Labour responsiveness to income tax changes: empirical evidence from a DID analysis of an income tax treatment in Italy

To conduct the empirical study by means of a DID model, we exploit a quasi-natural tax experiment that occurred in Italy at the end of 2006. The characteristics of the tax treatment (timing and treatment assignment) and the time pattern that response variables follow before the treatment, as well as the absence of confounding factors, have important implications for our empirical approach. They make the statistical framework of our analysis consistent with the identification requirements necessary for the efficient estimation of the Average Treatment Effect upon Treated (ATET) discussed by Roth et al. (2023, p. 2220) via DID under treatment homogeneity. The treatment affected a subset of Italian regions (treatment group) from 2007 onwards and consisted in an exogenously imposed increase in their regional income tax rates with respect to the rates existing in the other regions (control group) which remained unchanged. The new rates were imposed by the central government for reasons that we discuss later. Hence, the 2006/7 central government policy defines the framework of a quasi-natural tax experiment, fully described in Appendix 1, involving the taxpayers of five treated regions (not chosen at random) vs the taxpayers of the remaining fifteen untreated control regions. The goal of the estimation is to evaluate if exogenous differences in regional income taxation resulting from the central government income tax policy generated statistically significant responses of the labour supply of treated subjects. We also study the time adjustments of labour supply—and of other real variables such as the per capita regional income, its growth, and the family consumption—in the period following the income tax changes to evaluate how persistent the effect of income tax increase is over time. Results show that the negative adjustments of various response variables measuring the supply of labour services of treated taxpayers are statistically significant, fast, and strong but not long-time lasting. Important for tax policy are the results concerning the response to tax hikes of Self-Employed people—possibly the taxpayers mostly affected by the 2006/7 tax treatment. Our DID estimations show that the share of Self-Employed workers is reduced by the increase in income taxation. Not surprisingly, we also find that treated families reduce in a similar manner their consumption. Analogous adjustment responses to tax hikes characterise the growth of regional per capita GDP. Altogether, our results show that exogenous income tax hikes affect the behaviour of workers and their labour supply. The result is robust to changes in the response variables and to the use of a larger data set. For instance, using other labour response variables (e.g. the annual New VAT Certificates, necessary for conducting self-employment activities) we obtain identical findings. As for the above-mentioned persistency of the ATET over time, we find that the tax effect is not long-time lasting. After treatment, adjustments of the labour supply to a long-run trend are observable, on average, 3 years after the introduction of the treatment. We also find that anticipatory effects are generally absent which is another valuable element for policy considerations.

The paper is structured as follows. After a brief Introduction presented in Sect. 1, the relevant literature is discussed in Sect. 2. The tax factors considered in this study are described in Sect. 3, together with the short narrative of the Italian 2006/7 tax experiment. Details on the latter are discussed in Appendix 1. Section 4 contains the discussion of the DID identification hypothesis in terms of a TWFE model as well as a statistical survey of the data with pre- and post-treatment plots of the candidate response variables. In Sect. 5, we present the empirical results, which are commented upon in comparison with the results of the existing international literature. In Sect. 6, we conduct a robustness analysis by using provincial, rather than regional, response variables measuring other indicators of labour activities and a larger data set. We employ an augmented version of the previous basic TWFE DID model which includes cofactors clustered at both provincial and regional levels to evaluate the sensitiveness of ATET to regional and provincial cofactors. Results support the findings of our previous estimated model that tax hikes have a negative but not time lasting impact on labour supply. Policy implications are also discussed. Section 7 presents some additional test procedures permitting to implement what we call a partial falsification test of the parallel trend assumption of the DID model. The test is based on convergence analysis for detecting the presence of parallel trends during the pre-treatment period. Section 8 briefly concludes. The spatial–temporal patterns observed for every response variable in treated and untreated regions are shown in Appendix 2.

The empirical literature studying the relationships between income taxation and variables conducive to growth generally finds indications of heterogeneous aggregate effects of income tax changes. Zidar (2019, p. 1440). Some studies, for instance, provide evidence that lower-income households tend to have higher marginal propensities to consume (Dynan et al. 2004; Jappelli and Pistaferri 2010; Johnson et al. 2006; McCarthy 1995; Parker 1999; Parker et al. 2013) and so they respond to tax policy by increasing consumption more than high-income groups. In analogous way, labour supply and other growth-related variables of different income groups may respond to income tax policy in an opposite way.

