178 G. P. Kouretas and L. P. Zarangas
A second test deals with the null hypothesis of constancy of the cointegration space for a given cointegration rank. Hansen
and Johansen propose a likelihood ratio test that is constructed by comparing the likelihood function from each recursive subsample
with the likelihood function computed under the restriction that the cointegrating vector estimated from the full sample falls within
the space spanned by the estimated vectors of each individual sample. The test statistic is a x
2
distributed with p-rr degrees of freedom.
The third test examines the constancy of the elements of the cointegrating vectors. When more than one cointegration vector
is identified, then it is unclear which are the parameters whose time path should be plotted. The proposed test exploits the fact
that there is a unique relationship between the eigenvalues and the estimated eigenvectors, given the normalization from which
they have been obtained. Hansen and Johansen 1993 have de- rived the asymptotic distribution of the eigenvalues, which allows
us to plot the estimated eigenvalues along with their confidence bounds. This allows us to test whether the eigenvalue at the partic-
ular date differs significantly from its full sample value. The ab- sence of any particular trend or shift in the plotted values of the
estimated eigenvalues is considered to be an indication in favor of the constancy of the cointegrating vectors.
4. THE EMPIRICAL VECTOR AUTOREGRESSIVE MODEL 4A. The Data
The basic variables in z
t
suggested by the discussion in Section 3 are:
W 5 nominal wages
P 5 producer prices
N 5 employment
Q 5 output volume, which enters as a proxy for the demand shift factor, Z
l 2 t
a
5 average income tax rate 1 1 s 5 rate of employers’ social security contributions
CPIP 5 the price wedge, i.e., the consumer price relative to the producer price, which also proxies the effect of the indirect tax, y
P
m
5 price of imported raw materials including energy UNION 5
unionization rate, which is a proxy for the bargaining power of the unions.
All variables are in natural logarithms and the sample period is 1975:Q1 to 1992:Q2. Some of the data was collected from the
WAGE SETTING, TAXES, AND LABOR IN GREECE 179
Monthly Bulletin of the Bank of Greece and some from the Minis- try of Labour data bank. Furthermore, it has been shown that
there is a major difference between the behavior of the wage and employment series for the public sector and the rest of the
economy. We, therefore, exclude the public sector and focus on the analysis of the private sector. Moreover, because our main
interest is the study of real wage resistance, i.e., the impact of taxes on wage setting and demand for labor, we omit variables
of great interest such as the capital stock and the related user cost as well as unemployment benefits and strike allowances Tyr-
vainen, 1992, 1995; Lockwood and Manning, 1993. This is neces- sary to make our system manageable. Finally, we use a vector of
dummies D
t
that includes three centered seasonal dummies that account for the short-run effects that will otherwise violate the
Gaussian assumption about the residuals, and we use two interven- tion dummies as well. The first one captures the change in political
regime, and takes the value zero for the period 1974–81, when the conservative party was in power, and 1 for the period 1981
to 1989 when the socialist party was in power. The second policy dummy variable reflects the effects of political cycles and takes
the value of 1 during election years and zero otherwise.
4B. Testing the General Specification of the Model
All empirical models are inherently approximations of the ac- tual data generating process, and the question is whether the
benchmark model 6 is a satisfactorily close approximation. Two different aspects of the model will be investigated: 1 the stochas-
tic specification regarding residual correlation, heteroscedasticity, and normality; and 2 the relevance of the conditioning variables
in D
t
. These two aspects of the model are clearly related in the sense that residual misspecification often arises as a consequence
of omitting important variables in D
t
. According to the residual tests reported in Table 1, the bench-
mark model for k 5 3 and z
t
and D
t
as specified above, seems to provide a reasonably good approximation of the data-generating
process. The estimated residual standard deviations are generally very small, indicating that most of the variation of the vector
process can be accounted for by the chosen information set.
The estimates of the short-run effects of the conditioning vari- ables are also presented in Table 1. To facilitate readability, coef-
ficients with a t-value less than 1.0 have been set to zero. As
180
G. P.
Kouretas
and L.
