Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Tony OConnor
09383581
oconnot3@tcd.ie
November 28, 2014
Abstract
The objective of this paper is to investigate the change in the labour share of
income, or total value added, in common production functions. This paper finds
that the labour share, measured through the marginal productivity of production
labour, began to decline in the 1970s. This is contrary to the literature which date
the beginning of the decline to the 80s. A possible reconciliation here is that it took
time for the wage to adjust downwards to the marginal productivity of manual labour.
In contast, the marginal productivity of non-production labour has remained stable.
These findings seem to go against the hypothesis that technology and automation
was responsible for the decline, though not much new support can be found for the
idea that trade was responsible.
Introduction
Te objective of this paper is to analyse the shifts in the labour share of income,
measured as value added, using standard production functions.
Traditionally, the stability of the labour share has been regarded as a stylized
fact [Kaldor, 1961]. However, recent literature has indicated that the share of labour
in value added has been declining in the large majority of countries [Karabarbounis
and Neiman, 2013]. This paper will examine this question a panel of 473 industries,
over the years 1958-2009, to determine whether this occured at an industry level in
the U.S.
This paper will proceed as follows. First, we will examine the literature to date
detailing the trend in the labour share, and the factors that may be detrminging
its path. Second, the economic model and empirical approach of this paper will be
outlined. Third, a description of the data will be provided. Fourth, the estimations
results will be described and evaluated. Lastly, we will conclude.
Literature Review
Feenstra and Hanson [1999] note that not only did the wages of low-skilled workers
fell relative the the wages of high-skilled workers during the 1980 and the 1990s, but
the wages of the low-skilled fell in real terms. Using a modification of the conventioanl
price regression, they find that 35 percent of the incrase in the relative wage of
nonproduction workers is due to improvements in information technology, while a
further 15 percent is due to the international trade, principally outsourcing.
There are two dominant narratives explaining the decline in the labour share,
and accounting for why the wages of low-skilled workers experienced a steeper decline
than those of high-skilled workers. The first story is that improvements in technology
and automation is depressing the demand for low-skilled workers, leading to a fall
in their wages. However, under this story, one would expect the wages of those who
can understand and operate such machines to increase.
Alternatively, taking non-production workers as a grouping, then increases in
capital should increase their marginal product. Support for this is found is provied
in Autor et al. [2003], who find that computer capital sustitutes for workers performing routine, manual non-cognitive task and complements peforming non-routine
problem-solving and complex communications tasks. Therefore, we would expect to
see the wages of production workers falling, and the wages of non-production workers
increasing. However, this effect may be attenuated by the observation made by Autor
et al. [2013] whereby the target of automation moved from production tasks towards
non-manufactirung information-processing tasks.
The second dominant story is that the decline in wages is due to the increase
in international trade. Concretely, a large global workforce of unskilled workers
has entered into the global labour market, and tis may depress the demand for
unskilled workers in developed economies, due to phenomena such as outsourcing
and import competition in product where unskilled workers are an abundant factor.
For example, Acemoglu et al. [2014], taking account of input-output linkages and
general equilibrium effects, find that the accelerating U.S. imports from China from
1999 to 2011 was possible responsible for job losses in the range of 2.0 to 2.4 million,
which if true would have exerted significant downward pressure on wages.
Alternative explanatios for the declining capital share have been proposed, such
as that of Azmat et al. [2012], who find that privatisation is responsible for up to
one-fifth of the decline. However, the methods used in this paper are not capable of
assessing this finding.
This paper will also naturally link itself to the expanding range of literature that
examines whether the capital share of income has been rising over the last fifty years.
A last related strand in the literature that this paper will later examine is that of the
elasticity of substitution. In order for the capital share of income to rise continuously,
along with a rising capital-output ratio, an elasticity of substitution exceeding 1 is
necessary; less formally, diminishing returns will need to set in slowly.
Chirinko [2008] provides an overview of the literature, and finds that while the
estimates have a wide range, a value of 0.4 to 0.6 is most likely given the evidence.
