WPS6891 Policy Research Working Paper 6891 Episodes of Unemployment Reduction in Rich, Middle-Income, and Transition Economies Caroline Freund Bob Rijkers The World Bank Development Research Group Trade and International Integration Team May 2014 Policy Research Working Paper 6891 Abstract This paper studies the incidence and determinants of unemployment and, given unemployment, are more of episodes of drastic unemployment reduction, likely in countries with better regulation. An efficient defined as swift, substantial, and sustained declines in legal system that enforces contracts expeditiously is unemployment. Forty-three episodes are identified over a particularly important for reducing unemployment. The period of nearly three decades in 94 rich, middle-income, results imply that while employment is largely related and transition countries. Unemployment reductions to the business cycle, better regulation reduces the often coincide with an acceleration of growth and an likelihood of high unemployment and facilitates a more improvement in macroeconomic conditions. Episodes rapid recovery in the event unemployment builds up. are much more prevalent in countries with higher levels This paper is a product of the Trade and International Integration Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at brijkers@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Episodes of Unemployment Reduction in Rich, Middle-Income, and Transition Economies Caroline Freund Bob Rijkers Peterson Institute for International Economics World Bank Key words: unemployment, labor markets, regulation JEL codes: J23, J64, J65, K20, L20 * We would like to thank Davide Furceri, John Giles, Indermit Gill, Aaditya Mattoo, David Newhouse, Luis Serven, Beth Anne Wilson, seminar participants at the World Bank, the IMF, the University of Oxford, the Celebrating 10 years of Doing Business Conference, and especially Aart Kraay for constructive comments. Earlier versions of this paper appeared under the title “Employment Miracles”. Corresponding author: cfreund@piie.com, 202-454-1304, 1750 Massachusetts Avenue, NW, NW Washington, DC 20036. 1 Introduction Job creation is one of the most pressing priorities around the world. Despite a rich literature on the determinants of unemployment, 1 defining a set of policies to achieve enduring unemployment reductions remains elusive. This paper adopts an empirical approach to address this question by examining the incidence and the determinants of episodes of drastic unemployment reductions, defined as swift, substantial and sustained drops in unemployment. More specifically, an episode is defined as a reduction in the unemployment rate over a four-year period of at least three percentage points and at least a quarter of its initial level that persists for a minimum of three years. In addition, we insist that episodes start with a strict decline in unemployment and rule out overlap between declines to avoid counting the same unemployment reduction spell as two separate episodes. Episodes of unemployment reduction are quite common, and significantly more likely to occur in countries with better regulation. This result obtains using various alternative indicators of regulatory quality, and is robust to using alternative definitions of the filter used to identify reductions, alternative parameterizations of the filter, alternative estimation methods, and excluding episodes that are potentially driven by recovery from war or crisis. The results are also confirmed by Bayesian Model Averaging methods, which furthermore suggest that the efficacy of the legal system, enforcement of property rights, the control of corruption and business regulations are particularly important for reducing unemployment. Unemployment reductions tend to coincide with an acceleration of GDP growth, as well as a surge in trade, higher investment, and lower government spending as a share of GDP. The regulatory environment also tends to improve as unemployment declines. Although causation is difficult to establish, the coincidence of large reductions in unemployment with improving macroeconomic conditions is consistent with prudent macroeconomic management being conducive to employment creation. Sound regulation is associated with a double-dividend; countries with better regulation not only tend to have lower unemployment on average, but are also more likely to experience an unemployment drop when confronted with stints of high unemployment. The importance of an 1 See, for example, Blanchard and Wolfers, 2000, Nickell et al. 2005, and Djankov and Ramalho 2009. 2 impartial legal system, secure property rights, and efficient contract enforcement alongside lean regulation underscores the importance of implementation. The importance of economic reform is perhaps most evident in the transition countries, which experienced two waves of unemployment reductions. The first wave, in the mid-1990s, involved Hungary and Latvia, two of the early reformers. The second wave, around the millennium, involved Russia, Ukraine, Bulgaria, Croatia, Estonia and Lithuania. Focusing on the Baltics, Estonia, the lead reformer, maintained relatively low unemployment in the 1990s, averaging 7 percent, with little limited scope for a sharp reduction at that time. In contrast, Latvia and Lithuania built up high unemployment early on, topping 15 percent in the mid-1990s. In Latvia, unemployment began declining in the mid-1990s after business reforms were made. In Lithuania, despite similar external economic conditions and trade patterns, slower reform appears to have contributed to postponing a more sustained adjustment until 2001. 2 This paper contributes to several strands of literature. A large body of research examines the (cross-country) determinants of unemployment, focusing on the role of institutions, shocks, and crucially, the interaction between them as explanations for differences in the evolution of unemployment across countries (see, for example, Blanchard and Gali, 2010, Blanchard and Wolfers, 2000, Feyrer and Sacerdote, 2013, Nickell et al 2005, Helpman and Itshkoki, 2010). This paper aims to contribute to that literature by analyzing large and sustained declines in unemployment, focusing on their potential determinants and how they start. In doing so, we help differentiate between the drivers of secular changes in employment and those of cyclical labor adjustment (see also Davis et al, 2012, Elsby et al. 2009). In the process, we also provide evidence on the relationship between growth and employment creation. Moreover, by highlighting the association between business regulation and the reduction in unemployment, our results also contribute to the growing literature on their importance for development outcomes. 3 2 Privatization was relatively rapid and successful in Estonia and Latvia as compared with Lithuania (Havrylyshyn and McGettigan 1999), Latvia instituted the trade reforms required to accede to the WTO in 1998, as compared with 2000 for Lithuania. In 2004 (the earliest year for which data are available), the Doing Business indicators record Latvia ranked 14 as compared with 39 for Lithuania in days to start a business. 