On the political and fiscal determinants of income redistribution under federalism and democracy: evidence from Germany

Abstract

Several recent studies suggest that transfers from central to regional governments are motivated by political considerations. In this paper we examine if this is also the case for transfers from regional to central governments in the context of the German fiscal equalization system. We examine the factors that contribute to differences in tax revenues across German states. The evidence indicates that both fiscal incentives and political factors can explain these differences, although in Germany the former are more important. Moreover, accounting for fiscal institutions has important consequences for the empirical assessment of political influences on taxation. Overall we find that the political affiliation of the state governor is an important factor in explaining differences in state tax revenues. Thus, the right-wing party (CDU/CSU) is effective in relaxing the tax burden at the state level. In contrast, partisan alignment between the state government and the federal government loses its importance once fiscal conditions enter the empirical model.

Appendix: Model diagnosis

As argued, the GLS type model implementation takes into account both cross sectionally heterogeneous residual variances and first-order serial correlation. From diagnostic results provided in Table 3 it is evident that both distributional characteristics are relevant to a considerable fraction of cross-sectional entities. For instance, in the most comprehensive specification (Model 3), the residual variance ratio computed against the variance estimate for the first cross-sectional entity (Baden Würtemberg) varies between 0.315 and 8.211. A similar range of variance ratios (0.24 to 5.46) is documented for Model 2. Although we cannot provide critical values for these (joint) ratios under the assumption of cross-sectional homoscedasticity, it is most likely that variance heterogeneity is characteristic for our panel.¹³ Moreover, according to a rule of thumb threshold of $|2/\sqrt{T}|=|2/\sqrt {14}|=0.535$, up to eight out of 16 residual processes extracted from Models 1, 2 and 3 are characterized by significant first-order serial correlation. The evidence for both heteroscedasticity and serial correlation vanishes for the GLS residuals. First-order serial correlation might be present for only one (Hamburg) out of the 16 GLS residual processes used in Model 3. Consequently, we regard the estimation results from GLS regressions as the most reliable. In addition to GLS variants that take into account serial correlation and cross-sectional heteroscedasticity we have also implemented panel designs that are based only on first-order serial correlation. The results of this model specification are qualitatively identical to those obtained from the more general specification of distributional properties. We do not provide these results for space considerations.

Table 3 Cross section specific standard errors $(\hat{\sigma}_{i})$, variance ratios ($\hat{\sigma}_{i}^{2}/\hat{\sigma}_{1}^{2}$) and first-order autoregression coefficients ($\hat{\rho}$) are provided for panel residual processes and evaluated for GLS, AR1 implied estimates of fixed effects, and vectors γ and β. Estimates $\hat{\varepsilon}_{\mathit{it}}$ and $\hat {\varepsilon }_{\mathit{it}}^{*}$ refer to ordinary (unweighted) and GLS weighted residuals, respectively. Diagnostics for ordinary and GLS type residuals are shown for Model 1 to Model 3 and (only for) Model 3, respectively. Abbreviations are as follows Baden-Würtemberg (BW), Bavaria (BV), Berlin (B), Brandenburg (BB), Bremen (HB), Hamburg (HH), Hesse (H), Mecklenburg/Western Pommerania (MWP), Lower Saxony (LS), North Rhine-Westphalia (NRW), Rhineland-Palatinate (RP), Saarland (SL), Saxony (S), Saxony Anhalt (SA), Schleswig Holstein (SH), Thuringia (T)

A further distributional characteristic is that residuals extracted from the common fixed effect within regressions of Model 3 display contemporaneous correlation. For instance, in absolute terms five (out of 120) cross correlation statistics exceed a critical value of 2.57 and, thus, hint at significance at the 1 % level. The evidence for contemporaneous correlation in the GLS residuals from Model 3 drops to three significant correlation estimates. Given that the empirical rejection rate for GLS residuals (2.5 %) is markedly closer to the nominal test level (1 %), we do not consider neglected cross correlation to be a serious matter of concern. Furthermore, given that the cross-sectional dimension of the investigated panel (N=16) exceeds the time dimension (T=14), we cannot account for cross sectional correlation without further a priori assumptions. Qualitatively similar results on the prevalence of cross sectional correlation hold with regard to the more restricted specifications of Model 1 and Model 2.