This paper will seek to estimate the factor shares through the use of standard production functions. With the assumption of perfect competition and constant returns to
scale, then the respective share in income of a factor equals the marginal product of
that factor [Bertola et al., 2006]. For example, consider the standard Cobb-Douglas
production function:
Y = AK L1
. Under under our two assumptions, then the share of income rewarded to capital is
, while that going to labour is 1 .
To estimate the factor shares going to labour and capital, we will estimate the
logged Cobb-douglas production function which is:
(1)
Thus, we see that we can indirectly estimate the labour share of income, 1 ,
by running the above regression. Traditionally, these values have been assumed to
be in the range of 0.3 and 0.7, respectively.
3.1
Is = 1?
[2003] test 28 U.S. industries, and fail to reject the Cobb-Douglas specification if 20 of
the 28 industries. Therefore, in order to establish whether these regressions are valid
we will need to estimate the elasticity of substitution. We can do this by estimating
a Constant Elasticity of substitution (CES) production function, which allows to
vary, and which we can impute from the estimates. The CES production function
takes the form:
1
lnQi = 0 ln{L
i + (1 )Ki } + i
3.2
(2)
Marschak and Andrews [1944] noted that the input levels and unobserved productivity shocks may be correlated. More concretely, if a firm experiences a positive
productivity shock, they will raise output, and will achieve this through purchasing
more of the variable inputs such as labour, materials or energy. This will result
in the coefficient estimates on the variable inputs being biased upwards [Levinsohn
and Petrin, 2003]. Whether capital is upwardly or downwardly biased depends on
whether is is correlated or not with the productivity shock. Because the productivity
effect changes over time, it is not fixed and therefore using a fixed effect estimator
will lead to biased and inconsistent results. An possible instrument in this case would
be input prices, as we could see if changing inputs is due to changing input prices.
However, input prices are a weak instrument in this context, changes in the wage
being only weakly correlated with change is labour demanded. Lags are used instead.
Olley and Pakes [1996] attempts to correct this by using investment as a measure
for the unobserved productivity change; if a firm experienes a positive productivity
shock, then they ought to increase investment. However, Doms and Dunne [1998]
note that the time-series of plant-level investment exhibits lumpy behaviour, implying
adjustment costs may be convex. This finding leads Levinsohn and Petrin [2003] to
Data
All data is taken from the NBER-CES Manufacturing Industry dataset [Becker et al.,
2013]. This is an panel of 473 NAICS industries. The data covers 1958-2009 with
variables such as output, employment, payroll and other input costs, investment,
capital stocks, TFP, and various industry-specific price indexes. For our dependent
variable, we use total value added in all specifications. A number of variables are
generated from the data. For example, in order to distinguish between the factor
shares being allocated to production and non-production workers, we calculate nonproduction workers as being equal to Total Employment minus Production Workers.
Econometric Analysis
Mean
327009.662
1983.5
34.814
735.896
25.423
50.64
443.796
4799.495
2620.97
2190.409
156.655
585.429
2757.95
1664.517
1093.433
0.937
Std. Dev.
8889.717
15.009
45.053
1252.867
33.651
66.823
736.828
13196.309
9721.219
4710.187
462.108
1433.618
6388.03
4145.339
2418.355
0.257
N
24596
24596
24167
24167
24167
24167
24167
24167
24167
24167
24167
24162
24167
24167
24167
24167
VADD
INVEST
ENERGY
CAP
1.000
0.830
0.699
0.791
1.000
0.902
0.931
1.000
0.939
1.000
Labour is split into production and non-production workers, as this will allow us
to examine how their respective factor shares change over time. All results are called
with robust standard errors to correct for heteroskedasticity.