3 More cumbersome regulation can be damaging either because it is captured by incumbents or because it creates rent- seeking opportunities for government officials that are tasked with enforcing it. In either case, it hampers competition and impedes development. For example, De Soto (1990) shows that the economy of Peru, in the absence of property rights and well functioning legal systems, veered toward informality, creating many small producers that were not able to expand because they did not have legal rights to their property. As another example, Djankov et al (2002) show that more arduous entry regulation is associated with more corruption and informality across countries, as opposed to better 3 The remainder of this paper is organized as follows. The next section explains how drastic unemployment reduction episodes are identified. Section 3 examines the incidence, and correlates of, episodes of unemployment reductions. Section 4 assesses to what extent episodes can be predicted, and provides evidence of a robust association between the quality of the overall regulatory framework and the incidence of reductions. Section 5 examines which aspects of the regulatory framework matter most. Section 6 concludes. 2 Identification and Incidence of Episodes of Unemployment Reduction 2.1 Identification Following Hausmann et al. (2005) on growth accelerations and Freund (2005) on current account reversals, we use an event study approach. We define an episode of unemployment reduction to be a substantial reduction of unemployment that is sustained for a protracted period of time and rule out overlap between episodes. Specifically, a decline in the unemployment rate starting at period t qualifies as an episode if the following conditions are satisfied: 4 (i) Unemployment declines at least 3 percentage points over a 4 year period (ii) The decline in unemployment over this 4 year period is at least 25% of total initial unemployment. (iii) The drop in unemployment must remain below the critical unemployment reduction thresholds for at least another 3 years. (iv) Unemployment strictly declines in the first year of the episode. (v) An unemployment reduction episode did not commence in the previous 7 years (e.g. in the previous 7 years there was no year in which conditions i, ii, iii and iv were simultaneously met that was not itself preceded by the onset of an unemployment reduction in the previous 7 years); unemployment reduction episodes are not contiguous. quality products and improved competition as its proponents advocate. In follow-up work, Djankov et al (2008) also highlight the important of enforcement, by demonstrating that transparent contract enforcement is instrumental in preventing self-dealing—where agents exploit power to maintain excess share of profits, thus retarding investment and growth. 4 We also implemented two additional conditions, notably that unemployment declines were not driven by a single outlier alone, and that unemployment during the episode was always strictly lower than unemployment at the onset of the episode, but these conditions were not binding, in the sense that they were always satisfied when the other conditions were satisfied. 4 Conditions i and ii ensure that the decline in unemployment is substantial. A decline of 3 percentage points over a four year period is crudely equivalent to a standard deviation away from the mean 4 year change in unemployment. Condition ii ensures that the threshold for an unemployment reduction episode is higher for countries with higher levels of unemployment; for example, in a country with 40% initial unemployment, it needs to reduce by at least 10 percentage points for that reduction to qualify as an episode, whereas a country with an initial unemployment of 20% requires a reduction in unemployment of 5 percentage points or more. By construction, countries with lower than 3% initial unemployment cannot experience an episode. Condition iii requires that the decline in unemployment is sustained for at least three years, and does not merely reflect cyclical fluctuations. Condition iv ensures that unemployment reduction episodes commence with a decline in unemployment, and condition v rules out counting the same unemployment reduction spell as two separate events. One potential concern is that unemployment could decline due to falling labor market participation. We choose to focus on declines in unemployment, as opposed to increases in employment as a share of working age population since participation can fall as countries become richer, as more people of working age attend high school and university, or retire early. 5 Still, to control for this possibility, we also report results including only episodes where labor force participation remains constant or increases and results remain robust. Our primary data source is the World Bank’s World Development Indicators 1980-2008. The earliest year for which the onset of an unemployment event can be identified is 1980, whereas the last year is 2001. In case information on unemployment is missing for one year, it is imputed using the average of the unemployment rates in the preceding and the subsequent years for the purpose of identifying reduction episodes, but not in the subsequent analysis. Countries for which we do not have at least 8 consecutive annual observations on unemployment after imputing it are excluded. These restrictions reduce the sample to 94 countries. Arguably as a result of lower labor market monitoring capacity in poor countries, there are no low-income countries in our sample, only five countries from Sub-Saharan Africa and two from South Asia. Thus, our results are most relevant for middle income, transition, and OECD countries. See Table 1 for a description of all variables and their source. 5 In addition, employment to population ratios, typically measured as the population over the age of 15 (or 25) that is employed, are potentially (more) impacted by changing demographics, such as those due to increases in life expectancy. 5 2.2 Incidence Unemployment reduction episodes are common, as is demonstrated by Table 2 which lists all 43 events that we identify by region, country and year. Almost half of all the countries in our sample (40 out of 94) have experienced at least one episode. To arrive at the (unconditional) probability of experiencing an unemployment episode, we follow Hausmann et al. (2005) and divide the total number of events by the total number of country-years for which an unemployment episode could have been identified. The latter is calculated by the total number of country-observations eliminating the 7-year window after the onset of an event, since in this period unemployment reduction episodes are not allowed to initiate by construction. On average, approximately one in 20 countries which are not in an episode reduction event embark on one each year. Aside from the sheer number of unemployment events, the magnitude of the associated decline in unemployment is striking; unemployment declines from an initial average of 14.5% to 8.8% and then to 7.1% 4 and 7 years after the onset of the event respectively. Thus, at the end of the episode average unemployment was less than half of its initial value. Table 2 demonstrates that the incidence of unemployment reduction events does not vary dramatically across regions. Their incidence also does not vary strongly across income groups, yet appears to be increasing with the initial level of unemployment as is indicated by Table 3, which lists the incidence of episodes by initial unemployment quartile and income group. Only one country, Thailand, in the lowest unemployment quartile experienced an event. This is perhaps not a surprise once one considers that even supposedly well-functioning labor markets suffer some friction unemployment (Blanchard and Katz, 1997). For example, the natural rate of unemployment in the U.S., which arguably has one of world’s most flexible labor markets, has recently been estimated to be in the range of 5.6 to 6.9% (Daly et al., 2012). In what follows, we exclude observations with unemployment lower than 6% unemployment from the analysis for this reason. In section 4 we will show that our main results are robust to including them. The choice of parameters for the filter used to identify unemployment reduction episodes is inevitably arbitrary. For example, increasing the thresholds for unemployment declines by 10% and 20% reduces the number of events to 36 and 26, respectively. 