants aged
between 15 to 18
years. N = 120
Adolescents aged 16
to 19 years. N = 90
114
Research Question
Children
Adolescents
Table 4.35a & b (continued): Summary of childhood happiness studies from the economics literature (EconLit)
Author
Data
Sample
(Flouri, 2004)
Longitudinal study of
British children; aged
7-42 years N = 17,000
(Csikszentmihalyi &
Hunter, 2003)
Multi-year study of
American youth from the
Alfred P. Sloan Study of
Youth and Social
Development
Children aged 10 to 15
years. N = 160
Eurobarometer Surveys
115
Research Question
While many of the above-noted results mirror the findings from the study of adult
happiness, results from two studies do not. Blanchflower & Oswald's (1998) study of
European adolescents found that the happiness of those aged 15 years and older
increased in the period 1970 to 1990. Dockery (2005) also found evidence in the
Longitudinal Surveys of Australian Youth data that adolescent happiness increased
in age. These findings are consistent with Easterlin (2006) who stated that United
States happiness rises slightly, on average, from ages 18 to midlife. However, such
an increase conflicts with the Chapter 3 findings of a 6% decrease in the average
happiness of young Germans, the 7% decrease in the happiness of 15 to 23 year old
Australians, and, the 2.9% decrease in the happiness of 18 to 23 year-old Britons.
This chapter will not only consider the effects of family, school, interaction with
school friends and the environment on childhood happiness on Australian children,
the study will also seek to resolve whether the happiness of young Australians (aged
9 to 23 years) increases, decreases or is stable in age. Because the HILDA panel
data set only has data on individuals 15 years and older, data needs to be collected
from Australian children. First, an explanation of the HILDA data set used in this
study.
4.3
In seeking to explain the effects of family, school, interaction with school friends and
the environment on childhood happiness, and whether the happiness of the young is
increasing, decreasing or stable in age, I use two data sets. Data was collected data
from 9 to 14 year-old children and our view of lifetime happiness is extended beyond
the 18 to 93 year-olds we examined in the previous chapter by adding the data for 15
to 18 year olds from the Household, Income and Labour Dynamics in Australia
Survey (HILDA).
116
Table 4.36: Observations by age & year for the 15 to 23 year-old sample from HILDA waves 2-8
Age
Year of Wave
2005
2006
2002
2003
2004
2007
2008
Total
15
231
217
221
225
257
204
224
1579
16
218
223
203
223
219
259
191
1536
17
216
200
205
204
222
199
242
1488
18
201
206
180
203
199
221
201
1411
19
180
193
193
184
194
184
202
1330
20
165
175
178
201
174
192
188
1273
21
178
171
163
174
214
177
181
1258
22
169
183
158
175
189
193
175
1242
23
146
164
166
171
172
194
200
1213
1704
1732
1667
1760
1840
1823
1804
12330
Total:
50
Wave 1 was excluded because it does not include life event variables.
117
Table 4.37: Sample averages for the 15 to 23 year-old cohort and the entire HILDA sample
HILDA
HILDA
(entire sample
(15 to 23 year-olds)
15 to 93 year-olds)
Mean
s. d.
Mean
s. d.
7.94
1.47
8.07
1.38
43.74
17.84
18.76
2.59
2231.29
1690.90
358.69
98.13
5.26
3.16
10.85
1.22
male (1=yes)
0.47
0.50
0.48
0.50
12.71
1.80
11.83
1.07
0.73
1.10
0.06
0.30
married (1=yes)
0.51
0.50
0.03
0.17
employed (1=yes)
0.65
0.48
0.66
0.47
0.03
0.18
0.09
0.29
1117.62 1393.19
1158.13
age
age*age
unemployed (1=yes)
average regional income
44
1126.85
0.75
0.43
0.63
0.48
4.82
39.24
3.80
32.75
2.61
0.96
2.25
0.88
invalid (1=yes)
0.23
0.42
0.10
0.30
0.11
0.31
0.10
0.30
divorced (1=yes)
0.09
0.28
0.001
0.02
0.03
0.18
0.001
0.03
0.05
0.22
0.00
0.00
0.03
0.16
0.02
0.13
0.01
0.08
0.00
0.03
0.04
0.20
0.06
0.24
0.01
0.09
0.00
0.06
0.03
0.18
0.02
0.16
pregnant (1=yes)
0.05
0.22
0.04
0.20
0.03
0.17
0.04
0.21
imputed rent
51
Self-reported health
52
53
N=
77,132
12,330
51
118
Interpreting descriptive statistics from Table 4.37, the average happiness for 15 to 23
year-olds is 2% higher than the average happiness for the 15 to 93 year-olds in the
entire HILDA sample; the happiness of the young initially appears to decline as they
age. The average income of households with 15 to 23 year old children is higher than
for the entire HILDA sample. However, with self-reported pregnancy peaking at age
29 (Figure 4.32), Australian parents with 15-year-old children would be, on average,
44 years of age. Age 44 is when average annual household income peaks. With 15 to
23 year old children reliant on their parents income, the happiness of these children
could decline due to fewer resources being available to satisfy the childrens want
and needs. In addition, with the average number of children per household increasing
from 0.011 for households with a 15 year old to 0.19 for households with a 23 year
old, the happiness of those 15 to 23 years olds could be expected to decline. The
increase in the number of children per household could translate to fewer resources
per dependent child.
11.5
0.2
11
0.15
10.5
0.1
10
0.05
9.5
Figure 4.32: Log of average Annual Household Income and proportion of self-reported
pregnancy (self or partner) for Australians aged 15 to 93 years, 2002-2008 HILDA, N = 77,132
119
Life Satisfacition
8.5
7.5
6.5
AGE
Figure 4.33: Life Satisfaction of 15 to 93 year-old Australians (2002-2008 HILDA panel data)
Looking back the three extra years to age 15, we can see that, on average, the
happiness of 15 to 23 year-old Australians does decrease in age, just as we saw for
the German and British populations in Chapter 3. The decrease in the happiness of 15
to 18 year-old Australians is large, 4.7%. Added to the 2.5% decline for young adults
aged 18 and 23 years, this steep 7.2% drop in the happiness of young Australians is
120
54
While the HILDA extended our view of lifetime happiness to 15 to 23 year-olds, it does not include
data for young children. Young children are not usually included in large socioeconomic panel data
collection, probably because young children have difficulty, answering the survey questions either
because the questions are not relevant to children or the language too complex for the children to
understand (Gilman & Huebner, 1997). For example, the SF-36 health questionnaire included in the
HILDA survey is only suitable for those aged 15 years and over (Ware, 2009).
121
schools organize class groups to visit the train at their local railway station. The
children participate in exciting, interactive and informative teaching and learning
displays55 one of those was our Happiness display.
The Happiness teaching and learning display explained what makes individuals
happy, the importance of income, health and child-related factors such as school,
family, friends and the environment. As an extension to the Happiness teaching and
learning display on the Smart Train, we teamed with the Queensland
Government56 to anonymously collect the data from the children who visited the
Smart Train. To ensure the Smart Train data came from the target population,
each child visiting the Smart Train received a Happiness Postcard (Figures 34a &
34b) inviting them to participate in a Happiness Research project that included
responses to an online Happiness Survey survey. To maximize response rates and
to overcome technical impediments to the children responding, all the data collecting
steps the children had to participate in were Happiness themed and we integrated
the Happiness Survey response process into the childrens normal teaching and
learning program. We initially raised awareness of our Happiness Research project
by giving teachers Happy promotion posters to hang on classroom walls. Next, to
help with guiding the children through the survey response process, we provided
teachers with Happy Teaching Guides that included instructions on how the
children could use the computers in their classroom or school library to respond to
the online Happiness Survey.
To complete the online Happiness Survey the children needed the Happiness
Postcard given to them during their Smart Train visit. The postcard directed them
to the Smart Train website57 where they gained entry to the online survey by
clicking on the How happy are you? visual link, the same graphic we used for the
front of the Happiness Postcard (Figure 4.34a).
55
The 22 research displays on the 2008 Smart Train came from the QUT Faculties of Business, Built
Environment and Engineering, Creative Industries, Education, Health, Humanities and Human Services,
Information Technology, Law and Sciences; see http://www.train.qut.edu.au/ for details of what was in each
display.
56
Our thanks go to: Markus Schaffner, QUT post graduate student; Leesa Watkin, QUT Smart Train Project
Manager, and; Annie Harris, Senior Project Officer, Science Partnerships and Engagement, Department of
Tourism, Regional Development and Industry, Queensland State Government for their considerable financial and
technical assistance in making this research project a success.
57
More information on the Smart Train can be found on the website: http://www.train.qut.edu.au/
122
Figure 4.34a: Front side of the Happiness Postcard themed with the same graphics as the
Happiness promotional poster & the icon clicked to initiate the web-based Happiness Survey
Figure 4.34b: The obverse of the Happiness Postcard containing instructions on how to
respond to the online survey
123
58
After the three-week survey response period, the Department randomly drew a Happy Postcard for
the student and school prizes. The Queensland Government Department presented the lucky student
with an Apple iPod and their school received $1000 to spend on science resources.
59
The online Happiness Survey screens are in Figures 35a to 35l in Chapter 4 Appendix A.
124
For
ethical reasons, it was decided not to question the children on their family life
domain. Questioning young children on the marital status of their parents or whether
they live with both mum and dad was considered outside our universitys ethical
125
The past president of the American Psychological Association, Dr Martin Seligman (2011) states,
Life satisfaction essentially measures cheerful moods, so it is not entitled to a central place in any
theory that aims to be more than a happiology .... By that standard ... a government could improve its
numbers just by handing out the kind of euphoriant drugs that Aldous Huxley described in Brave
New World. While Dr Seligmans comments may be considered extreme, but they do amplify the
concern about happiness and its measurement.
126
1961); Temporal Satisfaction with Life Scale (TSWLS) (Pavot, Diener, & Suh,
1998); Watson, Clark, & Tellegen's (1988) Positive Affect/Negative Affect Scale
(PANAS); Delighted/Terrible Scale (Andrews & Withey, 1976), and; the
Experience/Day Reconstruction Method (DRM) which measures affective
experience in daily life (Csikszentmihalyi & Larson, 1987; Csikszentmihalyi &
Hunter, 2003). Michael Fordyce offers guidance in choosing a happiness scale.
In his (1988) paper, Michael Fordyce reviewed 18 years of the use of the Global life
Satisfaction Scale (Fordyce, 1973) and other measures of life satisfaction. Fordyce
asserted the validity of the Global Life Satisfaction Scale as a measure of
emotional well-being and global health. Fordyce supported his claim by citing the
collection of evidence from three independent research teams who examined the
accumulated findings from dozens of studies that compared happiness measures.
Generally, the independent research teams were of the opinion that Fordyces
Global Life Satisfaction Scale instrument showed good reliability and stability,
with a good record of convergent, construct, and discriminative validity. The
independent research teams found a high convergence between Fordyces Global
Life Satisfaction Scale and other happiness measures. In summary, Fordyce
supported the continued use of his Global Life Satisfaction Scale with this
statement: it has construct validity, the ability to discriminate between known happy
and unhappy groups, and it has association with widely accepted characteristics of
mental health. Accumulated findings on this instrument are believed to show good
reliability and stability, and a record of convergent, construct, and discriminative
validity. The Global Life Satisfaction Scale is suggested to be a potential
touchstone of measurement consistency in a field that generally lacks it.
Even if we recognise that Michael Fordyce may be somewhat biased towards
recommending the use of his own happiness measure, he does cite considerable
independent evidence to support its continued use. As to how economists first came
to use Fordyces Global life Satisfaction Scale, I found no firm evidence in the
economics literature. Nor could I find any justification for why the question was
selected for inclusion in socio-economic panel survey questionnaires. Whatever the
reason for originally choosing the question, and in the absence of further research
that justifies a question that is more acceptable to the economics discipline, there
appears to no evidence to compel economists to change from measuring global
127
happiness with Fordyces single Global Life Satisfaction Scale question. However,
economists have been asking adults to answer the Global Life Satisfaction Scale
question; is it suitable for children?
The Fordyces single Global Life Satisfaction Scale question is considered suitable
for our 9 to 14 year-old respondents. In their 2004 study, Kornilaki & Chlouverakis
(2004) found that children as young as 7 years understood the concept of happiness. I
am of the opinion that the 9 to 14 year-old children we surveyed could understand
the Global Life Satisfaction Scale question in the Happiness Survey. Together
with the other survey questions, the happiness question was pre-tested on six to
fifteen year-old children before the question wording was finalised. After simplifying
the wording for a small number of questions (but not the happiness question), the
children said they understood the words and the meaning behind all the questions in
our online Happiness Survey.
Prior to verifying whether the happiness and other questions were suitable for
children, questions were added to collect socio-demographic information about the
children. Careful to satisfy the strict ethical requirements for child anonymity, the
children were asked: which railway station they visited the Smart Train; their
gender; the postcode where they lived, and; their grade at school. Next, we sought to
include questions to collect data on the variables that have been found to have the
largest effect on adult happiness.
There is an extensive literature identifying variables with the largest effect on adult
happiness (see Clark, Frijters, & Shields, 2008 for a review). Table 4.38 summarises
the results from three national datasets used in this study and ranks the six variables
with the largest effect on overall happiness. The ranking varies across datasets, but
overall health or unemployment have a greater effect on the overall happiness of
adults followed by relationship status in varying orders depending upon the country
(married, separated & divorced) and finally income. Having identified the variables
with the largest effect on adult happiness, I sought to locate child-suitable questions
that could be incorporated into the online Happiness Survey.
128
Table 4.38: Cross-country OLS happiness results ordered (1) to (6) by the size of the
standardised beta coefficient
Australia61
Britain62
Germany63
income
(6)
0.032**
(6)
0.024**
(6)
0.04**
health
(1)
0.52**
(5)
0.05**
(4)
0.23**
unemployed
(2)
-0.31**
(2)
-0.35**
(1)
-0.92**
married
(4)
0.26**
(4)
0.20**
(5)
0.12**
separated
(3) 0.28**
(1)
-0.48**
(2)
-0.45**
divorced
(5)
-0.06*
(3)
-0.30**
(3)
-0.26**
Country:
Variable:
Number in brackets identifies the highest (1) to the lowest (6) standardised beta coefficient.
* significant at 5%; ** significant at 1%
In selecting questions suitable for children, I began by considering how we could ask
children about their familys annual household income. Children rely on parents or
guardians income to provide them with sustenance, shelter, education and
healthcare. With (relative) income being so important to happiness, how can we
accurately collect this data from children? Researchers have long been concerned
with inaccuracies in how adults report their income (Moore, Stinson, & Welniak,
2000). If adults misreport income we can hardly expect children to know what mum
or dad earns. However, while children aged 6 to 14 years may not completely
understand the economics of supply and demand, they do understand the concept of
money, its what they swap with others to get what they want (Leiser & Beth
Halachmi, 2006). Given children understand the concept of money, it is reasonable to
expect that a child would be able to gauge how well off (wellbeing) their family was
compared with their friends family. Children would be aware that their friend would
61
Pooled OLS regression results using the HILDA Panel 2002-2008; N = 75,529 (Beatton & Frijters,
2011, p.36).
62
Pooled OLS regression using the British Household Panel Survey (BHPS) waves 1 to 14;
N = 82,096 (Frijters & Beatton, 2011, p.39).
63
Pooled OLS regression results using the German SocioEconomic Panel 1984-2002; N = 176,770
(Frijters & Beatton, 2011, p.33).
129
be wealthier if that friend lived in a bigger house, had a more modern car, or had a
holiday home. Children understand that wealthier parents have more money to buy
better things. Children understand that poorer people do not have much money so
they cannot buy a big house with a swimming pool. Children should be able to gauge
whether their family is wealthier, or poorer, than their friends family. To gauge the
relative income of their parents we asked the children:
Would you say that your family is? (tick one box)
While we are of the opinion that children can gauge their familys relative wealth,
are the children capable of providing us with a self-assessment of what is arguably
the variable with the strongest effect on happiness, health. The authors of the SF-36
health questionnaire64 dont think so; their health questionnaire has only been
validated for use on individuals 15 years and older (Ware, 2009). In addition,
findings indicate that, unless they have psychological issues, the majority of children
self-report with excellent health (Flouri, 2004). In expectation of minimal variance in
health question response data, and because the health survey experts (the SF-36
authors) consider their health survey questions unsuitable for children, plus with a
requirement to minimise the number of questions in the survey65, I chose not to
include a health question in the Happiness Survey.
I next considered asking the children about their family life. The literature is clear in
identifying that a stable home environment and parental relationship status are
important to the happiness of children. For example, Ulker, ( 2008, p210) identified
64
For details of the SF-36 health scale, see section A of the Self Completion Questionnaire from the
Australian HILDA survey. It includes the SF-36 health items that gauge general health and well-being
(http://www.melbourneinstitute.com/downloads/hilda/Questionnaires/SelfCompletionQuestionnaireW
8.pdf).
65
Survey response levels reduce as the number of questions increase; respondents get bored and stop
answering or resort to the same response for each question (Cavana, Sekaran, & Delahaye, 2001).
130
(Huebner (1991) and Natvig, Albrektsen, & Qvarnstram (2003) found that the school
environment is an important domain of childhood happiness. In a subsequent study
involving 518 American elementary school students, Seligson, Huebner, & Valois
(2005) found a positive relationship between the environment at school and the
global life satisfaction of the children. School environment factors found to affect
childhood happiness include teacher support, student interaction, and competition in
class as well as a students psychological adjustment to school. The school
environment domain is central to childrens life satisfaction, particularly the support
of teachers and classmates (Suldo & Huebner, 2006).
Table 4.39: Personality factors, related behaviours, & Happiness Survey question numbers
Personality
Factor
Behaviours
Happiness Survey
Question Numbers
Extroversion
Agreeableness
Conscientiousness
Emotional stability
Openness to
experience
While the BFQ-C provided us with a personality scale suitable for children, we still
needed survey questions for our school environment and interaction with friends
domain factors. Looking closely at the question wording of Barbaranelli et al.s
(2003) BFQ-C personality scale, one notices that the questions are phrased66 to
measure the type of behaviours we would expect from children interacting with their
friends and classmates before, during or after school. I was of the opinion that the
wording of selected questions from the BFQ-C scale could be used to form our
school environment and interaction with friends scales.
Selected questions from the BFQ-C scale were chosen to form two of the three life
domain factors recommended by Gilman & Huebner (2003): 1) school
environment, and; 2) interaction with friends. The questions were chosen based on
the theoretical fit of the question wording and the extent to which the question(s)
converged to a single factor. To operationalize the questions, the children were asked
66
See the personality question wording in Figures 35g to 35k in Chapter 4 Appendix A.
