Meta-analysis approach
Given the relatively recent arrival of the meta-analysis technique in economics, there is no standardised empirical strategy. As previously explained, we restricted the sample to 124 observations (from a complete sample of 670 observations), although we will use the complete dataset to test the robustness of the results. Due to the presence of some extreme values in the sample, robust regression estimation is recommended.Footnote 10 We will also include some weighted least squares (WLS) estimates to control the quality of the studies.Footnote 11
We opt for a sequential approach, in line with Knell and Stix (2003), Evers et al. (2006) and Card et al. (2010). Firstly, we test the significance of geographical and temporal (by decades, from the 1950s to the 2000s) moderators, in the spirit of panel-data fixed effects estimation. This enables a basic specification to be obtained, in which we will include the main economic and methodological moderators as a second step (see Table 1, for the complete set). Instead of including all of these moderators simultaneously, we opt to include them one by one to avoid multicollinearity problems. Afterwards, we incorporate the significant moderators in a combined specification. As a final check, we perform an extreme-bounds analysis to evaluate whether results are influenced by the sequence of estimations.
In what follows, we will concentrate on reporting the significant results and focus on the role of economic moderators, which naturally suggest economic policy recommendations.Footnote 12 The meta-regression specification, based on Eq. 1, is straightforward
$$\begin{aligned} \varvec{\beta }_\mathbf s \mathbf =b+\alpha _\mathbf{s,i } \mathbf \,+\,\alpha _\mathbf{s,t } \mathbf \,+\,\gamma X _\mathbf s \mathbf \,+\,u _\mathbf s \end{aligned}$$
(2)
where \(\varvec{\beta }\)\(_\mathbf s \) is the reported wage elasticity to taxes in study s (defined as the impact of taxes on net wages), b is the ‘true’ elasticity, \(\varvec{\alpha }\)\(_\mathbf{s,i }\) and \(\varvec{\alpha }\)\(_\mathbf{s,t }\) are vectors with geographical and temporal ‘fixed effects’, X\(_\mathbf s \) is a vector of moderator variables, and u\(_\mathbf s \) is the error term. The independent variables are represented by dummies (for instance, if a study sample covers the period 1980–2008, D50 to D70 will be marked as a column of 0s for the particular study, while D80 to D00 will be represented by a column of 1s). X\(_\mathbf s \) will basically reflect imperfect competition variables stemming from the theoretical and empirical literature, such as the public sector effectiveness, labour market institutions (unions, wage bargaining), social security rules and tax wedge mix.
Results
Baseline specification (‘fixed effects’ moderators)
Firstly, we aim to obtain a baseline specification in the spirit of the fixed-effect panel estimation. Alternative estimations rejected robust geographical effects (we tested for differences in studies covering Latin America, Spain and the US vs. other OECD economies). For this reason, and since some of the economic moderators may be correlated, the basic specification does not include them. No significant temporal fixed effects, defined by the sample period in the data used (identified by the decades covered in case of time series analysis), are found either.
Another issue that should be controlled from the beginning in order to avoid spurious results is the presence of publication bias. It may arise from the tendency to report and/or to publish only the significant results, rejecting the null hypothesis of no effect. In order to reduce its potential influence, we included in our meta-database both published and unpublished papers. More than one third of the sample, 45 estimates, comes either from working papers or mimeos (see Table 2).Footnote 13
Based on the previous elements, in this initial benchmark specification the wage elasticity to taxes is \(-\)0.65, so workers bear almost two thirds of the tax burden (column I in Table 3).
Economic moderators
The attention devoted to the economic effects of labour taxation from international organisations, policy makers and academia ranks among the highest in applied macroeconomics, generating an enormous amount of research. However, the main issues affecting our results can be restricted to the following: the role of taxes under different institutional settings, the tax wedge composition, short- vs. long-run results and the role of the social security scheme.
Economic institutions matter. The theoretical and empirical literature shows that the impact of labour taxes on labour costs and unemployment is higher in economies with an intermediate centralisation of the wage bargaining process and a strong trade union presence (see Calmfors and Driffill 1988; Alesina and Perotti 1997; Daveri and Tabellini 2000). In this context, common to Continental and Mediterranean countries, the discipline effect of competition or the internalization of externalities are weak. This contrasts with the behaviour in decentralized Anglo-Saxon economies, and unionized centralized Nordic countries. Our analysis partially confirms this hypothesis. Column II in Table 3 shows that the average elasticity is \(-\)0.59. However, the degree of shifting is much higher in Nordic countries at \(-0.79 (-0.59- 0.20)\). This could be an indicator of the benefits of good governance, since Nordic public sectors are among the most effective (see Sapir 2006). By contrast, no significant differences were found between Anglo-Saxon and Continental-Mediterranean economies.
Throughout this paper we use the terms ‘tax wedge’, ‘labour taxation’ and ‘social security contributions’ in a flexible way, implicitly accepting the tax invariance theorem. However, there are several reasons to justify a differential impact of the fiscal wedge components. Focusing on social contributions, even though they are usually a payroll tax, their tax base and tariff usually differ from those of personal income taxes (not to mention the linkage effect), and consumption taxes (see OECD 1990, 2007). Additionally, the salary wedge includes the price wedge, that is, the gap between producer and consumer prices.