P. Zarangas
Table 1: Residual Misspecification Tests of Model 6 with k 5 3
Eq.
s
e
h
1
21
h
2
2 Skew.
Ex.kurt.
h
3
2
Dlw 0.017
35.0 1.7
0.05 3.09
0.73 Dlp
0.013 13.9
1.3 0.64
5.08 10.96
Dln 0.009
25.4 0.9
0.72 3.30
7.43 Dlq
0.019 29.9
1.2 0.07
2.42 0.53
Dl12t
a
0.004 17.3
0.0 0.52
5.91 19.69
Dl11s 0.001
23.5 0.8
1.17 6.57
14.74 DlCPIP
0.009 27.6
2.5 20.49
3.08 3.10
Dlp
m
0.011 30.0
1.1 0.01
3.28 1.60
Dlunion 0.014
19.8 2.6
20.26 7.80
45.01
Relevance of the Conditioning Variables in the Vector D
t
Eq. D
w D
p D
N D
Q D1 2 t
a
D1 1 s D
CPIP D
P
m
DUNION D1
0.011
1
0.001
1
0.005
1
0.018
2
0.023
1
D2 0.043
1
0.015
2
0.003
1
0.022
2
h
1
v is the Ljung-Box test statistic for residual autocorrelation, h
2
v is the ARCH test for heteroscedastic residuals, and h
3
v the Jarque- Bera test for normality. All test statistics are distributed as x
2
with v degrees of freedom.
1
Means significant at the 95 percent level, and
2
at the 99 percent level.
WAGE SETTING, TAXES, AND LABOR IN GREECE 181
Table 2: Testing the Rank in the l2 and the l1 Model
Testing the joint hypothesis Hs
1
r p-r
r Qs
1
rH
9 625.9
497.4 435.6
389.4 355.5
324.6 298.3
276.8 246.9
502.9 454.8
411.2 372.1
332.9 299.5
269.8 243.9
221.9 8
1 456.9
404.3 365.6
304.5 287.8
249.1 234.2
199.2 405.1
363.4 324.1
287.4 256.1
227.4 203.3
183.2 7
2 344.3
303.4 267.9
240.9 201.3
188.2 175.6
317.5 280.2
245.5 215.5
188.4 166.1
147.5 6
3 277.2
235.8 195.0
166.8 143.2
125.9 240.3
206.8 179.0
154.1 132.8
115.4 5
4 226.6
189.7 155.4
122.7 114.3
171.9 145.6
122.0 102.7
86.9 4
5 154.3
124.5 99.5
84.3 116.3
94.7 76.8
63.1 3
6 100.6
87.5 67.8
70.8 54.5
42.9 2
7 68.5
29.2 36.1
26.0 1
8 19.2
12.9 s
2
9 8
7 6
5 4
3 2
1 The numbers in italics are the critical values Paruolo, 1996; Table 5.
expected, both intervention dummies that refer to change in politi- cal regimes and to political cycles have their significant effect
mainly on the wage rate, the price level, and the tax rate, and a lesser effect on employment and output.
4C. Determining the Cointegration Rank and the Order of Integration
The two-step procedure analyzed in Johansen 1995a is used to determine the order of integration in a multivariate context and
the rank of the necessary matrices. The hypothesis that the number of L1 trends 5 s
1
and the rank 5 r is tested against the unrestricted H
model based on a likelihood ratio test procedure discussed in Paruolo 1996a. The test statistics reported in Table 2 have been
calculated for all values of r and s9
1
5 p 2 r 2 s
2
, under the assumption that the data contain linear but no quadratic trends.
1
Table 2 presents the results of the formal tests for the presence of I2 components in the multivariate context. The 95 percent
1
We would like to thank Katarina Juselius for providing us with her preliminary program for conducting the multivariate testing procedure for I2 components.
182 G. P. Kouretas and L. P. Zarangas
quantiles of the asymptotic test distributions are taken from Table 5 Paruolo 1996, and reproduced underneath the calculated test val-
ues. Starting from the most restricted hypothesis {r 5 0, s
1
5 0, and s
2
5 9} and testing successively less and less restricted hypotheses according to the Pantula 1989 principle shows that essentially all
I 2 hypotheses can be rejected on the 5 percent level.
In addition to the formal tests, Juselius 1995 offers further insight into the I2 and I1 analysis and the determination of
the correct cointegration rank. She argues that the results of the trace and maximum eigenvalue test statistics of the I1 analysis
should be interpreted with some caution: First, the conditioning on intervention dummies and weakly exogenous variables is likely
to change the asymptotic distributions to some unknown extent. Second, the asymptotic critical values may not be very close ap-
proximations in small samples. Juselius 1995 suggests the use of the companion matrix as an additional tool for determining the
correct cointegration rank.
