We estimate the distributional effects of fiscal consolidation measures over the short- and medium-run. In doing so, we follow the methodology proposed by Jorda (2005), who estimates impulse response functions (IRFs) from local projections. Jorda (2005) shows that the standard linear projection is a direct estimate of the typical impulse response, as derived from a traditional vector autoregression (VAR). In principle, there are other possibilities to measure dynamic effects; in particular, one could estimate a Panel Vector Autoregression (PVAR) or an Autoregressive-Distributed-Lag Model (ARDL). However, in our case both options are inferior to the local projections method. The VAR approach suffers from identification and size-limitation problems, which is not the case for the more flexible local projections method (Gupta et al. 2017, p. 18–19). And the stability of IRFs obtained from an ARDL is undermined by their lag-sensitivity (e.g. Ball et al. 2013). Moreover, Cai and DenHaan (2009) point out that when the dependent variable is very persistent (which is the case for Gini data), statistically significant long-run effects may result from “one-type-of-shock models”Footnote 2—a problem that does not haunt the local projections method since lagged dependent variables are not used to derive the IRFs, but only enter as controls. Another advantage of the Jorda (2005) method is that the uncertainty around the IRFs can be estimated directly from the standard errors of the estimated coefficients without any need for Monte Carlo simulations.
In the context of estimating the effects of fiscal adjustments on income inequality, we employ the local projections method introduced by Jorda (2005). The empirical investigation in this paper goes beyond previous attempts to study the dynamic effects of fiscal austerity on income inequality. First, we cover a longer time period, as we are able to include data on the crisis years 2010–2013, thereby covering the more extensive time period 1978–2013. Second, we account for a richer set of control variables. Third, we extend our analysis by various relevant aspects, thereby providing new insights into how the distributional effects depend on the size and duration of fiscal adjustments and the timing of the business cycle. Fourth, we employ a comprehensive set of robustness checks.
Econometric strategy
Our regressions are based on the following equation, which is estimated for each future time period k (with \(k = 1,\ldots ,8\)),Footnote 3 allowing us to obtain local projections on how income inequality changes following the start of a fiscal consolidation episode:
$$\begin{aligned} G_{i, t+k} - G_{i, t}=\beta _{k} D_{i, t}+\gamma _{k} Z_{i,t}+\sum _{j=1}^{l}\delta _{j}^{k}\Delta G_{i,t-j}+\zeta _{i}^{k}+\eta _{t}^{k}+\epsilon _{i, t}^{k} \end{aligned}$$
(1)
In Eq. 1, G represents our measure of income inequality, i.e. the Gini coefficient of (in most cases: disposable) income, where the data sources used throughout the analysis will be explained below (see Sect. 3.2); \(D_{i, t}\) is a dummy variable that takes the value of 1 for the starting year of each fiscal consolidation episode and 0 otherwise. \(Z_{i,t}\) is a vector of additional control variables that should be understood as pre-treatment variables (i.e. determined before the treatment of fiscal consolidation starts; see Nakamura and Steinsson (2018)). These pre-treatment controls will be introduced in Sect. 3.2. \(\Delta G_{i,t-j}\) are the lags in the change of the measure of income inequality, where we set the number of lags l to two,Footnote 4 although we show later on that the estimation results are robust to different numbers of lags. \(\zeta _{i}^{k}\) are country fixed-effects. \(\eta _{t}^{k}\) are period fixed-effects. And \(\epsilon _{i, t}^{k}\) represents the stochastic residual. Equation 1 is estimated by using the panel-corrected standard error estimator (PCSE). As shown by Beck and Katz (1995), the OLS-PCSE procedure is well-suited for time-series cross-section data, when the number of years covered is not much larger than the number of countries in the cross-sectional dimension of the data. The main reason for the superior performance of the OLS-PCSE estimation strategy—compared to the Parks estimator and other Feasible Generalized Least Squares estimators—is that the method proposed by Beck and Katz (1995) is well-suited to addressing cross-section heteroskedasticity and autocorrelation in the residuals, allowing us to avoid biased standard errors.