Focusing on labour supply, one should recall that the empirical literature studying the relationships between income taxation and labour supply is characterised by considerable controversy over how to estimate the responsiveness of labour supply to changes in wages and taxes. Keane (2011) and Stantcheva (2019) discuss the main econometric problems facing early studies and subsequent developments and the estimated values of the labour supply elasticities (Keane 2022). We refer the readers to their discussions. On the extensive margin for lower-income groups, Eissa and Liebman (1996) and Meyer and Rosenbaum (2001) show that in the USA the Earned Income Tax Credit (a refundable tax credit enacted in 1975 for low- to moderate-income working individuals and couples, particularly those with children) has increased labour force participation. Other authors (e.g. Romer and Romer 2010; Saez et al. 2012) show that for high-income earners there is some evidence that the efficiency costs of raising taxes on top-income taxpayers expressed in terms of labour supply and other margins may be limited or can be offset by shifting in the timing or form/source of income acquisition (Auerbach and Siegel 2000; Goolsbee 2000). Ziliak and Kniesner (2005) find that consumption and worked hours are direct complements in utility and increase (both) with a compensated increase in net wage. Other studies of labour response to tax changes analyse the potential adverse base effects of tax hikes such as tax avoidance, outright tax evasion, and a general reduction in economic activity (Piketty et al. 2014). They conclude that income tax can be responsible not only of a reduction in labour services supplied but also of a general contraction of the economic activity. As for self-employment (one of the response variables employed in this paper), Heim (2010), Appelbaum and Katz (1986), and Kihlstrom and Laffont (1979) consider explicitly how stochastic shocks in demand or cost functions affect the decision to become self-employed. In those models, an asymmetric tax reduction, favouring the self-employed relative to other workers, increases self-employment. We are interested in estimating if a negative effect of income tax hikes (a reduction in self-employed people after the tax increase) is consistent with the data.

When territorial aspects are accounted for, the analysis considers the tax-induced mobility of taxpayers and discusses whether income tax rates can be employed by local authorities to attract highly qualified foreigner taxpayers. The literature shows that local income taxes may affect local decisions (such as regional/local labour supply or residential location) and evaluates the possible consequences of these decisions on the level and geographical distribution of productive activities and on local technological spill over (Widmann 2023). For instance, Akcigit et al. (2016) find that “prolific” inventors migrate between countries in response to changes in personal top-income tax rates and obtain an estimate of the elasticity of the number of foreign inventors (in a country) with respect to the top net-of-tax rate close to one. Moretti and Wilson (2017) examine the mobility responses of inventors to changes in personal top-income tax rates across US states and find a corresponding elasticity with respect to the top net-of-tax rate of 1.8. In addition, since the market value of material goods, and its geographical distribution, may be affected by income taxation, other studies show that local income tax rates can generate tax externalities of various nature across local jurisdictions (Esteller-Moré and Solé-Ollé, 2002).

Implementing the DID approach, Jakobsen et al. (2019) provide causal evidence of the impact of net-of-tax rate on wealth accumulation in Denmark, whereas Kleven et al. (2013) study the causal relationship between tax rates and a particular subset of the population of migrant workers (football-players called super stars). Baskaran (2021) presents results from a set of municipalities who, in North Rhine-Westphalia, increased their local property and business tax rates. He studies the revenue and base effects of local property and business tax hikes in a generalised DID design. His results suggest that the property tax hikes had a revenue elasticity of unity even in the long run. He concludes that tax changes had no adverse effects on property tax bases. For the business tax, he finds no significant effects on revenues and tax bases.

