Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
In this paper we re-examine the relationship between inflation uncertainty and total output.
To properly specify an estimating equation, we investigate the time series properties of two
frequently used measures of inflation uncertainty. We fail to reject the hypothysis that each
series has a unit root. However, the uncertainty measures are not cointegrated with output
and relative oil prices. This means that the proper specification is in terms of differences.
With this specification we find that an increase in inflation uncertainty growth reduces real
GNP growth but, unlike earlier work, we find that this effect is temporary. It is also unlikely
that an inflation uncertainty shock on its own could produce a recession.
1. Introduction
IN HIS NOBEL Prize lecture Milton Friedman (1977) argued that an increase in
uncertainty about the future course of inflation reduces economic efficiency.
Costs originate from two sources. First, inflation uncertainty makes it more
difficult for consumers to accurately determine relative prices. Second, increased
inflation uncertainty increases the difficulty of entering into long-term contracts.
These costs hinder the efficient allocation of resources and reduce real output.
Some would argue that this decrease in real output is the major cost associated
with inflation. In the past 15 years economists have frequently subjected
Friedman's hypothesis to empirical tests. With few exceptions these studies have
found that inflation uncertainty has a statistically significant negative effect on
real economic, activity.1
Few of these papers have, however, attempted to examine either the
persistence or the magnitude of these costs. Examination of these issues
obligates us to consider the proper specification of the regression equation and
the nature of inflation uncertainty. We begin by discussing the implications of
the regression equation used in most previous studies. This discussion leads us
to an examination of the time series properties of survey-generated measures of
inflation uncertainty. To formally study these properties we carry out unit root
and cointegration tests. In contrast to earlier work, we conclude that it is
appropriate to first difference real G N P and uncertainty, and that, while
changes in uncertainty are very persistent, its effect on the growth rate of real
GNP is short-lived. In addition to being short-lived, the effect is not especially
large.
1
For exampes see Levi and Makin (1980), Mullineaux (1980), Amihud (1981), Makin (1982),
and Holland (1988). For an exception see Coulson and Robins (1985). Holland (1993) provides a
tabulation of various studies.
2. Specifications
where y, measures the level of real economic activity, for example real GNP or
the level of employment; u, measures the level of inflation uncertainty; z, is a
vector of stationary variables; sometimes including a price or money supply
surprise term; and e, is a random error term.2 Now, suppose that there is a
shock to u,. If this shock is temporary, the change in Ay, is also temporary.
For example, if shocks to u, are white noise, the effect lasts only one period.
At the other extreme, if u, follows a random walk, then eq. (1) means that a
shock to inflation uncertainty permanently changes real G N P growth. 3
The time series properties of u, are therefore essential in interpreting results
from equations like the one above. Below we provide new evidence that two
popular measures of inflation uncertainty generated from survey data have unit
roots. More specifically, Dickey-Fuller, Phillips-Perron, and augmented Dickey-
Fuller tests fail to reject the hypothesis that the logarithm of the level of either
the Livingston or Michigan-SRC measure have unit roots. This evidence
complements the results of Schwert (1989), who finds evidence for a unit root
in an econometrically generated measure of inflation volatility.
If inflation uncertainty has a unit root, this poses a problem in estimating eq.
(1). Earlier work concludes that measures of economic activity such as real
GNP, industrial production, and employment possess unit roots, and this holds
in our sample. Thus, eq. (1) involves regressing a series that is integrated of order
zero, 1(0), on a series that is integrated of order one, 1(1). Granger (1986)
observed that the results from such a regression are extremely difficult to
interpret because of the disparate time series properties of the two series. Indeed,
Granger states that if the left-hand-side variable is integrated of order zero, the
only theoretically admissible value for the coefficient on the right-hand-side
variable integrated of order one is zero. Therefore, given our finding that the
inflation uncertainty measures are integrated of order one, eq. (1) is inappropriate
for testing Friedman's hypothesis.
2
We measure uncertainty at time ( as the variance of forecasts about inflation for time f made
during time I — 1. In this way inflation uncertainty is not contemporaneous with real G N P at time
I. We have argued elsewhere (Davis and Kanago, 1994), that a relative measure of inflation
uncertainty is appropriate. However, to conform with the earlier literature we do not use a relative
measure here.