132
to tell us something about themselves and the extent to which a statement described
them. Responses were measured using a Likert scale of: definitely not true; probably
not true; dont know, maybe; probably true; definitely true). The complete survey
questionnaire is in Appendix A of Chapter 4; example questions include:
Example School environment questions:
Q35.
Q52.
Q55.
Q37.
Q56.
Principal factor analysis was used to maximise convergence and a Cronbachs alpha
test (Cavana, Sekaran, & Delahaye, 2001) was used as a measure of the internal
consistency and reliability of our school environment and interaction with friends
domain factors as a suitable psychometric test for our Smart Train subjects. Both
factors exhibited high scale reliability; Table 4.40 lists the question numbers for each
life satisfaction domain factor and their Cronbachs alpha test scores.
133
Table 4.40: Happiness Survey question numbers & behaviours for the school environment and
interaction with friends life satisfaction domains
Life Domain
(variable name)
Behaviours
Happiness Survey
Question Number
School environment1
(schoolenv)
q35
q36
q38
q43
q45
q48
q52
q53
q55
q60
q62
q30
q32
q37
q41
q42
q47
q56
q59
In addition to selecting questions from the BFC-Q scale to form our school
environment and interaction with friends domain factors, the children were
questioned on their natural environment life satisfaction domain. To gauge the
childrens attitudes to and the level of concern for their natural environment the
children were asked: if they were engaged in discussions on their natural
environment; if they were aware of environmental problems; what they were doing
about them; whether it was an acceptable behaviour to pollute their river or a river
in a neighbouring state, and; the importance of animals and plants in their lives67.
The questions on the natural environment were coded as dummy variables and
summed to form the natural environment life satisfaction domain factor; where 1 is
lowest level of concern for the natural environment and 13 the highest level of
concern for the natural environment.
67
The natural environment questions in the Happiness Survey are q13_1, q13_2, q13_3, q13_4,
q13_5, q17, q18, q21_1, q21_2, q21_3, q22, q23, and q24.
134
After including the questions for the natural environment domain factor, some fun
questions were added to amuse the children and encourage them to complete the
survey. The fun questions68 asked about, magic, belief in a greater being, lucky
charms, if they wrote with their right or left hand and a graphical question to help
the children gauge if their index finger was longer than their middle finger. With the
questions for our Happiness Survey finalised and the Smart Train beginning its
journey across the state of Queensland, we posted the Happiness Survey on the
Smart Train website and proceeded to collect the Smart Train data.
Average life satisfaction for our 9 to 14 year old sample is a very high 9.0; 14%
higher than the 7.91 for the 15 to 23 year-olds in the HILDA and 12% higher than
the 8.07 for the complete HILDA sample. Average happiness for female children in
the Smart Train data (9.31) is 8% higher than for male children (Table 4.41). There
was some variation in happiness by railway station and school grade (4 to 9) but this
variation arises mainly from the large sample size differences in the railway station
where the children visited the Smart Train and the childs school grade (50% of the
sample were in grade seven). At the more granular level, there was no significant
difference (ANOVA: F = 0.06, p = 0.82) between the happiness of urban (mean =
9.01, s.d. = 1.99) and regional (mean = 8.94, s.d. = 2.10) children.
68
The fun questions are q6, q7, q8, q10, q11, q19, q20, q21, q27, q28, q29, q63, and q64.
135
Table 4.41: Life Satisfaction for the 9 to 14 year old children in the Smart Train data
Count
69
Age
Dependent variable:
Life Satisfaction
All
9.00 (2.0)
9.31 (1.88)
8.60 (2.09)
62
5
5
12
59
16
4
58
41
4
6
60
3
3
41
8
8.94 (2.10)
9.68 (1.20)
9.24 (1.84)
8.98 (1.98)
8.50 (2.25)
8.80 (2.13)
8.80 (0.00)
8.99 (1.76)
8.85 (2.43)
10.00 (1.10)
8.07 (1.80)
9.61 (1.62)
8.80 (2.20)
7.33 (2.54)
9.28 (1.81)
8.25 (2.56)
9.19 (1.99)
10.00 (1.27)
6.60 (.)
10.00 (1.18)
8.51 (2.23)
9.63 (1.14)
8.80 (.)
9.75 (1.11)
9.28 (2.20)
10.00 (1.27)
8.80 (2.20)
9.62 (1.73)
9.90 (1.56)
4.40 (.)
9.40 (1.82)
7.92 (3.34)
8.50 (2.28)
8.80 (.)
9.90 (1.27)
7.48 (1.97)
8.49 (2.30)
7.98 (2.61)
8.80 (.)
8.17 (1.98)
8.31 (2.67)
10.00 (.)
7.33 (1.27)
9.90 (1.56)
6.60 (.)
8.80 (.)
9.15 (1.84)
8.80 (.)
Regional Children
Urban Children (Brisbane)
327
62
9.01 (1.99)
8.94 (2.10)
9.35 (1.85)
9.19 (1.98)
8.61 (2.28)
8.50 (2.07
School Grade
Grade 4
Grade 5
Grade 6
Grade 7
Grade 8
Grade 9
17
19
91
196
50
20
9.45 (1.51)
9.38 (1.43)
9.21 (2.08)
8.83 (2.04)
8.70 (1.68)
8.57 (2.73)
9.24 (1.84)
9.80 (1.51)
9.59 (1.84)
9.17 (1.89)
9.50 (1.42)
8.80 (2.81)
9.53 (1.43)
8.80 (1.18)
8.68 (2.32)
8.37 (2.16)
8.80 (1.83)
8.17 (2.76)
9
10
11
12
13
14
217
N = 389
69
172
For ethical reasons we could not collect identifying information other than gender from the
children; age was calculated from the mandated school starting age and the childs school grade.
70
Care should be taken when interpreting regression results by age & railway station; the minimum
sample size requirements were calculated using G*Power 3.08 software; with our low effect size, the
sample sizes at each age were too small, except for age 12 (Faul, Buchner, Erdfelder, Faul, & Lang,
2006). The sample sizes for analysis by gender are adequate.
71
Brisbane was the only urban centre; 84% of the respondents came from regional Queensland.
72
Some children did not visit the Smart Train but responded to the survey. To do this the children
needed the Happiness Postcard provided by their teacher.
136
As well as asking the children to answer the happiness question, the children were
asked how they perceived their natural environment (Table 4.42). The children had a
high awareness of environmental issues, they were aware of climate change (68%),
water restrictions (63%), native animals dying out (44%), declining fish stocks (22%)
and land salinity (19%), but ten per cent of the children had noticed none of these
environmental issues. The children perceived climate change (48%) then the loss of
native fauna (24%) as the worst environmental issues. These results are what we
would expect from our majority non-urban sample (327 of the 389). We would
expect that non-urban children would have a higher awareness of issues such as the
loss of native animals (81% of non-urban children said animals were an important
part of their lives), and salinity and climate change; all are current issues affecting
Australias agricultural communities.
Sixty-eight per cent of the children stated they had started a conversation about
climate change but, surprisingly only 50% of the children reported that their family
talked about the environment. Either the children are discussing the environment
among themselves or they are exposed to discussions on the environment in their
classrooms. Either way, Australias next generation does appear to have a higher
level of concern (73%) for the natural environment than the 53% of Australias
current adult population who show a concern for our natural environment (ABS,
2010).
The children in the Smart Train data are not just showing more concern for the
natural environment, they are acting on that concern. In example, sixty-one per cent
of the children are engaging in recycling. Sixty-eight per cent of the children have
tried to reduce their water consumption, and, 98% of the children said it was wrong
to pollute a river, even if that river was in another state. These environmental and the
other variables collected with our Happiness Survey form the cross-sectional
Smart Train data set, which I applied to the model of childhood happiness.
137
Table 4.42: Descriptive statistics for selected questions from the Smart Train cross-sectional data; N = 389
Mean (s.d.)
Female
Male
9.00 (2.0)
9.31 (1.88)
8.60 (2.09)
2.0 (0.45)
1.94 (0.40)
2.06 (0.51)
3.00 (1.10)
0.29 (0.46)
2.91 (1.21)
3.15 (1.04)
0.28 (0.44)
3.07 (1.16)
2.81 (1.14)
0.31 (0.47)
2.71 (2.81)
q9: Apart from weddings, funerals and christenings, how often do you attend religious services these days? 4
4.39 (2.72)
4.53 (2.65)
9.24 (1.84)
2.70 (1.22)
0.76 (0.43)
2.91 (1.21)
0.81 (0.40)
2.42 (1.18)
0.71 (0.46)
0.93 (0.25)
0.82 (0.39)
0.93 (0.26)
0.85 (0.36)
0.94 (0.24)
0.78 (0.42)
0.50 (0.50)
0.68 (0.47)
0.98 (0.98)
0.99 (0.10)
0.49 (0.50)
0.71 (0.45)
0.99 (0.10)
1.00 (0.00)
0.51 (0.50)
0.65 (0.48)
0.97 (0.17)
0.98 (0.15)
0.93 (0.26)
0.91 (0.30)
0.94 (0.25)
0.13 (0.33)
1.84 (0.88)
0.11 (0.31)
1.95 (0.88)
0.15 (0.36)
0.15 (0.36)
All
Dependent variable:
Life Satisfaction
q5: Would you say that your family is wealthier etc.?
138
4.4
(4.1)
where,
LSit
Constant
Xit
Sit
Fit
Nit
Zi
personality
it
error term
Childhood happiness (LSit ) is a function (4.1) of a constant (C), time-variant socioeconomic variables specific to the individual (Xit) and time-invariant individual fixed
effects (Zi). The childs happiness at any time is subject to changes in their school
environment (Sit), interaction with friends (Fit) and natural environment life domains
(Nit), with unobservables manifest in the error term (it).
139
4.5
Life Satisfacition
9.5
8.5
7.5
6.5
AGE
Australia (HILDA)
Life Satisfacition
9.5
8.5
7.5
AGE
Figures 4.35(a) & (b): Average Life Satisfaction for 9 to 14 year old Australian children in the
Smart Train data and 15 to 90 year-old Australians in the 2002-2008 HILDA panel data
73
There is a structural break between in the view of average happiness shown in Figure 4.35a & b.
The (1-5) scale for the Smart Train happiness question is different to the (0-10) scale for the happiness
question in the HILDA. The Smart Train data were rescaled (0-10).
140
The first thing we see from the above figures is that the steep fall in 15 to 23 yearolds continues back to 9 year-old Australians74. To the 7.2% (-0.73 unit) decline in
happiness we noted in 15 to 23 year-old Australians in the HILDA75 (solid line in
Figure 4.35 a & b), we see a further 9.3% decline (9.44 to 8.56) in the happiness of
Australians aged of 9 and 14 years76 (dotted line in Figure 4.35a & b). The total fall
in the happiness of 9 and 23 year old for young Australians aged is 16.5% (- 1.61
units). This 16.5% fall in the average happiness of 9 to 23 year-olds is much larger
than the 7.8% decline we see in older Australians (aged 85 to 93 years) whose
happiness we expect to decline due to failing health and the onset of death. So, why
does the happiness of 9 to 23 year-old Australians fall by 16.5%? To answer that
question I begin by seeking to explain why the childhood happiness of the 9 to 14
year-old Australians falls by the - 0.88 units we noted above.
74
It is possible that the very high level of life satisfaction we see in young children is due to social
desirability bias; the idea that nice kids are happy. The downward trend in happiness with age could
also partly be due to diminution of this bias as the children get older. In addition, a volunteer sample
(which this sample is not because it was compulsory for the school children to go to the Smart Train)
might have a higher mean level of happiness than a strictly representative sample but, results showing
a downward trend in happiness with age could not be due to volunteering.
75
Average life satisfaction for 15-year-olds in the HILDA data are: females, 8.35, and; males, 8.52.
76
Average life satisfaction for 9-year-olds in the Smart Train data are: females, 9.24, and; males, 9.53.
141
fixed effect is what many psychologists now consider to be stable over a lifetime,
personality (McCrae, Costa, Mroczek, & Little, 2006).
our model of childhood happiness, we see some expected, and some unexpected
results. As expected, extraverted individuals are happier and neurotics (those low on
emotional stability) are less happy. The unexpected results are for conscientiousness.
If we regress conscientiousness on the overall happiness of all Australians in the
HILDA, we get a significant negative effect77. However, we get an opposite effect
for the 9 to 14 year olds in the Smart Train data. Unlike their adult Australian
counterparts in the HILDA, Australian children who exhibit conscientious
behaviours (orderly, systematic, efficient, neat, organised, and efficient) are happier.
There could be many reasons why conscientious has the opposite relationship with
the happiness of Australian children than for Australian adults. The most obvious
reason is that conscientious, hardworking children would be more likely to complete
their work and get better grades. Academic achievement has been found to make
children happier (Huebner, 1991). Another reason could be the regimented and
procedural nature of the Australian grade school system. At lower grade levels (1 to
9), the school curriculum dictates the childrens behaviour; the children have little
scope for planning what they do at school. The children are told to be tidy (orderly,
neat) and the teacher sets the deadlines (goals) for work completion (systematic,
efficient). To encourage the above-bracketed conscientious behaviours in grade
school children, teachers reward the children (e.g. a gold star to attach to their good
work). This positive reward recognition would make the child happier. Such an
ordered and rewarding life could relieve the children from the stress of planning the
pursuit and attainment of their life goals (McKnight, Huebner, & Suldo, 2002), at
least at school. In their 30-year review of subjective well-being (happiness), Diener,
Suh, Lucas, & Smith (1999) listed stress as a major component of happiness. In
77
In example, the next chapter looks at the effect of personality on the happiness of all individuals in
HILDA panel; the effect is significant and negative (-1.40, t-value = 2.22).
142
discussing adult happiness, Diener stated that believing ones goals are important
has rewards but can also increase stress because of the increased pressure to
achieve those goals (Diener, et al., 1999, p.285) and increased stress can make us
unhappy78. The planned and rewarding nature of the grade school system rewards
children who complete their work; it is reasonable to expect that those children
should get better school grades, be less stressed and happier in a rewarding school
environment.
However, what would happen to childhood happiness if the school environment
changed as children moved up in grades? Recall that school environment was one of
the four life satisfaction domain factors that Gilman (2003 proposed as having a
major effect on childhood happiness. Our Happiness Survey collected data on three
of Gilmans four childhood life satisfaction domain factors, school environment,
natural environment and interaction with friends. Regression results for
specifications 2 (a) to 2 (e) that include combinations of these factors are shown in
Tables 4.40 & 4.41 (Chapter 4 Appendix B). Looking at the results for specification
2 (c), the largest effect on childhood happiness comes from the interaction with
friends domain factor; a 45% larger positive effect than the positive effect from
school environment and 55 times larger than the (non-significant79) positive effect
from the natural environment domain factor. To gain a clearer understanding of what
this means, we need to look back at the behaviours relating back to each life domain
factor (Table 4.40). Children with positive interaction with friends behaviours such
as sharing, being kind to others, and forgiving of their friends, are happier than
children with positive school environment behaviours such as understanding the
teacher, helping classmates, concentrating on class work or learning easily.
78
The effect of stress on happiness is explored in the next chapter (Chapter 5).
79
If we look at the effect from the individual natural environment questions in Table 37, we see that
only one environment question (q17) had a significant effect on happiness. Children who perceived
their family as wealthier than their friends families are more likely to discuss environmental issues
within their family (q17 was strongly positively correlated with wealth).
143
While the regressions results provide a holistic view across the six school years of
our 9 to 14 year olds, a decomposition of the prediction for each domain factor
facilitates an internal view of the happiness changes arising from each domain factor
as the children move up from lower school grade 4 to grade 9 in high school. The
decomposition takes the form:
(y
grade 4
grade 9
) = ( xku xkl ) * k + c
k =1
and the school grade transition is from one year of schooling (l ) to the next (u ):
l to u
4 to 5
5 to 6
l, u
6 to 7
7 to 8
8 to 9
144
PredictedlifeSatisfactionChange
0.15
0.1
0.05
SchoolEnvironment
0
4to5
5to6
6to7
7to8
8to9
Natural
Environment
0.05
Interactionwith
Friends
0.1
0.15
0.2
SchoolGrade
Figure 4.36: The predicted changes in childhood happiness from each domain factor as the
children move up in school grade
From the results of the decomposition in Figure 4.36, we see that predictions for the
school environment, interaction with friends, and, natural environment factors can
account for 44% (-0.39) of the -0.88 unit fall in childhood happiness for 9 to 14 yearold Australians we saw in the raw data depicted in Figure 4.35. The total change in
childhood happiness arising from the natural environment factor is flat between
grades 4 & 9 is a very small -0.001, just 0.1% of the -0.88 unit fall in childhood
happiness for 9 to 14 year-old Australians we saw in the raw data depicted in Figure
4.35. Childhood happiness is little affected by the childrens natural environment.
145
However, the predicted grade 4 to 9 changes in childhood happiness arising from the
school environment factor is statistically significant80 and much larger. Between
grades 4 and 9 an accumulated -0.31 unit drop in childhood happiness can be
attributed to the school environment factor, which accounts for 35% of the total fall
(-0.88) in childhood happiness we saw in the raw data (Figure 4.35). If we look back
at the decomposition predictions in Figure 4.36, we can see that the decline in
childhood happiness is not linear. Between grades 4 and 7, while the children are still
in the lower grades, the decline in the happiness of the children is reducing. By the
end of their time in the lower grades (grade 7), the happiness of the children is
actually increasing (+0.1) by a small amount; perhaps the children have positive
expectations of their forthcoming transition to high school. Alas, when the children
do transition from the lower grade school to the high school, their happiness falls
steeply.
In the first year of high school the happiness of the children is predicted to fall by
-0.06 and by grade 9 the predicted fall is -0.2 units. The children progressively
become more miserable the longer they are in high school; perhaps because the
schoolwork becomes more difficult, they have to allocate more time to study work
and less leisure time is available to interact with friends.
The predicted net change in childhood happiness arising from the interaction with
friends factor between grades 4 to 9 is -0.08, much smaller than the -0.31 fall from
the school environment factor. The interaction with friends factor accounts for just
8.8% of the -0.88 unit fall in childhood happiness for 9 to 14 year old Australians we
saw in the raw data depicted in Figure 4.35. Like the school environment factor, the
predicted changes arising from the interaction with friends factor are also non- linear.
Looking again at Figure 4.36, we can see that the changes in childhood happiness
arising from the interaction with friends factor mirrors that of the changes from the
school environment factor, they have a medium positive correlation (r = +0.59). In
Figure 4.36, we can see that the decline in happiness arising from the interaction
80
See the regression results for specification 2(c), Table 4.41 in Appendix B of Chapter 4.
