Fiscal Stimulus in a Monetary Union: Evidence from Eurozone Regions

Abstract

This paper contributes to the open economy local fiscal multiplier literature by estimating regional output and employment responses to federal expenditure shocks in the European Union. In particular, similarly to the literature on foreign aid and growth, I use shocks to the supply of federal transfers (European Commission commitments) of structural fund spending by subnational region as instruments for annual realized expenditure in a panel from 2000–2013. I find a large, contemporaneous multiplier of 1.8 which translates into a medium-run multiplier of 4.1 three years after the shock. Furthermore, using a novel dataset on bilateral trade between EU regions, I find evidence of demand-driven spillovers up to 2 years after a shock.

Appendices

Appendix 1: Tables

Table 1 Summary statistics all regions (NUTS 2 level), 2000–2011

Table 2 Correlation statistics Obj. 1 regions (NUTS 2 level), 2000–2011

Table 3 Panel IV—Objective 1 only (NUTS 2 level), 2000–2011

Table 4 Panel IV—Objective 1 versus 2 (NUTS 2 level), 2000–2011

Table 5 Panel IV—Objective 1 regions: all versus EU-12 (NUTS 2 level), 2000–2011

Table 6 Panel IV—Objective 1 only (NUTS 2 level), 2000–2011 Unemployment

Table 7 Panel IV—Objective 1 versus 2 (NUTS 2 level), 2000–2011 Unemployment

Table 8 Panel IV—Objective 1 regions: all versus EU-12 (NUTS 2 level), 2000–2011 unemployment

Table 9 Panel IV—Objective 1 regions (NUTS 2 level), 2000–2011 GDP Components

Table 10 Panel IV—All regions (NUTS 2 level, all objectives), 2000–2011 Import-weighted fiscal spillovers

1.1 Robustness checks

Table 11 Panel IV—Objective 1 only (NUTS 2 level), 2000–2011 RHS: Cumulative Expenditure Shock

Table 12 Panel IV—Objective 1 only (NUTS 2 level), 2000–2011 RHS: Spending levels

Table 13 Panel IV—Objective 1 only (NUTS 2 level), 2000–2011 LHS: Percentage output gap

Table 14 Panel IV—Objective 1 only (NUTS 2 level), 2000–2011 LHS: Percentage output gap and RHS: Spending levels

Table 15 Panel IV—Objective 1 only (NUTS 2 level), 2000–2011 Driscoll–Kraay robust standard errors

Appendix 2: Figures

Fig. 1

Sources (throughout, except as noted): European Commission—DG Regio, Eurostat, author calculations

Annual commitments as a share of GDP by NUTS 2 (representative years, Objectives 1+2).

Fig. 2

Key explanatory variable and instrumental variable for a representative year

Fig. 3

Correlation between key explanatory variable and average output growth

Fig. 4

Correlation between key explanatory variable and variance of output growth

Fig. 5

Correlation between key explanatory variable and magnitude of output contractions

Fig. 6

Correlation between key explanatory variable and average unemployment rate

Fig. 7

Sources: European Commission—DG Regio, Eurostat, Thissen et al. (2013), author calculations

GDP growth versus trade-weighted expenditure spillover shock: 2001–2011.

Appendix 3: Fuzzy Regression Discontinuity

Several papers have exploited the discrete jump in probability of receiving Objective 1 transfers in a regression discontinuity design (Kallstrom 2014; Pellegrini et al. 2013), though the most cited reference in this line of literature (and also the first one to really explore this design) remains Becker et al. (2010). As summarized earlier, the authors use a binary treatment variable (whether or not a NUTS 2 region was eligible for Objective 1 transfers) to estimate the local average treatment effect of structural funds on average GDP growth and employment within a programming period. The analysis below aims at mirroring the strategy used by Becker et al. (2010), but addressing a couple of its limitations which the data used in the panel instrumental variables exercise above enables me to do. In particular, instead of binary treatment, I use a continuous treatment variable (actual expenditures received), that is furthermore instrumented as in the panel IV above by corresponding commitments; the same rationale for instrumenting applies as before (endogeneity in expenditures with respect to aggregate demand conditions and measurement error). In addition, although the nature of the quasi-experimental design does arise from a discontinuity in eligibility for Objective 1 funds, I use the total amount of funds transferred to regions (including Objective 2 funds) as my treatment variable; the reason behind this choice is that using merely Objective 1 transfers implicitly attributes 0 valued transfers to most regions on the right-side of the discontinuity threshold, thus exacerbating the true treatment discontinuity¹⁷ and underestimating any existing treatment effect. As in robustness Table 12, I find some evidence of weaker treatment effects in the 2007–2013 period, due likely to the particular choice of fiscal shock identification in levels—a necessary evil of the regression discontinuity criterion.

Although I have data on annual commitments and expenditures by NUTS 2 region,¹⁸ the treatment assignment variable (running variable in the regression discontinuity created by the threshold of eligibility for Obj.1 funds, i.e., reference-period GDP per capita) is unique per programming period. Thus, in what follows I estimate the local average treatment effect of Obj.1 fund eligibility separately for each programming period (the samples can of course also be pooled, treating each region as a separate region-programming period unit).