3
In principle, including additional lags of the level of uncertainty in eq. (1) allows for a temporary
effect from a change in uncertainty. Holland (1988) extends eq. (1) in this way, and finds that a
permanent change in inflation uncertainty permanently changes real growth. A similar equation
using our data and method of lag selection fails to confirm this result.
G. DAVIS AND B. KANAGO 165
As Davidson and MacKinnon (1993, p. 723) point out the 'classical approach'
here would be to use the differences, which are stationary, in a regression.
However, the modern pratice is to first check for a cointegrating relationship
among the variables. If the variables are cointegrated, then an error correction
model is appropriate since these models take advantage of useful information
contained in levels. On the other hand, if the variables of interest are
not cointegrated, 'one simply estimates a model where all variables are
differenced' (Engle and Granger, 1991, p. 13), and so reverts to the classical
approach.
Having established that the statistical evidence indicates that uncertainty is
1(1), we test for cointegration below. Failing to find cointegration between real
GNP, uncertainty, and the relative price of oil, we adopt the classical approach
and estimate a model in first differences.
4
Appendix 1 contains a description of the uncertainty measures. In addition, details of all the
results are available from the authors on request.
5
The particular Phillips-Perron test statistic reported here is the transformation of the (-statistic
corresponding to the Dickey-Fuller test when a trend is included. Consistent estimation of the
variance for the purpose of constructing the Phillips-Perron test statistic requires choosing a lag
length for the order of autocorrelation in the residuals. The test reported here uses a lag length of
10. We tried some other smaller lag lengths, but they did not change significantly the test results.
166 ON MEASURING THE EFFECT OF INFLATION UNCERTAINTY
40
6 We follow Hamilton's (1983) example and use the producer proce index for crude petroleum.
The growth rate of the relative price is found using the G N P implicit proce deflator. The hypothesis
of a unit root in the relative price of oil is not rejected, so we use the growth rate.
7
See Hamilton (1994. p.63O)
G. DAVIS AND B. KANAGO 167
- 2 0 2 4
Current
FIG. 3. Logarithm of the Livingston variance
3.5 4
Current
Fie;. 4. Logarithm of the SRC variance
to reject the null of no cointegration for both series, except in the very restrictive
case of zero lags. The Engle-Granger tests therefore indicate no cointegration
among the variables.
The Johansen-Juselius test avoids the normalization problem by considering
a vector autoregression. For details of the testing procedure and its application
see Dickey et al. (1991), Johansen (1992), Hamilton (1994), and Enders (1995).
Test statistics are reported in Table 3. 8 We report both the maximal eigenvalue
and trace test statistics for a variety of lag lengths, a restricted constant (no
time trend), and an unrestricted constant (which allows for a time trend).
Tables for significance levels provided in Osterwald-Lenum (1992) and used
for Table 3 are based on large samples. Reinsel and Ahn (1988) argue that with
8
We use the same samples to generate the Johansen test statistics as we use in the FPE tests
and regressions below. For SRC the sample is 1969:3 to 1990:1. For LIV the sample is 1951:1 to
1990:1.
168 ON MEASURING THE EFFECT O F INFLATION UNCERTAINTY
TABLE 1
Unit root test statistics for the inflation uncertainty series
SRC Livingston
DF test /trend -3.048 [0.126] DF test/trend -2.733 [0.239]
PP test/trend -3.006 [0.138] PP test/trend -2.455 [0.383]
Lags Augmented DF test/no trend Augmented DF test/no trend
Notes: Approximate marginal significance levels are in brackets. These are calculated in TSP based on results
in MacKinnon (1990). Tests are run over the longest sample that the data will allow. With one lag the sample
for SRC is 1966:3 to 1990:1 and for LIV is 1947:4 to 1990:2.
a small sample size it is necessary to adjust the trace test in order to use these
tables. The adjustment is made by multiplying the trace test by (T — np)/T,
where T is the sample size, n is the number of lags, and p is the number of
variables in the system. Evidence for the accuracy of this adjustment is provided
in Reimers (1992). In Table 3 the statistics are reported without the small-
sample adjustment; however, we report on the significance of both the adjusted
and unadjusted test statistics.