146
with friends factor reduces as the children move up in grades towards their transition
to high school. In their last two years in the lower grades (grade 6 & 7), the
happiness of the children rises due to their long-term interaction with friends. Alas,
after the children transition from lower school to high school the change in childhood
happiness arising from the interaction with friends factor is again negative.
We could expect such a change. The children build friendships in the lower grades
until, by the time they are in the highest grade of the lower school they are enjoying
considerable satisfaction (utility) from the regular interactions they are having with
friends they have known for the past seven years. At the end of year seven those
friendships are fragmented, the children are sent (most probably due to parental high
school choices) to one of the many high schools, schools they no longer share with
their old friends; childhood happiness declines (-0.15). After a year at high school,
the children make new friends, and, their interaction with friends again contributes to
their happiness (+0.05). However, while their interaction with friends makes them
happy, the school environment continues to make the children unhappy.
4.5.2.1
To summarize the findings from the analysis of the Smart Train data, the natural
environment factor contributed minimally (0.1%) to the decline in the childrens
happiness. The interaction with friends factor accounted for 8.8% of the -0.88 unit
drop in childhood happiness for 9 to 14 year olds. By far the largest contributory
factor to the decline in happiness of 9 to 14 year olds was their school environment.
The school environment accounted for 35% of the decline in the happiness of 9 to 14
year olds. Together, the school environment and interaction with friends factors
accounted for nearly half, 44% (-0.39), of the -0.88 decline in the happiness of 9 to
14 year old children we saw in the raw data depicted in Figures 4.35 (a) & (b).
However, we still need to explain the (-0.73) decline in the happiness of young of
Australians aged 15 and 23 years.
147
LSit = C + 1 X it + 2 Leit + Z i + it
4.2
where,
LSit
Xit
Leit
Zi
it
148
The results from the Usual Suspects + Health specification with the 15 to 23 yearold cohort are mostly as expected, income, marriage, home ownership, and
employment all have a significant positive effect on the happiness. Results that are
not expected are the positive effects from a greater number of children in a family
and the positive, but non-significant, effect from more years of education. We get a
significant negative effect for these variables in the same regression with the entire
All HILDA sample (Table 4.46). For each additional sibling, the happiness of 15 to
23 year-olds household increases by 0.12 (1.9%) while it decreases (-0.06 for each
additional child) in the entire sample. Children are happier when they have siblings,
it just makes their parents less happy.
Another regression result difference between the 15 to 23 year-old cohort and the
entire HILDA sample is the size of the effect from unemployment and maleness,
both are much smaller. The maleness effect is one third the size and non-significant.
The negative effect from unemployment for 15 to 23 year -olds is half the size as it is
for the entire HILDA sample even though more of them self-report as unemployed.
A count of unemployment by age would seem to support this proposition.
Unemployment for 15 to 23 ranges from 12% for 15 year olds to 5% for 23 year olds
with an average of 8.8%. The average self-reported unemployment rate for 15 to 65
year olds is much lower, just 3.6%, probably because 15 to 23 year-olds expect to be
unemployed because they are still engaged in education and not yet seeking
permanent employment. However, none of the effects from the variables in the
Usual Suspects + Health specification explains the steep drop (-0.77) in the
happiness of Australians aged 15 to 23 years, except one: age.
The age coefficient (-0.332) for the 15 to 23 year old cohort is seven times larger
than it is (-0.046) for the entire HILDA sample , and it remains so, even when we
add life event shocks (see Chapter 4 Appendix B Table 4.46 specification 2(a) &
(b)). Negative life events like losing a job or a death in the family do not explain why
15 to 23 year olds progressively become less happy as they age. Other than age
effects, none of the explanatory variables in our regression can explain away the
steep -0.77 drop in the happiness of 15 to 23 year-old Australians. Adding fixed
effects to the Kitchen sink specification does not explain away age effects.
149
4.6
Chapter 4 Limitations
150
Independent of whether the children or the teacher chose to visit the Smart Train,
the childrens level of English comprehension could have biased the data. Depending
upon their age, children differ in how they respond to survey questions. Borgers, de
Leeuw, & Joop (2000) are of the opinion that children 8 years old and onwards can
be surveyed. However, the comprehension abilities of a 9-year-old respondent to our
Happy Survey could be different to a 14-year-old respondent. While the survey
questions were pretested (on 6, 10 & 14 year olds), the questions were not pretested
on all ages and the children we pretested could have had higher (or lower)
comprehension abilities than the average child of their age.
In addition to sample representativeness and the comprehension abilities of the
children, differences in the school systems across Australian states could change how
we interpret the decomposition graphic (Figure 4.36). Recall, we saw an increase in
the happiness decline arising from the school environment and interaction with
friends life satisfaction domain factors as the children transitioned from grade 7 (the
final grade in lower school) to grade 8 (the first grade of high school). In other
Australian states, children go to high school a year earlier, in grade 7. It would be an
interesting natural experiment to collect data from Queensland children in 2015 after
Queensland changes to the same school system as the other eastern Australian states.
As a robustness test of the results of this study, the pre and post data could be
compared to see if changes in childhood happiness still correlated with the grade
when the children transition from the lower grade school to high school.
151
4.7
Chapter 4 Summary
Chapter 4 was an exploratory study that began by revisiting the steep decline in the
happiness of young 18 to 23 year old Germans that we saw in the Chapter 3 study.
Chapter 4 progressed by focussing on the Australian population and sought to clarify
confusion in the literature as to whether there was a similar steep decline, or increase,
in the happiness of young Australians. The Chapter contributed to our understanding
of lifetime happiness by extending our view of lifetime happiness back to childhood.
After developing scales to measure individual characteristics, personality and life
domain factors proposed by school psychologists to explain childhood happiness, an
Internet-based survey was developed to collect data from 9 to 14-year-old children
on their overall life satisfaction and on factors considered to affect childhood
happiness.
81
Domain models of wellbeing have been used for a considerable time in the scientific discipline of
psychology.
152
153
154
Figure 4.37a: The online Happiness Survey: initial screen and question q1
155
156
157
158
159
82
160
83
161
162
163
164
84
165
Figure 4.33l: The online Happiness Survey: concluding screen and questions q63and q64
166
2 (b)
Demographics
2 (a)
Personality
School environment
School environment
coefficient
t-value
coefficient
t-value
-0.023
1.72
-0.014
1.05
0.695
3.44
0.709
3.51
1.52
-0.178
1.84
-0.137
0.057
0.26
-0.066
1.69
0.102
2.72
0.055
1.50
0.142
4.54
coefficient
t-value
coefficient
t-value
0.732
5.56
0.429
2.66
0.617
3.17
4.903
7.45
Personality factors
extraversion,
agreeableness
-0.053
1.50
conscientiousness
0.072
2.85
emotional stability
-0.099
4.17
openness to experience
0.004
0.14
constant
8.906
0.063
Adjusted R
10.54
6.981
0.1969
0.1757
167
6.39
6.286
0.0739
12.63
0.0974
0.0927
Table 4.44: The determinants of Life Satisfaction for children aged 9 to 14 years in the Smart Train dataset; OLS regression, N = 389
2 (c)
3 (a)
Demographics Personality
School environment
School environment
2 (d)
Natural environment
Natural environment
Natural environment
Natural environment
School environment
Variable:
2 (e)
Demographics
coefficient
t-value
coefficient
t-value
coefficient
t-value
coefficient
t-value
0.011
0.79
0.013
0.95
0.532
2.66
0.701
3.46
-0.149
1.59
-0.129
1.43
0.023
0.11
-0.088
0.41
0.057
1.50
0.052
1.40
0.105
2.08
Personality factors
extraversion,
agreeableness
-0.123
1.44
conscientiousness
0.033
0.61
emotional stability
-0.097
4.02
openness to experience
-0.027
0.63
0.421
2.54
0.612
3.11
0.011
0.23
0.111
2.45
constant
4.877
7.30
8.172
23.24
5.867
Adjusted R
0.415
2.46
0.482
0.84
0.463
2.22
0.514
0.82
-0.003
0.07
0.035
0.75
5.71
6.884
6.26
0.0975
0.0153
0.1251
0.2001
0.0905
0.0127
0.1067
0.1724
168
Table 4.45: Other determinant variables of Life Satisfaction for the children aged 9 to 14 years
in the cross-sectional Smart Train dataset; OLS regression, N = 389
Variable:
coefficient
t-value
-0.018
1.31
0.733
3.40
-0.166
1.63
0.111
0.48
0.091
2.30
-0.107
0.83
Magic:
q6: good luck charms do bring good luck (1 definitely not true to 5 definitely true)
q7: Do you have a lucky charm such as a mascot or a talisman? (yes = 1)
0.129
0.52
-0.032
0.25
q10: Some fortune tellers really can foresee the future (1 definitely not true to 5 definitely true)
0.053
0.53
q11: Is there someone who cannot be seen by others watching over you? (yes = 1)
0.032
0.12
0.534
1.23
q8: Do you believe that a lucky charm can protect or help you? (1 definitely not true to 5
definitely true)
Environment:
q15: Are animals an important part of your life? (yes = 1)
q16: Are plants an important part of your life? (yes = 1)
q17: Does your family talk about the environment much? (yes = 1)
q18: Have you ever started a conversation about nature or the environment? (yes = 1)
q22: Let's say that in your neighbourhood everyone throws their garbage in the river; would
that be all right? (no = 1)
q23: Let's say that in New South Wales, a whole neighbourhood throws its garbage in the
river. Do you think it is all right or not all right for them to throw their garbage in the river?
-0.089
0.30
0.396
1.78
0.034
0.14
-0.579
0.68
1.029
0.92
-0.467
1.13
-0.177
0.56
-0.032
0.27
6.723
2.75
(no = 1)
q24: Do you think that throwing garbage in the river is harmful to the birds that live around
the river? (yes = 1)
Handedness:
q63: What hand do you write with? (left = 1)
q64: Which finger is longer? (1 my ring finger is longer; 2
constant
2
0.089
Adjusted R
0.042
169
Table 4.46: The determinants of Life Satisfaction; Pooled OLS regression results for 15 to 23 year-old cohort and All (15 to 92 year-olds) in the HILDA; N = 12,330
1 (a) 15 to 23 year-olds
Usual Suspects+Health
Variable:
1 (b) All
Usual Suspects+Health
2 (a) 15 to 23 year-olds
Kitchen Sink
2 (b) All
Kitchen Sink
coefficient
t-value
coefficient
t-value
coefficient
t-value
coefficient
t-value
-0.322
-4.31
-0.046
-28.73
-0.315
-4.2
-0.038
-21.8
age*age
0.007
3.29
0.001
35.61
0.006
3.22
0.001
29.38
ln household income
0.035
3.3
0.033
6.37
0.034
3.27
0.032
6.09
-0.031
-1.35
-0.092
-9.34
-0.032
-1.38
-0.088
-8.89
age
male
education
0.011
0.76
-0.061
-21.16
0.008
0.55
-0.064
-22.06
number of children
0.113
2.71
-0.063
-12.27
0.105
2.04
-0.061
-11.43
15.41
married
0.419
6.06
0.362
30.57
0.324
3.41
0.259
employed
0.076
2.58
-0.126
-9.26
0.084
2.84
-0.112
-8.27
-0.190
-4.27
-0.371
-12.74
-0.172
-3.85
-0.315
-10.76
regional income
0.000
4.27
0.000
6.65
0.000
4.23
0.000
6.01
home owner
0.093
3.47
0.148
12.15
0.086
3.2
0.135
11.09
unemployed
imputed rent
0.000
-0.09
0.000
-0.53
0.000
-0.1
0.000
-0.55
health
-0.528
-39.43
-0.527
-91.72
-0.522
-38.97
-0.524
-91.63
invalid
-0.187
-4.92
-0.108
-8.18
-8.03
-0.185
-4.88
-0.105
family death
-0.027
-0.72
0.016
1.03
divorced
-0.445
-0.94
-0.063
-2.89
0.023
0.75
partner dead
(dropped)
just married
0.103
0.83
0.101
3.21
just divorced
-0.825
-1.41
-0.192
-2.84
just separated
-0.267
-5.48
-0.419
-15.71
-0.266
-1.4
-0.249
-4.47
-0.070
-0.68
0.120
3.6
pregnant
just fired from job
constant
12.322
Adjusted R
16.99
10.246
149.430
0.145
2.09
0.125
4.51
-0.171
-3.08
-0.288
-9.89
12.282
16.9
10.189
147.13
0.1704
0.1672
0.1795
0.1750
0.1687
0.1670
0.1769
0.1748
170
Chapter Five
Do changes in the lives of our peers make us
unhappy?
85
This chapter is a peer-reviewed paper (Beatton & Frijters, 2009) presented at the 2009 HILDA
Conference at the University of Melbourne, Australia (HILDA, 2009). In 2011, this chapter was
submitted, as the joint paper with my supervisor Professor Paul Frijters, to the Journal of Happiness
Studies. The paper is titled Do changes in the lives of our peers make us unhappy? and is currently
with the reviewers. We would like to thank conference attendees, anonymous referees, and, seminar
participants for useful comments and suggestions.
171
5.1
Introduction
Whilst we have now had over 30 years of experience in running regressions on selfstated happiness, our ability to predict happiness has so far been rather poor. Recent
economics papers usually manage to explain about 15% of the cross-sectional
variance. For instance, Di Tella, et al. (2001, p.340), using over 250,000 observations
from twelve Western European countries, found that age together with demographic
variables like gender, education, employment status, income, marital status and
number of children explained 17% of the variance. Blanchflower & Oswald (2004),
using sixteen socio-demographic variables, explain only 9% of the happiness of
individuals in the US General Social Survey. Frijters & Beatton (2008), using a
kitchen sink set of nineteen socio-demographic variables and eight life event
dummies, explain just 8% of the happiness of individuals in the German
Socioeconomic Panel (GSOEP).
Only when one includes other subjective variables does the percentage of variance
really go up. Ferrer-i-Carbonell & Frijters (2004) thus add subjective health which,
added to the usual socio-economic variables, explained 26% of West German
happiness. Personality factors and mood are able to increase this to about 60%. Yet,
for economists, the explanation of one fairly subjective question by another is rather
disappointing and raises the spectre of endogeneity (Powdthavee, 2007). We would
prefer to be able to explain happiness by variables we can interpret as prices,
constraints and consumption. It is remarkable that, after more than a decade of
intense economic research in this area (see Clark, et al. 2008) for a meta-analysis),
we have come no further than explaining 15% with socio-economic characteristics,
just as Argyle, et al. (1999) reported that psychologists and demographers managed
to explain in the decades prior.
In this chapter, I hypothesise that there are two main problems with analysing
happiness at the individual level: unmeasured peer effects and measurement error.
Almost no dataset is able to track all the peers of an individual and all the subtle
interactions between them. As a result, we do not measure all the influences that
friends, family, and neighbours exert on us daily. A well-known example of missing
172
peer effects arises when considering income. Results from happiness regressions
consistently show a positive and significant coefficient for income; an increase in
income makes us happier (Clark & Oswald, 1996; Di Tella, et al., 2001; Winkelmann
& Winkelmann, 1998). However, is the magnitude of the regression coefficient and
the level explanation truly reflective of the effect of the income increase? Easterlin
(2001) reminds us of the importance of peer influences. Whilst we get happier if our
own income goes up, we get unhappier if the income of our peers goes up. We
adjudge our happiness relative to the peers we compare ourselves with (Falk &
Knell, 2004); all our friends, family and acquaintances. For a model of individual
level happiness to truly reflect the effect of the change in income on happiness, we
need happiness and income data for an individual and all that individuals peers. Of
course, this is not possible because panel surveys like HILDA and GSOEP follow
families, not individuals and all their peers; we have missing variables. This problem
may also hold for other variables like marriage events and children. Whilst the
individual who marries is happier during the wedding, those attending may feel
jealous and be unhappier. Childless individuals might be more miserable when their
friends have many children. Regression results for models of individual level
happiness that lack peer variables will suffer from bias if the observed characteristics
correlate with the unobserved peer effects.
As to measurement error, no dataset that we know of is capable of perfectly
measuring all the consumption variables economists think of as being important to
the utility of individuals. Indeed, no variable we usually put on the right-hand side
can be unequivocally interpreted as a certain unit of consumption of something. For
example, income is nearly always included in the list of explanatory variables but is
known to be measured with a great degree of error (due to recall bias, missing
compensation wage variation, contingent in-kind welfare, etc.) Even if it were
perfectly measured, it would still only be a proxy for what economists theoretically
think is really important, consumption. Another such example is marriage.
Researchers routinely add a marriage indicator in regressions, but not all marriages
are the same. Some marriages work well and produce lots of unmeasured
household goods, whilst others can be virtual prisons with negative production. Yet
all that remains of this heterogeneity in actual married life is a single marriage
dummy that is implicitly hypothesised to have exactly the same effect on everyone.
173
What holds for marriage could be argued to hold for every variable we routinely
include on the right-hand side: we ignore the measurement error involved in our
variables. Ignoring measurement error is almost unavoidable in any applied empirical
work using many variables, but it may be one of the key reasons for our inability to
explain more of the variation in happiness.
The approach, which I believe to be entirely novel in the economic happiness
literature, is to focus not on explaining individual happiness but on the aggregate
happiness of individuals of approximately the same age. We call this an aggregate
model of happiness. The advantage of this aggregation is two-fold. Firstly, when one
uses averages, measurement errors are dampened because the signal to noise ratio
increases. Secondly, and perhaps much more importantly, the average characteristics
by age are likely to coincide with the average characteristics of the peers. Hence,
peer effects that are almost impossible to identify at the individual level, because of
the inability to include all the relevant peers, come within reach when one averages.
Note that this does not mean we assume that the peers of an individual are others of
the same age. Rather, it means that the average peer of the average individual is of a
similar age86.
Drawing on the psychological literature, I then test the possibility that nearly all the
cross-sectional variance in happiness is due to stress. I measure stress not by using
subjective questions, but rather by measuring the believed cause(s) of stress: a
weighted average of the frequency of life events that psychologists have argued are a
major cause of stress. I test this hypothesis on the Household Income Dynamics for
Australia, progressively expanding our aggregate model of happiness. By comparing
the effect of stress at the aggregate level with the effect of the same life events at the
individual level, we can also say something about the likely direction of peer effects
and hence something about the data we are missing at the individual level. After
testing the basic idea, which holds remarkably well (over 80% of variance is
explained by a simple weighted average of life events), I expand the basic model to
allow the effect of stress to differ by personality, but first let us look at the data.
