Has the Euro increased risk sharing across Euro Area countries? Alessandro Ferrari∗ 1 Anna Rogantini Picco∗ Introduction The aim of this paper is to evaluate the level of risk sharing across Euro Area countries over the last 25 years and to address the following two questions: i) whether risk sharing has changed; ii) whether the introduction of the Euro has had any impact on it. We begin our analysis by measuring risk sharing by means of the BrandtCochrane-Santa-Clara (BCS) Index, which relies on how much countries’ Stochastic Discount Factors (hereafter SDF) are close to each other. This index helps us to get a rough idea on whether and how risk sharing has changed over our sample period. In order to decompose the channels through which risk is shared across countries we use the GDP decomposition introduced by Asdrubali, Sorensen, and Yosha (1996). What we find is that risk sharing has certainly not increased after the introduction of the Euro and, in fact, it has decreased. To tackle the second question we want to build counterfactual time series to reproduce what would have happened in terms of risk sharing had the Euro not been introduced. In order to do so we use the Synthetic Control Method (SCM). This method allows us to build synthetic time series with which to estimate a counterfactual level of risk sharing with the absence of the Euro. A comparison of risk sharing estimations between actual and synthetic data allows us to infer whether the Euro has had an impact on the level of risk sharing across Euro Area countries. We find evidence suggesting that the introduction of the common currency limited the ability to smooth income variations, in particular by reducing the scope for smoothing through international capital markets and fiscal transfers from governments. Related Literature. In order to get a counterfactual of some countries’ macro variables for the period after the introduction of the Euro we resort to SCM (for a more thorough discussion on SCM see Section 2.1). The method has been firstly introduced by Abadie and Gardeazabal (2003) to test for the impact of terrorism in the Basque Country in the late 60s. Building on that seminal paper, it has been further employed by Abadie, Diamond, and Hainmueller (2010) to estimate the effect of a large-scale tobacco control program that California implemented in 1988. ∗ European University Institute, Villa San Paolo, Via della Piazzuola, 43, I-50133 Florence, Italy; Email: [email protected] and [email protected]. 1 Billmeier and Nannicini (2012) use SCM to investigate the impact of economic liberalization on real GDP per capita in a worldwide sample of countries. Closer to our focus, Campos, Coricelli, and Moretti (2014) make use of SCM to evaluate the benefits from being part of the European Union. For our risk sharing assessment, we refer to two methodologies. A first measure is the the BCS index, proposed by Brandt, Cochrane, and Santa-Clara (2006). This is an indicator of bilateral risk sharing which relies on the similarity of pricing kernels. BCS index is computed by Rungcharoenkitkul (2011) to assess risk sharing among some Asian countries in the first decade of the 2000s. In order to better understand the channels through which risk sharing is accomplished, we adopt a second methodology, which has been introduced by Asdrubali et al. (1996) to identify risk sharing channels in the US over the period 1963-1990. Later on, their methodology has been used by Furceri and Zdzienicka (2015) to analyze and compare risk sharing among Euro Area countries with that across US states. However, even though the US are the closest currency union to the Euro Area, we do not believe that risk sharing estimations for the US are a good counterfactual for the Euro Area, as the Euro Area does not share fiscal policy as opposed to the US, where transfers can be used to smooth GDP shocks across states. The analysis of Furceri and Zdzienicka (2015) is updated with more recent data by van Beers, Bijlsma, and Zwart (2014), who try to assess the functioning of insurance mechanisms in the Euro Area, and by Kalemli-Ozcan, Luttini, and Sørensen (2014) who consider separately countries hit by the sovereign debt crisis in 2010. 2 Methodology 2.1 The Synthetic Control Method One purpose of this paper is to assess whether the introduction of the common currency had an effect on the level of risk sharing between member states. In order to meaningfully address this question one would need to have estimates of the economic performance of the member states under the alternative scenario in which the currency area had not been established. Since it is not possible to have a real counterfactual for the treatment “introduction of the Euro”, we resort to the SCM by Abadie and Gardeazabal (2003), which allows the evaluation of a synthetic counterpart. This method is a data driven procedure that has been used to estimate the effect of policy interventions in absence of a natural counterfactual.1 Using Abadie and Gardeazabal’s notation, let X1 be a set of determinants of variables of interest for the Eurozone member states before the introduction of the Euro and let X0 be the values of the same variables in all non Eurozone countries during such a period. In addition, let V be a diagonal matrix representing the relevance of 1 For reference, see Abadie et al. (2010), Billmeier and Nannicini (2012), Campos et al. (2014), and Saia (2016). 