(2)
1979
(3)
1989
(4)
1999
(5)
2009
0.287
(8.89)
0.203
(6.49)
0.162
(5.27)
0.152
(5.29)
0.111
(3.16)
LOG(NPRODE) 0.402
(13.44)
0.421
(14.98)
0.458
(17.26)
0.413
(15.05)
0.411
(12.78)
LOG(CAP)
0.297
(17.26)
0.365
(20.47)
0.391
(21.45)
0.462
(23.48)
0.542
(21.01)
Constant
2.253
(24.80)
2.539
(25.55)
3.001
(28.98)
2.998
(25.30)
2.695
(17.90)
462
0.939
462
0.934
462
0.926
462
0.929
473
0.921
LOG(PRODE)
Observations
Adjusted R2
t statistics in parentheses
Robust Standard Errors
In this table, five regressions are run, with each variable taking the logged value of
the previous 10-year average to eliminate any autocorrelative errors. Thus, for 1969,
in column 1, we have the natural log of the 10-year averages of capital, production
and non-production workers, regressed on the natural log of the 10-year average of
value added, where each industry forms one observation. In this manner, we estimate
an aggregate prodution function that is representative of the decade from 1959-1969.
A similar regression is run for each of the following four decades.
Firstly, in Table 3 the regression has strong explanatory power; more than 92% of
the variation in value added is explained by variation in capital and labour. Secondly,
all variables are significant at the 1% level, most at the 0.1% level. The coefficients
in this regression illustrate marginal productivites, for example in the 60s, a 1%
5.1
In this table, we see that the share of value added going to production labour (also
known as blue collar labour) has been in continuous decline, declining every decade.
In the 00s, it is just over one-third its value in the 60s. Interestingly, our finding here
may shed new light on the view that the income share of labour began to decline in
the 80s. In our results, nearly half of the 61% drop in marginal productivity occured
during the 70s. Interestingly, there may have been a lag before declines in marginal
productivity crossed over into declines in the labour share.
Interestingly, over the entire period the share of income going to non-production
workers has remained stable, even though it does exhibit a concave trend, peaking
in the 80s.
The income share going to capital has also greatly increased over the periods
analysed, nearly doubling from 0.3 in the 60s to 0.54 in the 00s.
(2)
1979
(3)
1989
(4)
1999
(5)
2009
LOG(PRODE)
0.225
(7.79)
0.153
(5.26)
0.123
(4.27)
0.114
(4.23)
0.0121
(0.40)
LOG(NPRODE)
0.388
(13.42)
0.413
(14.79)
0.435
(15.59)
0.367
(12.28)
0.429
(12.54)
LOG(CAP)
0.205
(6.69)
0.208
(5.41)
0.250
(5.71)
0.376
(7.95)
0.259
(5.92)
0.0147
(0.61)
0.0208
(0.78)
-0.00602
(-0.19)
-0.0562
(-1.68)
0.0722
(2.75)
LOG(MATCOST) 0.166
(6.53)
0.209
(8.91)
0.217
(8.52)
0.222
(9.08)
0.278
(10.99)
Constant
2.082
(13.89)
2.383
(13.84)
2.639
(14.22)
2.365
(11.90)
2.615
(11.91)
462
0.948
462
0.947
462
0.936
462
0.940
473
0.940
LOG(ENERGY)
Observations
Adjusted R2
t statistics in parentheses
Robust Standard Errors
The regression in Table 3 is expanded to include more control variables for each
of the five decades in Table 4. In order to check the robustness of these trends, an
extended Cobb-Douglas regression is run, including the energy and materials factor
input. Tis increases the Adjusted R2 of the model slighly, and has a far-reaching effect
on the coefficients, but does little to change the trends. The variable representing
materials and fuels is significant, though that of energy only become significant for
the 00s.
We see that the share of income going to production labour declines over the
entire period, even becoming insignificantly different from zero in the 00s.
The share of income going to non-production labour over the period increased,
10
though the trend is non-monotonic. Somewhat similarly, the share of income going
to capital increases over the period, but it finishes well beow the peak attained in
the 90s.
5.2
CES Regression
Addopting a cobb-Douglas functional form carries the assumption that the elasticity
of substitution is equal to 1. This is a strict assumpition that we will relax using the
a CES functional form. In this functional form, the elasticity of substitution must
be constant, though it may take on a value different from 1.
Two sets of CES regression are estimated, in line with Equation 2; therefore, delta
is the coefficient on total labour hours in production. In Table 5, we see that delta
declines from 0.723 in the 60s to 0.612 in the 90s.
When we look at the snapshot regressions in Table 6, we see that the share of
income going to labour experienced a steeper decline, going from 0.731 in 1969 to
0.554 in 1999. These estimates are significant at the 0.1% level.