86
174
5.2
The Data
In this chapter I use the first six waves of the Household, Income and Labour
Dynamics in Australia (HILDA) Survey. This is a householdbased panel study
which began in 2001 (HILDA, 2008b). It has the following key features:
Interviews are conducted annually with all adult members of each household.
Wave 6 (2006) tracks 12,905 individuals with 95% retention from Wave 5.
The happiness question is based on the Fordyce (1973) Global Happiness Scale87. It
asks All things considered, how satisfied are you with your life? with the ordinal
responses ranging from 0 (very unhappy) to 10 (very happy). It seeks to measure the
aggregate utility from all the good and bad things that occur throughout our lives
(Fordyce, 1988). Table 5.47 shows the sample averages for the 55,177 person-year
observations we have available. Average life satisfaction is 7.94, which is relatively
high compared to other Western countries (see Clark et al. 2008).
87
The HILDA user manual (HILDA, 2008a; p.144) notes that the HILDA questions relating to life
satisfaction domains is based upon the work of (Cummins (1996). Fordyce (1998, p.357) notes that
the Global Happiness Question arises from the early work of Wessman & Ricks (1966).
175
s.d.
Min
Max
11,035
155.16
10,869
11,255
happiness
7.94
1.49
10
age
43.61
17.67
15
93
age*age
2214
1672
225
8649
time* time
2858.5
2048.03
225
7056
5.188
3.145
9.195
1054.29
1050.01
9845
97.28
162.67
3000
female
.530
.499
education years
12.68
1.785
18
married
.520
.500
separated
.035
.183
never married
.231
.421
divorced
.087
.282
widowed
.048
.214
employed
.646
.478
unemployed
.034
.180
disability
.234
.423
health
3.391
.959
3.075
.688
Variable:
Tables 5.44 & 5.45 show the sample averages of the life-events used to construct the
measure of stress. As one can see, the HILDA includes more self-reported life events
(21, versus 7 in the GSOEP). In addition, there are many recorded life events by
category. For instance, per person-year observation, 0.059 change jobs. That is
almost 3200 job changes over the 6 years of the sample. Similar high numbers of life
events hold for all the other categories.
176
Table 5.48: Sample averages for life events affecting individuals in the HILDA; N = 55,177
Mean
s.d.
Min
Max
spouse/child death
.007
.078
.87
death of a relative
.086
.246
.79
personal injury
.067
.218
.78
jailing of self
.002
.034
.76
.121
.296
.72
.039
.160
.70
victim of violence
.012
.091
69
just separated
.027
.131
.66
just reconciled
.007
.070
.66
.019
.108
.64
worsening finances
.018
.104
.62
death of friend
.066
.189
.61
friend jailed
.007
.064
.56
just married
.011
.068
.43
.059
.148
.43
just pregnant
.021
.089
.41
moving house
.061
.133
.35
improved finances
.011
.058
.33
promoted at work
.021
.081
.33
birth of child
.011
.060
.33
just retired
.006
.041
.28
Variable:
177
5.3
5.3.1
Theoretical framework
Suppose the true model of the happiness (GSit) of individual i in period t is given by:
(5.1a)
Here, GSit is affected by an individuals own circumstances (Xit) and those of the
peers (Xpeerit), as well as random errors (uit). Suppose now that what is usually
estimated in empirical happiness regressions is the following:
GSit = Z it + vit
(5.1b)
Z it = X it + eit
Here, Zit is now a noisy measure of (Xit) that includes random measurement error
(eit). In the absence of peer effects, it is well known that the estimate of will be a
downward biased estimate of the true s because of the presence of measurement
error. In our case though, there is both measurement error and missing variables.
What we then get as the asymptotic estimate of is (neglecting errors that go to zero
as i and t go to infinity):
cov( Z it , GSit )
X2 s
cov( X it , Xpeerit )
=
+
=
2
2
2Z
X2 + e2
X +e
which is biased in two directions: biased towards zero because of measurement error
and biased in an unknown direction (because we have no a priori expectation of the
sign of ) due to the correlation between individual characteristics and the peer
characteristics.
178
1
Nt
iS ( t )
GSit =
1
Nt
iS ( t )
Z it + wt
(5.1c)
Where S(t) is the set of individual-year combinations of approximately the same age t
and Nt is the number of observations on age t. If we now presume that:
1
Nt
1
X it =
Nt
i =1
Xpeer
i =1
it
then both the measurement error problem and the peer effects issue get solved in
the sense that the asymptotic estimate of the parameter now becomes:
= s +
which occurs because averaging gets rid (asymptotically) of the measurement error,
and the assumption that the average peer is the same as the population average of the
same age means we obtain a coefficient whose estimate we can interpret as the sum
of the individual and peer effect. It is important to point out that this procedure gives
very different results to simply including the average X by age in equation (1a)
because the correlation between the characteristics of the actual peers of each
individual and our artificial aggregate peer may be very small88. It is only for the
aggregate of individuals of the same age that we assume the aggregate peer has the
same characteristics as themselves.
88
The obvious disadvantage of averaging is that we have far fewer observations than before: from
55,000 person-year observations, the data is reduced to a mere 70 different age-happiness points. This
means we should apply extreme care when choosing which variables we wish to include and hence we
wish immediately to choose a variable that can be argued to be responsible for a lot of variation in
happiness.
179
Typically, our peers are those with whom we have something in common. Those
referent few with whom we compare ourselves; our friends, family, neighbours and
work colleagues. In models of individual of happiness, peers are often identified as
those who have a similar income (Becchetti, Trovato, & Bedoya, 2011), education
level (Mora & Oreopoulos, 2011; Ng, 2002) or age (Glaeser, Laibson, & Sacerdote,
2002) to ourselves. However, the model in this study is an aggregate model of
happiness and the peer effects emerge from the aggregate societal peer. This study
takes the position that the aggregate peer is someone from the same age-band. We
grow up in age-bands, armies are organised in age-bands, old-age care is roughly
organised in age-bands, and many laws are age-specific (e.g. when we need to start
school). Changes in the socio-economic situation of peers in our age-band have an
effect on our happiness throughout our lives.
Let us look at some examples of how change in the socio-economic variables that are
known to affect happiness can occur in age bands. To begin, we have laws that
dictate when we have to start school; in Australia, this is age five. During our school
years, our happiness can be affected by what happens to us and our classmates; who
must be in the same age-band as ourselves. Our peer in our age-band gets high
marks, they are happy, we are less happy. Let us look at what happens as we age.
According to The Australian Council for Educational Research (Marks, 2007), young
adults leave high school, enter then complete university or tertiary training within the
first three years after leaving high school; in an age-band from 20 to 23 years of age.
As a person in the 20 to 23-year age-band you and your peer look for your first job.
As an average graduate with average ability, your peer gets an average job with an
average income, but you get a higher paying job. You are happy but your peer in the
same age-band is less happy.
We move through the working period of our lives in age-bands. Most people get
work promotions and income increases in a certain age-band. For Australians in the
HILDA, on average, income increases occur in the age-band of 38 and 43 year of
age. We also retire with in age-bands. The Australian Bureau of Statistics reports that
on average Australians retire at age 58 (ABS, 2008). Your wealthy work peer, who
on average is most likely to be in our age-band because you both started school at the
same time, retires early but you have to and continue to work because you dont have
180
the resources to retire; she is happy and you are not. Peers in our age-band continue
to have an effect on our happiness throughout our lives.
It is not just work peers in the same age-band that affect our happiness. Our peers are
often our neighbours, and, neighbourhoods often include people in the same life
stage who are of similar ages (Feld, 1984, p.641). We saw in (Figure 4.32, p. 119)
that Australian women have babies in the age-band of 27 and 36 years. If a
neighbour (a peer that is most probably in our age-band) were to become pregnant
that would make them happier but it could make you unhappy if you could not get
pregnant. Similarly, our happiness is affected by failing health within particular ageband(s) because health conditions such as cardiovascular disease, diabetes and
arthritis or osteoporosis particularly are associated with people at particular ages
(ABS, 2008). On average, changes in happiness affecting socio-economic variables
like education, income, job status, retirement or our health situation among others are
more likely to occur in age bands across our lives.
To capture this age-band, happiness and independent variable data have been
smoothed about age. The happiness and independent variable(s) at each age is an
equally weighted smoothed average of the variable two years prior to the age, the
age, and three years after the age; a six year equally weighted smoothed average. The
obvious disadvantage of averaging is that we have far fewer observations than
before: from 55,000 person-year observations, the data gets reduced to a mere 70
different age-happiness points. This means we should apply extreme care when
choosing which variables we wish to include and hence we wish immediately to
choose a big variable that can be argued to be responsible for a lot of variation in
happiness.
The proposed big variable that is constructed to explain happiness change over the
lifetime is the stress arising from life events. Richard Easterlin (ed., 2002; 2006)
argues that life is a succession of little mishaps and triumphs that determine how we
feel in the short-run, and big events that determine how we feel in the medium term
(say, a year). Imagine the difference in our response on an average day versus the
response on a rainy day where we had missed our train and walked to work sans
umbrella. We may have been perfectly happy before all this happened and self-rated
181
a happiness level of seven. However, having been asked the happiness question after
getting wet and missing our train, and because we now feel miserable, we record a
happiness level of five. Similarly, how we feel about a whole year will depend on the
various positive and negative life events our peers and we have experienced.
Therefore,
The SRRS levels emerged from a US study of 3,122 individuals and the review of the results by a
panel of 30 professionals from the behavioural, medical, and social sciences (Hobson, et al., 1998).
The professional review panel was representative of the gender and ethnic diversity of present day
U.S. society: 15 (50%) panel members were female; 15 (50%) male; 3 (10%) were African American;
3 (10%) Hispanic; 2 (7%) Asian, and; 22 (73%) white.
182
1950). The more salient, and unexpected, the life event the greater the level of stress
from that event (Hobson et al., 1998). Also, stress is believed to be cumulative so the
more events affecting us at a particular time in our lives, the greater our aggregate
level of stress (Carlopio, Andrewartha, Armstrong, & Whetten, 2001) and the less
happy we are. For now, we take the SRRS weights as given, though we will return to
the issue of whether these weights are really reasonable later. Tables 5.46 & 5.47
show all the life events considered by SRRS theory, highlighting those life events
available in our HILDA data set.
Table 5.49: Stress levels defined by the Social Readjustment Rating Scale90
Stress Level
90
Life Event
.87
Death of a spouse
.79
.78
.76
.72
.71
Foreclosure on a loan/mortgage
.71
Divorce
.70
Victim of crime
.69
.69
Infidelity
.69
.66
.66
.64
Being fired/laid-off/unemployed
.62
.61
.59
Surviving a disaster
.59
.56
.56
.56
The highlighted and italicised life events are in the HILDA panel data waves 2 to 6.
183
Table 5.50: Stress levels defined by the Social Readjustment Rating Scale (continued)
Stress Level
Life Event
.53
.53
.51
.50
.48
.47
.46
.45
Employer reorganising/downsizing
.44
.43
Getting married/remarried
.43
Changing employers/careers
.42
.41
Pregnancy of self/spouse
.39
.39
.38
.37
.35
Change in residence
.34
Our measure of Stress by age is the smoothed simple average of the life events,
weighted by the SRRS-based stress level impact estimates (Table 5.49 & 5.50):
Stress t =
SR R S s * L e st 1
Which defines Stress as the sum of, the life events (Le) occurring in the previous
period (year91) weighted by the stress level (SRRS) for each type (s) of life event92.
91
While stress is calculated at each age, the resultant time series data (Figure 5.39) are smoothed with
a simple moving average that equally weights t-2 to t+3; therefore, a peer is defined as someone two
years younger to three years older than self.
92
Tables 5.46 & 5.47 highlight the stress level for each life event based on SRRS theory and the
highlighted life events are the twenty-one life events used from the HILDA.
184
GS t = C + ( Stress t ) + t
(5.2)
8.67
-1.18
(215.43) (17.21)
R2 = .81
empirical fit between happiness and stress by showing happiness by age and stress by
age; stress over a lifetime looks like the inverse of lifetime happiness. At age fifteen,
we have a higher level of happiness because we have been exposed to less stress. A
steep increase in stress between the ages of 15 to 19 years negatively correlates with
the steep decline in happiness we saw in 15 to 23 year olds in Chapter 4. Perhaps
stress really is explaining some of the change in happiness over a lifetime. As we age
further into our mid-years, we are exposed to more stress-creating life events and
subsequently become less happy. As we grow older, we are subject to less stress and
this leads to an increase in our happiness. Interestingly, over a lifetime, stress and
raw happiness appear as a U-shaped inverse of one another (Figures 5.38 & 5.39).
185
8.6
Average Life Satisfaction
8
8.2
8.4
7.8
7.6
20
40
60
80
Age
.2
Stress
.4
.6
.8
20
40
60
Age
186
80
These two figures are also informative in the sense of the time series properties of the
two variables. It is known that if one regresses two lines with strong trends on each
other that one gets a high spurious relationship. This is clearly not the case with our
data , life satisfaction is much the same at age 18 as it is at age 80. Stress goes up,
plateaus, and then almost linearly reduces93.
A peculiar, so far implicit, aspect of the regression results for equation (5.2) is that
all life events affect aggregate life satisfaction negatively. This view is consistent
with Social Readjustment Rating Scale (SRRS) theory from psychology, but differs
considerably from what we see from economic models of individual happiness. In
economics, negative life events like unemployment or declining health have been
shown to decrease our happiness (Clark & Oswald, 1994; Wilson, 1967) while
positive life events like marriage or the birth of a child lead to increased happiness
(Frey & Stutzer, 2005). Essentially, equation (5.2) presumes all those events that
seem to be positive at the individual level are in fact still negative at the aggregate
level due to missing peer effects. Even though Table 5.54 (Chapter 3 Appendix B)
shows that marriage, promotion and financial improvements increase life satisfaction
at the individual level, equation (5.2) presumes they decrease life satisfaction at the
age level because of the negative effect of these events on peers.
I test if this really holds at the aggregate level by applying model (5.3) and splitting
the Stress variable into Positive_Stress and Negative_Stress where positive stress is
made up of those life events with a positive effect on the individual (see fixed effect
regression results in Table 5.53 in Chapter 5 Appendix B). Figure 5.40 shows the
evolution by age of this positive stress and negative stress.
93
The impact of life-events is not cumulative: there is adaptation to each life vent, implying that new
shocks are needed to sustain a low level of happiness. When we get older, the lack of new shocks
allows us to retain high levels of happiness.
187
.5
.4
.3
.2
.1
0
20
40
60
80
Age
Negative Life Events
Figure 5.40: Average stress from positive and negative life events; Australians aged 15 to 84
Looking at Figure 5.40, we see that positive events happen more in mid-life
(promotions, marriages, income increases), whilst negative events are more
concentrated earlier on (injuries to family members, crime).
When we look at how positive stress and negative stress affect happiness, we get:
-1.76
(5.13)
R2 = .82
(5.3)
-0.57
(1.58)
When disaggregated into positive and negative events we find that both positive and
negative life events reduce happiness, but that the effect of positive life events is
stronger and more significant. This is quite revealing. Why do life events that are
positive at the individual level, suddenly become negative in the aggregate? Within
the context of equation (5.1), the reason is the peer effects: what makes us happier at
the individual level can increase the jealousy, frustration, and hence stress levels of
188
our peers. On aggregate, it is clear that the peer effect dominates the individual
effect. More unhappiness is created by promotions, marriages, births, etc., via our
peers than we gain personally. This obviously has very strong policy ramifications
since it would mean nearly all life events not essential for our continued survival
should be reduced to a minimum, ceteris paribus.
So far, I have relied on objective variables to explain life satisfaction. These life
events do not suffer from endogeneity problems to the same extent that, for instance,
health or mood does: it is not our unobserved individual proclivity to be happy that
causes our friends to marry and get promotions. We now introduce more subjective
variables and turn to the hypothesis that stress may not be equally bad for everyone
and that the importance of personality for life satisfaction is mainly in terms of how
personality allows us to cope with stress.
Psychologists have long argued that the level of stress is not only affected by the
number of life events, but is directly affected by personality (Mroczek & Kolarz,
1998). I test this in steps. Firstly, I look to see if there is any residual effect of
personality on happiness, after which I test the mediating effect of personality on
stress. The direct effect of personality is usually argued to hold mainly for
extraversion and emotional stability. Costa & McCrae (1980) and Headey (2008)
identified an increased variance in the happiness of extraverts (talkative, outgoing,
lively) and neurotics (moody, touchy, jealous, temperamental) who exhibit lower
levels of emotional stability. In (5.4), we look at the direct effects of personality
traits on happiness (GS):
-0.21
(0.29)
3.84
(5.62)
R2 = .88
189
-0.74
(1.03)
-1.40
(2.22)
-1.53
(2.43)
The personality traits are the average at each age for the individuals measured in the
HILDA on a scale of 1 (lowest) to 7 (highest), using Goldbergs Big-Five personality
factors:
Pa
agreeableness
Pex
extraversion
Pem
emotional stability 94
Pc
conscientiousness
Po
openness
The thirty-six items tapping personality in the HILDA are based on Saucier's (1994)
edited version of Goldberg's (1990) Big-Five personality factors.
Wave 5 of the HILDA (2008b, p. 10) measured personality traits on a seven-point
scale and the five trait factors are composed by taking the average of the items
(Losoncz, 2007). The higher the score from the items in Table 5.51, the better that
personality trait describes the respondent. The Big 5 personality traits and related
behaviours are:
Extroversion talkative, bashful (reversed), quiet (reversed), shy (reversed),
lively, and extroverted.
Agreeableness - sympathetic, kind, cooperative, and warm.
Conscientiousness - orderly, systematic, inefficient (reversed), sloppy
(reversed), disorganised (reversed), and efficient.
Emotional stability - envious (reversed), moody (reversed), touchy (reversed),
jealous (reversed), temperamental (reversed), and fretful (reversed).
Openness to experience - deep, philosophical, creative, intellectual, complex,
imaginative.
94
190
191
Explaining each personality trait in more detail, openness refers to the extent to
which people are sensitive, flexible, creative or curious. Low scored individuals tend
to be more resistant to change and less open to new ideas, they are more fixed in
their ways. Agreeableness refers to traits where we are courteous, good-natured, kind
and considerate of others. This personality trait develops trust between individuals.
People with low agreeableness tend to be uncooperative, short-tempered and
irritable; they are hard to deal with. Conscientiousness refers to people who are
careful, dependable and self-disciplined; they have a will to achieve. Low
conscientiousness tends to predict carelessness, disorganisation and sloppy work.