From the graphs below, it is apparent that we are dealing with a fuzzy discontinuity, since there are a few Obj.1 regions above the eligibility threshold. However, most regions receiving positive funds above that threshold are Obj.2 regions. Thus, a simple binary treatment regression discontinuity local average treatment effect estimation would vastly underestimate any actual effects, by imputing zero treatment on Obj.2 regions which actually get some federal funding (albeit significantly smaller than the Obj.1 sample on average). In addition, it is worthwhile noting that several Obj.1 regions are allocated funds at per capita levels comparable to Obj.2 regions—in other words, Obj.1 treatment is highly heterogeneous, and comparisons should be made between those with high levels of transfers and regular Obj.2 regions, conditional on similar reference-period GDP per capita, rather than all regions within a given band of the threshold (Figs. 8, 9).

Fig. 8

Intended treatment given reference-period GDP per capita (forcing variable)

Fig. 9

Realized treatment given reference-period GDP per capita (forcing variable)

For the “normal” 2000–2006 programming period, we can see the graphical evidence in Fig. 12 in support of a regression discontinuity design below: commitments (per capita and as a share of initial GDP) show a discontinuity and change in slope (interaction effect) at the eligibility threshold with respect to the assignment variable (reference-period GDP per capita). However, when looking at the outcome variable (average real GDP growth), it is difficult to see such a discontinuity—if anything, my evidence goes in the opposite direction from that found by Becker et al. (2010), in that growth outcomes are slightly lower immediately before the threshold relative to those immediately after. This is apparent using linear fits or 5th order asymmetric polynomials on both sides of the threshold. Note in this graphical evidence we are using commitments, rather than expenditures as treatment, since the latter can be driven by local demand conditions (this was in fact the core motivation underlying my earlier broad sample IV estimation). Graphically, however, expenditures and commitments behave alike with respect to the assignment variable (Figs. 10, 11, 12, 13).

Fig. 10

Funds committed (Euros per capita) by treatment eligibility: 2000–2006 programming period

Fig. 11

Funds committed (for entire programming period) as a share of initial annual GDP: 2000–2006 programming period

Fig. 12

Average annual GDP per capita growth by treatment eligibility: 2000–2006 programming period, EU-12 only

Fig. 13

High-order polynomial approximation to the conditional expectation of average GDP per capita growth given the forcing variable

Since I have a continuous treatment variable with non-compliers to the treatment assignment rule (fuzzy design), the RD naturally has to be estimated in two stages. In addition to using a binary indicator of whether a region is above or below the threshold (the typical exclusive instrument for treatment in this setup), I add the respective commitments series as an additional instrument. In particular, the second stage specification is

$$\begin{aligned} \frac{Y_{i,t}-Y_{i,t-1}}{Y_{i,t-1}}=\gamma _{t}+\alpha _{j}+\phi G_{it}+g\left( F_{it}\right) +\xi _{it} \end{aligned}$$

where in the first stage I estimate

$$\begin{aligned} G_{it}=\theta _{t}+\nu _{j}+\delta \mathbb {1}\left[ F_{it}>0\right] +\beta C_{it}+f\left( F_{it}\right) +\epsilon _{it} \end{aligned}$$

$g\left( F_{it}\right)$ and $f\left( F_{it}\right)$, functions of the forcing variable $F_{it}$, are approximated by asymmetric polynomials on each side of the threshold, including interaction terms. In what follows, I treat the data as a pooled cross section for each programming period, although year and country fixed effects are used, just as for the panel estimates above.¹⁹ I estimate the treatment assignment function in the first stage via 3rd/5th degree asymmetric polynomial approximation of the total funds spent in a given programming period around the threshold.²⁰ The outcome variable then on the second stage is the annual GDP per capita growth rate over the respective years.

While I find a large LATE consistent with the IV results above for the 2000–2006 period [consistent with the findings of Becker et al. (2010)], the same is not true of the 2007–2013 period. Looking at the yearly cross section of GDP versus forcing (or assignment) variable, we see that the 2000–2006 negative correlation pattern persists through 2007, starts shifting toward no correlation during 2008–2009 as most regions in the Eurozone head toward negative growth territory, and is inverted in 2010 and 2011, as regions in Greece, Portugal and Spain remain at negative growth rates (trapped at the epicenter of the then emerging sovereign debt crisis), while other countries start to recover from the 2008 recession. In other words, there were asymmetric macroeconomic shocks at the country level affecting these regions over time which blur the normally observed positive correlation between the forcing variable and GDP growth.²¹ Furthermore, some of the reasons related to the particular interference of the recent financial and sovereign debt crisis with the allocation mechanisms of these transfers may also contribute to explain the unusual results. This contrasts with the stronger-in-recession measured multipliers in the panel IV estimation core section of this paper using year-on-year changes in expenditures as the fiscal shock of interest, in line with the cross-country findings of Auerbach and Gorodnichenko (2012).

It is worthwhile noting that, in contrast to the earlier panel IV, the coefficient estimates on Table 16 cannot be directly interpreted as fiscal multipliers. Rather, they reflect an average response over the medium-run (the length of a programming period, 6–7 years) to ongoing treatment levels. Recall the independent variable of interest in the RD is the level of transfers (as a ratio to population or GDP), rather than a change in levels, or a shock per se whose propagation can be estimated in subsequent years.

Table 16 Regression Discontinuity IV (NUTS 2 level), 2000–2011