6 Conversely, relaxing the thresholds 6 A 10% (20%) higher threshold implies that unemployment should decrease by at least 3.3% (3.6%) and 27.5% (30%) of its initial value over the initial four year period. Conversely, a 10% (20%) lower thresholds implies that unemployment 6 by 10% and 20% increases the number of events to 49 and 54, respectively. As another example, adopting a more stringent definition of sustainability by requiring that unemployment 7 years after the onset of the episode cannot be higher than its level 4 years after onset (that is, not allowing for rebounding of unemployment in this period) reduces the number of events to 34. Ruling out the possibility that unemployment reduction episodes are driven by declines in labor force participation, by insisting that the labor force participation rate four years after the onset of the event is at least as high as at the beginning of the event, leaves 33 episodes. In section 4.2 it is shown that adopting these and other alternative parameterizations of the filter does not alter the qualitative pattern of results. It is important to recognize that our identification comes from the timing of episodes within countries and also across countries, between those that have and those that fail to have episodes of unemployment reduction. An example is provided in Figure 1, showing unemployment in Bulgaria and Macedonia, two neighboring countries, with similar populations and trade patterns. While Bulgaria was undertaking reforms starting in the mid-1990s in order to join the European Union, reform in Macedonia did not take hold until a decade later. Their respective patterns of unemployment are quite different. While Bulgaria had an episode of unemployment reduction starting in 2001, Macedonia did not, despite much higher unemployment, which typically makes sustained reduction easier. 3 Correlates and Antecedents of Unemployment Reduction Episodes Now that we have defined and identified episodes of unemployment reduction, we next characterize them by the evolution of macro- and institutional variables. 3.1 Initial Conditions Table 4 presents descriptive statistics on initial conditions at the time of onset of an episode (columns 1) and compares those with the conditions prevailing in countries that are not currently in should decrease by at least 2.7% (2.4%) and 22.5% (20%) of its initial level during the four years after the onset of the episode. 7 an episode and in which no event occurs (column 2). Coefficients in the table with an asterisk indicate that differences in initial conditions are statistically significant at the 5% level. Consistent with the results demonstrated in Table 3, countries that experience unemployment reduction episodes have significantly higher initial unemployment, notably 14.7% on average, than countries in which episodes do not occur, with average unemployment at 11.6%. These numbers may seem high, but recall that we have confined the sample to country-year observations experiencing unemployment levels of at least 6%. In spite of this differential in average initial unemployment, countries embarking on episodes of unemployment reduction are remarkably similar in other macroeconomic domains to those that do not. 7 They neither have significantly different initial GDP levels, nor record significantly different growth rates at the time of onset. They also do not have a significantly higher propensity to export, import and to receive FDI inflows. Moreover, they do not differ from those that do not embark on episodes in terms of average inflation and government spending as a share of GDP. In addition, they are not significantly more or less democratic on average. Countries that have unemployment reductions also do not score higher on indicators of the overall regulatory framework. Our preferred proxy for overall regulatory quality is the economic freedom index of the Fraser Institute, as it is time-varying. Since this indicator and its subcomponents are only available every 5 years from 1980 until 2010, we linearly interpolate scores for the intervening years. No statistically significant differences between countries that embark on episodes and ones that do not are detected in any of the sub-components or the aggregate index itself. We complement these broad indicators with specific indicators of labor regulation and financial openness from the IMF, as well as with data on governance and regulation from the Worldwide Governance Indicators (Kaufmann et al., 2010) and the Doing Business Indicators. Since these last two data sets are only available from 1996 and 2004 onwards, respectively, we follow Collier and Goderis (2008) and extrapolate them backwards for years in which they were not available. Note that this procedure implicitly assumes that these policies do not vary over time, whereas in reality they may well have changed in response to (the absence of) unemployment reduction events. In other words, they are potentially endogenous. Collier and Goderis (2008), however, argue that because these indicators tend to capture structural policies that change only slowly over time, the 7 We also examined differences in average debt, savings, real interest rates, and the terms of trade and did not find any statistically significant differences in these either. 8 magnitude of potential endogeneity bias is likely limited. Bearing in mind this caveat, table 4 suggests that countries embarking on episodes are not characterized by more flexible labor laws, greater financial openness, a policy environment especially conducive to trade, or superior governance. However, they are characterized by expedited enforcement of contracts, which, on average, takes roughly 23% longer in countries that do not embark on episodes. Overall, the descriptive statistics suggest that countries embarking on episodes are not different in terms of initial conditions, except for their business regulations and their initial level of unemployment. 3.2 The Evolution of Key Explanatory Variables Examining the evolution of key explanatory variables sheds light on the likely determinants of unemployment reduction episodes, which is important in view of the stark similarity in initial conditions reported in the previous section. Table 5 presents regressions in which key explanatory variables are regressed on dummies that indicate whether the country is currently in the first four years of an event (labeled “Beginning”), or the subsequent three-year period (labeled “End”) to determine if there are significant changes in these variables. 