• Keine Ergebnisse gefunden

Part I: Political Economy of Climate Policy

3 Public Interest vs. Interest Groups: Allowance Allocation in the EU Emissions

4.2 Estimation results

Paying the Piper and Calling the Tune? 79 4.4.2. Table 20 and 5 of Appendix 4.4.2 report the summary statistics for metric and dummy variables.

Besides assessing these observable determinants of publication outcomes, we additionally control for unobservable study characteristics. Following Nelson and Kennedy (2008), we include dummy variables for those studies that feature multiple results via alternative model assumptions and thus have a relatively high weight in our sample of publications (yielding in total 15 study dummies).40 These dummy variables take over the value 1 for each simulation result of a study and 0 otherwise. They capture all explanatory information of the publication outcome that is not covered by our core regressors, thus accounting for omitted variable bias and representing unobservable study characteristics not reported consistently across the respective articles. We test for the joint significance of these additional dummy variables using the Wald test for the parameters of the correspondingly fitted regression model.

Finally, Stanley and Jarrell (1989) emphasize that meta-regression errors are likely to be heteroscedastic41 because studies may differ in the employed datasets and other characteristics. Thus, we will test for heteroscedasticity and employ robust estimation techniques.

Paying the Piper and Calling the Tune?

80

Stanley and Jarrell (1989) regarding heteroscedasticity: For the standard OLS model, the Breusch-Pagan test (χ2(1) = 13.73) indicates that we must reject the null hypothesis of equal variance. As a consequence, we account for White’s heteroskedasticity-robust coefficient estimators. Furthermore, due to differing units of our regression variables, coefficients have been standardized yielding so-called Beta coefficients.

Table 17 presents the OLS parameter estimation for the linear regression model with the simulated employment effect as the dependent variable. It shows that if we account for observable study characteristics, employment effects are significantly determined by the magnitude of emission reduction: the higher the emission reduction due to environmental taxation, the smaller the prospects for an employment dividend (i.e. higher emission reductions increase employment losses or likewise decrease employment gains). Table 17 further indicates that for our data set the employment effects of double-dividend studies are not significantly influenced by the underlying simulation period. The same finding is true both for the regional focus of the study as well as the choice of the model type.

Paying the Piper and Calling the Tune? 81 Table 17: Observable study characteristics: parameter estimation by OLS

Dependent variable Independent

variables

Employment

Emissions 0.333 ***

(3.54)

Simulation period -0.041

(-0.39)

Germany -0.047

(-0.31)

Austria_Switzerland -0.135

(-1.17)

Model_Macro -0.222

(-1.42)

Labor_Fixed 0.328 *

(1.97)

Labor_Bargaining 0.136 **

(2.03)

Labor_WageCurve 0.039

(0.40)

Contract_Environment -0.040

(-0.42)

Contract_Industry -0.278 **

(-2.23)

Constant 0.013 **

(2.46) Number of observations

Goodness of fit F-test

Ramsey RESET test

73 R2 = 0.41 F(10, 62) = 4.08 ***

F(6, 56) = 1.72

T-statistics in parentheses. * (**, ***) implies that the null hypothesis of the respective parameter being zero can be rejected at the 10% (5%, 1%) level of significance (according to the corresponding two-tailed test).

We further find that employment effects are driven by labor market assumptions: Table 17 reveals that the assumption of imperfect labor markets – either represented by fixed real wages or a unionized labor market with a bargaining process – leads to significantly larger simulated employment gains (or smaller losses) than the assumption of a perfect labor market (i.e. the reference category). This result confirms the theoretical arguments that (i) in the presence of (long-term) real wage rigidities a cut in labor taxes or a reduction of social security contributions may reduce labor costs and increase employment and (ii) in

Paying the Piper and Calling the Tune?

82

bargaining models the chance for an employment dividend increases, as tax shifting from workers to the unemployed becomes possible, leading to wage moderation and thus to lower producer wages. On the contrary, the simulated employment effects of modeling labor market imperfections via a wage curve mechanism do not differ significantly from assuming a perfect labor market.