Emotional stability and extraversion are the two traits that are most considered to
impact on happiness (Costa & McCrae, 1980; Diener, Sandvik, Pavot, & Fujita,
1992; William Pavot, Diener, & Fujita, 1990; Sahoo, Sahoo, & Harichandan, 2005).
Individuals exhibiting a low level of emotional stability (high in neuroticism) suffer
from negative affect and dissatisfaction while those high in extraversion exhibit
positive affect, satisfaction, and higher levels of happiness (Costa & McCrae, 1980;
Furnham & Petrides, 2003).
Contrary to early opinion of those like Freud who stated that our personality is fully
formed by adolescence, others have argued that our personality traits, particularly in
younger individuals, can change over time (see McCrae, Costa, Mroczek, & Little
,2006 for a review). Such a change manifests in the HILDA (Figures 3a to 3e), albeit
very small personality trait changes. On average95, Australians become more
agreeable (+2.4%), less extraverted (-1.6%), more emotionally stable (+7.9%), more
conscientious (+5.9%), and less open (-5.2%) to changes over their lifetime96.
One needs to be careful when interpreting this evidence of changes in personality
traits over time. There is a large body of opinion in the psychology literature that the
changes in the personality traits over time arise from measurement error (Ehrhardt,
Saris, & Veenhoven, 2000) and thus, personality is thought to be stable over age and
gender (McCrae, et al., 2002), others disagree (Rantanen, Metsapelto, Feldt,
95
While there is no such thing as an average person, we are individuals; Table 5.49 in Chapter 5
Appendix A provides personality trait averages, by age, for Australians in the HILDA.
96
Assuming the five pooled years of the 2002-2006 HILDA panel data are indicative of a typical
lifetime.
192
Pulkkinen, & Kokko, 2007). Over time, subjects respond differently when asked to
respond to the same Big-5 personality questionnaire (this could not be the case in our
dataset because personality was measured in just one year of the six waves in our
panel dataset (wave 5, 2005). Other psychologists still assert that an individuals
personality traits do change over time. In adults aged 33 to 42, Rantanen, Metsapelto,
Feldt, Pulkkinen, & Kokko (2007) found that neuroticism decreased over time and
extraversion, openness to experience, agreeableness, and conscientiousness increased
from age 33 to 42 (we can see some of these finding in Figures 5.41a to 5.41e). The
jury is still out on whether personality traits change over time, are subject to cohort,
or peer effects97. Given that the personality trait data in the HILDA are crosssectional (wave 5 only); researchers could better pursue the question of personality
change over time if data were available tracking individuals over a lifetime. To do
this, the socio-economic panel surveys would need to include the personality
questions in every annual survey.
97
It is possible that the personality of a particular age cohort could have been formed by events
particular to that cohort; in example, children exposed to food scarcity or traumatic events during
world war two ( todays 70 year olds).
193
4.05
4.22
4.3
4.15
Figures 5.41a to e: Change in personality factors over time for Australians aged 15 to 84; scale is 1 to 7
20
40
60
20
80
40
60
80
Age
Age
4.15
4.1
4.35
4.4
20
40
60
80
Age
194
20
40
60
Age
80
3.9
4.1
Figure 5.41a to e (continued): Change in personality factors over time for Australians aged 15 to 84; scale is 1 to 7
20
40
60
80
Age
195
Looking back at our regression results in Table 5.53 (Chapter 5 Appendix B),
personality has a direct effect on happiness and increases the level of explanation
from R2 = .82 to .88 whilst the effect of Stress reduces by thirteen-per cent. Stress
remains by far the most important variable, which, from an economists point of
view, is heartening because it suggests real events can trump subjective perceptions
in the ability to explain life satisfaction.
Looking at each trait, extraversion has a significant positive direct effect on
happiness (Table 5.53). Extraversion and neuroticism impact life satisfaction through
daily emotional experiences (Howell, 2006). Extraversion is associated with a
positive outlook on life, better health, higher levels of success in marriage, work and
other aspects of our lives; this positive reinforcement makes us happier
(Lyubomirsky, et al., 2005). Conscientiousness and openness make us unhappier.
Conscientiousness impacts happiness through daily behavioural choices (Howell,
2006). We can imagine the systematic, procedural organised person getting unhappy
when life event changes upset the equilibrium of their orderly environment. The
openness items in the personality measure tap the notion of intelligence (creative,
intellectual, imaginative). Cognitive level (intelligence) is highly positively
correlated with education level (Rindermann, 2008) and higher levels of education
translates to reduced happiness (Clark & Oswald, 1996). We can imagine a creative
intellectual worrying about the problems of the world (like the environment) and
seeking answers to improve the situation, at the expense of their happiness. The
personality variables that have an insignificant direct effect on happiness are
emotional stability and agreeableness.
Agreeableness relates to those who exhibit sympathy and are warm, kind and
cooperative towards others. The coefficient is non-significant but negative, probably
because agreeableness acts indirectly through daily behavioural choices (Howell,
2006). The non-significance and negative coefficient for emotional stability defies
current literature. Costa & McCrae (1980) and Furnham & Petrides (2003) found that
emotional stability positively affects happiness. Perhaps this is because emotional
stability impacts happiness through daily emotional experiences (Howell, 2006). We
need to examine the indirect effect of personality on the life events that impact our
lives and lead to the stress that affects our happiness.
196
where
and
1
2
= 3
4
5
1
2
= 3
4
5
Pa
Pex
P = Pem
Pc
Po
.
197
(5.5)
The first striking aspect is that happiness is almost completely explained (R2 = .95)
by this set of variables98. Since we are now up to 11 variables explaining 70 data
points, a high R2 was to be expected, but 0.95 is simply a novelty in this literature.
The inclusion of the indirect effects of personality (P) and Stress has reduced the
direct effect of Stress by 56%, which suggests strong intermediate effects of
personality on the experience of stress. In addition, the direct effect of the personality
traits has changed; openness and conscientiousness have both become insignificant.
Only extraversion remains as a mildly significant personality factor directly affecting
happiness. Otherwise, the effect of personality is entirely through the life events
leading to stress, with the main interactions being for the direct positive effects of
extraversion and the negative effects from stress (Figure 5.42).
Big5
Personality
Traits
Agreeableness
+
Extraversion
Emotional
Stability
Happiness
Conscientiousness
Openness/Intellect
LifeEvents
Stress
Figures 5.42: The role of the direct and indirect effects from personality on life events and
the stress arising from those life events
98
Personality traits should be viewed as causally antecedent to life events, they are partly genetic.
Traits predispose people towards experiencing specific patterns of events. In example, those who
are high in the neuroticism trait experience more negative events than those low in neuroticism.
As supported by the regression results (Table 5.50, p.195), the total effects (both direct & indirect
via life events) are strong;, and explain much of the variance in happiness.
198
199
5.3.2.1
In order to see whether the main results are highly dependent on the SRRS scales, we
ran equation (5.2) in Table 5.54 with the 11 life events that made the biggest
contribution to aggregate stress, adding them in sequentially. When we include all 21
life events (not shown), standard deviations become very large and all significance is
lost. Table 5.54 shows that there is general non-robustness of the effects of individual
life events. Personal injury for instance has a strongly negative effect in the first four
specifications, with a coefficient of -0.014 if it is included as the only life event.
When eleven life events are included, personal injury has a coefficient of -0.001 and
is non-significant. Similar parameter instability holds for being a victim of violence
(which has a positive coefficient!) and financial stress, which we attribute to the
strong multi-collinearity between the frequencies of these life events.
Despite the multi-collinearity problem, which makes it difficult to take the relative
magnitudes at face value, Table 5.54 does confirm that positive life events can have
strong negative effects on aggregate happiness. Having just married and an
improvement of finances all have significant negative aggregate effects, whereas they
are strongly positive at the individual level. Indeed, the negative effect of
improvement in finances is the single highest coefficient in the final specification.
Interestingly, the other significant negative variables are separated and worsening of
finances. It is tempting to think of this group of variables as highly visible variables
that are likely to affect friends and families. Given that we do not want to put too
much emphasis on these results due to the multi-collinearity problem, we do not
want to overplay this interpretation and merely note that the main thrust of the
analyses based on a particular weighting of the life events is also evident if we use
unweighted aggregate life events.
200
5.4
Chapter 5 Limitations
The first limitation of the Chapter 5 study applies to all studies in this thesis. The
studies use panel data from western countries. There are cultural differences
between Western and Asian countries (Hofstede, 1983). While an extraverted
person may be more desirable and therefore happier in an individualistic western
culture like Australia, an extraverted person in a collective culture (e.g. China)
may be ostracised and therefore unhappier because they are not complying with
the cultural norm of their (Chinese) group; generally introverted. It would be
interesting to analyse how cross-cultural personality differences affect happiness.
Research that considers happiness, adaption, as well as the cultural norms may
help to explain if and why some immigrants contribute more slowly (faster) to
the economic growth of their adopted country.
The second limitation of the Chapter 5 study arises from a claim that the dependent
variable of the study is not necessarily life satisfaction but the age profile of life
satisfaction. The aggregate happiness variable was constructed from the individual
happiness question data. This constructed aggregate variable is the dependent
variable in the regressions that seek to explain happiness change over a lifetime,
happiness change as we age. The left-hand side variable is the age profile of
happiness and abstracts from the vast majority of variation in happiness. As such, all
the statements in the study apply to this aggregate variable, not interpersonal
difference.
The next limitation concerns whether the age-profile in happiness would be well
explained by anything that looks U-shaped or inverted-U shaped. The selection of
stress as the independent variable did not emerge from a process of test and selection.
The decision to use stress as the independent variable is theory-based and its
construction is consistent with the application of the well-accepted Social
Readjustment Rating Scale Theory from psychology (Holmes & Rahe, 1967; Hobson
et al., 1998). The stress variable was not selectively chosen because it looked Ushaped or inverted-U shaped, its choice was founded upon a large existing
psychological literature.
201
Another limitation is the acceptance of the scaling of the life events with the SRRStheory-based weights to form the aggregate stress variable. Interacting personality
directly with non-SRRS weighted life event dummies produces a very different stress
profile in age. Ideally, econometricians would prefer that weights come from the data
and not be imported from the theory of another scientific discipline. Frijters,
Johnston & Shields (2009) used this approach when they sought to contribute to our
understanding of how to weight the effect of each life event type on the overall
wellbeing of individuals. They used quarterly life event data to model anticipation
and adaption to a small number of life events that included changes in financial
situation, marital status, death of a close relative, and becoming a victim of crime.
Speaking with the authors, the reason they chose a small number of life events from
the 21 available in the HILDA was because some excluded events replicated the
included events (e.g. pregnancy, death of other close relative). The ten events they
did use were the only ones that occurred often enough to provide significant
regression results. The authors were not only concerned about the small number of
observations per life event type; they also voiced their concern about measurement
error in the self-reported quarterly life event data.
HILDA subjects respond to the survey annually but are asked to recall whether a life
event occurred one, two, three or four quarters ago. Can subjects really recall how
many quarters ago a life event occurred? Certainly, we could recall a major event
like the death of a spouse. However, as we see in Figures 5.43a to 5.43k, the total
stress at each age arises from the accumulation of stress from a large number of
minor individual life events. Can we truly remember when minor individual life
events occurred? Was it three or four quarters ago that my finances worsened, or did
I acquire that injury on the 28th of June (Quarter 2) or on the 3rd of July (Quarter 3)?
Measurement error in quarterly life event data is a major limitation to a happiness
study. Looking at the effects of a partners life events on individual happiness,
Mervin & Frijters (2011) uncovered so much noise in the life event data that the
causal effects were very difficult to identify. Measurement error in life event
reporting is of particular concern to this economics of happiness researcher because I
need to know when an event occurred if I am to accurately gauge the happiness
change from each life event category. As to there not being enough of each life event
202
category in the data to identify their effect on happiness, we will just have to wait for
additional waves of the HILDA panel data. In the meantime, researchers might
consider using the aggregate approach from this study.
5.5
Chapter 5 Summary
This chapter sought to account for life event shocks by gauging if their stressful
effect explained changes in aggregate happiness over the life cycle. The advantage of
looking at the aggregate level of happiness was that it solved the problems of missing
peer effects and measurement error that plague models of individual level happiness.
The key assumption under which this aggregation allowed us to say something about
peer effects is the assumption that the aggregate peer of the aggregate individual is
someone of approximately the same aggregate age.
It was hypothesised that happiness is almost entirely explained by the direct effects
of the stress from the life event shocks, mediated by personality. We found that the
use of the stress variable could indeed explain over 90% of the variation in aggregate
happiness. Of particular interest is the finding that both negative and positive life
events bring about aggregate unhappiness. It might not be surprising that negative
events stress us. Aristotle already said that we humans focus our energies in pursuit
of virtuous happiness and hence become unhappy at negative events: we plan, set
expectations, and are delightedly happy when our plans are achieved and our
expectation met; yet we are disappointedly unhappy when they are not (Aristotle,
1819, p. 234, 254, 257). However, the finding that positive events (the fruits of our
planning, if you like) bring aggregate unhappiness makes no sense at the individual
level. Our interpretation is that the positive feelings we individually get from
promotions, pay rises, births, etc., are swamped by the stress this causes amongst our
family, friends and workmates.
The happiness maximising policy recommendation, ceteris paribus, is that we should
minimise life event shocks on society. All changes that are not essential to
procreation and minimum needs appear to lead to net loss of life satisfaction. At face
value, this would mean that divorcees should be taxed because their actions have
203
negative effects on their peers; people who go to jail or move house should also be
taxed to compensate the misery they are causing their neighbours and friends, etc.
These fairly radical conclusions require deeper examination99. Replication of our
results to other countries would generate more variation and would allow richer
specifications to be run. In addition, should these results turn out to be robust, we
might have to reconsider whether happiness is such a good measure of social utility.
If, in order to be happy, we have to force everyone to lead exceedingly dull lives,
perhaps happiness is not everything after all.
99
The application of a tax is a typical economic policy response to a negative externality, but, not
necessarily the only effective policy response. The consideration of incentives that lead to lower
rates of divorce or jailing etc. is more worthy of future research.
204
s.d.
Min
Max
8.041
.317
7.646
8.556
.531
.248
.076
.836
.315
.130
.049
.495
.216
.123
.026
.417
agreeableness
4.275
.037
4.199
4.402
extraversion
4.082
.030
4.025
4.164
emotional stability
4.254
.082
4.132
4.519
conscientiousness
4.228
.059
4.084
4.366
openness
4.030
.050
3.842
4.099
agreeableness * stress
2.310
1.024
.396
3.555
extraversion * stress
2.221
.997
.373
3.419
2.284
.1
.402
3.491
conscientiousness * stress
2.277
1.001
.398
3.506
openness * stress
2.197
.987
.361
3.391
Variable
Positive_Stress/1000
(average sum of
205
206
(2)
Stress
Variable:
Stress/1000
Positive_Stress/1000
Negative_Stress/1000
coefficient
-1.18
(3)
Stress Valency
t-value
(17.21)
(4)
Stress + Direct Personality
coefficient
t-value
-1.76
-.572
(5.15)
(1.58)
Average Personality
Agreeableness
Extraversion
Emotional Stability
Conscientiousness
Openness/Intellect
Stress * Personality
Agreeableness * Stress
Extraversion* Stress
Emotional Stability * Stress
Conscientiousness * Stress
Openness/Intellect * Stress
constant
R2
8.67
0.81
207
(215.43)
8.73
0.82
(177.83)
(5)
Stress + Direct Personality
+ Indirect Stress*Personality
coefficient
t-value
-0.68
(2.54)
coefficient
-1.54
t-value
(10.91)
-0.21
3.84
-.074
-1.40
-1.53
(0.29)
(5.62)
(1.03)
(2.22)
(2.43)
-0.55
1.62
-0.17
0.004
-0.39
(1.07)
(2.64)
(0.34)
(0.01)
(0.73)
(2.54)
4.15
10.43
0.44
1.83
-8.65
6.63
(0.95)
(5.71)
(0.22)
(0.79)
(5.17)
(2.42)
9.35
0.88
0.95
Table 5.54: OLS regressions for the Aggregate Model of Happiness (5.2) with the eleven most
important life events; N = 70
Model (2)
Variable:
personal injury
just separated
just reconciled
victim of violence
worsening finances
Model (2)
coefficient
t-value
coefficient
t-value
coefficient
t-value
-0.014
(6.16)
-0.011
-0.012
(6.96)
(9.39)
-0.010
-0.008
-0.011
(6.87)
(2.56)
(1.07)
constant
R2
8.76
0.36
(73.15)
Model (2)
Variable:
Model (2)
coefficient
t-value
8.38 (110.87)
0.72
Model (2)
coefficient
t-value
8.83 (110.14)
0.73
Model (2)
coefficient
t-value
personal injury
-0.011
(8.92)
-0.005
(3.75)
-0.001
(0.61)
just separated
-0.017
(5.97)
-0.012
(4.44)
-0.008
(2.46)
just reconciled
-0.007
(0.88)
-0.004
(0.57)
-0.001
(0.08)
victim of violence
0.018
(6.58)
0.014
(5.49)
0.009
(3.77)
-0.015
(5.53)
-0.009
(3.36)
improved finances
-0.034
(5.02)
0.001
(1.18)
death of a spouse/child
0.007
(1.20)
just married
-0.007
(2.03)
just pregnant
-0.003
(0.69)
birth of child
0.005
(0.81)
8.55
(147.7)
worsening finances
constant
R
8.83
(141.10)
0.84
8.67
(145.70)
0.88
208
0.94
20
40
20
muleseprsrrs
muleinsrsrrs
60
40
80
60
Figure 5.43a to k: Graphics of the stress at each age arising from the eleven most important life events
20
40
60
80
20
age
40
60
80
age
10
mulerclrsrrs
10
muleviorsrrs
20
30
15
40
20
20
40
60
80
20
40
60
age
age
209
80
10
mulefnirsrrs
mulefnwrsrrs
20
10
30
15
40
Figure 5.43a to k (continued): Graphics of the stress at each age arising from the eleven most important life events
20
40
60
20
80
40
60
80
age
age
10
mulefrdrsrrs
20
muledscrsrrs
5
30
10
40
20
40
60
80
age
210
20
40
60
age
80
10
20
mulemarrsrrs
20
muleprgrsrrs
40
30
60
40
80
Figure 5.43a to k (continued): Graphics of the stress at each age arising from the eleven most important life events
20
40
60
80
40
mulebthrsrrs
20
30
10
0
40
40
60
50
20
20
age
age
60
80
age
211
80
212
.0080
22.24
-.0451
.0005
26.01
31.29
-.0478
.0007
22.81
31.05
-.0450
.0006
21.33
29.32
.0158
5.70
. 0114
4.13
-.0001
. 0853
-.0622
. 1620
-.6360
-.2031
-.2288
-.1641
-.1323
-.3826
-.0798
. 5140
. 1252
2.30
7.16
17.99
6.76
16.59
8.36
7.56
4.24
7.45
11.12
5.07
72.02
13.98
-.0001
. 0857
-.0616
.1148
-.4891
-.1701
-.2128
-.1687
-.0881
-.2706
-.0591
.4923
.1233
- .2995
.0184
-. 1268
. 0379
-. 0876
-. 2560
1.72
7.24
17.85
4.73
12.50
6.91
7.06
4.35
5.20
7.85
3.77
68.89
13.85
4.01
0.77
4.62
0.23
4.03
7.08
These pooled OLS regression results are for Australians aged 15 to 84 in the HILDA panel data waves 2 to 6 for the period 2002 to 2006.