The measures adopted by the Italian government at the end of 2006 are a good example of a possible exogenous income tax treatment. In December 2006, the Italian central government decided to impose an increase—automatic and compulsory—of two regional income tax rates, namely IRAP tax rates and regional Income Tax Surcharge to regions who had increased, with respect to 2005, the deficit of their regional health care systems above a certain percentage. The other regions were not affected by the tax measure.¹ The measure was adopted to force those regions to collect additional local resources and use them exclusively to reduce the deficits of their health system. The government finally decided that the percentage increase to be used as a threshold for the inclusion in the treatment group was the 7% of the previous year deficit. Five out of twenty Italian regions (Abruzzo, Campania, Lazio, Molise, and Sicily) fell into the treatment group. The content of the 2006 policy (described in details by Caruso and Dirindin (2019) is summarised below; a more complete description of the taxes involved in the treatment is shown in Appendix 1):

For both the above taxes, the increase in the tax rates was compulsory, quantitatively relevant and finalised to collect some additional tax yield that the involved regions had to employ for the reduction in the prolonged budget unbalances of their regional health care systems. No other uses were permitted. At the end of 2006, the law introducing the new tax regimes for the above five regions indicated that the measures had to be adopted for a 3-year period, but the measures remained in force for the entire sample period used in this study (see below). Then, the condition of Irreversibility of Treatment applies to our case because once a region became treated in 2007 that region remained treated in the next periods.

Relevant for the applicability of our method is the circumstance that almost any Italian region could had been assigned to the treatment group. The health deficit was high and time increasing in the overwhelming majority of the regions, and it was completely unknown ex ante at what level the central government was going to set the above cut-off threshold determining a non-penalising annual tax increase. Some regions (e.g. Liguria) were expecting to be included in the treatment but remained out, whereas some other regions were expecting to remain out but were included (e.g. Lazio). The central government determined the above-mentioned 7% value in the last days of 2006 when the budget law for 2007 was finally approved by the Parliament and the 7% value arrived quite unexpected, according to national newspapers. As a result, at the end of 2006 each region had a positive probability of being eligible as a member of the treatment group based on her previous and irreversible behaviour. Therefore, in the Italian tax experiment the randomness in the data does not come from a pure stochastic assignment of regions to treatment or from drawing regions from an infinite super-population of regions. Actually, the Italian tax experiment is a variant of a “design-based” experiment with (almost) random participation based on past behaviour of treated and untreated units. Past behaviour qualified regional taxpayers as future treated or untreated units and for this reason the inference drawn from our DID analysis can be seen as a “design-based inference”. In this connection, Roth et al. (2023, p. 2219) recall that methods that are valid from the canonical sampling-based view are typically also valid from the design-based view as well, with the recommendation of clustering standard errors at the level at which the treatment is independently assigned (the regional level).

Finally, it is worth noting, as a complication, that the design of the tax measure introduced with the 2006/7 treatment does not allow to exclude that some interactions in the response variables could be possible as well as special correlation in the treatment status among individuals.²

We follow Roth et al. (2023, p. 2220) and adopt the identification hypothesis of a TWFE panel data DID method to conduct the causal–effect study of the impact of a central government exogenous income tax policy. We employ response variables related to labour supply and other growth-related variables (such as family consumption) to evaluate the short- and long-run impact of income tax hikes taking both time and individual effects into account.³ To illustrate the estimated model, we introduce the following notations:

The total sample period is T = (1995… 2021) with t = (1995 … 2006) being the non-treatment period (dummy variable D2 = 0) and t = (2007 … 2021) being the treatment period (dummy variable D2 = 1).

As noted above, we define the target or response variable \(y\) in terms of potential outcomes and estimate the average effect of treatment on the treated units. This compares the potential outcomes with treatment to the potential outcomes with no treatment, in the treated group. Written mathematically, the estimated effect of the treatment is the β coefficient of panel data OLS equation:where \({\varepsilon }_{it}\) is an idiosyncratic error term. The crucial parameter to estimate is β. If we call R a binary income tax treatment indicator, we have \(\widehat{\beta }\equiv \text{CATET}={\mathbb{E}}\left({y}_{i,t}^{1}-{y}_{i,t}^{0}|{X}_{it};{R}_{i,t}\right)\) or \(\widehat{\beta }\equiv \text{ATET}={\mathbb{E}}\left({y}_{i,t}^{1}-{y}_{i,t}^{0}|{R}_{i,t}\right)\) if cofactors X are included or excluded. CATET (Conditional Average Effect upon Treated) or ATET will be interpreted as the mean effect (conditional or unconditional) of the “tax treatment” for those taxpayers who were compelled to participate in the central government program of income tax changes (i.e. the residents in the treated areas).