The roots of the companion matrix are calculated from the estimation of the model without allowing for I2 trends, i.e., we
estimate the standard I1 model. This estimation shows that the first six unit roots are very close to unity, ranging from 0.99 to
0.91, while the next three are well inside the unit circle, ranging from 0.75 to 0.45. This result seems to support the choice of six
unit roots in the time series process and three long-run relation- ships, and it is in line with the results of the formal test of I2
components. The estimated roots are shown in Figure 1.
Table 3 reports the results of the Johansen-Juselius I1 cointe- gration analysis.
2,3
We have included in the estimation procedure three centered seasonal dummies and two shift dummies, which
account for the change in political regime and for the effects of political cycles. To take into consideration the two issues raised by
Juselius 1995 we have simulated the critical values for the trace test using the DisCo routine developed by Johansen and Nielsen
1993, and we have made a small sample adjustment to these statistics according to Reimers 1992. Both the trace and maximum
2
The calculations of the eigenvectors as well as the stability tests have been performed using the program CATS in RATS 4.20 developed by Henrik Hansen and Katarina Juselius
1995, Estima, IL.
3
A small sample adjustment has been made in all likelihood ratio statistics equal to 22lnQ 5 2T 2 kp
o
k i
5
r0
1
1
ln1 2 l, as suggested by Reimers 1992.
WAGE SETTING, TAXES, AND LABOR IN GREECE 183
Figure 1. The eigenvalues of the companion matrix.
eigenvalue statistics reject the existence of zero cointegrating vectors or nine common trends. The two statistics suggest that we can accept
three significant cointegrating vectors. These results reconfirm the decision made based on the roots of the companion matrix.
Table 3 also reports the results of several significant tests. Ac- cording to the first test, none of the variables of the system is
nonrelevant and, hence, could be excluded from the cointegrating vector the exclusion test. The second test shows that all series
are in a nonstationary conditional upon three cointegration vectors the multivariate stationarity test. Finally, weak exogeneity is
tested with the last test, and it is rejected for all variables of our system. In addition, Table 3 reports the multivariate residual
statistics, because the Gaussian assumption is violated in the pres- ence of nonnormality, serial correlation, and conditional hetero-
skedasticity in the residuals of the VAR. No evidence of serious misspecification was detected.
4
4
Gonzalo 1994 shows that the performance of the maximum likelihood estimator of the cointegrating vectors is little affected by nonnormal errors. Lee and Tse 1996 have
shown similar results when conditional heteroskedasticity is present.
184 G. P. Kouretas and L. P. Zarangas
Table 3: Johansen-Juselius cointegration tests
95 Percent critical values r
Trace lmax
Trace lmax
r 5 312.41
1
78.03
1
223.35 66.89
r 1
234.39
1
64.34
1
168.98 60.34
r 2
170.05
1
50.14
1
135.78 49.12
r 3
99.87 38.59
100.60 42.33
r 4
81.31 29.75
84.98 37.45
r 5
51.57 22.26
60.09 33.88
r 6
29.31 18.01
33.43 31.56
r 7
11.30 9.75
14.67 12.56
r 8
1.55 1.55
3.45 3.45
Exclusion Stationary
Weak Variable
restrictions test
exogeneity
lw 9.81
1
26.40
1
9.73
1
lp 25.21
1
25.96
1
10.08
1
ln 8.44
1
29.08
1
9.72
1
lq 21.93
1
20.98
1
8.22
1
l 1 2 t
a
9.65
1
28.22
1
9.05
1
l 1 1 s
8.20
1
32.90
1
10.02
1
l CPIP
10.34
1
30.55
1
8.93
1
lp
m
32.01
1
25.53
1
14.38
1
lunion 17.01
1
26.08
1
14.90
1
Multivariate residual statistics
LB17 5 1461.630.06 x
2
18 5 68.980.02 Notes: r
denotes the number of eigenvectors. Trace and lmax denote, respectively, the trace and maximum eigenvalue likelihood ratio statistics. The 95-percent critical values
have been simulated using the Johansen-Nielsen 1993 DisCo program for a model with two intervention dummies. In performing the Johansen-Juselius tests, a structure of three
lags was chosen according to a likelihood ratio test, corrected for the degrees of freedom and the Ljung-Box Q statistic for detecting serial correlation in the residuals of the
equations of the VAR. A model with an unrestricted constant in the VAR equation is estimated according to the Johansen 1992 testing methodology.