213
100
Variable:
victim of violence
just separated
just reconciled
fired from job
worsening finances
death of friend
friend jailed
just married
start new job
just pregnant
moving house
improved finances
promoted at work
birth of child
just retired
constant
R2
Age
coefficient t-value
7.5938
451.84
0.01
Age + Age2
coefficient t-value
8.6336
0.03
214
232.26
+ Demographics
coefficient t-value
7.2077
0.17
102.42
+ Life Events
coefficient t-value
- .5945
- .5987
- .1893
- .2474
- 1.2115
.0938
.0880
.3015
- .1504
.3192
.0411
.5446
- .0497
.2479
.7762
7.2998
0.19
9.20
12.40
2.20
4.46
21.54
3.04
0.97
3.48
3.55
3.93
0.89
5.47
0.68
2.09
5.49
102.49
Table 5.56: The determinants of Life Satisfaction for Australians; Fixed-effect regression results for individuals in the balanced HILDA panel; N = 55,177 101
Variable:
age
age*age
ln (weekly household
income)
pension Income
female
education years
married
separated
never married
divorced
widowed
employed
unemployed
disability
health
health a year ago
spouse/child death
death of a relative
personal injury
jailing of self
injury to a family member
property crime victim
victim of violence
101
Age
coefficient t-value
-.0233
7.16
Age + Age2
coefficient t-value
-.0488
.0003
5.30
2.96
+ Demographics
coefficient t-value
+ Life Events
coefficient t-value
-.0501
.0004
5.09
4.18
-.0450
.0004
4.55
3.63
.0115
3.13
. 0094
2.58
-.0001
1.45
-.0001
1.17
-.0310
-. 0501
-.6000
-.2228
-.3025
-.6463
-.0095
-.1840
-.0255
. 2479
. 0890
2.01
1.06
9.02
5.68
4.36
7.19
0.42
5.30
1.55
25.84
10.36
-.0362
-.1217
-.4815
-.1304
-.2922
-.6206
-.0027
-.1571
-.0182
..2394
.0844
- .1721
-.0205
-. 0937
. 0092
-. 0424
-. 1507
- .2574
2.35
2.37
6.96
3.30
4.17
6.68
0.12
4.52
1.11
24.99
9.81
2.45
0.97
3.72
0.05
2.12
4.59
4.10
These fixed-effect regression results are for Australians aged 15 to 84 in the HILDA panel data waves 2 to 6 for the period 2002 to 2006.
215
Variable:
Age
coefficient t-value
Age + Age2
coefficient t-value
+ Demographics
coefficient t-value
just separated
just reconciled
fired from job
worsening finances
death of friend
friend jailed
just married
start new job
just pregnant
moving house
improved finances
promoted at work
birth of child
just retired
constant
R2 (overall)
7.5938
451.84
0.01
9.443
0.00
216
45.53
8.605
0.04
33.72
+ Life Events
coefficient t-value
- .4554
- .0156
.0276
- .6187
.0366
.0138
.2884
.0417
.2944
.2533
.3299
.0497
.3624
.0364
9.98
0.20
0.55
11.75
1.29
0.15
3.36
1.06
4.07
5.87
3.74
0.728
3.70
0.28
8.624
33.84
0.04
Chapter Six
Summary of Findings
Chapters 3 to 5 of this thesis have taken us on a journey of discovery along the road of
lifetime happiness. The three empirical studies sought to inform current econometric
models of happiness by considering happiness from multi-disciplinary perspectives. The
research questions addressed in each study emerged from gaps between the economic
and other scientific literatures.
The gap considered in the first study (Chapter 3) analysed a puzzle in the relationship
between age and happiness. A review of the literature revealed a U-shape finding of
happiness in age in the economic literature that is not shared by other scientific
disciplines. To begin resolving this difference in theoretical opinion, the U-shape of
happiness in age was initially replicated for 18 to 93 year-olds using three oft-used panel
data sets; the German Socio-Economic Panel, the Household, Income and Labour
Dynamics in Australia, and, the British Household Panel Survey. The U-shape arising
from age effects (the age & age2 variables) was evident in all three data sets and
remained so even when socio-demographic and life event variables were included in the
regression specifications. The findings were typical of other scholars in the economics
literature and confirmed the presence of a U-shape of happiness in age in all three data
sets.
After confirming the presence of a U-shape of happiness in age in all three data sets, the
study sought to explain it. Several explanations were offered for the U-shape of
happiness in age. The initial explanation pursued a line of enquiry that considered
average happiness changes at young and old ages; there is a sharp decline in raw average
happiness from 18 years to 22 years and in those close to death (where there are fewer
individuals). Excluding the young and old (22 to 80 year olds) from all three data sets,
the regressions were repeated. There was no clear qualitative difference between the
previous results and those excluding the very old and the very young. The U-shape could
217
not be explained by the extremities of the age range. It was proposed that the U-shape of
happiness in age must be due to relationships in large parts of the age range.
With the age and age2 coefficients remaining strongly significant across all
specifications with all three data sets, the search refocussed back on the full sample and
considered omitted variables and reverse causality. This line of enquiry was based upon
important findings in the literature, that happiness is strongly affected by stable
personality traits. In an econometric model of individual happiness, these fixed
individual traits are usually part of the error term. The line of enquiry pursued a stylised
finding from both the economic and the psychological literature that accounting for
fixed traits has a very strong impact on the coefficients found for socio-economic
variables. A leading explanation for this is the possibility of reverse causality arising
from unobserved heterogeneity.
Pursuing the line of enquiry that reverse causality caused by unobserved fixed traits
explained the U-shape, the same OLS regressions were rerun but this time with fixed
effects. With all three data sets the age effects disappeared, the U-shape was indeed due
to reverse causality. In seeking robustness in this result, we considered more deeply how
the unobserved heterogeneity could bias the pooled regression results. Consistent with
the findings from other studies, the coefficients of most of the socio-economic variables
became much smaller when fixed effects were included. Next, we sought to identify if
changes in the coefficients of these non-age variables lead to a difference in the
predicted age-profile. The OLS prediction showed a clear inverted u-shape for all three
data sets and, with fixed effects added, we saw that the inverted U-shape was much less
pronounced. The inclusion of fixed-effects reduces the coefficients of variables that
themselves systematically vary by age (incomes and marriage peak in middle age) and
that this in turn reduces the predicted inverted U-profile of their effects.
218
Robustness checks provided weight of evidence that the U-shape of happiness in age did
indeed emerge from reverse causality caused by unobserved fixed traits. When the
analysis was repeated with multiple latent-variable techniques, the results were
consistent with the OLS findings. The highly significant and positive effect on agesquared found in the cross-section disappeared with the inclusion of fixed-effects.
Robustness checks went further. Instead of including self-reported health as a
continuous variable, the 5 possible health states (from very bad to very good) were
included as separate dummy variables (as recommended by Terza, 1987). Again, this
made almost no difference to the age-squared effects. Adding an additional robustness
check using age-bands (Clark, 2006) supported the initial finding; the U-shape of
happiness in age did indeed emerge from reverse causality caused by unobserved fixed
traits.
In the process of using the German Socio-Economic Panel to explain the U-shape of
happiness in age, a new puzzle emerged. For some reason, even when we use controls in
our regressions, the happiness of Germans declines with age (but not to the same extent
for British and Australians). Tests for time or cohort effects did explain a small amount
of the decline in German happiness, but the predicted size of the effects over a lifetime
left much unexplained. Perhaps there was something wrong with the German panel data.
Regressions including a time in panel variable revealed there was a large decline in
reported satisfaction the longer an individual remained in the panel. Various reasons
were discussed for why the amount of time in the German panel would bias happiness
results. Perhaps it is an artefact of the GSOEP, perhaps it is in the nature of the
respondents, perhaps we need a country specific variable to account for the bias. Either
way, further research into these time and cohort effects is required otherwise findings
from GSOEP-based cross-sectional or panel analyses may become highly suspect.
After beginning our journey into happiness over a lifetime in Chapter 3, Chapter 4
pursued a cohort little visited by economists, children. Examining childhood happiness
was considered study-worthy because there is some evidence from other scientific
disciplines that the happiness of individuals in their childhood can affect the happiness
219
of those same individuals in later life. After extending our view of lifetime happiness to
15 to 93 year olds in the Household Income & Labour Dynamics in Australia panel data
set, we saw a steep 7.2% decline in the happiness of 15 to 23 year-old Australians. This
steep decline in the happiness of young Australian was twice the size of the 3.6%
happiness decline we see in 75 to 86 year old Australians who we expect to have
declining happiness due to their falling incomes, failing health and the onset of death.
We questioned when this happiness decline began in young Australians.
To reveal when this happiness decline began in young Australians, we collected data
from 9 to 14 year-old Australian children. Drawing on happiness-life domain theory
from the school psychology scientific discipline, measurement scales were developed
and data were collected with an Internet-based Happiness survey that questioned the
children on three life satisfaction domain factors (the childs natural environment, their
school environment, and, their interaction with friends). This additional Smart Train
data initially revealed a further 9.3% decline in the happiness of 9 to 14 year-old
Australians. Subsequent analysis of the Smart Train data with a model of childhood
happiness and a subsequent decomposition of the prediction for each domain factor
revealed that two of the life domain factors explained almost half of this steep decline in
the happiness of 9 to 14 year-old children.
Children are much affected by their school environment, it accounts for 35% of the
happiness decline and much of that happiness decline occurs when the children
transition from lower school grades to high school. This transition would also seem to
affect the childrens school friendships. We find that the childrens interaction with
friends positively contributes to their happiness right up until they transition to high
school where its effect goes negative then recovers somewhat as the children make new
friends at high school. However, the overall effect of the interaction with friends life
domain factor is negative, accounting for 8.8% of the decline in happiness of Australian
children aged 9 to 14 years.
220
The childhood happiness not only revealed the steep decline in childhood happiness, the
study revealed one expected and one unexpected finding concerning personality and
happiness. The expected finding was that extraverted, talkative, gregarious, outgoing,
children are happier. Extraversion has the same effect in adulthood. The unexpected
finding was the effect of conscientiousness on happiness. Conscientious adults are
unhappier. This is not the case with children. Conscientious children are happier.
Perhaps conscientious children are happier because they do better at school. Perhaps
conscientious children enjoy more attention from peers who seek out the conscientious
students so they can benefit from their greater knowledge. These propositions are worthy
of future research. In the mean time, let us review the findings from the 15 to 23 year old
cohort in the Australian HILDA data.
With much of the decline in the happiness of young children explained, Chapter 4
returned our focus to explaining steep 7.2% decline in the happiness of 15 to 23 year-old
Australians we noted earlier in this findings summary. The 15 to 23 year old cohort from
the HILDA panel data were analysed with a model of individual happiness and the
results were compared from those from the entire HILDA panel. The most obvious
finding was that age effects in regressions on the 15 to 23 year old cohort were seven
times larger than for the same regression with the entire HILDA sample. Other than this
overly large negative effect from 15 to 23 year-old unemployment, none of the sociodemographic or life event variables that are usually included in happiness regressions
nor the inclusion of fixed effects adequately explained away the steep decline in the
happiness of 15 to 23 year old Australians. We were left to ponder as to why this steep
decline in happiness continued through adolescence and into young adulthood.
A review of the literature provided some direction into explaining why happiness
continues to decline for adolescents and young adults.
negative change in our happiness can be a function of our happiness expectations and
how those expectations are affected not only by the changing economic circumstances
that were included in the Chapter 3 & 4 analysis, our happiness is also affected by the
stressful life events that change our everyday lives. Chapter 5 pursued this path of
221
reasoning by taking advantage of the additional fourteen life events in the HILDA panel
data and we asked if changes in our lives and the lives of our peers make us (un)happy.
Chapter 5 accounted for changes in overall life satisfaction by considering the effects
from life event shocks over time and sought to gauge if their stressful effects explained
the changes in aggregate happiness over the life cycle. Panel data from the Household
Income and Labour Dynamics for Australia was used to construct an index of the
severity of the stress from life changes at each age. This single-variable Stress Index
explained over 94% of the variation in happiness over time. Unexpectedly, aggregate
positive stress (such as marriage rates by age or levels of job promotion) has a greater
negative effect on aggregate life satisfaction than negative stress (such as negative
financial events or deaths of spouses), We interpreted this as a strong indication that
what is deemed a positive event by the person involved is a highly negative event for his
or her peers. We saw some evidence that extraverted individuals are affected less
negatively by stress.
222
Chapter Seven
Discussion, Policy Conclusions and Future Research
This thesis pursued a multidisciplinary line of inquiry into happiness over a lifetime.
Recent advances in the content of socio-economic data sets created research
opportunities for economists to consider happiness from multidisciplinary perspectives.
These opportunities emerged from socio-economic panel data surveys that now include
the psychology-derived scales that allow us to econometrically consider and test how
behavioural factors contribute to the theme of this study, lifetime happiness.
Of course, one could ask why economists should be concerning themselves with
happiness. Shouldnt we leave happiness research to other sociological sciences that
have decades of experience in the research of overall wellbeing. Some political leaders
would disagree. President Sarkozy of France and David Cameron, the Prime Minister of
Great Britain, are of the opinion there is more to maximising the overall wellbeing of
society than GDP growth. Certainly, money is important, and its absence definitely
makes us unhappy. However, if western societies are to become better places in which to
live, the people of those societies expect not only health and wealth, they expect
happiness as well.
If economists are to provide political leaders with advice on how to best allocate scarce
resources and thereby maximise the overall wellbeing of individuals and society, we
need to consider happiness from all theoretical perspectives. We need to include in our
econometric models of happiness all the variables that could affect individual happiness,
including those variables considered important by other scientific disciplines. Clearly,
this has not happened to date because the vast majority of the variance associated with
individual happiness remains unexplained. At the onset, this thesis did not purport to
include all relevant variables that could affect individual happiness, but it did consider
some variables that have not been considered before and one that warranted
reconsideration, age.
223
The first study in this thesis sought to explain happiness change over time by
reconsidering what was becoming a stylised fact in the economics literature, the U-shape
of happiness in age; individuals get happier as they age. The empirical analysis began by
seeking to confirm the presence of the U-shape of happiness in age. Regression results
provided strong evidence that the U-shape clearly existed in the German data and was
evident in the British and Australian socio-economic panel data. However, No U-shape
can be seen when we look at an age-happiness prediction or a scatter plot of the age
happiness relationship. There was no U-shape to be seen in the graphical depiction of
the raw average happiness data. There was no visually discernable U-shape for the
Germans, or for the Australians, or for the British. Looking at age cohorts, we could see
no strong age-happiness relationship for the age range 22 to 60 years. We did see a steep
decline in raw average happiness for 18 years to 23 year-olds (perhaps the left-hand side
of a u-shape), but, we also saw a steep decline in the happiness of those in old age (more
like an inverted U-shape). For all three data sets, no U-shape could be seen in the agehappiness relationship.
However, for some countries like Australia it could be argued that the raw average
happiness across age cohorts does look somewhat like a U-shape. There is a steep
decline in the happiness of 15 to 23 year old Australians, a midlife low period of
happiness before happiness increases from about 50 to 60 years of age. Removing the
effects of the steep happiness decline of the young and the old could have eliminated the
U-shape of happiness in age finding; it persisted. In addition, the U-shape finding
continued to persist even when we tested for cohort effects that could emerge from
newcomers to the panel data sets. The results were robust across all three data sets; it
would appear that we do get happier as we age.
Given the predicted happiness depictions in Figures 3.20, 3.21 & 3.22, a happiness
decline in old age is more believable. Old age is that period in our life when we are
starved of resources, we no longer enjoy the high income of our middle years and our
health is failing, those we grew up with are dying around us and we see the imminent
224
onset of our own demise. With such a miserable outlook, it is difficult to believe that the
U-shape of happiness actually holds in reality. We cannot honestly believe that we just
have to wait around until we get old to be happy. Clearly, something else is happening.
Nevertheless, evidence from existing economics of happiness studies and the OLS
regressions in this study, revealed that the U-shape of happiness in age did exist in the
data, it just must be a proxy for something else.
The first study pursued this proxy notion and revealed that the U-shape of happiness in
age could be explained by time invariant unobserved fixed traits. Running the same
regressions with fixed effects eliminated the U-shape. This finding was robust across all
three data sets, the German GSOEP, the British BHPS and the Australian HILDA.
Before concluding that the U-shape emerged from fixed effects, the choice of analytical
methods was considered.
One could argue that the findings had more to do with the selection of analytical
methods than the data or any real difference in the theoretical perspective of economists
or those from other scientific disciplines. To allay this possibility, robustness tests were
completed using three additional analytical methods. The same results emerged from the
latent variable analysis (ologit) that some in the economics literature demand as the
preferred method of happiness analysis. Coding health as ordinal did little to change the
results, and, the results from the recent conditional fixed effect ordered logit method
mirrored the other results. The results were consistent across all analytical methods and
with all three data sets. The U-shape of happiness in age was robustly explained by
reverse causality arising from unobserved fixed traits.
For the economics literature, this empirical finding was new, but how did it contribute to
explaining the changes in happiness over a lifetime? From a theoretical perspective, I
argue that it should wake us from the complacency of following a single analytical
method. There is strong evidence that the U-shape emerged from fixed effects and any
econometrician could have found this if only they had pursued analytical methods
beyond the latent variable analysis that has recently pervaded the economics of
225
226
finding expands the range of variables that we can now incorporate into our economic of
happiness research. We now have strong evidence of the presence of reverse causality
caused by unobserved fixed traits and therefore need to include personality and variables
affected by personality in our happiness models. The second study did just this by
including personality-affected variables in a model new to economics, a model of
childhood happiness.