Unless it is not otherwise specified, in all estimates the average treatment effect is estimated by adjusting for both cross-sectional and time effects. Notice that, given the design of the tax treatment, the response variables used in the estimations will all be independent on treatment conditional upon the X that will be used. Then in addition to the assumption of the parallel trend and the absence of anticipatory effects, our DID specification relies also on the unconfoundedness hypothesis (treatment participation is uncorrelated with the realisations of response variables). For Eq. (1) to identify the parameters of interest, tax shocks need to be exogenous conditional on fixed effects and controls. Intuitively, this identifying assumption is that the national tax shock of 2007 is not favouring regions that are doing poorly relative to how fast they normally perform in terms of response variables. Then, the validity of comparing outcomes of regions having different distributions of the response variables relies on three key assumptions: (1) the national tax shock is exogenous, (2) the targeted tax shocks are unrelated to any possible targeted level of the response variables, and (3) the outcomes from untreated regions provide a reasonable counterfactual. Since we control for region and year fixed effects in Eq. (1), the first assumption maintains that Italian policy makers of the time were not systematically setting income tax policy to respond to idiosyncratic regional shocks other than the budget unbalance of the health care systems mentioned in Sect. 2. Notice that the central government income tax policy introduced a contemporaneous treatment to a fixed group of regions (treatment effect homogeneity). Then, our TWFE DID model (1) does not make “forbidden comparisons” between already treated units (Roth et al. 2023, p. 2228).

Yet, the fact that assumptions (1), (2), and (3) permit to use TWFE regression specification for the estimation of the treatment parameter does not imply that other conditions for efficient parameter estimations of an OLS-based model are met to obtain asymptotically valid inference, particularly when the sample period is as large as ours. Stationarity and cross-sectional dependence are two issues. The former is not an absolute requirement for panel data analysis, but it is preferred in many cases. With a short time series of, for instance, up to 10 years, stationarity may not be critical. In that case, nonstationary of the data still allows using fixed effects (or first differences) models to control for unobserved heterogeneity and trends over time. Yet, with a sample like ours (27 years), it is necessary to conduct tests to check for the presence of panel unit root because, in case of non-stationarity, the use of OLS-like procedures can produce invalid estimates. Granger and Newbold (1974) called such estimates "spurious regression" results: high R² values and high t-ratios yielding results with no real meaning. As it is shown in Table 1 (see Sect. 4.1), we reject the panel unit root hypothesis at usual significance levels and interpret the result as evidence that a statistically significant proportion of the units (both treated and untreated) are stationary.

TWFE panel DID results are presented in Table 2. The ultimate causal parameters of interest (ATET) are reported in the first row. Tests of parallel trend and anticipation effects are reported in the last rows and commented in Notes in the table. We refer the reader to Table 7 in Appendix for a detailed summary of the estimated coefficients, including those associated with temporal and cross-sectional fixed effects. The plots of the ATET evolution over the period following the tax shock are shown in Fig. 2 after the sections containing the discussion of the estimations of the tax effect on each response variable.

Before commenting the specific results for each response variable, we discuss the test for the presence of the parallel trend and anticipation effects. The last two rows of Table 2 report the estimated F-tests for parallel trend hypothesis and the test for absence of anticipation effects. The former is based on the results of an augmented linear DID model based on our Eq. (1) with additional terms that capture the differences in slopes between treated and controls. As for anticipation effects, Table 2 also reports the results of a Granger type test obtained from a linear DID model based once again on Eq. (1) augment in this case by interacting the dummy variable that marks treated observations with dummy variables for time periods prior to the treatment to capture any potential anticipatory treatment effects.

The augmented DID model employed for parallel trend test is the followingwhere S₀ is a time dummy indicating pre-treatment years, S₁ indicates treatment years, and w indicates either treated regions (i = 1) or untreated regions (i = 0). If the estimated ψ₁ ≠ 0, then a difference in the slope of the time trend between treated and untreated units in pre-treatment years is statistically consistent with the data. In Table 1, we report the Wald test of the null hypothesis H₀: ψ₁ ≠ 0 versus ψ₁ = 0. The null is that of linear parallel trends (ψ₁ = 0). This means that the test statistics is obtained by dividing the maximum likelihood estimate of the slope parameter ψ₁ by the estimate of its standard error. Under the above null, this ratio follows a standard normal distribution.