2 these predictors in determining the outcome variables, estimated via a factor model. Abadie and Gardeazabal (2003)’s algorithm looks for a vector of weights W ∗ that P minimizes (X1 −X0 W )0 V (X1 −X0 W ), subject to wi ≥ 0 and i wi = 1 for i = 1...N , N being the number of countries in the potential counterfactual pool. Finally, let Y1 and Y0 be the outcome variables for the treated (Eurozone) and untreated (Rest of the world) countries, then the method uses Y1∗ = Y0 W ∗ as counterfactual for the outcome of the treated country. This method relies on two identification assumptions: 1) the choice of the pre treatment covariates on which the matching is carried out should be such that the variables that are able to mimic the pre treatment path are included, but should not rely on observables that anticipate the effect of the treatment itself; 2) the units in the donor pool should not be affected by the treatment. For the latter reason the matching is carried out for one treated unit at the time, meaning that we iteratively drop all but one Eurozone member states, so that the procedure always involves one treated unit and N control units. The choice of the covariates in matrix X is such that it maximizes the ability of the synthetic control to reproduce the pre treatment behaviour of the treated unit. The baseline matching function will always take the past value of the variable we investigate, meaning that if we are evaluating what consumption would have been in Germany without the Euro we will always start by matching on the consumption of Germany in every pre treatment year. Further discussion on the advantages and drawbacks of this strategy will follow in Section 3. Our matching strategy reduces to matching on all the available lags of the dependent variable and the pre-treatment period average of the other covariates we analyse. These variables include, all in per capita terms, GDP, household final consumption, government expenditure, national income and disposable national income. To give a more precise example when we want to match the behaviour of per capita GDP ∗ we match on the following list of values: {GDPt }Tt=0 , C, G, DN I, N I where T ∗ is the treatment year and bar on variables represents the pre-treatment average. The intuition behind this method is that one can use the best linear combination of untreated units, in terms of matching pre treatment behaviour, as a counterfactual for the treated unit after the policy. It is worth mentioning that the evaluation of the robustness of these estimates has been discussed in the literature but no analytical result is available to compute the standard deviation of these estimates, in particular because the estimated component is the weighting vector; robustness checks are then carried out via bootstrap, randomly resampling the donor pool (see Saia (2016)), or via estimation of a difference in difference and testing for difference from zero of the outcome (see Campos et al. (2014)). A third way to check robustness is to run placebo studies on units in the donor pool to assess whether the method delivers spurious effect of the treatment. A relevant assumption for the correct use of the Synthetic Control Method is that the control group is unaffected by the treatment. This assumption can be trou- 3 blesome since, given the magnitude of the potential effect of the Euro, one might indeed think that the introduction of the common currency indirectly affected all the countries in the world, particularly the ones in our control group being OECD countries with strong trade and financial linkages with our treated sample. This concern is legitimate if we define as the treatment the introduction of the Euro itself. However one can think of the total effect of the Euro for member states as being the composition of two effects: i) the effect of the existence of the Euro; ii) the effect of being a member of the currency union. Under this decomposition all countries in the world are subject to the first effect, but only Euro Area member states are subject to the second one. Hence, one could interpret our treatment as being the membership of the Euro, conditional on the Euro existing. 2.2 Risk Sharing: the BCS Index and Consumption Correlation In this paper we adopt SCM in order to produce a synthetic dataset of macroeconomic indicators. The first analysis we carry out is the estimation of consumption behaviour under the two scenarios. Hence, we use SCM to predict what consumption would have been in absence of the common currency. With this preliminary analysis we want to check the direction of the change, if any, of risk sharing as measured by the degree of similarity of SDF, denoted by M . Being this a preliminary analysis carried out solely to inspect how risk sharing changes, we take a relatively agnostic stance and assume that all countries in the Eurozone have the same deep parameters. This allows us to evaluate the two following objects, where subscripts T and S denote treated, meaning with the Euro, and synthetic, meaning the counterfactual: !−σ cTi,t+1 T (1) Mi,t = β cTi,t !−σ S c i,t+1 S Mi,t =β (2) cSi,t Where we assume that preferences are CES with risk aversion σ, i denotes a country index, and β is the discount factor. Economic theory, under the assumption of no arbitrage and complete markets, predicts that, when two countries fully share risk, their stochastic discount factors should be the same. Following this intuition Brandt et al. (2006) propose an indicator of bilateral risk sharing that relies on similarity of pricing kernels. In particular the index takes the following form: BCSi,j = 1 − var(logMi,t+1 − logMj,t+1 ) var(logMi,t+1 ) + var(logMj,t+1 ) (3) This metric ranges between −1 and 1 with a higher number meaning a higher degree of risk sharing. 