8 The sample is confined to events for which information on unemployment was available three years prior to the onset of the episode. All regressions are estimated using a standard Fixed-Effects estimator to remove time-invariant country characteristics. Although this procedure controls for country-differences that are constant over time, it does not control for global shocks, or the possibility that some variables, such as trade flows, might share a common trend. To negate the possible impact of these, we also present regression estimates where we use as dependent variable explanatory variables demeaned by the sample average of observations available in the relevant year. This demeaning removes the impact of both covariate shocks, such as global growth booms and crises, as well as time trends; the resulting coefficient estimates thus provide information on how countries in episodes fare relative to other countries in the sample. Turning to the results, the reduction in unemployment that characterizes episodes coincides with an acceleration of growth. This is illustrated by Figures 2 and 3, which plot the evolution of unemployment and GDP growth, respectively, demeaned by the sample average. Unemployment 8Note that we confine the sample to country-year observations that are either experiencing an episode or about to embark on one within at most three years. 9 tends to increase in the build-up to the event (Figure 2), reflecting the fact that episodes often involve a reversal of fortune, which is typically accompanied by a distinct jump in GDP growth (Figure 3) at the onset of the unemployment episode. The increase in GDP growth is accompanied by a significant increase in investment as a share of GDP, whereas government spending as a share of GDP declines, suggesting that overall macroeconomic conditions improve. In addition, countries experiencing an unemployment reduction episode are significantly less likely to be experiencing a crisis than they were before the onset of the episode; some events concur with recovery from crisis. Moreover, episodes are accompanied by a significant surge in trade (see also Freund and Pierola, 2008), exports and imports both increase significantly and roughly by the same magnitude. The regulatory environment appears to improve during unemployment reduction episodes, as is indicated by the positive and significant improvement in the Economic Freedom Index. This improvement persists in the last three years of the episode and is predominantly driven by significantly improved regulation, improvements in the legal system and property rights and access to sound money (reflecting inflation, its volatility, money growth, and the ability to own foreign currency bank accounts – see Table 1). These beneficial changes appear to persist throughout the second phase of episodes. By contrast, no significant changes in financial openness and labor regulations as proxied by advance notice and severance pay requirements as well as the generosity of unemployment benefits are detected during either the first or the second phase of episodes. 4 Predicting Unemployment Reduction Episodes We now turn to potential predictors of successful episodes of reduction in unemployment. We estimate a probit model, where the dependent variable takes the value 0 if there is potential for an unemployment reduction episode to start in country i at time t but none has, and 1 if it starts in year t. Years in which an episode cannot begin (e.g. the seven years after onset) are excluded. The estimating equation is: Pr( | ) = Φ( ′ β) where is the vector of explanatory variables. 10 Table 6 presents various specifications of this model for countries with unemployment rates of at least six percent that are not currently experiencing an episode. All specifications control for initial unemployment and include year dummies. Column 1 additionally controls for GDP per capita and its growth, whereas column 2 instead controls for a host of macroeconomic indicators, notably investment, openness, FDI, government consumption, inflation, as well as democracy, proxied by the polity indicator (from Marshall et al., 2011). Column 3 simultaneously controls for both. Column 4 instead controls for GDP per capita, its growth, initial unemployment and whether or not a country is in crisis or at war. Column 5 examines the impact of regulation using the EFW indicator of economic freedom as a proxy for the overall quality of regulation, controlling for GDP and its growth as well as initial unemployment. Column 6 includes all additional explanatory variables. Overall, while unemployment episodes are difficult to predict, there are some important regularities. The models we present explain between 10-18% of the observed variance, which is not low in these type of event studies. 9 Of greater interest, the difficulty in predicting episodes is reflected in the fact that few explanatory variables are statistically significant. Macro-conditions other than unemployment do not predict the onset of episodes; the other macro variables, including GDP per capita and its growth, are never statistically significant, neither individually nor jointly. Simultaneously controlling for both, as is done in column 3, does not overturn this conclusion. Column 4 demonstrates that countries at war are more likely to embark on an episode. However, this effect is only significant at the 10% level, and, moreover, not robust to including additional control variables (see column 6). In contrast, initial unemployment is a strong and significant predictor of the onset of unemployment reduction episodes. On average, a 1% point increase in initial unemployment increases the likelihood of the incidence of an episode by 0.7%-0.9%. Although this effect may not seem large in absolute terms, one has to bear in mind that the unconditional probability of an episode taking off is 5.8% for countries with unemployment in excess of 6%. To put this into perspective, ceteris paribus, the odds of an episode happening in a country with 20% unemployment are approximately twice as high than it happening in the average country in our sample, even though, to qualify as an episode, its unemployment must decrease by almost a full percentage point more due to the criterion that unemployment must decrease by at least a quarter of its pre-event level (condition ii). 9For example, the models Hausmann et al. (2005) use to predict growth accelerations explain between 5 and 8% of the observed variance. The low pseudo R2 also could be driven by rare-events bias (King and Zeng, 2001), which is addressed in the next section. 