Comparative fiscal illusion: a fiscal illusion index for the European Union

In the 1995, the European Union was enlarged to 15 member states.

Following Weck-Hannemann (1983), other economists have used this approach in the statistical analysis of the shadow economy, including Frey and Weck-Hannemann (1984), Schneider (2005), Dell’Anno (2007), Dell’Anno et al. (2007) and Dell’Anno and Solomon (2008).

In the standard MIMIC model (Jöreskog and Goldberger 1975), the measurement errors are assumed to be independent of each other, but this restriction could be relaxed (Stapleton 1978). In this analysis, several covariances between indicators are relaxed since they are empirically and theoretically plausible. Figure 1 shows some of these estimated covariances.

An alternative is to set the variance of the unobserved variable \(F\) at unity.

We find that estimated parameters are robust.

This evidence may suggest that tax revenue should be included in the model among both the causes and indicators of the latent variable. However, this cannot be done due to perfect collinearity. Thus, we checked information loss as consequence of excluding tax revenue as potential indicator of fiscal illusion. Regressing the tax burden on the existing six indicators, we found that four of six indicators have statistically significant coefficients and explain about 20 % of the tax revenue variance. In this sense, the exclusion of tax revenue from measurement model is partly accounted for by these variables.

According to Oates (1988), the results from the various studies of the revenue-complexity hypothesis are mixed. For instance, Clotfelter (1976) found no evidence that the Herfindahl index affects fiscal illusion. Moreover, Pommerehne and Schneider (1978), Dollery and Worthington (1999), Sausgruber and Tyran (2005), Mourão (2008), and Dell’Anno and Mourão (2012) find evidence supporting this hypothesis.

We also consider an interaction term \((\text{ Y }_{5}\times \text{ Y }_{6})\) instead of two separate variables. The estimates of the MIMIC models with multiplicative term do not converge towards satisfactory solutions. Thus, to take into account this interaction, we specify all the MIMIC models relaxing the assumption that the measurement errors of \(\text{ Y }_{5}\) and \(\text{ Y }_{6}\) are orthogonal.

The arrows that link some indicators and some causes indicate that the MIMIC model estimates covariances among structural and measurement errors. The rationale for these covariances is theoretically based.

To support this choice we apply the following stepwise approach: (1) factor analysis is applied to estimate the scores of latent variable on the measurement model; (2) the factors used to explain the covariance structure of the observed data are unobserved, but we estimated them from the loadings and observable data. These factor score estimates are used as substitutes for the higher-dimensional observed data; (3) estimated factor scores are used as dependent variables for three panel regressions in which the independent variables are the four “causes” of MIMIC model (structural equation). For each of these regressions we performed redundant fixed effects tests. According to Likelihood tests, the cross-country fixed effects are statistically significant. This implies that cross-unit fixed effects have to be included in the model specification.

We estimated several MIMIC specifications to check the robustness of the index with different sets of observed variables. The estimates proved robust. Omitted outputs will be provided by the authors upon request.

In regression analysis, the omission of relevant variables can lead to biased coefficient estimates. To take into account this issue, the correlations between errors of the different equations of the measurement and variables of the structural model are included in the model estimation. It implies we specify a non-diagonal covariance matrix of the measurement errors \(\Theta _\varepsilon \) and structural disturbances \(\Psi \).

Data on the revenue from each tax are not available. Therefore, as is usual in the literature on the Herfindahl index of tax revenue, we use diverse types of taxation instead of the number of different taxes. This procedure makes Herfindahl index far from being a perfect measure of tax complexity, since it does not account for the effective sources of tax complexity, but measures complexity through the proportions of revenue collected by diverse types of taxation (e.g. direct, excise, capital). It implies that two countries (A and B) with the same level of tax revenue, but different levels of tax complexity, have the same values of the Herfindahl index. It occurs, for instance, if country A collected tax through a single direct tax (or excise), while country B levies several direct taxes (or excises). It can explain the ‘gap’ between theory and empirical evidence.

With the exception of MIMIC 4-1-6a RML.

Including outliers, \({R}^{2}=0.41\).