4 As a preliminary check we compute, under the same parametrization, the BCS indices for both treated and synthetic samples to assess whether consumption fluctuations got closer with the introduction of the common currency. As a robustness check we do the same exercise on the pre intervention period in order to check the soundness of our matching. A sound matching would result in relatively similar indices in both samples for the pre treatment period.2 In addition, we take another route to evaluate the SDF under the two regimes. In the previous discussion we matched over consumption and computed the pricing kernels with treated and synthetic data. The alternative strategy is to compute the SDF on actual data and only then generate a synthetic counterpart. This competing procedure is potentially more convoluted because normally the matching is carried out on levels, but the pricing kernel is a function of gross growth rates of consumption. To exemplify why this difference may be troublesome, consider you want to check the stochastic discount factor of Germany under the two policy regimes. Then, matching on consumption levels would optimally put weight on countries with similar levels of per capita consumption, and in particular it is likely that the counterfactual is mostly made as a linear combination of developed countries. If instead we match on the pricing kernel directly, which ultimately boils down to matching on consumption growth, we may have that the counterfactual is made by countries with completely different fundamentals which happened to display similar dynamic behaviour as pre treatment Germany. Ultimately both strategies are econometrically correct, though their outcomes may vary considerably and one may have different preferences on the two competing procedures. We do not take a stand on which of the two is more advisable, though it is worth mentioning that, precisely because of this reason, they may deliver very different results. In order to get an idea of how risk sharing might have changed due to the introduction of the common currency we also compute bilateral correlations of consumption across Euro Area members. Economic theory would suggest that a higher level of risk sharing should increase correlation of consumption between countries while their GDP correlation still being low. We are well aware that, in fact, economic theory is not supported by the data and that often GDP correlation is much higher than consumption correlation even across countries which are known to share risk. However, what we are interested in is not much the level of correlation in itself, but the difference in correlation obtained from the actual and the synthetic data. If the introduction of the common currency has had an impact on risk sharing, we should find that the difference in consumption correlation between actual and synthetic data should be significantly different from zero. 2 Note that one of the assumptions of the SCM is that there was no anticipation effect for the treatment. 5 2.3 Risk Sharing: GDP Decomposition Given the preliminary inspection on how risk sharing changed due to the introduction of the Euro, the naturally ensuing question is whether we can track back this change to different channels through which risk is shared. We carry out this analysis following a methodology proposed by Asdrubali et al. (1996). The idea of this analysis is to check which of the potential risk sharing channels absorb output shocks. In particular this is implemented by decomposing GDP into the following national account aggregates: Gross Domestic Product (GDP), Net National Income (NI), Disposable National Income (DNI), and Private and Government Consumption (C+G). According to this decomposition, GDP can be disaggreagated as this accounting identity: GDP = GDP NI DNI DNI+G (C+G) NI DNI DNI+G C+G (4) The ratios on the RHS can be given a very specific interpretation as channels through accounts for income insurance stemming from which risk is absorbed. Namely, GDP NI internationally diversified investment portfolios. NI measures the income (net of depreciation) earned by residents of a country, whether generated on the domestic territory or abroad, while GDP refers to the income generated by production activities on the economic territory of the country. Therefore, the first ratio captures the private insurance channel due to private cross-border investments or, as KalemliOzcan et al. (2014) refer to, holding of claims against the output of other regions. NI , instead, can be interpreted as the public insurance channel due to The ratio DNI government taxes and transfers. DNI is the income that households are left with DNI after subtracting taxes and adding transfers. The ratios DNI+G and DNI+G account C+G for smoothing through public and private saving channels respectively. In order to measure how variations in output is absorbed by each channel we proceeds as in Asdrubali et al. (1996). We first take logs of equation 4, we difference the series, and we multiply by the change of log GDP. Then, we regress the resulting equations on ∆ log GDP, which allows us to estimate how much of a country GDP shock is absorbed by each of these components. A zero coefficient in the regression of consumption on GDP (equation 9) means that a shock to GDP is fully absorbed through capital markets, fiscal transfers, public and private savings, thus leaving consumption unchanged. A high coefficient of consumption in the same regression, instead, means that only a minor part of the shock is absorbed through risk sharing, while a significant part stays unsmoothed. 