11 The most important finding is that regulatory quality is positively correlated with the incidence of episodes as is evidenced by the strongly statistically significant coefficient on the index of economic freedom in columns 5. This finding is robust to controlling for macro-variables, democracy, and whether or not a country is in crisis or at war (column 6). Moreover, the magnitude of this association is remarkable; a one-standard deviation improvement in regulatory quality (i.e. an increase of 1.08 in the Economic Freedom Index) is associated with an increased probability of incidence of 3.6%-5.0%. Note that this is a conditional association; recall that we did not find a significant positive bivariate correlation between the incidence of employment episodes and indicators of overall regulatory quality because countries with better regulation tend to have lower initial unemployment to start with. Tables 7 and 8 present robustness checks using specifications that replicate those in columns 5 and 6 in Table 6; that is, one that controls for initial unemployment, GDP per capita and GDP per capita squared (referred to as the initial conditions, “IC”, specification), and one that includes all explanatory variables (referred to as the “Full” specification), which we can only estimate on a smaller sample for which all of these variables are available. To conserve space we only present the coefficient estimates associated with our key variable of interest, notably economic freedom. To start with, as alternative proxies for overall regulatory quality, the ease of doing business rank from the Doing Business Indicators and the regulatory quality index from the World Governance Indicators are used. The results presented in Row A of Table 7 are robust to using these alternative proxies; the ease of doing business rank is consistently negatively correlated with the incidence of episodes, indicating that worse regulation renders unemployment reduction less likely. The indicator of regulatory quality from the Worldwide Governance Indicators assigns higher scores to countries with better regulatory frameworks and is significantly positively correlated with the incidence of unemployment reduction episodes. Second, we examine the robustness of our results to using more stringent definitions of an unemployment episode by imposing additional criteria for the identification of an episode. We begin by ruling out the possibility that reductions in unemployment are driven by declines in labor force participation. In particular, we impose as an additional criterion for the identification of an episode that labor force participation after the first four years of the unemployment episode is at least as high as it was at the start. While this reduces the number of episodes to 33, the Economic Freedom Index remains a strongly significant predictor of the onset of episodes (see columns 1 and 2 in Row B). Next, we impose a stricter definition of sustainability and require that unemployment does not 12 increase between four and seven years after onset of the episode. This leaves 34 events. If anything, adopting this more stringent definition of sustainability leads to a stronger correlation between regulation and the incidence of episodes (see columns 3 and 4 in Row B). Third, we assess the robustness of the results to using higher thresholds for unemployment declines, which we increase and decrease by 10% and 20% respectively; the results, which are presented in row C, are qualitatively robust to using these alternative thresholds, although changing the thresholds by a wide margin diminishes the statistical significance of the conditional correlation between regulation and the incidence of employment episodes. Fourth, we examine the robustness of our results to alternative sample restrictions. To start with, the requirement that initial unemployment must be in excess of six percent is dropped. The qualitative pattern of results does not change (see columns 1 and 2 in row D). Finally, we rule out episodes being driven by recovery from war or crises by excluding countries which were at war or in a crisis at any point during the past four years. Discarding these observations substantially strengthens the association between regulation and the incidence of episodes. Fifth, Table 8 examines the robustness of the results to using alternative estimations methods that are better equipped to deal with unobserved heterogeneity and rare-occurrence bias. The latter bias may arise because we are focusing on the onset of episodes; even though, at any given time a substantial number of countries are experiencing an unemployment episode, onset is of course more rare. This may result in bias, which is typically downwards, due to both small sample selection bias (which tends to bias coefficient estimates downwards) and not explicitly accounting for estimation uncertainty (which tends to reduce the estimated variance, which in turn result in underestimation of the likelihood of the occurrence of rare events). To address these issues, we re- estimate our models using the modified logistic regression models proposed by King and Zeng (2001) that correct for these potential problems. The results are presented in the top row of Table 8. Rare-events bias does not appear to affect the coefficient estimate associated with regulation which, if anything, is lower than in standard logistic regressions, which are presented in columns 5 and 6 for purposes of comparability. Nonetheless, accounting for estimation uncertainty strengthens our results somewhat; the attributable risk 10 associated with a 1 point increase in the Economic Freedom Index is estimated to be 3.9% in the specification that controls for initial conditions only and 6.0% in the specification that includes all control variables. These increases in probability are a 10Attributable risk (or the first difference risk) is defined as the change in the probability as a function of a change in a covariate; see King and Zeng (2001). 13 bit higher than the corresponding marginal effects obtained using a standard probit model (recall the results presented in columns 5 and 6 of table 5 which yielded marginal effects of 3.6% and 5.0% respectively). Finally, we attempt to address potential bias due to unobserved heterogeneity, starting with random effects probit models. The specifications presented in columns 1 and 2 of the bottom row of Table 8 show that the results are robust to controlling for such random effects, and that the null hypothesis that they should not be included is not rejected. A well-documented drawback of the random effects estimator is that it imposes that unobserved country-specific effects are not correlated with the explanatory variables. To allow for the possibility that they are, we re-estimate these models using Chamberlains fixed-effects logit, at the cost of having to exclude countries that never experienced episodes from our sample. This in turn leads to convergence problems in the model that includes all explanatory variables (the full specification). We also present standard linear fixed effects estimates, which obviously do not appropriately account for the binary nature of the data yet help shed light on the likely impact of unobserved time-invariant heterogeneity by using information from countries that did not witness an episode. Overall, our results suggest that such heterogeneity is important, as is evidenced by F-tests that reject the null that country-fixed effects do not matter in the linear fixed effects specifications. Nonetheless, such heterogeneity is unlikely the key driver of the results we observe; the positive association between regulation and the incidence of episodes remains significant both in the conditional logit and linear fixed effects models, albeit at the 10% level in the latter specifications. This is a strong result, since regulation evolves only slowly over time, and because a one year horizon over which to identify the beneficial impacts of reforms is fairly short. To summarize, the relationship between regulation and unemployment reduction episodes appears robust. 