Formally, we define pit as the personal income tax effective rate, essc the employee social security contributions rate, fssc the firm social security contributions rate, ctthe consumption tax rate, pp the domestic output deflator, pp* international prices deflator, \(a\) the share of domestic products in the consumption basket, and \(e\) the exchange rate. The ‘salary wedge’ (wedge between the nominal labour costs and the net real wage) can be defined as \(\left[ {\frac{( {\text{1+ct}})( {\text{1+fssc}})}{( {\text{1-pit}})( {\text{1-essc}})}} \right]\left[ {\frac{\text{ epp*}}{\text{ pp}}} \right]^\mathrm{{1-a}}\). The ‘fiscal wedge’ includes exclusively the tax components of the salary wedge and can be expressed as \(\left[ {\frac{( {\text{1+ct}})( {\text{1+fssc}})}{( {\text{1-pit}})( {\text{1-essc}})}} \right]\). The ‘direct tax wedge’ only focuses on taxes on income as \(\left[ {\frac{( {\text{1+fssc}})}{( {\text{1-pit}})( {\text{1-essc}})}} \right]\) and, finally, the ‘employer social taxes wedge’ can be written as (1+fssc).
Therefore, we tested the impact of alternative fiscal wedge definitions, starting from the general salary wedge (fiscal wedge plus price wedge), down to employer social security contributions. Results confirm that different taxes have different effects on wages. While the elasticity of net wages to labour taxes is \(-\)0.59, the salary and fiscal wedge impact (inclusive of indirect taxes and of price wedge between output and consumption in the first case) is close to 80 % (\(-0.79 = -0.59-0.20\), as shown in column III in Table 3).
Moreover, as highlighted in Hamermesh (1993), the presence of lags in the shifting of social contribution is usually significant (on average, long-run shifting equilibrium takes more than a year) due to the presence of nominal rigidities. Our estimation confirms this point (column IV in Table 3). The short-run elasticity \(-0.43 (-0.74 + 0.31)\) is lower than the long-run counterpart, \(-\)0.74. Therefore, workers bear less than half of the tax burden in the short run (generally represented by results during the year of the tax change).
Finally, a key institutional issue that has to be tested is related to the social security scheme. In particular, the literature highlights a potential role of the ‘linkage effect’, that is, the perception of the link between contributions and social benefits (see Gruber and Krueger 1990; Gruber 1994a; Gruber 1994b; Disney 2004). Social contributions would have no effect on the equilibrium employment rate if agents perceive a full linkage effect, since they become a deferred salary (positively shifting the labour supply curve). This tends to be the case in contributory Bismarckian social security systems, in contrast to redistributive pension systems à la Beveridge. Our empirical analysis obtains the expected sign, since shifting is lower in more redistributive systems, but the effect is not significant (coefficient 0.10, column VIII in Table 6).Footnote 14
Other moderators
A second set of moderators tries to reflect methodological differences, from both the data (economy-wide or private sector, and its frequency) and the estimation techniques (cross-section, time series or panel, and selected estimator). Our analysis suggests that data coverage is relevant. Studies that refer to the entire economy, and not just the market economy, tend to show higher negative elasticities, \(-0.73 (-0.52-0.21\); column IX in Table 6). We are not convinced of its implications, since wage bargaining in the public sector is not competitive. Finally, more than half of the papers include labour costs, rather than the net wage, as the dependent variable. Those studies tend to obtain a higher negative elasticity (\(-0.80 = -0.43-0.37\); column X in Table 6), although no economic implication seems straightforward either.
Combined results
To conclude, we incorporated all the individually significant moderators into the specification (column V in Table 3). Every coefficient keeps its sign, but only the temporal focus and the dependent variable definition remain significant. Finally, column VI (Table 3) presents our baseline specification, focused on the moderators that have a more solid economic foundation. In this case, the net wage elasticity is \(-\)0.70, so workers bear 70 % of the tax burden via lower wages (or lower wage increases), somewhat lower than the starting estimate surveyed in Fuchs et al. (1998). In the Nordic economies the degree of shifting is higher, close to 90 % (\(-0.88 = -0.70-0.18\)), so nearly all the tax changes are offset by negative wage variations. Finally, we confirm that shifting takes time, and in the short run workers bear less than 40 % of the tax burden (\(-0.39 = -0.70 + 0.31\)).Footnote 15
Robustness checks
In order to further check the robustness of the results, we first tested whether the main results hold for the whole sample (670 elasticities from the 52 papers). Results, reported in columns XI to XV in Table 6, are in line with the ones reported for the core sample. These additional tests confirm that shifting is higher in Nordic economies, in studies that include indirect taxes in the tax wedge and in the long run. However, only the latter is significant. The publication bias hypothesis seems also present, with the aforementioned caveats on its downsizing effects to the sample. Taking all these considerations on board, our baseline specification for the complete sample is reported in column XIV. The net wage elasticity is \(-\)0.85, so workers bear 85 % of the tax burden. In the short term the backward shifting process is lower as well (\(-0.51 = -0.85 + 0.34\)).
Additionally, we performed a quality control study of the literature other than the publication format (books, journals, working papers and mimeos). Specifically, we used Google Scholar to weight each estimate for the number of citations.Footnote 16 The weighted least square estimation of the baseline sample is reported in columns XVI to XVIII in Table 6. In comparison to our baseline specification, estimates confirm that taxation shifting is significantly lower in the short run. However, the shifting degree seems higher, and does not depend on the economic model.
Finally, we tested how sensitive results are depending on the specification selection procedure. In particular, we calculated the confidence intervals of the impact of taxes on wages, based on the ‘extreme-bounds analysis’.Footnote 17 The lowest bound is obtained including fixed effects and controlling for the study coverage (ECO) and the publication format (BOOK, WP + MIM): \(-\)0.10 (\(-\)0.45 minus two standard errors). The highest one stems from a specification with the social security model (BEVSS) and the time horizon (DEPVAR): \(-\)1.09 (\(-\)0.77 plus two standard errors). Therefore, this test robustly rejects the no-shifting hypothesis.