Childhood happiness is little studied by economists. This is understandable, because,
while children could be considered consumers, it is the childrens parents who produce,
make choices, purchase and contribute to the economic wellbeing of society. However,
there is evidence that unhappy children make unhappy adults and unhappy adults can be
a burden on society, e.g. increasing health costs. The unhappy get sick more often, are
more likely to be absent from work, are less productive and can make those around them
unhappy. Therefore, it is in the long-run best interests of society if we study the
happiness of the children in expectation that they can become happier and more
productive adults in our societies of the future.
To begin our study into the happiness of the young, we initially saw evidence of
increasing unhappiness as Australians progressed from adolescence into adulthood.
There was a 7.2% decline in the happiness of Australians between the ages of 15 and 23
years. This steep happiness decline is twice what we saw in 75 to 86 year old Australians
who we expect to be unhappy due to declining incomes, failing health and the onset of
death. The Smart train data collected with our online Happiness Survey uniquely
extended our lifetime view of happiness to children and revealed a happiness decline
even more startling than what we saw in 15 to 23 year olds. Between the ages of 9 and
14 years, the happiness decline in Australian children is bigger, 9.3%. Added to the
7.2% decline we saw in 15 to 23 year olds, the happiness of 9 to 23 year old Australians
declines by a very large 16.2%; five times the happiness decline we saw in 75 to 86 year
old Australians. The economics of happiness literature focuses on adults. With evidence
that the happiness decline in the young can be five times what we see in the old, it is
time we extended our study of happiness beyond adulthood to children.
227
To extend the study happiness back to childhood, the life satisfaction domain approach
of Frijters (1999c) was adopted. Frijters found that if we are happy in all the domains of
our life then overall we are happy. The review of the literatures revealed that domain
approach methods are commonly used by other scientific disciplines but not by
economics of happiness researchers. We see that economists use the raw data that
emerges from individual questions in the panel surveys. In recent times, these panel
surveys have begun to personality factor questions. This has made it easier for
economists to include personality in happiness regressions because the providers of
panel data like the HILDA have done the hard work of pre-constructing and validating
the factors. This is not to say that economists lack the skills to construct and test the
reliability and validity of measures. Kapteyn, Smith, Van Soest, & Vonkova (2011)
developed survey instruments for sleep, concentration and memory domains in their
study of the importance of vignette equivalence and response consistency when
individuals are asked to describe and rate their health. However, the domain approach
remains little used in economics of happiness studies. This is an opportunity forgone; the
breadth and depth of the questions in panel data surveys like the HILDA provide us with
rich data that lends itself to the use of a life satisfaction domain approach.
It was the use of the life satisfaction domain approach in the Chapter 4 study that
contributed to our understanding of childhood happiness. Three life satisfaction domain
factors were constructed and their reliability and validity tested before the factors were
incorporated into a unique model of childhood happiness. The three life domains were
identified by taking a multidisciplinary approach to the literatures. The school
psychology literature proposed that a childs school environment, their childs natural
environment, and the childs interaction with friends were domain factors that could
affect childhood happiness. The three childhood happiness domains incorporated into
the new model of childhood happiness were constructed from a survey instrument
employed by school psychologists, the BFC-Q scale. Again, a multi-disciplinary
approach to the literatures prompted the development of reliable instruments that proved
suitable for this economics of happiness research. The instrument is considered suitable
for constructing the life satisfaction domains used in this study because the children
228
were questioned on how they behave at school and how they behave with their friends.
After construction, the three childhood happiness domain factors exhibited good internal
reliability and consistency. The BFQ-C scale and the childhood happiness domain
factors are recommended for use by other economics of happiness researchers.
The BFC-Q personality scale is considered suitable for economics of happiness research
for two additional reasons. Firstly, the adult personality scales usually included in socioeconomic panel data surveys are unsuitable for children. However, there have been
numerous studies where the BFC-Q personality scale was successfully used to measure
the personality of children (Barbaranelli at al., 2003; del Barrio Carrasco Miguel &
Holdago, 2006). To add robustness to the validity of the BFC-Q scale, these researchers
cross validated the scale by comparing the childrens responses to the BFC-Q
questionnaire with the parents evaluation of their childs personality using the adult
personality scale; the two sets of responses were strongly positively correlated.
The second reason the BFQ-C scale is appropriate for economics of happiness research
comes from empirical evidence in this study. Analysis of the data collected with the
scale revealed regression results consistent with those from the adult happiness
literature; extraverted adults (children in this study) are happier and neurotic adults
(children) are unhappier. Of course, one could argue that childhood personality operates
differently from adult personality, childhood happiness research that reveals the same
personality results as adult happiness research could emerge from an unsuitable
childhood happiness scale because extraverted children are not really happier. Decades
of findings from psychologists generally do not support this reverse argument. The
consensus in the psychology literature is that personality remains stable from a very
early age and that any change in adolescent or adult happiness emerges from
measurement error. Therefore, this researcher is of the opinion that the BFC-Q scale and
the life satisfaction domain factors developed for this research are suitable instruments
for use by other economics of happiness researchers.
229
Analysis using life satisfaction domain factors provided interesting results. The school
environment factor explained 35% of the steep decline in happiness of 9 to 14 year old
Australian children. Much of this happiness decline occurs as the children transition
from lower grade school to high school. One could argue that this decline in happiness is
predominantly due to physiological and emotional changes taking place in 9 to 14 year
old children. Hormonal changes could be contributing to the steep happiness decline in
children. In the absence of hormonal change data we have to rely on the data we do
have. Analysis of the data collected with our online Happiness Survey revealed, that
children had: difficulty understanding their teacher (or the material taught); the children
had to concentrate more in class; they have to forego leisure because they have to work
harder, and; the children find it more difficult to learn. It is no wonder that children get
unhappier as they move from lower grade school to high school. The question is; how
can we help children to be happier as they make this compulsory transition from lower
grade school to high school.
The second finding in the study provides some direction. We could help schoolchildren
with their transition from the lower grade school to high school by making them more
extraverted. Children high on extraversion are happier. Looking at the questionnaire
items used to construct the extraversion trait factor reveals the extraverted behaviours
that correlate with happier children. Extraverted children are more talkative, they are
more assertive and participative. Extraverted children make friends easier, they like to
share their thoughts and experiences with others. If we could make children more
extraverted this would make them happier. However, changing the personality of 9 to 14
year-old children is not considered feasible because personality is formed much earlier,
during the childs formative years. While personality may not be changeable, the
findings of Lischetzke & Eid (2006) provide some direction on what we could do to
make children more extraverted and happier. Lischetzke & Eid find that extraverts have
better mood regulation; we could teach the children how to maintain a more positive
attitude in their day-to-day lives, but how.
230
One way to teach the children how to maintain a more positive attitude in their day-today lives is through behaviour modification, or learning through reinforcement. Operant
and reinforcement theory takes the view that learning is dependent upon the
environment (McShane & Travaglione, 2003, pp. 45-50). With behaviour modification
theory, thinking is not considered part of the learning process but an intermediate step
between behaviour and the environment. Our experience with the environment teaches
us to alter how we behave so that we maximise positive and minimise adverse
consequences (Miltenberger, 1997). A law of cause and effect is enacted whereby an
operant behaviour will be repeated dependent upon how behaviour is reinforced.
Positive reinforcement is provided for preferred behaviours and negative reinforcement
for non-preferred behaviours (Connellan, 1978). With children, the behaviour
modification process could manifest as positive-reinforcing praise when the child
exhibits extraverted behaviours like being more assertive and participative in class or
when the children share their thoughts and experiences with others. Negative
reinforcement (sometimes called avoidance learning) would involve not criticising
children when they do not exhibit extraverted behaviours. By withholding criticism, the
children are more likely to repeat the extraverted behaviours for which they received
praise. Behaviour modification would not change the personality of the schoolchildren;
it would just provide feedback that incentivizes the schoolchildren to behave in a
happiness-maximising way.
There are examples of the successful use of behaviour modification with children. In
1978 Wolraich, Drummond, Salomon, O'Brien, & Sivage found that behaviour
modification assisted with the classroom behaviour and academic performance of 6 to 9
year olds. Behaviour modification has also been successfully applied to treat psychotic
children (Meyers & Craighead, 1979); to increase school attendance and improve
behaviour in school cafeterias (Fabiano, et al., 2008); to help disabled children learn
(Routh, 1979), and; reduce youth delinquency and violence in the classroom (Eddy,
Reid, & Fetrow, 2000). Behaviour modification has even been used by parents reduce
the amount of time their children spend viewing television (Jason & Fries, 2004).
However, the use of behavioural modification with children is controversial. As far back
231
as 1986, Boivin, Sewell, & Scott (1986) questioned whether behavioural modification
was an ethically appropriate procedure for children. In spite of this early criticism,
behavioural modification continues to be used; particularly on students with extreme
behavioural problems arising from conditions such as ADHD; Attention Deficit
Hyperactivity Disorder (Miranda, Presentacin, & Soriano, 2002; Waxmonsky et al.,
2008).
Behavioural modification has also been effectively used on students by school teachers.
After being taught behavioural modification skills teachers could better cope with
keeping problem children in the classroom (Schiff & BarGil, 2004). Dua (1970)
investigated the effectiveness of behavioural orientated therapy programs used to treat
introversion and extraversion in the classroom. Dua found that a behavioural
modification program was more effective than a re-education program in inducing
attitudinal change in students. Lowenstein (1983) took shy 9 to 16 year old students and
used behavioural modification therapy to increase extraverted behaviours. In doing so,
he improved the students reading, spelling and math outcomes. More recently, Nelson
(2010) explored classroom participation in the presence of a token economy. He found
that undergraduate students participated more when they received bonus points and
extraverted students participated more than non-extraverted students. Behavioural
modification programs that sought to increase extraverted behaviours in school children
and young adults have been shown to improve classroom behaviour and educational
outcomes.
A cursory search of the Queensland State Government Department of Education
teaching methods revealed little mention of the general use of similar behavioural
modification programs in Queensland State schools. This is not to say that behavioural
modification is not on the agenda of the Department of Education; the 2011 School-wide
Positive Behaviour Support conference is scheduled to discuss how to apply behavioural
modification methods to children with ADHD (Riffel, 2011). One wonders whether the
ethical issues have forced behavioural modification off the general teaching agenda or
whether the method is only considered appropriate for use on children with extreme
232
behavioural problems. With evidence from this study that children with extraverted
behaviours are happier and do better at school, one wonders whether it is time to put
behavioural modification for schoolchildren back on the research agenda.
If we are to have a highly educated, productive and happy labour force in the future,
perhaps now is the time for economists to contribute to an argument that hitherto has
been the property of educationalists and school psychologists. What would happen to
long-run educational outcomes, the level and quality of human capital, and GDP growth
if schoolchildren were incentivized to engage in extraverted behaviours? On the other
hand, perhaps the effect of personality has a general equilibrium element in that the
effect would change if we change the distribution (if everyone is an extravert, it might
be less fun to be one). Such questions lend themselves to collaborative research between
economists, educationalists and school psychologists.
With the first part of the second study finding that extraverted 9 to 14 year old
schoolchildren are happier, the research refocussed on the 15 to 23 year old cohort in the
HILDA. Recall, we saw a steep 7.2% decline in happiness of 15 to 23 year-old
Australians. A happiness fall of twice what we see in older Australians who we expect
to be unhappier due to their declining incomes, failing health, and the imminent onset of
their own demise. So, what is it about young adulthood that makes us unhappier than the
threat of imminent death?
Analysis with the Australian HILDA panel data using the Kitchen Sink specification
from the first study did not provide an answer. None of the demographic variables
(health, wealth, gender, home ownership, education etc.) nor the six life event variables
(death of spouse, job loss, pregnancy, marriage, divorce etc.) could adequately explain
the steep decline in the happiness of young Australians. We did see a higher than
expected effect for unemployment. This is to be expected because the average
unemployment level for 15 to 23 year olds in the HILDA is 8.8% versus 3.6% for the
Australian population. In addition, most 15 to 23 year olds are either still in school or
undertaking technical or tertiary education and 35% of also have a part-time job at the
233
same time. A leisure reduction and increased workload from holding down a part-time
job while still studying may partly explain why young Australian adults progressively
become unhappier. However, neither the larger than expected negative effect from
unemployment nor the effect of any of the other variables in the Kitchen Sink
specification could adequately explain the steep decline in the happiness of young
Australians.
However, recall the Kitchen Sink specification only included life event variables
common to both the GSOEP and the HILDA. The HILDA has an extra fourteen life
event variables that could potentially extended the level of happiness explanation.
Unfortunately, including the extra fourteen life event dummies minimally increased the
level of happiness explanation from 17 to 19 percentage points, a mere 2%. The
additional life event variables in the HILDA panel data did not explain the steep decline
in the happiness of young Australians. A new theoretical and or analytical approach was
needed if happiness over a lifetime was to be adequately explained.
A review of the findings from the first two studies in the context of theory from another
scientific discipline revealed a new approach to explaining happiness over a lifetime.
Recall the first study found that time invariant fixed traits (like personality) explained
the U-shape happiness change over a lifetime and the second study showed that a steep
decline in the happiness of schoolchildren correlated strongly with the life event change
of the children transitioning from lower grade school to high school. What would
happen if we incorporated personality and life event changes into a model of happiness?
The literature from another scientific discipline provided direction. An individuals
personality can affect their happiness (Diener, et al., 1999, p.214). In addition, happiness
is affected not just by economic circumstances it is also affected by life events
(Easterlin, 2002). Easterlin stated that the degree of positive or negative change in our
happiness is a function of an individuals consideration of their happiness expectations
(their aspirations) with the changing economic circumstances and the life events that
affect them (Easterlin, 2002, p 214). The third study leveraged from the findings of the
234
first and second studies and sought to extend those findings by proposing that the stress
arising from the interaction between personality and life events can explain the change
in happiness over a lifetime.
A review of the stress literatures from psychology and organisational behaviour revealed
that, on average, individuals are change averse and that the stress arising from
unexpected life event shocks can make us happy. Too much stress can make us feel
physically unwell, causing the onset of hypertension, heart attack and even death. In
addition, while small amounts of stress can motivate some people, too much stress
negatively affects the overall wellbeing of all individuals. The magnitude of the stress
from life events is dependent upon the personality of an individual. When confronted
with unexpected life event shocks extraverted individuals look on the bright side, but
neurotic individuals view the same change negatively and experience wide swings in
their happiness. Because personality mediates the level of stress from a life event shock,
the life event dummies were not just added to the usual model of individual happiness
(as is usual in the literature). The stress arising from the interaction between personality
and the life events was incorporated into a new model of happiness.
Of course, just including a scaled life event dummy into a happiness model just changes
the size of the regression coefficients. The Aggregate Model of Lifetime Happiness
provided a new methodological approach to overcome this impediment. This timeseries-based methodological approach drew on findings from other sociological
literatures (the Social Readjustment Rating Scale). Stress is cumulative; individuals
become more stressed and unhappier as they are confronted with additional life events.
In addition, stress arises not only from changes in our own lives, but also from changes
in the lives of our peers. For example, if our fellow worker receives a pay rise and we do
not, they are happier but we are less happy because we did not get a pay rise. This notion
of relative happiness is familiar to economists; it is the theoretical proposition at the
heart of the Easterlin paradox. Happiness is a relative self-measure, we adjudge our
happiness relative to those around us, our peers. The Aggregate Model of Lifetime
235
Happiness incorporates the cumulative effect of stress from changes in our lives as well
as the stress from the changes in the lives of our peers.
Construction of the stress variable from the interaction between the life events affecting
an individual and that individuals personality provided a most surprising revelation. A
graph of average stress and average happiness at each age appeared as a U-shaped
inverse image of one another. The regression results from the Aggregate Model of
Lifetime Happiness supported this view. The aggregate stress coefficient was negative
and highly significant, explaining over 94% of the change in happiness over a lifetime.
Increased stress in the young and through midlife significantly explained the decrease in
happiness that occurs in the young and on into mid and later life. Stress has an inverse
U-shape in age. Perhaps this research has revealed one of the variables that lie behind
the age effects we see reported in so many happiness studies.
If we accept that the stress from life events does reduce individual happiness, then it
would be reasonable to ask; what is the cost of stress to an individual and society. A few
economic researchers have sought to price the affect of life event changes; Frijters et al.
(2008) used quarterly life event data to calculate that the loss of a partner could be offset
by a windfall income gain of $US 200,000. However, they were constrained by
insufficient life events of each type and measurement error in the quarterly life event
data (probably arising from recall bias). As was argued in this study, the use of average
stress from the life event shocks overcomes these impediments and the Aggregate Model
of Happiness offers a new approach to pricing the cost of stress on the average
individual in society. As to whether the use of stress as the independent variable in the
Aggregate Model of Happiness similarly prices the loss of a partner at two-hundred
thousand dollars, we will just have to wait for the results of future research.
This thesis has taken us on a journey that sought to explain changes in happiness over a
lifetime. Along the way, this study has contributed to the economics of happiness
literature and illustrated the benefits of using multiple econometric methods. By taking
existing econometric models of happiness and complementing them with theory from
236
other scientific disciplines, this thesis has provided an explanation for the U-shape of
happiness in age. This explanation provided the theoretical foundation for developing
two new models of happiness that provided explanations to changes in the happiness of
children as well as changes in happiness over a lifetime. Both models extended the level
of happiness explanation in the economics literature. Along the way, this study revealed
opportunities for future research, the opportunity for economists to study a little-visited
cohort, childhood happiness; an opportunity to build on the stress approach of the final
study and price the impact that life event changes have on the happiness of society.
237
238
Chapter 1
Verse 1
Confucius said,
To learn and to practice what is learned time and again is pleasure, is it
not? To have friends come from afar is happiness, is it not? To be
unperturbed when not appreciated by others is gentlemanly, is it not?
(Cheung, 2010)
239
REFERENCES
ABS. (2008). Health of Mature Age Workers in Australia: A Snapshot, 2004-05 Quality
Declaration. Canberra, Australia: Australian Bureau of Statistics.
ABS. (2010). 4102.0 - Australian Social Trends, Jun 2010: Australian Bureau of
Statistics.
Alesina, A., Di Tella, R., & MacCulloch, R. (2004). Inequality and Happiness: Are
Europeans and Americans Different? Journal of Public Economics, 88(9-10),
2009-2042.
Andrews, F., & Withey, S. (Eds.). (1976). Social Indicators of Well-being: Americans'
Perceptions of Life Quality. New York, USA: Plenum Press.
Argyle, M., Kahneman, D., Diener, E., & Schwarz, N. (1999). Causes and correlates of
happiness Well-being: The foundations of hedonic psychology. (pp. 353-373).