In order the examine the applicability of the Cobb-Douglas regressions, we use
the CES parameter estimates to impute the elasticity of substitution, as =
1
1+ .
As
we see from Table 5, the elasticity of substitution ranges from 1.18 to 0.97 (the 00s
estimation is invalid). In Table 6, it ranges from to 1.07 to 1.2 (th 2009 regression
is invalid). Thus, we see that the estimated elasticity is not too far from what is
required to assume the functional form is a Cobb-Douglas function.
5.3
In order to deal with the simultaneity bias outlined in the literature review, we use
the Levinsohn-Petrin estimator, outlined by Levinsohn and Petrin [2003]. the essence
of this is that we use an intermediate input energy as a variable that can control for
any productivity shock. The estimates are presented in Table 7 and Table 8. The
11
b0
Constant
rho
Constant
delta
Constant
Observations
Adjusted R2
(1)
1969
(2)
1979
(3)
1989
(4)
1999
(5)
2009
1.111
(11.83)
1.438
(9.90)
1.775
(8.98)
2.243
(9.93)
0.221
(9.25)
-0.155
(-1.75)
-0.0997
(-1.01)
0.0293
(0.28)
-0.162
(-1.41)
-18.26
(.)
0.723
(12.91)
0.634
(7.86)
0.496
(4.81)
0.612
(5.56)
0.775
(.)
462
0.898
1.18
462
0.887
1.11
462
0.874
0.97
462
0.889
1.19
473
0.817
-0.05
t statistics in parentheses
sigma
b0
Constant
rho
Constant
delta
Constant
Observations
Adjusted R2
(1)
1969
(2)
1979
(3)
1989
(4)
1999
(5)
2009
1.286
(11.82)
1.702
(10.33)
2.147
(10.40)
2.293
(8.09)
4.662
(149.17)
-0.170
(-1.78)
-0.104
(-1.02)
-0.221
(-1.93)
-0.0680
(-0.53)
13.94
(.)
0.731
(11.84)
0.623
(7.04)
0.661
(6.58)
0.554
(4.09)
0.908
(.)
462
0.893
1.2
462
0.880
1.11
462
0.877
1.28
473
0.883
1.07
473
0.756
0.06
t statistics in parentheses
Robust Standard Errors
12
regression is run across all observations in the decade, as is done in in Levinsohn and
Petrin [2003].
In Table 7, we see that the contribution of production workers to value-added declines across all decades, with the exception of the 90s when it temporarily increases.
The contribution of non production labour increases slightly over the period.
However, an important anomaly with this regression is the capital coefficient.
With capital exceeding 1 in every decade, we end up rejecting the null hypothesis of
constant returns to scale in every decade. The fact that the capital coefficient sometimes takes a value exceeding 2 leads one to doubt the consistency of this estimator
when applied to an industry panel. It may be the case that the grater persistence of
productivity shock in an industry renders this estimator biased and inconsistent.
Nevertheless, it is useful to note that even using this estimator the coefficient
declines in every decade, in both specifications of the model.