1
Indicates statistical significance at the 5 percent critical level. Notes:
For the test of exclusion restrictions figures are x
2
statistics with three degrees of freedom, for the stationarity test figures are x
2
with five degrees of freedom, and for the weak exogeneity test figures are x
2
with three degrees of freedom. Notes:
LB 5 Ljung-Box statistic for serial correlation with 17 degrees of freedom. x
2
is the Bera-Jarque test for normality distributed with 18 degrees of freedom.
WAGE SETTING, TAXES, AND LABOR IN GREECE 185
Figure 2. The Trace tests.
Figures 2–4 report the results of applying the recursive tests for parameter stability of Johansen’s results. Figure 2 shows that the
rank of the cointegration space does not depend on the sample size from which it has been estimated, because we are unable to
reject the null hypothesis of a constant rank—in our case three accepted eigenvectors. Furthermore, Figure 3 shows that we al-
ways accept the null hypothesis of the constancy of the cointegra- tion space for a given cointegration rank. Finally, Figure 4 shows
that the estimated coefficients do not exhibit instability in recursive estimation, because each corresponding eigenvalue has no signifi-
cant trend or shift during the period under investigation.
4D. Structural Identification
The model in Section 2 was designed for analysis of wage setting and the demand for labor. As we explained in Section 3, when
more than one significant vector has been found, then we must
186 G. P. Kouretas and L. P. Zarangas
Figure 3. Test of known beta eq. to betat.
impose independent linear restrictions based on structural hypoth- eses implied by economic theory so that we assign each vector to
a specific structure. Given that we have three significant vectors, it is expected that one will be the wage-setting equation, the second
the demand for labor, and the third can take alternative economic structures. Below, these restrictions are made explicit. We first
consider the following relationship:
logW 5 b
P
logP 1 b
N
logN 1 b
Q
logQ 1 b
t
a
log1 2 t
a
1 b
s
log1 1 s 1 b
y
logCPIP 1 b
P
m
logP
m
1 b
U
logUNION 11
For this relation to be considered as a wage-setting schedule the signs should be:
b
P
0, b
Q
0, b
t
a
0, b
s
0, b
y
0, b
U
0, and possibly b
P
m
0.
There are two extreme hypotheses that could be considered regarding wage setting. The first one takes the unions as having
WAGE SETTING, TAXES, AND LABOR IN GREECE 187
Figure 4. Test for the eigenvalues.
the dominant role. In the second the firms are the dominant player. These hypotheses also indicate how taxes and relative prices enter
the wage equation. There are several alternative restrictions that we may impose on the wage relation. One possibility is that we
assume wage resistance; then we could impose a wedge-restriction like:
2b
t
a
5 b
y
5 1 1 b
s
5 v 12
in Equation 11, which is a standard way to proceed when no distinction is made between proportional and marginal income
tax rates. If v 5 1, the taxes fall fully on the firm. If v 5 0, they fall fully on the worker. In Table 4 we give the full set of restrictions
imposed on the wage equation.
Equation 11 can also be used to describe the demand for labor relation because this equation is in fact a vector of nine endogenous
variables. Thus, if we now normalize this vector with respect to logN, then we receive a demand for labor equation. In this case,
188
G. P.
Kouretas
and L.
P. Zarangas
Table 4: FIML Estimated Coefficients and Identification
Wage setting Labor demand
Output Coeff.