5 Which Policies Matter Most? Bayesian Model Averaging The regulatory proxies used thus far are quite broad, and the positive association between regulation and the incidence of episodes prompts the question: which aspects of regulation matter most? To help answer this question and to assess the robustness of our previous results, we employ Bayesian Model Averaging methods using subcomponents of the Economic Freedom Index as well as 14 alternative, more detailed, proxies for labor, trade, financial and business regulation, and indicators of governance as potential predictors of the onset of episodes. Bayesian Model Averaging offers a systematic method to deal with the uncertainty inherent in model selection by allowing one to assess the posterior likelihood of models and coefficients, thereby helping us assess which variables are most relevant (see Hoeting et al., 1999, for an introduction to Bayesian Model Averaging). To formalize the notion of model uncertainty, let X denote a n x p matrix of potential predictors of outcome variable Y that is assumed to be a binary indicator of a latent variable ∗ that follows a logistic density with mean =Xβ. Given the number of potential explanatory variables p, there are q=2p possible different models which, following Raftery (1995), we shall assume to be a priori equally likely. The marginal distribution of the data for a given model is given by: (| ) = � (| , ) ∗ ( | ) ∗ where ( | ) is the prior distribution 11 of the parameters of model . The posterior model probability for any given model can in turn be computed as: (| ) ∗ ( ) ( |) = . ∑=0 (| ) ∗ ( ) where ( ) is the prior probability that model is true. The posterior distribution of a quantity of interest can now be computed as the weighted average of each of the models considered, where weights are given by the posterior model probabilities. For example, the posterior expected value for the coefficient vector β after averaging across models is: (|) = � ( |) ∗ ( | , ). =0 We implement the Bayesian Model Averaging technique three times. 12 The results are presented in Table 9. To start with, as a robustness check, we use the Bayesian Model Averaging procedure using all variables included in our most general model, notably specification 6 in Table 6, as potential explanatory variables. The Bayesian analysis, which is presented in the first column 11 In our application, these prior distributions are computed using the BIC approximation, which is akin to the Unit Information Prior (UIP) (see Raftery, 1996). 12 To implement the BMA method, we use the “BMA” package in R (version 3.15.1 by Raftery et al. (2012)) which uses approximate Bayes factors and Occam’s window algorithm to reduce the model space to a set of parsimonious models that have decent explanatory power. One general caveat to bear in mind when interpreting the results of BMA procedures is that their results can be very sensitive to measurement error (see Ciccone and Jarocinski, 2010). 15 (labeled model set 1), corroborates our results; the most potent predictor of the incidence of episodes is initial unemployment. This is evidenced by its extremely high posterior inclusion probability (PIP) (99.9). 13 The next best predictor, albeit with a substantially lower posterior inclusion probability (35.3), is the Economic Freedom Index. The other explanatory variables considered do not appear to be useful predictors. 14 Second, to assess which aspects of policy matter most we replicate this analysis, but now replace the Economic Freedom Index by its component indicators, notably indices of the regulation of business, credit and labor, the freedom to trade internationally, the size of the government, legal system and the security of property rights, and access to sound money. The results, which are presented in the second column (labeled model set 2), suggest that the legal system and property rights and rules governing credit, business and labor, are the most important aspects of the business environment. Note that while the posterior inclusion probability of the indicator of the legal system is higher than that of regulation, the coefficient associated with regulation is much higher. Third, to validate these results, and to further probe which aspects of regulation matter most, we run a Bayesian Model Averaging analysis where we use as explanatory variables initial unemployment and a host of alternative indicators of regulation. To proxy labor regulation we use indicators of the severance pay and advance notice requirements as well as the generosity of unemployment benefits from the IMF (Aleksynska et al., 2011) and the rigidity of employment index from the World Bank Doing Business Indicators. The time it takes to export as recorded in the Doing Business Indicators is used as a proxy for trade regulation, whereas the time it takes to open and close a business and the time it takes to enforce a contract are used as proxies for business regulation. We also include an indicator of financial regulation from Abiad et. al. (2008), as well as indicators of political stability, the control of corruption and the rule of law from the Worldwide Governance Indicators. The Bayesian Model Averaging analysis suggests that amongst them the time it takes to enforce a contract and the control of corruption are the most important predictors of the incidence of episodes, albeit at much lower posterior inclusion probabilities than initial unemployment. In particular, protracted contract enforcement is associated with a reduced 13 The PIP is a measure of how important a predictor a variable is and is defined as the sum of the posterior probabilities of all models that include the variable in question; if models that include this variable are more likely, one can infer that the variable in question has predictive power. 14 In robustness checks not presented to conserve space but available upon request we also experimented with including additional explanatory variables, notably, literacy, mortality, demographic indicators, the exchange rate and domestic credit as a share of GDP, but none of these had substantial explanatory power. 16 likelihood of unemployment reduction events commencing. These findings are consistent with the results we obtained when we used the subcomponents of the Economic Freedom Index, which also pointed towards the importance of an efficacious legal system and secure property rights as important enabler of employment growth. They also resonate with the descriptive statistics recorded in Table 4, which demonstrated that countries that embark on episodes are characterized by both higher initial unemployment and expedited contract enforcement relative to countries that do not. Interestingly, the stringency of labor regulations and the generosity of unemployment benefits do not appear to help predict the onset of unemployment reduction episodes. Financial openness does not appear an important determinant of drastic unemployment reductions either. 6 Conclusions We examine how countries across the world have generated sustained reductions in unemployment by using an event study analysis. The frequency with which such unemployment reduction episodes occur is encouraging. Each year approximately 1 in every 20 countries not already in an episode embarks on such an unemployment reduction episode. Moreover, the associated decline in unemployment is typically large, since, average unemployment seven years after onset stood at less than half its initial level. In spite of their prevalence, unemployment reduction episodes are difficult to predict ex ante, reflected in the low predictive power of models of their onset and the fact that countries embarking on episodes are characterized by initial conditions very similar to those which do not in terms of growth, GDP, FDI inflows, exports, imports, investment, government spending, inflation, democracy, and various proxies for regulatory quality. Nonetheless, countries that embark on episodes tend to suffer higher unemployment and have de jure policies dictating more prompt enforcement of contracts. Episodes tend to coincide with an acceleration of growth, an overall improvement in macroeconomic conditions manifested, inter alia, in higher trade flows, high investment and lower government spending, as well as improvements in the regulatory framework. In addition, the incidence of crises reduces significantly, suggesting that episodes sometimes concur with recovery. Although we are not able to attribute causation, these findings point towards the importance of prudent macroeconomic management in fostering sustainable employment growth. 