6 2.3.1 Estimation methods We start our analysis by estimating the system of equations proposed by Asdrubali et al. (1996): ∆ log GDPi,t − ∆ log N Ii,t = β m ∆ log GDPi,t + m i,t g ∆ log N Ii,t − ∆ log DN Ii,t = β ∆ log GDPi,t + p (5) gi,t ∆ log DN Ii,t − ∆ log(DN Ii,t + Gi,t ) = β ∆ log GDPi,t + (6) pi,t ∆ log(DN Ii,t + Gi,t ) − ∆ log(Ci,t + Gi,t ) = β s ∆ log GDPi,t + si,t u ∆ log(Ci,t + Gi,t ) = β ∆ log GDPi,t + ui,t (7) (8) (9) where each β coefficient represents the share of the income shocks smoothed by a given channel. In particular, β m accounts for the share of GDP shocks smoothed by capital markets, β g by fiscal transfers, β p by public savings, β s by private savings. What is left, β u , is the unsmoothed part of the GDP shock. By construction β m + β g + β p + β s + β u = 1. The estimation of coefficients for the above system is carried out using the following methods: OLS with time fixed effects, OLS with panel correlated standard errors, generalized method of moments, and seemingly unrelated regressions. In the rest of the paper our baseline estimation for the analysis of risk sharing channels will be an OLS estimation with time fixed effects. We also perform OLS with panel correlated standard errors, seemingly unrelated regression and GMM. In particular in GMM, we separately estimate the above described relations using lags of GDP growth as an instrument. In the reported results we use up to 3 lags of the instrument. Also note that the estimation procedure, which follows Arellano and Bond – see Roodman (2009) – automatically includes past values of the dependent variable as instruments. We show the results of these estimation strategies as computed in a difference in difference model, which is equivalent to separate estimation. Namely we stack together our treated and synthetic samples and include the independent variable interacted with the 4 possible combinations of treated/synthetic and euro/no euro. In particular our results should then be interpreted as follows: the coefficient associated to the independent variable interacted with the treated dummy and the no euro dummy, both taking value 1, represents the share of GDP variation smoothed by a given channel for our actual data before the introduction of the euro; this coefficient should be compared with its synthetic counterpart, meaning the coefficient of the ∆ ln GDP when the data is synthetic and before the Euro. If our matching is successful we should not find a statistical difference between these two estimates. For our treatment period we should then compare the coefficient associated with ∆ ln GDP of actual data with the one of the synthetic data, which tell us the share of income variation smoothed by the given channel after the Euro. If the Euro had an effect on this channel the two estimates should be statistically different. All our specifications have time fixed effects, unless specified otherwise. Provided we have a good match for pre-treatment period, we can be relatively sure 7 that the common trend assumption is fulfilled. We provide an example of this in Figure 3 which shows the last dependent variable, ∆log(C + G) (the one that delivers us the coefficient of the unsmoothed component) for both the treated and the synthetic group over the whole sample period. 2.4 Data The data that we use for our analysis are taken from the OECD National Account Statistics. In particular, we use household final consumption expenditure for C, general government expenditure for G, gross domestic product computed following the so called output approach for GDP, net national income for NNI, and net disposable income for DNI. Our dataset covers 31 countries from 1960 to 2014. As the Synthetic Control Method requires the data not to display missing values, in order to keep for the matching all the countries in our sample, we limit our matching window to the period 1990-1998.3 This limitation leaves us with 21 countries for which we have a complete set of data for the variables we need. Out of these 21 countries 11 are Eurozone member states, hence treated, while 10 are OECD countries that are not in the currency area. The weighting matrices resulting from the synthetic control method are displayed in Tables 1 and 2. 3 Results 3.1 BCS Index and Consumption Correlations As discussed in the methodology section we perform two different approaches to first inspect the potential change in risk sharing generated by the introduction of the common currency: the first implements the matching on the consumption data and then computes the SDF out of the actual data and the synthetic series; the second, instead, relies on the computation of the SDF from the consumption data and then on the application the SCM to generate a synthetic version of the SDF directly. As mentioned above the difference between these two approaches boils down to using the SCM on levels of consumption or on growth rates. Both methodologies produce a very good matching on the pre-treatment window, even though, given that we aim at generating a series that closely resembles the level of consumption, the first approach performs poorly when we compute the ensuing SDF. On the other hand, we are able to match relatively well the dynamics of the SDF by matching on consumption growths and then computing our object of interest. Figure 1 show the actual and synthetic SDF for Greece over the period 1990-2011. 3 With the aim of increasing the number of countries in the donor pool for the synthetic control matching, we also tried to use World Bank data series. However, we encountered some problems in accounting divergence with the OECD data, hence for the moment we limited our analysis to the OECD data series. 