Variables b
i
a
i
b
i
a
i
b
i
a
i
b
w
W 1.000
0.003 20.163
20.100 1.260
20.008 b
P
P 21.616
0.022 0.523
20.014 21.912
20.002 b
N
N 1.045
0.015 1.000
20.012 20.414
0.018 b
Q
Q 21.955
0.041 21.142
0.125 1.000
0.009 b
ta
1 2 t
a
2.283 20.007
20.032 20.004
22.854 0.005
b
s
1 1 s 2.840
20.001 25.401
0.009 1.640
0.001 b
y
CPIP 22.704
20.011 20.338
0.045 2.300
20.009 bP
m
P
m
2.007 20.036
20.506 0.033
1.153 20.001
b
U
UNION 20.116
20.064 0.063
21.267 0.011
20.259
Hypothesis testing Wage setting
Labor demand Output
1 2
3 4
5
b
w
5 2b
P
5 b
w
5 2b
P
b
w
5 2b
P
b
w
5 2b
P
5 b
w
5 2b
P
5 2b
Pm
5 b
ta
, 2b
Pm
5 b
N
5 2b
Pm
5 b
s
, b
s
, b
ta
5 2b
y
, b
N
5 2b
Q
, 2b
Q
5 2b
ta
, b
ta
5 b
ye
5 b
ta
5 b
y
5 b
N
5 2b
Q
, b
ta
5 1 2b
s
5 b
ta
5 1 2b
s
5 b
U
5 0 b
Pm
5 b
U
5 0 b
s
5 b
Pm
5 2b
y
2b
y
b
U
5 0 x
2
5 1.03 x
2
5 5.03 x
2
5 7.33 x
2
5 3.99 x
2
5 5.78 Q
16 5 11.59 Q
16 5 10.45 Continued
WAGE SETTING,
TAXES, AND
LABOR IN
GREECE
189
Table 4: Continued
Estimates of the impact of cumulated innovations in variable j, on variable k kj
w P
N Q
1 2 t
a
1 1 s CPIP
P
m
UNION
w 20.082
20.610 0.171
0.129 20.958
4.049 0.300
20.2 0.018
P 0.021
0.477 0.176
0.142 0.035
0.604 0.202
0.4 0.008
N 20.009
0.230 0.431
20.135 0.405
20.379 0.803
20.1 0.001
Q 20.011
20.041 20.006
0.177 0.236
0.707 0.176
0.04 0.009
1 2 t
a
0.012 20.100
0.073 0.079
20.171 20.352
0.002 0.1
0.002 1 1 s
0.011 0.006
0.014 20.005
0.017 0.113
0.038 20.0
0.000 CPIP
0.035 20.111
0.381 0.162
20.254 21.335
0.331 0.2
0.006 P
m
20.015 0.592
0.089 0.289
0.152 21.101
20.073 0.7
0.009 UNION
0.584 25.291
23.975 7.262
23.659 17.710
25.972 4.2
0.243
Notes: All tests are x
2
distributed with 5, 6, 6, and 8 degrees of freedom respectively. Q. is the Johansen and Juselius 1994 test for overidentifying restrictions and is a x
2
distributed with 16 degrees of freedom. Notes:
The t-values for the impact-matrix for the MA representation can be found in Paruolo 1997.
190 G. P. Kouretas and L. P. Zarangas
the expected signs of the coefficients should be b
w
0, b
P
0, b
Q
0, b
ta
5 0, b
s
0, b
y
5 0, and possibly b
Pm
0. The impacts of both the income tax and the indirect tax derive from their wage
effect. If the union has a direct influence on employment for a given wage, then b
y
? 0. The restriction b
w
5 2b
P
5 2b
Q
5 b
s
. 0 implies that it is the real labor cost for a produced unit that is
important. Again, if it is the relative raw material price that is important, we write b
w
5 2 b
P
2 b
Pm
2 b
Q
5 b
s
. We concluded above that there are probably three cointegrating
vectors in our data space. Two of them are well specified, i.e., a wage-setting schedule and demand-for-labor condition. The third
eigenvector could describe either the demand side or the supply side of the variables concerned. Thus, we expect the following to
be possible candidates: 1 the supply of output, 2 the demand for output, 3 the constant mark-up pricing rule, and 4 price
setting conducted by the product-market demand conditions. How- ever, the resulting “semirelations” may also be mixtures of two
or more competing but misspecified relations. Hence, one should not put too much emphasis on the interpretation of the remaining
vectors. Table 4 introduces the three nonrestricted eigenvectors.
The strategy applied below is as follows. First, we do partial identification for each of the three vectors based on the wage setting,
demand for labor and the other “semirelations.” This sort of identification is done through the imposition of linear and homoge-
neous restrictions on each vector Johansen and Juselius, 1990, 1992. Second, following Johansen and Juselius 1994 and Jo-
hansen 1995a and the discussion in Section 3C, we provide formal identification of the joint hypotheses of the complete model.
5. TESTING STRUCTURAL HYPOTHESES