17 Our most important finding is that unemployment reduction episodes are much more likely in countries with better regulation. This relationship is robust to using alternative proxies for regulatory quality, the imposition of additional criteria to identify employment episodes, alternative parameterizations of the filter used to identify episodes, alternative estimation methods and various different sample restrictions, and is crucially conditional on initial unemployment. On average, countries embarking on episodes do not outperform countries that do not in terms of overall regulatory performance. The reason is that countries with good regulation are less likely to have high unemployment in the first place and consequently less likely to experience an episode. However, if they do end up with high unemployment, they are much more likely to escape from it. The results are validated by Bayesian Model Averaging procedures which point towards contract enforcement and the security of property rights as critical components of the business environment and important enablers of employment growth. The relatively strong role for contract enforcement relative to other indicators of governance, labor regulation and access to finance is perhaps not too surprising if we consider that net job creation is typically accounted for by (young) small firms (Haltiwanger et al, 2010), which are disproportionately reliant on lean regulation and consistent implementation thereof in order to expand (Beck et al., 2005). The importance of averting corruption is not entirely unexpected either, since excessive regulation goes hand in hand with graft (Djankov et al 2002), taxing employers and making expansion more difficult, especially for firms lacking connections to government officials. 18 References Abiad, Abdul, Enrica Detragiache and Thierry Tressel, 2008. A New Database of Financial Reforms, IMF Working Paper, WP/08/266. Aleksynska, Mariay and Martin Schindler, 2011. Labor Market Regulations in Low-, Middle- and High-Income Countries: A New Panel Database, IMF Working Paper, WP/11/154. Beck, Thorsten, Asli Demirgüç-Kunt and Vojislav Maksimovic, 2005. Financial and Legal Constraints to Growth: Does Firm Size Matter?, Journal of Finance 60, 137-177. Blanchard, Olivier and Lawrence Katz, 1997. What we know and do not know about the natural rate of unemployment, Journal of Economic Perspectives 11, 51–72 Blanchard, Olivier and Jordi Gali, 2010. Labor Markets and Monetary Policy: A New Keynesian Model with Unemployment, American Economic Journal: Macroeconomics 2, 1-30. Blanchard, Olivier and Justin Wolfers, 2000. The Role of Shocks and Institutions in the Rise of European Unemployment: The Aggregate Evidence, Economic Journal 110, C1-33. Ciccone, Antonio and Marek Jarociński, 2010. Determinants of Economic Growth: Will Data Tell?, American Economic Journal: Macroeconomics 2(4), 222-46. Collier, Paul and Benedikt Goderis, 2009. Structural policies for shock-prone developing countries, Oxford Economic Papers 61, 703-726. Daly, Mary, Bart Hobijn, Ayşegül Şahin, and Rob Valletta, 2012. A Search and Matching Approach to Labor Markets: Did the Natural Rate of Unemployment Rise?, Journal of Economic Perspectives 26, 3- 26. Davis, Steven J., Jason Faberman and John Haltiwanger, 2012. Labor market flows in the cross section and over timem Journal of Monetary Economics 59, 1-18. De Soto, Hernando, 1990. The Other Path, Harper and Row, New York, NY. Djankov, Simeon, Rafael La Porta, Florencio Lopez-De-Silanes and Andrei Shleifer, 2002. The Regulation Of Entry, The Quarterly Journal of Economics 117, 1-37. Djankov, Simeon, Rafael La Porta, Florencio Lopez-de-Silanes, and Andrei Shleifer 2008, The law and economics of self-dealing, Journal of Financial Economics 88, 430-465. Djankov, Simeon and Ramalho, Rita, 2009. Employment laws in developing countries, Journal of Comparative Economics 37, 3-13. Economic Freedom House, 2011. Economic Freedom of the World, Annual Report. 19 Elsby, Michael, Ryan Michaels,and Gary Solon, 2009. The ins and outs of cyclical unemployment, American Economic Journal: Macroeconomics, 1(1), 84-110. Feyrer, James and Bruce Sacerdote, 2013. How Much Would US Style Fiscal Integration Buffer European Unemployment and Income Shocks? The American Economic Review: Papers & Proceedings 13, 125-128. Freund, Caroline, 2005. Current Account Adjustment in Industrial Countries Journal of International Money and Finance 24, 1278-1298. Freund, Caroline and Martha-Denise Pierola 2012, Export Surges Journal of Development Economics 97, 387–395. Gwartney, John, Robert Lawson and Joshua Hall 2010. Economic Freedom of the World: 2010 Annual Report. Hausmann, Ricardo, Lant Pritchett, and Dani Rodrik, 2005. Growth Accelerations Journal of Economic Growth 10(4), 303-329. Haltiwanger, John, Ron S. Jarmin and Javier Miranda, 2010. "Who Creates Jobs? Small vs. Large vs. Young," NBER Working Papers 16300. Havrylyshyn, Oleh and Donal McGetttingan, 1999. Privatization in Transition Countries: A Sampling of the Literature, International Monetary Fund Working Paper 99:6. Helpman, Elhanan and Oleg Itshkoki, 2010. Labour Market Rigidities, Trade and Unemployment Review of Economic Studies 77, 1100-1137. Hoeting, Jennifer, David Madigan, Adrian Raftery, and Chris Volinsky, 1999. Bayesian model averaging: A tutorial, Statistical Science 14, 382-401. Kaufmann, Daniel, Aart Kraay and Massimo Mastruzzi 2010. The Worldwide Governance Indicators : A Summary of Methodology, Data and Analytical Issues. World Bank Policy Research Working Paper No. 5430. King, Gary and Langhe Zeng, 2001. Logistic Regression in Rare Events Data, Political Analysis, 137- 163. Lacina, Bethany and Nils Petter Gleditsch, 2005. Monitoring Trends in Global Combat: A New Dataset of Battle Deaths, European Journal of Population 21, 145–116 Laeven, Luc, and Fabian Valencia, 2010. Resolution of Banking Crises: The Good, the Bad, and the Ugly, IMF Working Paper, WP/10/146. Marshall, Monty G., Keith Jaggers and Ted Robert Gurr, 2011. Polity IV Project: Dataset Users’ Manual. Arlington: Polity IV Project. 