As for the anticipation effects we adopt a procedure similar to a Granger model, but with a different parameterisation that includes lags in addition to leads. Recall that 2006 is the last year before treatment. Then, let m be the number of periods prior to 2006, q ≥ 0 be the number of periods after 2006, and b is the baseline period. The fitted model is (using MOD1 of Table 2 as reference; no cofactors included)where \({w}_{i}\) has the same meaning as above (see fn. 4) and \({B}_{it}^{k}=({t}_{it}={I}_{i}+k)\) because we use all available leads and lags. Table 2 reports tests that coefficient referring to years prior 2007 are significantly different than zero.

The above Wald test for parallel trend and the Granger type test for anticipation effects are adopted in all the estimation procedures presented in the paper as it is recalled in Notes below each table.

Yet, one should recall that the Granger type test is less robust than the above parallel trend F-test since it requires more degrees of freedom because it estimates a higher number of parameters. Hence, the parallel trends test has higher statistical power. The results allow us to conclude (at any significance level) that the data generation process is consistent with the hypothesis of parallel trend but not always with the absence of anticipation effects hypothesis. The latter result may be informative for the policy design of future tax treatments. In the following sections, the estimation results of each response variable are commented with greater emphasis on those related to labour supply.

We replicate the TWFE panel DID estimation using Rate of Employment (male and female, age from 15 to 64 years, Eurostat NUTS-3 data) recorded at the provincial level in treated and untreated regions from 2004 to 2021. Altogether, we use data of 103 provinces distributed among 20 regions: 5 treated and 15 untreated. The model is identified by the same basic TWFE DID assumptions of Eq. (1), plus the requirement that each province belonged to the same region at the beginning and at the end of the sample period 2004–2020 (no province abandoned the original region). We also include cofactors relative to both provincial and regional levels. Equation (1) rewrites aswhere \(t=(2004, \dots , 2021)\) and \({X}_{ijt}\) is the year value of each of the possible \(P\) cofactors existing at the provincial level \(i\) (when \(i\) belongs to region j) and \({W}_{jt}\) is the year value of each possible \(R\) cofactor existing at the regional level j and associated with the provinces nested in that region (for each \(i\in j\)). Finally, \({u}_{ijt}\) is an error term normally distributed. Clearly, the Tax Treatment Component is the same as in Eq. (1).

All estimated versions reported below include both unit fixed effects and time fixed effects in ordinary least squares estimation. Table 3 reposts results of various version of the model.

Results reported in Table 3 indicate that the exogenous increase in the income taxation affects negatively the rate of employment in any specification of the estimated model. As expected, the presence of cofactors affects the magnitude of the estimated ATET (or CATET) coefficients. In MOD1 for example, a one point percentage increase in the income tax rate leads to a 1.7% decrease of the ongoing mean employment rate in treated provinces, whereas in MOD3 and MOD4 the effect reduces to a 1.3%. Province level cofactor (GDPs) is strongly statistically significant (for the interpretation of the magnitude, see Notes in Table 2) and has the expected sign. As for regional level cofactors, GDP has effects analogous to the provincial GDP, but its rate of growth has no statistically significant effect. Controlling for the National Mean Employment Rate has no consequences. Graphical illustration of the existence of parallel trends is reproduced in Fig. 3. The plot supports the results of the tests reported in the last rows of Table 2 and those commented about in Notes in the table and in the footnote. Yet, both procedures lead us to conclude that the pre-treatment parallel trend assumption is consistent with the data.