8 The results of the latter procedure are displayed in Tables 3-6. The first three tables represent the bilateral differences between the BCS index computed from actual data and its synthetic counterpart for the full sample, the matching window and the treatment period respectively. In order to better summarize our results, in Table 3 we provide summary statistics for the three matrices in Tables 4-6. In particular, we observe that the full sample result points towards a reduction of the ability to share risk due to the introduction of the Euro, though the coefficient is not significant using standard inference. The same object for the matching period is extremely close to zero, suggesting that our matching procedure does reasonably well in generating the synthetic series. Finally, for the treatment period we find again a negative coefficient, suggesting that risk sharing decreases, though the size of the standard errors implies that this result is not significantly different from zero. A comparison between consumption correlations obtained from actual and synthetic data leads us towards a similar conclusion. Tables 7-10 show the difference between bilateral consumption correlation computed from the actual and the synthetic data. Table 8 displays the difference for the sample period 1990-1998, i.e. before the introduction of the Euro. The fact that all the differences are not significantly different from zero confirms the good quality of our match. Table 9 and Table 10 show the difference in consumption correlations for the sample periods 1999-2007 and 20082011, respectively. Since many are significantly different from zero and negative means that consumption correlation computed with the actual data is lower than that computed with the synthetic data. A lower consumption correlation means that consumption smoothing happens at a lower degree with the Euro than without. One might think that during the crisis the general confidence loss in the economy led to a lower cross-country risk sharing. However, 9 shows that even in the pre-crisis period after the introduction of the Euro consumption correlations computed with actual data are lower than consumption correlations computed with synthetic data (the difference in negative and significantly different from zero). Summary statistics of Tables 8-10 are reported in Table 7. Even though in the sample periods 1999-2007 and 2008-2011 the difference between correlations of the actual and synthetic data are not significant on average, there are still many bilateral correlation difference which are significantly different from zero. 3.2 Risk Sharing Channels Figure 2 shows the actual and synthetic series of household and government consumption expenditure for Finland. The two series are very close in the matching period spanning from 1990 to 1998, and then start to diverge over the treatment period. Even if only household and government consumption series for Finland are displayed here, actual and synthetic series for the national account aggregates and the countries that we considered look similar to those reported. Given that the aim of synthetic control method is to get a synthetic series which is as close as possible to the actual one for the matching window, our matching proves to be successful. 9 The obtained actual and synthetic series are used to estimate Equations 5-9. Table 11- 15 display the results of our estimations for the full sample period, i.e. 1990-2011. Four specifications of the model are shown, both for the actual (Treated) and the synthetic (Control) series. Each table shows both the estimations for the treated and the control group in the period before (Pre-tr) and after (Post-tr) the introduction of the Euro (Treatment). The third and the sixth row of each table exhibits the p-value of a Wald test on whether the treated and control group coefficients are significantly different from each other. Table 11 shows the OLS estimates. In the pre-treatment period coefficients of the treated and control group are never significantly different from each other, implying that the quality of our match is good. In the pre-euro subsample (1990-1998), most of the risk sharing happens through private savings, while a smaller part goes through public savings and fiscal transfers. The unsmoothed portion of risk is around 47%. In the post treatment period (1999-2011), the difference between the coefficients of the capital market channel and the unsmoothed portion becomes significant. In particular, while the coefficient of capital market channel is insignificant for the treated sample, it is significant and equal to 18.9% for the control group. This means that without the Euro, capital markets would have had a remarkable role is absorbing the risk. Even more significant is the difference between the unsmoothed coefficients. Namely, the the unsmoothed portion of risk in the treated group is 73.9% against only 36.8% for the control group. This implies that the Euro has significantly lowered the level of risk sharing across Euro Area countries, at least in comparison to our counterfactual experiment. The higher degree of absorption would have happened according to our estimates through capital markets. Estimates obtained with OLS and panel-correlated standard errors as well as with SUR confirm the results of the baseline model. GMM estimates, instead, are somewhat different. While the coefficients of public and private risk sharing channels are similar to the estimates obtained with the other methods both in terms of value and significance, what changes is the coefficient of capital market channel. While with OLS the control group coefficient after the treatment becomes significant, with the GMM it is smaller and insignificant. Apart from GMM the other method consistently deliver the result that our control group after the treatment would have guaranteed a statistically significantly higher level of risk sharing and this would have happened mainly through capital