In interpreting the test statistics we follow the strategy outlined by Johansen
(1992). First, we consider the null of zero cointegrating vectors with a restricted
constant (no trend). If this null is rejected, then we turn to the null of zero
cointegrating vectors with an unrestricted constant (trend). If this null is
rejected, then the case of one cointegrating vector with a restricted constant is
taken up; and so on. We stop the first time that the null is not rejected.
For the Livingston data, we do not reject zero cointegrating vectors at any
lag length. At a lag length of two, zero cointegrating vectors and no time trend
G. DAVIS AND B. KANAGO 169
TABLE 2
Engle-Granger tests for cointegration between uncertainly measures, real GNP and the relative price
of oil
Notes: significance at the 5",, level is noted by *t and significance at the 10"o level by *. The sample is 1951:1
lo 1990:1 for LIV and 1969:3 to 1990:1 for SRC.
is rejected, but zero cointegrating vectors and a time trend is not. So there may
be a time trend, but no cointegration.
Results are a bit less clear for the SRC data. For a lag length of one, the
maximal eigenvalue test rejects zero cointegrating vectors, but does not reject
one cointegrating vector. The trace test leads to the same conclusion, if the
small sample size adjustment is ignored and the 10% significance level is used.
For longer lags, the tests fail to reject the null of zero cointegrating vectors.
We can choose across lag lengths for the SRC equations by giving preference
to those that eliminate serial correlation. The smallest lag length for which the
170 O N M E A S U R I N G THE EFFECT O F I N F L A T I O N UNCERTAINTY
TABLE 3
Johansen testsJor coinlearation between uncertainty measures, real GNP and the relative price oj oil
Noies: Significance at the 5",, level is noted by + and significance at the 10"o level by *. The significance levels
are based on tables from Oslerwald Lenum (1992). We use Table I for the unrestricted case and Table I.I* for
the restricted case. The sample is 1951:1 to 1990:1 for LIV and 1969:3 to 1990:1 for SRC. 5 indicates significance
at the 5",, level, and J at the 10",, level after the Reinsel Ahn small sample adjustment.
Lagrange multiplier F-tests for serial correlation in the SRC equations fail to
reject the null hypothesis of no serial correlation is 2.9 At this lag length there
11
We test for first, second, fourth, and eighth order serial correlation.
G. DAVIS AND B. K.ANAGO 171
is no evidence for cointegration. Further, Gonzalo (1990) finds that when using
the Johansen-Juselius test, it is better to err on the side of long lag lengths
rather than short ones. Thus, on the whole, the data fails to support the
hypothesis of a cointegrating relationship between the variables.
3. Estimation
Given our findings that the uncertainty measures have unit roots, and that
there is no cointegration among the variables, the appropriate equation to
estimate is
'"The sample mean of uncertainty is lower during 1950-70 than it is for 1970-90. Our
results indicate that this rise in the level of uncertainty is not a major factor of the growth
slowdown.
172 ON MEASURING THE EFFECT OF INFLATION UNCERTAINTY
TABLE 4
Regression results for the growth rate of real GNP
0 -0.478(1.52) -0.305(0.609)
1 -0.283(0.794) -1.30(2.58)
2 -0.701(1.95) -0.607(1.20)
3 0.259(0.739) -1.41(2.71)
4 -0.536(1.65) -0.843(1.57)
5 -0.528(1.01)
6 -0.678(1.29)
7 -0.142(0.263)
8 -1.81(3.49)
Notes: the I statistics of the coefficients are in parentheses. The sample size is 1951:1 to 1990:1 when the
uncertain!) measure is ALivingston. and 1969:3 to 1990:1 when the SRC measure is used. The F Test that all
the coefficients on the uncertainty measure are zero is 2.48 with a marginal significance level of 0.039 for
ALiungsion. and 2.89 with a marginal significance level of 0.006 for ASRC.