(2)
70s
(3)
80s
(4)
90s
(5)
00s
0.293
(8.74)
0.195
(9.56)
0.141
(5.08)
0.188
(7.61)
0.0930
(2.59)
LOG(NPRODE) 0.385
(12.15)
0.461
(16.79)
0.452
(19.10)
0.356
(13.18)
0.441
(11.54)
LOG(CAP)
1.332
(8.20)
1.201
(3.86)
2.212
(5.67)
2.251
(8.49)
1.980
(4.06)
Observations
Adjusted R2
4620
4620
4620
4653
4729
LOG(PRODE)
t statistics in parentheses
Robust Standard Errors
13
(2)
70s
(3)
80s
(4)
90s
(5)
00s
LOG(PRODH) 0.571
(23.22)
0.514
(19.20)
0.477
(16.76)
0.460
(17.50)
0.457
(19.12)
LOG(CAP)
1.842
(8.66)
1.988
(3.99)
4.590
(15.67)
3.222
(7.67)
3.635
(4.42)
Observations
Adjusted R2
4620
4620
4620
4653
4730
t statistics in parentheses
Robust Standard Errors
5.4
IV Regressions
14
(2)
70s
(3)
80s
(4)
90s
(5)
00s
LOG(PRODH) 0.590
(96.00)
0.496
(46.52)
0.511
(28.93)
0.462
(65.71)
0.456
(65.56)
LOG(CAP)
0.377
(61.75)
0.465
(43.04)
0.482
(31.08)
0.551
(104.06)
0.612
(53.01)
Constant
1.070
(46.55)
1.341
(28.76)
1.813
(49.25)
1.864
(68.12)
1.608
(31.59)
4620
0.887
4620
0.843
4620
0.848
4642
0.871
4730
0.875
Observations
Adjusted R2
t statistics in parentheses
Driscoll-Kraay Standard Errors
(2)
1979
(3)
1989
(4)
1999
(5)
2009
LOG(PRODH) 0.580
(23.60)
0.509
(19.29)
0.452
(17.82)
0.511
(17.47)
0.458
(11.87)
LOG(CAP)
0.378
(19.45)
0.477
(21.83)
0.544
(25.18)
0.544
(21.33)
0.650
(18.00)
Constant
1.286
(14.56)
1.570
(14.50)
1.784
(15.35)
1.859
(14.50)
1.298
(6.96)
462
0.894
462
0.880
462
0.876
473
0.885
473
0.861
Observations
Adjusted R2
t statistics in parentheses
Driscoll-Kraay Standard Errors
15
5.5
For completeness, a fixed effect estimator is also run on the data. We find that
the coefficients on this to be somewhat erratic. Consider the coefficient on capital,
which rises from around0.6 in the 60s to 1.75 in the next decade, only to fall to 1
in the decade after. In contrast to most of the other estimators, there is no clear
trend exhibited. This is inline with with what Levinsohn and Petrin [2003] found;
they found that the fixed effect estimator is the most incompatible with all other
estimators for firm data. This is because there is no fixed effect; in our case the
productivity shock varies within industry over time.
(2)
70s
(3)
80s
(4)
90s
(5)
00s
LOG(CAP)
0.642
(16.87)
1.754
(22.94)
0.993
(10.10)
0.821
(12.61)
0.616
(6.31)
LOG(PRODH)
1.029
(34.31)
0.542
(6.27)
0.503
(6.00)
0.665
(17.70)
0.634
(11.60)
Constant
-2.199
(-12.36)
-7.794
(-20.60)
-1.848
(-4.03)
-0.812
(-1.93)
1.019
(1.62)
4620
4620
4620
4653
4730
Observations
Adjusted R2
t statistics in parentheses
Driscoll-Kraay Standard Errors
Interpretation of Results
16
added going to labour declining in turn, a fact that has been mostly verified by the
literature.
Interestingly, when we decompose labour down into production and non-production
workers, we see that the share of income (adn the marginal productivity) of nonprduction workers has remained approximately stable, while that of production wrokers
has declined relatively steeply.
Recalling the literature review, it was stated that if technology was the dominant
story behind this structural change, then it would have been likely that the marginal
productivity of non production workers would have increased. As this does not seem
to have occurred, our results would seem to favour the trade story behing the decline
of labour.
However, perhaps the most interesting result is that the decline in marginal productivity seems to have begun before the 80s, when the labour share started declining. This also favours the trade story, given that there was a well-known productivity
slowdown in the 70s.
Some important caveats need to be atttached. Firstly, nonproduction labour is a
broad category in and of itself; it is simply everyone who does not work directly with
production, and so would include secretaries to executives. Therefore, the stability o
returns to this broad category could mask a great deal of heterogeneity in marginal
productivity underneath.
Secondly, the proxy was simply the number of production and non production
workers. It is possible that the number of workers would not change, but one group
could work much more intensively withing a given year. This is unlikely for nonproduction workers, and in various specification the number of production hours
was included explicity as a variable. It is also worth noting that in the correlogram
production hours and number of production workers were extremenly correlated 0.996 - such that one could almost say they are the same variable.
17
Conclusion
18
cannot be immediately ruled out. However, the wherewithal to answer this question
are not in this dataset, given that there is no information on international trade.