markets, public and private savings. A general and legitimate concern regarding our estimates is that they may be prone to measurement error driven bias. This may be particularly worrisome given that we are estimating our parameters on data we may have generated with error. As it is well known, random measurement error generates attenuation bias, which would bring our risk sharing channel for the counterfactual data closer to zero than the true parameter. Firstly, this cannot be the case for all the parameters given the identity nature of our problem. In particular, assuming that we generate our series with random error, we can only have that the first 4 parameters suffer from attenuation bias, while the last one in fact can be computed as a residual. If the first 4 10 parameters are closer to zero than their true counterpart, this implies that the the unsmoothed share must be higher than the true value. Since we consistently find that the unsmoothed parameter is lower in the counterfactual experiment than in the true data and we have no reason to believe that the true data is subject to the same measurement error, then our estimated difference in smoothed income variation can only be a lower bound to the actual value. By the same token, our estimated changes in the risk sharing channels can be viewed as lower bounds since we consistently find that the channels would be more effective in the counterfactual and given the potential attenuation bias we may find we may be underestimating this change. The only case in which this measurement error in our synthetic data can be alarming is for the pre treatment period of the international capital markets and fiscal transfer channels, in which we find a higher estimate for the synthetic data than for the actual one, implying that if we were to measure the coefficient without bias the two would be further apart. In particular, the difference between treated and control estimate for the international capital markets channel, which is already statistically different from zero, would be even larger. 4 Conclusion This paper assesses the effect of the introduction of the common currency on the ability to smooth consumption for Eurozone member states. We do so by building a dataset of counterfactual macroeconomic variables for the Eurozone countries without the Euro via Synthetic Control Method. We run a number of econometric procedures, including the evaluation of bilateral correlations of consumption, the Brandt-Cochrane-Santa Clara Index, and the decomposition introduced by Asdrubali et al. (1996) to evaluate the existence of this effect and the channels through which it may have occurred. Our results show evidence of a decrease in risk sharing across Euro area countries for the period after the introduction of the Euro. Bilateral consumption correlations calculated from synthetic data are higher than those computed from actual data, indicating that with the introduction of the Euro consumption smoothing has decreased. In particular, we find that international capital markets and fiscal capacity of governments would have had a larger ability to absorb income variations. In order to find additional evidence on what might have caused the reduction in risk sharing we would like to proceed as follows. First of all, we conducted our analysis using OECD data. In order to see how sensitive the matching to produce synthetic data is to the countries’ donor pool that we use, we plan to conduct the same analysis using data from the World Bank. Furthermore, we plan to improve our measures of risk sharing. In particular, we would like to get a better estimate of SDF. As we are well aware, there are several difficulties related to SDF estimation. In this paper we resorted to a consumption based estimation, but in further research we would like to use the term structure of 11 interest rates as in Rungcharoenkitkul (2011). Finally, given our findings of a decrease in risk sharing over the period following the introduction of the Euro, it would be of primary interest to inquire which could be the determinants of this phenomenon. Indeed, a higher degree of financial and goods market integration across Euro area countries would prompt us to think that risk sharing should increase. Taken our results into account, we would like to get a deeper understanding of why this is not the case. 12 A Synthetic Control Method Figure 1 – Actual and synthetic series of Greek SDF Note: The matching window is 1990-1998. Figure 2 – Actual and synthetic series (a) Household Consumption (b) Government Consumption Note: The matching window is 1990-1998. 13 B Weighting matrices Table 1 – Weighting matrix OECD Control Australia Canada Japan Korea Mexico NewZealand Sweden Switzerland UK US Austria 33 2 Belgium Finland France 25.70 33 39.40 Germany Greece Ireland 44.10 Italy 26.70 Netherlands Portugal Spain 2.500 50.40 2.400 32.90 24 2.900 1.800 14 9 11.30 12.20 40.20 33.20 28.90 5.700 8.300 6.700 35.10 44.50 7.100 0.500 12 40.30 27.70 10.40 46.40 17.40 16.20 47.70 38.80 48.50 5.600 12.90 36.70 45.10 56 27.30 33.40 Table 2 – Weighting matrix World Bank Control Brazil Cameroon Central African Republic Chile Comoros Costa Rica Denmark Japan Jordan Lebanon Madagascar Mexico Rwanda Senegal Sweden Switzerland Turkey C C.1 Austria Finland 14.30 France Germany Italy Netherlands Portugal Spain 1.600 13.30 1 65.50 5.900 42.60 31.30 19 3.300 0.100 3 3.100 1.200 4.900 59 14.50 50.70 2.700 55.50 26.40 3 1.600 3.800 18.30 2.200 70.70 15 6.800 3.700 32.40 12.80 1.700 2.200 2.400 85.50 13.20 18.80 17 6.100 4 9.900 48.20 0.500 1.700 BCS With total consumption Table 3 – BCS index differences (BCSt − BCSs ): summary statistics 1990-2011 Mean t-stat BCSt − BCSs -0.265 -.379 1990-1998 Mean t-stat -.105 14 -.522 1999-2011 Mean t-stat N -0.316 -.355 110 Table 4 – BCS index differences (BCSt − BCSs ): sample period 1990-2011 AT BE FI FR DE GR IE IT NL PT ES AT BE FI FR DE GR IE IT NL PT ES . . -.399 (-.939) -.398 (-.945) -.634 (-1.464) -.510 (-1.170) -.492 (-1.375) -.373 (-1.038) -.413 (-1.001) -.260 (-.661) -.436 (-1.102) -.327 (-.821) . . -.449 (-1.038) -.245 (-.569) -.517 (-1.204) -.272 (-.689) -.230 (-.580) -.162 (-.379) -.087 (-.207) -.172 (-.405) -.127 (-.302) . . -.576 (-1.332) -.484 (-1.137) -.285 (-.729) -.179 (-.458) -.439 (-1.009) -.156 (-.374) -.303 (-.721) -.290 (-.677) . . -.657 (-1.512) -.294 (-.790) -.113 (-.303) -.272 (-.639) .042 (.105) -.159 (-.391) -.101 (-.244) . . -.722 (-1.969) -.454 (-1.234) -.323 (-.773) -.370 (-.922) -.264 (-.655) -.443 (-1.093) . . -.085 (-.195) -.418 (-1.046) .036 (.086) -.480 (-1.133) -.252 (-.610) . . -.087 (-.220) .069 (.163) -.038 (-.091) .256 (.621) . . .165 (.395) -.217 (-.515) -.265 (-.611) . . .110 (.253) .308 (.730) . . -.165 (-.387) . . Table 5 – BCS index differences (BCSt − BCSs ): sample period 1990-1998 AT BE FI FR DE GR IE IT NL PT ES AT BE FI FR DE GR IE IT NL PT ES . . -.431 (-.655) -.683 (-1.189) -.088 (-.130) -.306 (-.463) .404 (.800) .131 (.266) -.226 (-.350) -.083 (-.152) .010 (.023) -.004 (-.007) . . -.432 (-.644) -.527 (-.836) .012 (.017) .396 (.719) .042 (.071) .107 (.152) -.188 (-.302) .060 (.109) .138 (.218) . . -.670 (-1.154) -.942 (-1.40) -.175 (-.287) .143 (.210) -.367 (-.540) .202 (.297) -.555 (-.872) -.214 (-.313) . . -.644 (-1.014) .115 (.218) .201 (.395) -.246 (-.395) -.003 (-.006) -.327 (-.729) -.144 (-.249) . . .034 (.062) -.142 (-.237) -.453 (-.642) -.340 (-.544) .175 (.322) -.345 ( -.545) . . .500 (.759) .169 (.303) .308 (.454) -.357 (-.536) .021 (.032) . . .298 (.484) .078 (.113) -.284 ( -.410) .335 (.486) . . -.111 (-.176) -.280 (-.499) .074 (.116) . . -.103 (-.154) .257 (.364) . . -.347 (-.527) . . 15 Table 6 – BCS index differences (BCSt − BCSs ): sample period 1999-2011 AT BE FI FR DE GR IE IT NL PT ES D AT BE FI FR DE GR IE IT NL PT ES . . -.381 (-.699) -.191 (-.350) -.702 (-1.284) -.557 (-1.007) -.699 (-1.537) -.507 (-1.092) -.418 (-.791) -.283 (-.555) -.629 (-1.220) -.408 (-.834) . . -.431 (-.799) -.184 ( -.335) -.712 (-1.322) -.430 (-.885) -.331 (-.662) -.167 (-.307) -.022 (-.041) -.290 (-.539) -.197 (-.387) . -.564 (-1.022) -.264 (-.490) -.320 (-.682) -.259 (-.558) -.311 (-.573) -.061 (-.120) -.260 (-.492) -.284 (-.547) . . -.676 (-1.252) -.406 (-.838) -.239 (-.492) -.249 (-.455) .082 (.156) -.202 (-.374) -.121 (-.232) . . -.898 (-2.036) -.543 (-1.202) -.192 (-.371) -.308 (-.618) -.465 (-.924) -.465 (-.979) . . -.403 (-.742) -.531 (-1.039) -.069 (-.132) -.516 (-.978) -.238 (-.450) . . -.228 (-.448) .056 (.104) -.139 (-.265) .105 (.206) . . .193 (.359) -.237 (-.430) -.342 (-.630) . . .180 (.331) .327 (.627) . . -.152 (-.279) . . Consumption correlations Table 7 – C correlation differences (ρt − ρs ): summary statistics 1990-1998 Mean t-stat ρt − ρs -.0183 -.061 1999-2007 Mean t-stat -.410 16 -1.03 2008-2011 Mean t-stat N -0.316 -1.56 110 Table 8 – Difference in correlation between actual and synthetic consumption: sample period 1990-1998 AT BE FI FR DE GR IE IT NL PT ES AT BE FI FR DE GR IE IT NL PT ES 0 . -.019 (-.06) .025 (.13) .033 (.12) -.014 (-.05) -.042 (-.14) .126 (.41) -.037 (-.05) -.054 (-.16) -.162 (-.43) -.017 (-.07) 0 . -.118 (-.44) .043 (.10) .041 (.09) -.186 (-.51) -.069 (-.24) .013 (.02) -.028 (-.20) -.421 (-.97) -.035 (-.08) 0 . .146 (.45) .121 (.32) -.009 (-.04) .020 (.09) .001 (.00) -.173 (-.54) -.112 (-.44) .090 (.33) 0 . -.036 (-.12) .042 (.15) .298 (.76) -.024 (-.04) .002 (.00) .015 (.03) -.040 (-.20) 0 . -.020 (-.04) .272 (.56) -.225 (-.35) .043 (.09) .002 (.00) -.075 (-.26) 0 . .043 (.14) .185 (.28) -.266 (-.61) -.073 (-.30) .022 (.09) 0 . .002 (.00) -.163 (-.47) -.104 (-.28) .234 (.64) 0 . .009 (.01) .114 (.16) .094 (.14) 0 . -.484 (-1.03) -.104 (-.21) 0 . -.040 (-.10) 0 . Note: t statistics are in parentheses. Table 9 – Difference in correlation between actual and synthetic consumption: sample period 1999-2007 AT BE FI FR DE GR IE IT NL PT ES AT BE FI FR DE GR IE IT NL PT ES 0 . -.797 ( -.49) -.116 (-2.11) -.130 (.01) .123 (-.37) -.304 (-.70) -1.271 (-.51) -.220 (-.35) -1.342 (-.32) -1.554 (-2.86) -1.380 (-.10) 0 . -1.326 (-3.11) -.082 (.24) -.541 (-.95) -1.779 (-.41) -.199 (-.14) -1.404 (.07) -.177 (-.33) -1.471 (-2.61) -1.731 (.27) 0 . -.672 (-2.19) -.229 (-.90) -.006 (-3.46) -1.654 (-3.27) -.004 (-2.31) -1.737 (-2.40) -1.275 (-2.42) -.991 (-2.39) 0 . -.234 (-1.15) -1.390 (-.14) -.635 (.13) -.801 (.05) -.485 (.05) -1.765 (-2.03) -1.887 (-.13) 0 . -.922 (-1.07) -1.263 (-1.00) -.243 (-1.23) -.822 (-.84) -1.991 (-2.24) -1.905 (-1.33) 0 . -1.911 (-.48) .163 (-.08) -1.888 (-.49) -.933 (-2.13) -.621 (-.12) 0 . -1.617 (.31) -.077 (-.25) -.816 (-2.13) -1.152 (.24) 0 . -1.724 (-.19) -.976 (-1.75) -.661 (.05) 0 . -1.049 (-2.68) -1.326 (.03) 0 . -.056 (-1.98) 0 . Note: t statistics are in parentheses. 17 Table 10 – Difference in correlation between actual and synthetic consumption: sample period 2008-2011 AT BE FI FR DE GR IE IT (-.31) NL PT ES AT BE FI FR DE GR IE 0 . -.0922 (-.77) -.9490 (-.25) .0049 (-.12) -.1401 (.15) -.2559 (-.28) -.1616 (-1.17) -.0994 (-1.42) -.0509 (-1.57) -1.262 (-1.76) -.0315 (-1.35) 0 . -1.2882 (-1.42) .0458 (-.13) -.4475 (-.67) -.1078 (-2.97) -.0256 (-.29) .0132 (-.01) -.0483 (-.28) -1.125 (-1.83) .0498 (-2.71) 0 . -1.1148 (-.58) -.4461 (-.23) -1.3714 (-.00) -1.285 (-2.10) -1.1618 (-.65) -1.0626 (-2.98) -.9908 (-1.20) -1.2058 (-.87) 0 . -.4113 (-.40) -.0453 (-1.88) .0372 (-.88) .0100 (-.21) .0118 (-.48) -1.0378 (-3.10) -.0127 (-4.36) 0 . -.5676 (-1.12) -.5386 (-1.69) -.5210 (.21) -.3614 (-.80) -1.1774 (-15.33) -.5115 (-5.57) 0 . -.0998 (-5.02) -.0265 (-2.28) -.1464 (-4.13) -.8560 (-1.08) -.0386 (-.76) 0 . .0707 . -.0605 (-.09) -.8784 (-1.07) .0620 (-1.49) IT NL PT ES 0 . -1.1782 (-.99) .0081 (-1.36) 0 . -1.0174 (-.22) 0 . 