Time
Fie;. 5. Effects of an increase in Livingston
G. DAVIS AND B. KANAGO 173
Time
FIG. 6. Effects of an increase in SRC
4. Conclusion
Earlier work strongly suggested that real GNP growth permanently changed
when there was a permanent change in the level of inflation uncertainty. Our
analysis finds that once the time series properties of inflation are taken into
account, this is not the case. Therefore, we estimate real GNP equations that
allow for the possibility of temporary effects on real GNP—even when changes
to uncertainty are permanent. Empirical tests based on these equations indicate
that changes in uncertainty have transitory effects lasting about two years. We
find that while these effects are significant, it would take an unusually large
shock to uncertainty to produce a recession.
ACKNOWLEDGEMENTS
We would like to thank the referees for very useful comments. We would also like to thank Bill
Even and Prosper Raynold for their help. All, however, are absolved from guilt.
REFERENCES
AMIHUD, Y. (1981). 'Price—Level Uncertainty, Indexation and Employment', Southern Economic
Journal, 47, 776-87.
CARLSON, J. A. (1977). 'A Study of Price Forecasts', Annals of Economic and Social Measurement,
6, 27-56.
COULSON, E. N. and ROBINS, R. P. (1985). 'Aggregate Economic Activity and the Variance of
Inflation: Another Look'. Economics Letters, 17, 71-5.
CUKIERMAN, A. and WACHTEL, P. (1979). 'Differential Inflationary Expectations and the Variability
of the Rate of Inflation: Theory and Evidence', American Economic Review, 69, 595-609.
DAVIDSON. R. and MACKINNON. J. G. (1993). Estimation and Inference in Econometrics, Oxford
University Press, New York.
DAVIS. G. and KANAGO, B. (1994). 'Contract Duration, Inflation Uncertainty, and the Welfare
Effects of Inflation", Working Paper No. 1013, Department of Economics, Miami University,
Miami, Ohio.
DICKEY, D. A., JANSEN, D. W., and THORNTON, D. (1991). 'A Primer on Cointegration with an
Application to Money and Income', Federal Reserve Bank of St. Louis Review, 73, 58-78.
ENDERS, W. (1995). Applied Econometric Time Series, John Wiley, New York.
174 ON MEASURING THE EFFECT O F INFLATION UNCERTAINTY
APPENDIX I
I. Description of uncertainty measures.
The standard deviation of inflation forecasts across respondents to surveys is a widely-used measure
of inflation uncertainty. This sort of measure is justified on the grounds that a high variance across
the forecasts of respondents indicates a lack of confidence in their predictions. Cukierman and
Wachtel (1979) formalize this idea in a model of expectation's formation. Zarnowitz and Lambros
G. DAVIS AND B. KANAGO 175
(1987) provide some empirical support. In any case, this type of measure has been the proxy of
choice for inflation uncertainty in the majority of the literature. We consider the variance across
respondents from two different surveys.
The Livingston survey began in 1946 and is conducted every six months; it is the most often
used series in studies of the effects of inflation uncertainty. The survey asks economists for their
predictions of price indices over the next six, 12, and 18 months." We use the Livingston variance
for the 12-month-ahead prediction of the CPI. Our data, provided by the Philadelphia Fed, runs
from 1946.2 to 1989.4.12
The second measure is from the Michigan-SRC survey. It surveys households regarding their
expectations for key economic variables, and in particular the survey asks questions about their
inflation expectations. From 1946 through 1965 consumers were asked whether they thought
inflation would rise or fall. From 1966 to 1977.2 consumers were asked to choose one of a number
of given ranges for the inflation rate. Since 1977.3 consumers have been asked what they expect
the inflation rate over the next 12 months will be. Prior to 1978 the survey was usually conducted
in February. May. August, and November. Since then the survey has been conducted monthly.
Juster and Comment (1980) constructed a quarterly series from this data. We use an updated
version of their series over the period 1966 to 1989.4.l3
" Carlson (1977) argues that these are actually eight-, 14-, and 20-month forecasts. Also see
Carlson for additional details about the collection and interpretation of the Livingston data.
12
Here the dates given are the survey dates. For the econometric work we treat uncertainty for
a given date as the uncertainty at the time of the previous survey about the current period.
13
Although Juster and Comment (1980) construct a series of the expected inflation rate and the
standard deviation across respondents for observations prior to 1966, we choose not to use this
data which is based on an up/down response. It is common in the literature not to use the series
before 1966.