Our meta-regression analysis yields an interesting result concerning the contracting body: a significantly negative impact of the dummy variable Contract_Industry on the employment effect. While studies with an environmental contracting body do not show significantly different labor market impacts than non-commissioned studies, it implies that those publications commissioned by an industrial contracting body identify larger employment losses (or smaller gains) induced by the environmental tax reform. Given that employment losses are a popular and effective argument against environmental regulation in the context of high unemployment rates, this result seems to back the reservation of industrial interest groups towards environmental tax reforms. Besides looking at the order of magnitude of employment changes, we run additional regressions to assess the sign of employment effects in order to reflect the “existence debate” of the double dividend debate.42

The lower part of Table 17 provides additional diagnostics for our linear regression model:

Besides a goodness of fit of 0.41, it shows a high overall significance of the included independent variables (see F-test). Moreover, we employ the RESET test using the powers of the independent variables in order to test for potential specification errors. It shows that we are not able to reject the null hypothesis of no specification error for our regression model (in other words, we do not find significant evidence for misspecification).

4.2.2 Controlling for unobservable characteristics

Our second regression model additionally controls for the role of unobservable study characteristics. As discussed in the previous section, the additional dummy variables capture all explanatory information that is not covered by our core regressors representing observable study features. As before the regression diagnostics supports the concerns by Stanley and Jarrell (1989) regarding heteroscedasticity: For the standard OLS model, the Breusch-Pagan test (χ2(1) = 57.86) indicates that we must reject the null hypothesis of equal variance. We

42 Employing a logistic regression model with the simulated employment effect as a dichotomous dependent variable largely confirms the above results (the corresponding estimation results are available upon request from the authors).

Paying the Piper and Calling the Tune? 83 therefore take into account White’s heteroskedasticity-robust coefficient estimators and present Beta coefficients. Table 18 presents the corresponding estimation results.

Table 18: Controlling for unobservable study characteristics: parameter estimation by OLS Dependent

variable Independent

variables

Employment

Emissions 0.354 ***

(2.68)

Simulation period 0.211

(0.71)

Germany 0.210

(1.10)

Austria_Switzerland -0.051

(-0.24)

Model_Macro -0.198

(-1.02)

Labor_Fixed 0.304

(1.15)

Labor_Bargaining 0.152

(0.78)

Labor_WageCurve 0.168

(1.15)

Contract_Environment 0.024

(0.20)

Contract_Industry -0.376

(-1.35)

Constant 0.515

(0.66) Number of observations

Goodness of fit F-test

Ramsey RESET test

73 R2 = 0.61 F(25, 47) = 9.33 ***

F(6, 41) = 0.97

T-statistics in parentheses. * (**, ***) implies that the null hypothesis of the respective parameter being zero can be rejected at the 10% (5%, 1%) level of significance (according to the corresponding two-tailed test).

The table presents an interesting finding of the extended regression model that controls for unobservable study characteristics: Compared with the initial regression model that explained the simulated employment effect by observable study features alone, only the coefficient of

Paying the Piper and Calling the Tune?

84

the emissions variable is equally significant. This emphasizes the crucial importance of the stringency of environmental taxation for the prospects of obtaining a double dividend. When we account for unobservable study characteristics, neither the specification of imperfect labor markets, nor the industrial contracting body exerts a significant stand-lone impact on the simulated employment effect.43 Nevertheless, the three model assumptions that entered significantly in our first regression (Emissions, Labor_Fixed and Labor_Bargaining) still play a joint role for the publication outcome: for our second regression, the corresponding Wald test is highly significant (F( 3, 47) = 3.79***). At the same time, also the Wald test for the parameters of those dummy variables representing unobservable study characteristics shows a high joint significance of the corresponding parameters (F(15, 47) = 4.11***), which substantiates the high relevance of implicit publication features. We conclude that when controlling for non-observable study characteristics, both the contracting body and specific model assumptions do no longer play a significant role for the simulated employment effect of environmental taxation. In contrast, the average publication outcome of our sample is determined by a joint set of modeling features as well as implicit study characteristics.

The lower part of Table 18 provides additional diagnostics for our linear regression model:

Besides an improved goodness of fit of 0.61, it shows a higher overall significance of the included independent variables as the previous model accounting for observable study features only (see F-test). While we do not find significant evidence for misspecification either, the extended regression model controlling for the role of unobservable study characteristics features an even lower F-statistics of the RESET test.