0 -.04256624 (-2.50) -.8967 (-.80) .0105238 (-.52) Note: t statistics are in parentheses. E Risk sharing channels Figure 3 – ∆ln(C + G) for treated and control group Note: The matching window is 1990-1998. 18 Table 11 – OLS estimated risk sharing channels - sample period 1990-2011 Capital markets Pre-tr Control Treated Post-tr Control Treated Fiscal transfers Public savings Private savings Unsmoothed -.022 (-0.32) -.054 (-0.67) .040 (1.72) -.005 (-0.20) .139∗∗∗ .345∗∗∗ ( 4.56) .001 ( 0.05) (4.42) .058 (0.63) .496∗∗∗ (6.59) -.000 ( -0.00) .124 (1.52) -.003 (-0.03) -.012 (-0.47) -.015 (-0.52) -.112∗∗ (-3.11) .009 ( 0.24) -.031 ( -0.34) -.173∗ ( -1.69) .032 (0.37) .183∗ (1.85) Note: t statistics are in parentheses. ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Table 12 – OLS with PCSE AR(1) estimated risk sharing channels - sample period 1990-2011 Pre-tr Control Treated Post-tr Control Treated Capital markets Fiscal transfers Public savings Private savings Unsmoothed -.021 (-0.27) -.054 (-0.58) .041∗ .154∗∗∗ (1.84) -.006 (-0.23) (5.36) -.005 (-0.14) .346 (4.21) .058 (0.56) .496∗∗∗ (5.63) -.002 (-0.02) .122 (1.28) -.002 (-0.02) -.014 (-0.56) -.013 (-0.45) -.127∗∗∗ (-3.53) .020 ( 0.45) -.031∗∗∗ (-0.30) -.174 (-1.44) .026 ( 0.24) .189 (1.47) Note: t statistics are in parentheses. ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Table 13 – OLS PCSE HET estimated risk sharing channels - sample period 1990-2011 Pre-tr Control Treated Post-tr Control Treated Capital markets Fiscal transfers Public savings Private savings Unsmoothed -.022 (-0.30) -.054 (-0.59) .040∗ (1.87) -.005 (-0.20) .139∗∗∗ (5.28) .001 (0.05) .345∗∗∗ (5.21) .058 (0.63) .496∗∗∗ (7.18) -.000 (-0.00) .124 (1.46) -.003 (-0.03) -.012 (-0.49) -.0159 (-0.53) -.112∗∗∗ (-3.49) .009 (0.24) -.031 (-0.39) -.173∗ (-1.70) .032 (0.39) .183∗ (1.73) Note: t statistics are in parentheses. ∗ p < 0.1, ∗∗ p < 0.05, 19 ∗∗∗ p < 0.01. Table 14 – GMM estimated risk sharing channels - sample period 1990-2011 Capital markets Pre-tr Control Treated Post-tr Control Treated Fiscal transfers Public savings Private savings Unsmoothed -.040 (-0.35) -.045 ( -0.31) .060 (1.63) -.085 (-1.86) .212∗∗∗ (4.94) -.062 (-1.16) .298 (2.42) .194 (1.26) .469∗∗∗ (4.07) -.000 (-0.01) .045 (0.36) .046 (0.30) -.044 (-1.09 ) .069 (1.42) -.157∗∗ ( -3.31) .060 (1.05) .108∗∗ (0.80) -.382∗∗ (-2.32) .047 (0.37) .205 (1.33) Note: t statistics are in parentheses. ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Table 15 – SUR estimated risk sharing channels - sample period 1990-2011 Pre-tr Control Treated Post-tr Control Treated Capital markets Fiscal transfers Public savings Private savings Unsmoothed -.022 ( -0.33) -.054 (-0.69) .040∗ .139∗∗∗ .345∗∗∗ ( 1.77) -.005 ( -0.20) (4.70) .001 (0.06) (4.56) .058 (0.65) .496 -.000 - .124 ( 1.56) -.003 (-0.03) -.012 (-0.48) .009 (-0.54) -.015 ( -3.21) -.112∗∗ (0.25) -.031 (-0.35) -.173∗ ( -1.75) .032 .183∗ (3.64) Note: t statistics are in parentheses. ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01. The bottom right number in brackets (3.64) is the test statistic of a Wald test performed on the coefficient. 20 References Abadie, A., Diamond, A., & Hainmueller, J. (2010, June). Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California’s Tobacco Control Program. Journal of the American Statistical Association, 105 (490), 493–505. doi: 10.1198/jasa.2009.ap08746 Abadie, A., & Gardeazabal, J. (2003). The Economic Costs of Conflict: A Case Study of the Basque Country. The American Economic Review , 93 (1), 113– 132. Asdrubali, P., Sorensen, B., & Yosha, O. (1996, November). Channels of Interstate Risk Sharing: United States 1963-1990. The Quarterly Journal of Economics, 111 (4), 1081–1110. Billmeier, A., & Nannicini, T. (2012, October). Assessing Economic Liberalization Episodes: A Synthetic Control Approach. Review of Economics and Statistics, 95 (3), 983–1001. doi: 10.1162/REST a 00324 Brandt, M. W., Cochrane, J. H., & Santa-Clara, P. (2006, May). International risk sharing is better than you think, or exchange rates are too smooth. Journal of Monetary Economics, 53 (4), 671–698. doi: 10.1016/j.jmoneco.2005.02.004 Campos, N. F., Coricelli, F., & Moretti, L. (2014, May). Economic Growth and Political Integration: Estimating the Benefits from Membership in the European Union Using the Synthetic Counterfactuals Method (SSRN Scholarly Paper No. ID 2432446). Rochester, NY: Social Science Research Network. Furceri, D., & Zdzienicka, A. (2015, May). The Euro Area Crisis: Need for a Supranational Fiscal Risk Sharing Mechanism? Open Economies Review , 26 (4), 683–710. doi: 10.1007/s11079-015-9347-y Kalemli-Ozcan, S., Luttini, E., & Sørensen, B. (2014, January). Debt Crises and Risk-Sharing: The Role of Markets versus Sovereigns. The Scandinavian Journal of Economics, 116 (1), 253–276. doi: 10.1111/sjoe.12043 Roodman, D. (2009). How to do xtabond2: An introduction to difference and system GMM in Stata. Stata Journal , 9 (1), 86–136(51). Rungcharoenkitkul, P. (2011, October). Risk Sharing and Financial Contagion in Asia: An Asset Price Perspective (SSRN Scholarly Paper No. ID 1956389). Rochester, NY: Social Science Research Network. Saia, A. (2016). Choosing the open sea:the cost to the uk of staying out of the euro. van Beers, N., Bijlsma, M., & Zwart, G. (2014). Cross-country Insurance Mechanisms in Currency Unions: an Empirical Assessment. Bruegel Working Paper (04). 21
© Copyright 2026 Paperzz