New York, NY US: Russell Sage Foundation.
Aristotle. (1819). A new translation: The Nichomachean Ethics. In R. Pearson (Eds.)
Available from http://www.archive.org/details/newtranslationof014417mbp
Baetschmann, G., K. E. Staub, et al. (2011). Consistent Estimation of the Fixed Effects
Ordered Logit Model. Discussion Paper Series. P.O. Box 7240, 53072 Bonn,
Germany, The Institute for the Study of Labor (IZA).
Barbaranelli, C., Carpara, G., Rabasca, A., & Pastorelli, C. (2003). A questionnaire for
measuring the Big Five in late childhood. Personality and Individual
Differences, 34, 645664.
Bassi, M., & Delle Fave, A. (2004). Adolescence and the Changing Context of Optimal
Experience in Time: Italy 1986-2000. Journal of Happiness Studies, 5(2), 155179.
Beatton, T., & Frijters, P. (2009). Do changes in the lives of our peers make us
unhappy? . Working Paper Series. Queensland University of Technology:
Brisbane. Australia.
BHPS. (2010). British Household Panel Survey: Institute for Social & Economic
Research, University of Essex.
240
Baetschmann, G., K. E. Staub, et al. (2011). Consistent Estimation of the Fixed Effects
Ordered Logit Model. Discussion Paper Series. P.O. Box 7240, 53072 Bonn,
Germany, The Institute for the Study of Labor (IZA).
Becchetti, L., Trovato, G., & Bedoya, D. A. L. (2011). Income, Relational Goods and
Happiness. Applied Economics, 43(1-3), 273-290.
Blanchflower, D. G. (2008). International evidence on well-being. National Bureau of
Economic Rseearch.
Blanchflower, D. G., & Oswald, A. J. (1998). The Rising Well-Being of the
Young.Unpublished manuscript.
Blanchflower, D. G., & Oswald, A. J. (2001). Well-Being Over Time in Britain and the
USA (pp. p1359, 1328p1347 pages): University of Warwick, Department of
Economics, The Warwick Economics Research Paper Series (TWERPS).
Blanchflower, D. G., & Oswald, A. J. (2004). Well-being over time in Britain and the
USA. [Article]. Journal of Public Economics, 88(7/8), 1359.
Blanchflower, D. G., & Oswald, A. J. (2007). Is Well-being U-Shaped over the Life
Cycle? (pp. 36 pages): University of Warwick, Department of Economics, The
Warwick Economics Research Paper Series (TWERPS).
Blanchflower, D. G., & Oswald, A. J. (2008). Hypertension and happiness across
nations. Journal of Health Economics, 27 218233.
Boivin, M., Sewell, R., & Scott, K. (1986). Attitudes Toward Behavior Modification.
Behavior Modification, 10(4), 435-456.
Borgers, N., de Leeuw, E., & Joop, H. (2000). Children as Respondents in Survey
Research: Cognitive Development and Response Quality. Bulletin de
Methodologie Sociologique, 66(April 2000), 60-75.
Bruhin, A., & Winkelmann, R. (2009). Happiness Functions with Preference
Interdependence and Heterogeneity: The Case of Altruism within the Family.
Journal of Population Economics, 22(4), 1063-1080.
Campbell, A. (1976). Subjective measures of well-being. American Psychologist, 31(2),
117-124.
Campbell, A., Converse, P. E., & Rodgers, W. L. (1976). The Quality of American Life.
New York: Russell Sage Foundation.
241
Cantril, H. (1965). The pattern of Human Concerns. New Brunswick, NJ: Rutgers
University Press.
Carlopio, J. R., Andrewartha, G., Armstrong, H., & Whetten, D. A. (2001). Developing
management skills : a comprehensive guide for leaders (2nd ed.). Frenchs Forest,
N.S.W: Pearson Education.
Cavana, R. Y., Sekaran, U., & Delahaye, B. L. (2001). Applied business research :
qualitative and quantitative methods. Milton, Qld.: John Wiley & Sons Australia.
Chamberlain, K., & Zika, S. (1988). Religiosity, Life Meaning and Wellbeing: Some
Relationships in a Sample of Women. Journal for the Scientific Study of
Religion, 27(3), 411.
Chamberlain, K., & Zika, S. (1992). Stability and change in subjective well-being over
short time periods. Social Indicators Research, 26(2), 101-117.
Cheng, H., & Furnham, A. (2004). Perceived Parental Rearing Style, Self-Esteem and
Self-Criticism as Predictors of Happiness. Journal of Happiness Studies, 5(1), 121.
Cheung, W. (2010). The Lun Yu in English. Retrieved from
http://www.confucius.org/lunyu/ed0101.htm
Clark, A. E. (2006). Born to be mild? Cohort effects don't explain why well-being is Ushaped in age (Vol. Working Paper N 2006 - 35): Centre National de la
Recherche Scientifique cole des Hautes tudes en Sciences Sociales, cole
Nationale des Ponts et Chaussees cole Normale Superiure.
Clark, A. E., Frijters, P., & Shields, M. A. (2008). Relative Income, Happiness, and
Utility: An Explanation for the Easterlin Paradox and Other Puzzles. Journal of
Economic Literature, 46(1), 95-144.
Clark, A. E., & Oswald, A. J. (1994). Uhhappiness an unemployment. [Article].
Economic Journal, 104(424), 648-659.
Clark, A. E., & Oswald, A. J. (1996). Satisfaction and comparison income. [Article].
Journal of Public Economics, 61(3), 359.
Connellan, T. K. (1978). How to improve human performance. New York: Harper &
Row.
242
243
Diener, E., Emmons, R. A., Larsen, R. J., & Griffin, S. (1985). The Satisfaction With
Life Scale. Journal of Personality Assessment, 49(1), 71-75.
Diener, E., Lucas, R. E., & Scollon, C. N. (2006). Beyond the Hedonic Treadmill:
Revising the Adaptation Theory of Well-Being. American Psychologist, 61(4),
305-314.
Diener, E., Sandvik, E., Pavot, W., & Fujita, F. (1992). Extraversion and subjective
well-being in a U.S. national probability sample. Journal of Research in
Personality, 26(3), 205-215.
Diener, E., Sapyta, J. J., & Suh, E. (1998). Subjective well-being is essential to wellbeing. Psychological Inquiry, 9(1), 33-37.
Diener, E., Suh, E. M., Lucas, R. E., & Smith, H. L. (1999). Subjective well-being:
Three decades of progress. Psychological Bulletin, 125(2), 276-302.
Dockery, A. M. (2005). The Happiness of Young Australians: Empirical Evidence on
the Role of Labour Market Experience. Economic Record, 81(255), 322-335.
Dockery, A. M. (2005). Happiness, life satisfaction and the role of work: Evidence from
two Australian surveys. from
http://www.melbourneinstitute.com/hilda/Biblio/cp/conf-hd05.pdf
Dua, P. S. (1970). Comparison of the Effects of Behavioral Oriented Action and
Psychotherapy Reeducation on Introversion Extraversion, Emotionality, and
Internal External Control: Journal of Counseling Psychology.
Duncan, E., & Grazzani-Gavazzi, I. (2004). Positive Emotional Experiences in Scottish
and Italian Young Adults: A Diary Study. Journal of Happiness Studies, 5(4),
359-384.
Easterlin, R. A. (1974). Does Economic Growth Improve the Human Lot ? . Paper
presented at the Nations and Households in Economic Growth: Essays in Honor
of Moses Abramovitz.
Easterlin, R. A. (1995). Will raising the incomes of all increase the happiness of all?
[Article]. Journal of Economic Behavior & Organization, 27(1), 35.
Easterlin, R. A. (2001). Income and Happiness: Towards a Unified Theory. The
Economic Journal, 111(473), 465-484.
244
245
Falk, A., & Knell, M. (2004). Choosing the Joneses: Endogenous Goals and Reference
Standards: CEPR Discussion Papers: 4459.
Faul, F., Buchner, A., Erdfelder, E., Faul, F., & Lang, A.-G. (2006). G*Power version
3.08 Software. Kiel, Germany.
Ferrer-i-Carbonell, A. (2005). Income and Well-Being: An Empirical Analysis of the
Comparison Income Effect. Journal of Public Economics, 89(5-6), 997-1019.
Ferrer-i-Carbonell, A., & Frijters, P. (2004). How Important Is Methodology for the
Estimates of the Determinants of Happiness? Economic Journal, 114(497), 641659.
Flouri, E. (2004). Subjective Well-Being in Midlife: The Role of Involvement of and
Closeness to Parents in Childhood. Journal of Happiness Studies, 5(4), 335-358.
Fogle, L. M., Scott Huebner, E., & Laughlin, J. E. (2002). The Relationship between
Temperament and Life Satisfaction in Early Adolescence: Cognitive and
Behavioral Mediation Models. Journal of Happiness Studies, 3(4), 373-392.
Fordyce, M. W. (1973). Measuring happiness. Edison Community College.
Fordyce, M. W. (1988). A review of research on the happiness measures: A sixty second
index of happiness and mental health. Social Indicators Research, 20(4), 355381.
Frey, B. S., & Stutzer, A. (2000). Happiness, Economy and Institutions. Economic
Journal, 110(466), 918-938.
Frey, B. S., & Stutzer, A. (2002). What Can Economists Learn from Happiness
Research? Journal of Economic Literature, 40(2), 402-435.
Frey, B. S., & Stutzer, A. (2005). Happiness Research: State and Prospects. Review of
Social Economy, 63(2), 207-228.
Frey, B. S., & Stutzer, A. e. (2007). Economics and Psychology: A Promising New
Cross-Disciplinary Field: CESifo Seminar Series.Cambridge and London:MIT
Press.
Frijters, P., & Beatton, D. (2011). The mystery of the U-shaped relationship between
happiness and age. QUT Working Papers.
246
Frijters, P., Haisken-DeNew, J. P., & Shields, M. (2004). Money Does Matter! Evidence
from Increasing Real Income and Life Satisfaction in East Germany Following
Reunification. [Article]. American Economic Review, 94(3), 730-740.
Frijters, P., Haisken-DeNew, J. P., & Shields, M. A. (2004). Investigating the Patterns
and Determinants of Life Satisfaction in Germany Following Reunification.
Journal of Human Resources, 39(3), 649-674.
Frijters, P., Johnston, & Shields, M. (2008). anticipation & adaption effect of life events,
shocks. QUT Working Paper.
Frijters, P., Johnston, D., & Shields, M. (2008a). What is the shadow price of crime and
bereavement to well-being? QUT Working Paper.
Frijters, P., Johnston, D., & Shields, M. (2009). Life Satisfaction Dynamics with
Quarterly Life Event Data. Scandinavian Economic Journal.
Frijters, P. (1999c). Explorations of welfare and well-being. Unpublished Article,
University of Amsterdam, Amsterdam.
Furnham, A., & Petrides, K. V. (2003). Trait emotional intelligence and happiness.
Social Behavior & Personality: An International Journal, 31(8), 815-823.
Gardner, J., & Oswald, A. J. (2006). Do Divorcing Couples Become Happier by
Breaking Up? Journal of the Royal Statistical Society: Series A (Statistics in
Society), 169(2), 319-336.
Gerdtham, U.-G., & Johannesson, M. (2001). The relationship between happiness,
health, and social economic factors: Results based on Swedish microdata.
Journal of Socio-Economics, 30(6), 553-557.
Gilman, R., & Huebner, E. S. (1997). Children's reports of their life satisfaction:
Convergence across raters, time and response formats. School Psychology
International, 18(3), 229-243.
Gilman, R., & Huebner, E. S. (2000). Review of life satisfaction measures for
adolescents. Behaviour Change, 17(3), 178-195.
Gilman, R., & Huebner, S. (2003). A review of life satisfaction research with children
and adolescents. School Psychology Quarterly, 18(2), 192-205.
Glaeser, E. L., Laibson, D., & Sacerdote, B. (2002). AN ECONOMIC APPROACH TO
SOCIAL CAPITAL. The Economic Journal, 112, F437-F458.
247
250
Lucas, R. E., Clark, A. E., Georgellis, Y., & Diener, E. (2003). Reexamining adaptation
and the set point model of happiness: Reactions to changes in marital status.
Journal of Personality and Social Psychology, 84(3), 527-539.
Lyubomirsky, S., King, L., & Diener, E. (2005). The Benefits of Frequent Positive
Affect: Does Happiness Lead to Success? Psychological Bulletin, 131(6), 803855.
Magnus, K., Diener, E., Fujita, F., & Pavot, W. (1993). Extraversion and Neuroticism as
Predictors of Objective Life Events: A Longitudinal Analysis. Journal of
Personality & Social Psychology, 65(5), 1046-1053.
Marks, G. N. (2007). Completing University: Characteristics and Outcomes of
Completing and Non-completing Students. Camberwell, Victoria, Australia: The
Australian Council for Educational Research Ltd.
Mawson, T. (2002). Mill's Proof. Cambridge Journal, 77(3), 375-405. Retrieved from
http://journals.cambridge.org
McCrae, R. R., & Costa, P. T. (1995). Positive and Negative Valence within the FiveFactor Model. Journal of Research in Personality, 29(4), 443-460.
McCrae, R. R., Costa, P. T., Jr, Terracciano, A., Parker, W. D., Mills, C. J., De Fruyt, F.,
et al. (2002). Personality trait development from age 12 to age 18: Longitudinal,
cross-sectional and cross-cultural analyses. Journal of Personality and Social
Psychology, 83(6), 1456-1468.
McCrae, R. R., Costa, P. T., Jr., Mroczek, D. K., & Little, T. D. (2006). Cross-Cultural
Perspectives on Adult Personality Trait Development Handbook of personality
development. (pp. 129-145). Mahwah, NJ US: Lawrence Erlbaum Associates
Publishers.
McKnight, C. G., Huebner, E. S., & Suldo, S. (2002). Relationships among stressful life
events, temperament, problem behavior, and global life satisfaction in
adolescents. Psychology in the Schools, 39(6), 677-687.
McShane, S., & Travaglione, T. (2003). Organisational Behaviour on the Pacific Rim.
Sydney, Australia: McGraw-Hill Australia Pty Limited.
Mervin, C., & Frijters, P. (2011). Shared Misery Double Misery? The Effects of the
Partner's Life Events on Own Mental Health. School of Economics Discussion
Paper, The University of Queensland. Australia.
251
253
254
Schiff, M., & BarGil, B. (2004). Children with behavior problems: improving
elementary school teachers' skills to keep these children in class. Children and
Youth Services Review, 26(2), 207-234.
Seifert, W. (2003). Subjective economic well-being in Eastern Europe. Journal of
Economic Psychology, 24(3), 329-348.
Seligman, M. (2011). Flourish A Visionary New Understanding of Happiness and Wellbeing New York: Free Press
Seligson, J., Huebner, E., & Valois, R. (2005). An Investigation Of A Brief Life
Satisfaction Scale With Elementary School Children. Social Indicators Research,
73(3), 355-374.
Senik, C. (2004). When information dominates comparison: Learning from Russian
subjective panel data. [Article]. Journal of Public Economics, 88(9/10), 2099.
Shields, M. A., & Price, S. W. (2005). Exploring the economic and social determinants
of psychological well-being and perceived social support in England. Journal of
the Royal Statistical Society: Series A (Statistics in Society), 168(3), 513-537.
Shinn, K. (2009). Confucianism: A Brief Summary of Confucius and His Teachings.
Retrieved from
http://www.csuchico.edu/~cheinz/syllabi/asst001/fall97/11kshinn.htm
Singer, E., von Thurn, D. R., & Miller, E. R. (1995). Confidentiality assurances and
response: A quantitative review of the experimental literature. Public Opinion
Quarterly, 59(1), 66-77.
Srivastava, A. K., & Misra, G. (2003). Going beyong the model of economic man: An
indigenous perspective on happiness. Journal of Indian Psychology, 21(2), 1229.
Stutzer, A., & Frey, B. S. (2006). Does Marriage Make People Happy, or Do Happy
People Get Married? Journal of Socio-Economics, 35(2), 326-347.
Suldo, S., & Huebner, E. (2006). Is Extremely High Life Satisfaction During
Adolescence Advantageous? Social Indicators Research, 78(2), 179-203.
Terza, J. V. (1987). "Estimating Linear Models with Ordinal Qualitative Regressors."
Journal of Econometrics 34(3): 275-291.
255
256
Veenhoven, R. (1994). Is happiness a trait? Tests of the theory that a better society does
not make people any happier. Social Indicators Research, 32(2), 10.
Veenhoven, R. (2007). Healthy happiness: effects of happiness on physical health and
the consequences for preventive health care. Journal of Happiness Studies, 1-21.
Veenhoven, R. (2008). Healthy happiness: effects of happiness on physical health and
the consequences for preventive health care. Journal of Happiness Studies, 9(3),
449-469.
Wagner, G. G., Burkhauser, R. V., & Behringer, F. (1993). The English language public
use file of the German Socio-Economic Panel. Journal of Human Resources,
28(2), 429-433.
Ware, J. E. (2009). SF-36 Health Survey Update. Retrieved January, 21, 2010, from
http://www.sf-36.org/tools/sf36.shtml
Watson, D., Clark, L. A., & Tellegen, A. (1988). Development and validation of brief
measures of positive and negative affect: The PANAS Scales. Journal of
Personality and Social Psychology, 47, 10631070.
Watson, N., & Wooden, M. (2010). Data Survey: The HILDA Survey: Progress and
Future Developments. Australian Economic Review, 43(3), 326-336.
Waxmonsky, J., Pelham, W. E., Gnagy, E., Cummings, M. R., O'Connor, B., Majumdar,
A., et al. (2008). The Efficacy and Tolerability of Methylphenidate and Behavior
Modification in Children with Attention-Deficit/Hyperactivity Disorder and
Severe Mood Dysregulation. Journal of Child and Adolescent
Psychopharmacology, 18(6), 573-588.
Wessman, A. E., & Ricks, D. F. (1966). Mood and personality. Oxford England: Holt,
Rinehart, & Winston.
White, N. P. (2006). A brief history of happiness. Malden, MA ; Oxford: Blackwell Pub.
Wilson, W. R. (1967). Correlates of avowed happiness. Psychological Bulletin, 67(4),
294-306.
Winkelmann, L., & Winkelmann, R. (1998). Why Are the Unemployed So Unhappy?
Evidence from Panel Data. [Article]. Economica, 65(257), 1-15.
Wolff, H. G., Wolf, S. G., & Hare, C. C. (1950). Life stress and bodily disease.
Baltimore: Williams and Wilkins.
257
258