Data
As consolidation data from the ’conventional approach’ imply the risk of obtaining biased estimates on the macroeconomic effects of fiscal austerity (e.g. Perotti 2013; Guajardo et al. 2014), the empirical analysis in this paper builds on fiscal consolidation measures that were identified according to the ’narrative approach’. Cyclically-adjusted data based on the ’conventional approach’ used by Afonso (2010) will only be used for the purpose of checking the robustness of our results. The ’narrative’ fiscal consolidation data includes 17 OECD countries (Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, Netherlands, Portugal, Spain, Sweden, the United Kingdom, and the United States of America). The size of the country group and the years that we cover in our data are dictated by data availability on this ’narrative’ fiscal consolidation variable. To identify episodes of cuts in government spending and/or increases in taxes which aim at reducing the budget deficit, we obtained annual data from DeVries et al. (2011) for the time period 1978–2009. By using the same ’narrative methodology’ as DeVries et al. (2011), Alesina et al. (2015a) have extended this dataset for the years 2010–2013. There are 60 consolidation episodes in total, covering 214 years with fiscal consolidations over the time period 1978–2013. The average size of the 60 fiscal consolidation programs amounts to 4.2% of GDP.
Table 2 summarizes the occurrence of fiscal consolidation episodes in our dataset. It can be seen that many fiscal consolidations actually come in the form of packages that span two or more years. The longest adjustment period was started by Canada in 1984 and lasted until 1997. For the period after the financial crisis, it can also be seen that fiscal adjustments were in general bundled into multi-year packages. For example, Ireland’s consolidation lasted from 2009 to 2013, and countries such as Austria, Denmark and Germany consolidated from 2011 to 2013. Table 2 presents the 60 fiscal consolidation episodes (that were bundled into packages of one or several years). The average duration of the fiscal consolidation programmes is 3.5 years.
Table 2 Fiscal consolidation episodes, based on DeVries et al. (2011) and Alesina et al. (2015a) For the measure of income inequality (G), we obtained data on Gini coefficients for market income and disposable income from Version 5.1 of the Standardized World Income Inequality Database SWIID (Solt 2016). Gini coefficients are bounded between 0—each reference unit receives exactly the same share of income—and 100, which would imply that a single reference unit gets all the income. The average Gini of disposable market income in our data set is 28.4, with a minimum of 16.5 and a maximum of 37.6. It has to be noted that the panel data are unbalanced, since Solt (2016) does only provide Gini data for 550 out of 612 possible observations (T = 36; N = 17). In the robustness check section, we will show that results do not change markedly when we interpolate the data in order to balance the panel. In the baseline regressions, however, we take the data from Solt (2016) as they are—without interpolating the missing values.Footnote 5 As we are mainly interested in how income inequality changes after taxes and transfers, the baseline results will be based on net (disposable) Gini data. Figure 2 shows the evolution in the disposable Gini coefficient for the 17 OECD countries in our country sample.
There are three main advantages of using the SWIID dataset. First, the data ensure that income inequality across countries is measured in a harmonized way. Second, the data include a large group of countries and allow us to obtain long time-series on Gini coefficients of market and disposable income for all the 17 OECD countries in our sample. Third, comparability across countries is enhanced by a transparent procedure of how the data were collected. As a robustness check, we will later on also use the database provided by Milanovic (2016a), who offers an ’All the Ginis’ index for disposable income by merging several data sources.
We control for five additional variables that function as pre-treatment controls (see Vector \(Z_{i,t}\) in Eq. 1): First, to consider possible effects of international trade on (future) income inequality, we include the change in trade openness (measured as the sum of imports and exports in relation to GDP), where data were obtained from the European Commission’s AMECO database. Consistent with Woo et al. (2017) and other studies, we use trade-to-GDP as a proxy to control for trade globalization, which might explain parts of the increase in income inequality in developed countries by affecting wages for skilled and low-skilled labor through various channels (e.g. Bensidoun et al. 2011). Second, we control for the change in the average years of schooling (Barro and Lee 2013)—here data come from the International Human Development Indicator—in order to capture possible effects of education on future income inequality. Third, we include GDP growth (AMECO data), as a decrease in economic activity may lead to an increase in the debt-to-GDP ratio via the workings of automatic stabilizers, so that the probability of fiscal consolidation measures might increase. Fourth, we account for the change in the unemployment rate (OECD data), where the rationale is the same as for considering GDP growth. Fifth, we control for the growth rate in Total Factor Productivity as a proxy for capturing the effects of technological change on income inequality (e.g. Acemoglu 2003; Roser and Crespo-Cuaresma 2016), where data were obtained from the AMECO database. These additional regressors are typically considered in the empirical literature on the determinants of income inequality (e.g. Acemoglu 2003; Ball et al. 2013; Roser and Crespo-Cuaresma 2016; Woo et al. 2017). Hence, the model presented in Eq. 1 covers the most relevant control variables from the empirical literature. With another robustness check, we will introduce additional lags for GDP growth and the change in the unemployment rate, since it might be argued that these variables have lagged impacts on the relation between fiscal consolidation and future income inequality.