19
References
Arpad Abraham and Kirk White. The Dynamics of Plant-Level Productivity in U.S.
Manufacturing. Working Papers 06-20, Center for Economic Studies, U.S. Census
Bureau, July 2006. URL http://ideas.repec.org/p/cen/wpaper/06-20.html.
Daron Acemoglu, David Autor, David Dorn, Gordon H. Hanson, and Brendan Price.
Import competition and the great u.s. employment sag of the 2000s.
Work-
ing Paper 20395, National Bureau of Economic Research, August 2014. URL
http://www.nber.org/papers/w20395.
David H. Autor, Frank Levy, and Richard J. Murnane. The skill content of recent technological change: An empirical exploration. The Quarterly Journal of
Economics, 118(4):12791333, 2003. doi: 10.1162/003355303322552801. URL
http://qje.oxfordjournals.org/content/118/4/1279.abstract.
David H. Autor, David Dorn, and Gordon H. Hanson. Untangling trade and technology: Evidence from local labor markets. Working Paper 18938, National Bureau
of Economic Research, April 2013. URL http://www.nber.org/papers/w18938.
Ghazala Azmat, Alan Manning, and John Van Reenen. Privatization and the decline
of labours share: International evidence from network industries. Economica,
79(315):470492, 2012. ISSN 1468-0335. doi: 10.1111/j.1468-0335.2011.00906.x.
URL http://dx.doi.org/10.1111/j.1468-0335.2011.00906.x.
Edward
An
J.
Balistreri,
Estimation
Computational
of
Christine
U.S.
Economics
A.
McDaniel,
Industry-Level
0303001,
and
Eina
Vivian
Capital-Labor
EconWPA,
March
Wong.
Substitution.
2003.
URL
http://ideas.repec.org/p/wpa/wuwpco/0303001.html.
Randy Becker, Wayne Gray, and Jordan Marvakov. NBER-CES Manufacturing Industry Database: Technical Notes, 2013.
20
Giuseppe Bertola, Reto Foellmi, and Josef Zweimller. Income Distribution in Macroeconomic Models. Princeton University Press, Princeton, 2006.
Robert
S.
Working
Chirinko.
Paper
2234,
Series
The
Long
CESifo
And
Group
Short
Of
Munich,
It.
2008.
CESifo
URL
http://ideas.repec.org/p/ces/ceswps/ 2234.html.
Mark E. Doms and Timothy Dunne. Capital Adjustment Patterns in Manufacturing Plants. Review of Economic Dynamics, 1(2):409429, April 1998. URL
http://ideas.repec.org/a/red/issued/v1y1998i2p409-429.html.
Robert C. Feenstra and Gordon H. Hanson. The impact of outsourcing and hightechnology capital on wages: Estimates for the united states, 19791990. The Quarterly Journal of Economics, 114(3):907940, 1999. doi: 10.1162/003355399556179.
URL http://qje.oxfordjournals.org/content/114/3/907.abstract.
N. Kaldor. Capital accumulation and economic growth. In F. A. Lutz and D. C.
Hague, editors, The Theory of Capital, page 177. St Martins Press, 1961.
Loukas Karabarbounis and Brent Neiman. The global decline of the labor share.
Working Paper 19136, National Bureau of Economic Research, June 2013. URL
http://www.nber.org/papers/w19136.
James Levinsohn and Amil Petrin. Estimating production functions using inputs to
control for unobservables. The Review of Economic Studies, 70(2):pp. 317341,
2003. ISSN 00346527. URL http://www.jstor.org/stable/3648636.
Jacob Marschak and Jr. Andrews, William H. Random simultaneous equations
and the theory of production. Econometrica, 12(3/4):pp. 143205, 1944. ISSN
00129682. URL http://www.jstor.org/stable/1905432.
21
URL
http://www.nber.org/papers/w10950.
G Steven Olley and Ariel Pakes. The Dynamics of Productivity in the Telecommunications Equipment Industry. Econometrica, 64(6):126397, November 1996. URL
http://ideas.repec.org/a/ecm/emetrp/v64y1996i6p1263-97.html.
22