The presence of different trends between treated and untreated units may jeopardise the parallel trend assumption of the DID. Given the macroeconomic nature of our regional data set, the presence of converging trends of some response variables cannot be excluded a priori. When the time period tested is short (for example a 2-year period), the assumption is the more likely to hold. Yet with a period of 27 years the possibility that treated and control groups have different outcome trends (which may generate time convergence or time divergence among them) cannot be excluded. Then, in addition to the tests reported in the last rows of Table 2 and the information provided by the visual inspection of the plots of Fig. 1, in this section we propose a falsification test that at least excludes that the response variables of treated and untreated units follow some converging trend. Time convergence, i.e. converging trends of the response variables for treated and control groups (employment, labour supply, consumption, etc.), would then be at odds with the hypothesis of parallel trends. This is so because the change (or “path”) in response variables over time that regions in the treated group would have experienced if they had not been included in the treatment may not be the same as the path of response variables that regions in the control group actually experienced. In the case of a quasi-natural experiment where assignment is possibly not entirely random, the treatment assignment must be mean independent of factors that affect the trend in the response variables. Roth et al. (2023, p. 2222) clarify that if the bias for selecting into the treatment is the same each year, then parallel trends allow for selection bias. Excluding converging trends is not synonymous of validating parallel trends, however. Trends can be non-parallel because they might diverge. Also discovering that the trends diverge would be a sign that the treatment assignment may not be mean-independent of factors that affect the trend in the response variables and that there is absence of parallel trends. Hence, discovering that response variables do no converge is not necessarily a sign of parallel trends.

Bearing in mind the above limitation, we propose an application of the log(t) convergence test of Phillips and Sul (2007, 2009), which proposed the “log(t)” regression test approach to test for the convergence hypothesis based on a nonlinear time-varying factor model. The proposed approach has the following merits: first, it accommodates heterogeneous agent behaviour and evolution in that behaviour. Second, the proposed test does not impose any particular assumptions concerning trend stationarity or stochastic non-stationarity, and consequently, it is robust to the stationarity property of the series. In other words, it could be used whatever the unit root results reported in Table 1. Yet, to make the test more adherent to the time span of the Italian tax treatment we apply separately the test to the entire sample period (1995–2021), to the pre- and post-treatment periods, and to the two subsets of data (treated and control units) separately.

Results of log(t) tests are reported in Tables 4, 5, and 6.

Results show that during the pre-treatment period the hypothesis of convergence is not consistent with the data and this finding supports the DID parallel trend hypothesis. In one case (i.e. per capita GDP growth), the null hypothesis of convergence cannot be rejected at usual significance levels for the entire time period and for the post-treatment period (this contradicts the Ptrend test reported in Table 2 for that variable), but no convergence is found for the pre-treatment period.

Obtaining the rejection of the null of convergence of the response variables when pooling treated and control observations together just indicates that the series do not converge. However, it does not exclude that trends could be not parallel because they diverge. Yet, a different result (not rejecting the null hypothesis of convergence) would be a clear contradiction of the parallel trend hypothesis and that is why we propose this test (which gives possible indeterminate outcomes) in addition to those reported in Table 2.

We have analysed the effects of the Italian binary income tax treatment that was adopted at the end of 2006 and remained on afterwards. It affected taxpayers in a treated group of regions with no variation in treatment timing and assignment. TWFE panel DID results provide robust evidence that labour supply and other regional growth-related variables are responsive to income tax hikes. Robustness analysis conducted by using a multi-level TWFE DID panel, which includes provincial and regional level cofactors, supports the results.

We found that the presence of anticipatory effects is not consistent with the data, and this may be relevant in terms of evaluation of tax policy measures. The absence of anticipatory effects somehow implies that the specific tax bases of the two income taxes analysed in this paper may be more rigid to anticipatory adjustments on the part of taxpayers than the tax base given by the general personal income tax.

Finally, our DID estimates of the ATET of the income tax increase on regional per capita GDP and its growth indicate that they react negatively to a heavier income taxation. Results of Table 7 in Appendix 2, where time effects are reported, indicate that the ATET value of the tax treatment on the per capita GDP is 630 euros which is interpretable as the average reduction—across time and treated regions—of the net-of-tax personal disposable income. Per capita personal income growth is negatively affected by an income tax increase, too. Both results are consistent with the estimated contraction of labour supply and family consumption. Yet, estimates show that the negative effects of the tax increase on per capita GDP growth loses significance at the end of the sample period. This finding shows that per capita GDP is more sensitive to long-term than to short-term fluctuations, including income tax hike. We obtained the opposite result for labour supply. In the long run, the dynamic of regional per capita income and its growth is affected by a larger set of factors, such as those entering the local multiplier of the public expenditure financed with the additional tax yield. Then, long-run effects of income tax increase should be evaluated by weighting the effects of taxes, on the one hand, and public expenditures and cyclical components on the other.