20 Nickell, Stephen Luca Nunziata and Wolfgang Ochel, 2005. Unemployment in the OECD Since the 1960s. What Do We Know?, Economic Journal 115, 1-27. Raftery, Adrian, 1995. Bayesian model selection in social research, Sociological Methodology 25, 111-163. Raftery, Adrian, 1996. Approximate Bayes Factors and Accounting for Model Uncertainty in Generalised Linear Models, Biometrika, 83, 251-266. Raftery, Adrian, Jennifer Hoeting, Chris Volinsky, Ian Painter and Ka Yee Yeung, 2012. BMA: Bayesian Model Averaging. R package, version 3.15.1.URL: http://CRAN.R- project.org/package=BMA 21 Table 1: Variable Description and Sources Variable Name Source Description Unemployment WDI Log GDP per capita WDI Natural logarithm of real GDP per capita based on purchasing power parity (PPP). Data are in constant 2005 international dollars. GDP per capita growth WDI GDP per capita growth (% terms) Exports WDI Exports of goods and services (% of GDP) Imports WDI Imports of goods and services (% of GDP) Openness WDI (Exports+Imports) FDI WDI Foreign direct investment, net inflows (% of GDP) Investment WDI Gross fixed capital formation (% of GDP) Government Spending WDI General government final consumption expenditure (% of GDP) Inflation (log) WDI log((100+annual inflation(%))/100) Democracy Marshall et Polity2: the combined polity score which is the difference between the al. (2011) democracy and autocracy indicators War Lacina and Indicator variable taking value 1 if the country was engaged in a war (i.e. a Gleditsch conflict with at least 1,000 battle related death in a given year), and 0 otherwise (2005) Crisis Laeven and Indicator of banking crises taking the value 1 if the country was experiencing a Valencia banking crisis and 0 otherwise. (2010) Broad Indicators of Regulatory Quality Economic Freedom EFW Composite index of economic freedom Ease of Doing Business Rank Country’s rank score in the ease of doing business indicators in 2008 DB (extrapolated backwards over time) (1=most business friendly regulations) Regulatory Quality Measures the ability of the government to formulate and implement sound policies and regulations that permit and promote private sector development (with higher values corresponding to better outcomes). For years in which WGI information on this indicator was missing, we use the average of the preceding and the subsequent years if those are available (1997,1999 and 2001) and the earliest available year otherwise (i.e. for years preceding 1996 we take the 1996 value). Dimensions of Economic Freedom Government Size EFW Indicator of central government involvement in the economy (comprising general government consumption spending, transfers and subsidies, government enterprises and investment and the top marginal tax rate) Regulation EFW Indicator of credit (ownership of banks, foreign bank competition, private sector credit), labor (minimum wages, hiring and firing regulations, centralized collective bargaining, mandated cost of hiring and worker dismissal and conscription) and business regulation (price controls, administrative requirements, bureaucracy costs, starting a business, bribes, licensing restrictions, costs of tax compliance). Legal System EFW Indicator of legal structure and the security of property rights, taking into account judicial independence, whether courts are impartial, property rights, military interference in the political process, integrity of the legal system, legal enforcement of contracts, and regulatory restrictions on sale of real property Money EFW Indicator of the efficacy of money as a medium of exchange (comprising measures of money growth, inflation and its volatility, as well as the freedom to access foreign bank accounts). Free Trade EFW Indicator of the ease with which goods can be traded across borders (comprising taxes on trade, regulatory trade barriers, size of the trade sector relative to expected, black-market exchange rates, capital market controls) Labor Regulation Rigidity of Employment DB Measures flexibility in the regulation of employment in 2008, specifically as it 22 affects the hiring and redundancy of workers and the rigidity of working hours Unemployment Benefits Aleksynska Generosity of unemployment benefits measures by gross replacement rate et al. (GRR), that is, the ratio of unemployment insurance benefits a worker receives (2011) relative to the worker’s last gross earnings after being unemployed for one year Severance Pay Aleksynska Index of legally mandated severance payments for workers with 9 months of et al. service (2011) Advance Notice Aleksynska Index of legally mandated advance notice requirements for workers with 9 et al. months of service (2011) Finance Financial Openness Abiad et Financial liberalization index (rescaled) (comprised of 8 sub-components) al. (2008) Business Time to Enforce a Contact DB Log time to enforce a contract in 2004 Starting a Business DB Log time required to start a business in days in 2004 Closing a Business DB Log time to resolve a bankruptcy in 2004 (in years) Trade Time to Export DB Log time to export in days in 2004 Governance Political Stability WGI Measures the likelihood that the government will be destabilized or overthrown by unconstitutional or violent means, including domestic violence and terrorism. For years in which information on this indicator was missing, we use the average of the preceding and the subsequent years if those are available (1997,1999 and 2001) and the earliest available year otherwise (i.e. for years preceding 1996 we take the 1996 value). Rule of Law WGI Measure of the extent to which agents have confidence in and abide by the rules of society, in particular the quality of contract enforcement, the police, and the courts, as well as the likelihood of crime and violence. For years in which information on this indicator was missing, we use the average of the preceding and the subsequent years if those are available (1997,1999 and 2001) and the earliest available year otherwise (i.e. for years preceding 1996 we take the 1996 value). Control of Corruption WGI Measure of the extent to which public power is exercised for private gain, including petty and grand forms of corruption, as well as capture of the state by elites and private interests. For years in which information on this indicator was missing, we use the average of the preceding and the subsequent years if those are available (1997,1999 and 2001) and the earliest available year otherwise (i.e. for years preceding 1996 we take the 1996 value). Note: WDI=World Development Indicators, WGI=Worldwide Governance Indicators, DB=Doing Business Indicators, EFW=Economic Freedom of the World 23 Table 2: List of Unemployment Reduction Episodes by Region, Year, and Country Region Episode Year Country Unemployment ∆Unemployment After 4 After 7 First 4 All 7 (countries) % (No./Obs) At onset years years years years East Asia and 6.3% (5/80) 1986 Malaysia 8.3 5.1 3.0 -3.2 -5.3 the Pacific (8) 1986 Singapore 6.5* 2.1* 2.7* -4.5* -3.8* 1987 Thailand 5.8 2.7 1.3 -3.1 -4.5 1988 Fiji 9.4 5.4 5.4 -4.0 -4.0 1998 Korea, Rep. 7.0 3.3 3.7 -3.7 -3.3 Europe and 6.3% (8/127) 1993 Hungary 12.1 9.0 6.6 -3.1 -5.5 Central Asia 1996 Latvia 20.2 13.9 10.5 -6.3 -9.7 1999 Russia 13.5 8.2 7.2 -5.3 -6.3 2000 Ukraine 11.6 8.6 6.4 -3.0 -5.2 2001 Bulgaria 19.4 10.1 5.7 -9.3 -13.7 2001 Croatia 20.5 12.6 8.4 -7.9 -12.1 2001 Estonia 12.6 7.9 5.5 -4.7 -7.1 2001 Lithuania 16.8 8.3 5.8 -8.5 -11.0 Industrial (26) 3.6% (12/332) 1983 Canada 12.0 8.8 8.1 -3.2 -3.9 1983 United States 9.6 6.2 5.6 -3.4 -4.0 1984 Netherlands 14.2 9.1 6.9 -5.1 -7.3 1986 Portugal 8.6 4.7 5.5 -3.9 -3.1 1993 Denmark 10.7 5.4 4.5 -5.3 -6.2 1993 Ireland 15.6 10.2 4.3 -5.4 -11.3 1993 Britain 10.3 7.1 5.6 -3.2 -4.7 1994 Finland 16.4 11.4 9.1 -5.0 -7.3 1995 Netherlands 7.0 3.5 3.1 -3.5 -3.9 1995 Spain 22.7 15.6 11.4 -7.1 -11.3 1997 France 12.6 8.6 9.2 -4.0 -3.4 1997 Sweden 10.0 5.0 6.5 -5.0 -3.5 Latin America 4.6% (15/329) 1982 Chile 19.6 8.7 5.3 -10.9 -14.3 and the 1982 Costa Rica 9.4 6.3 3.7 -3.1 -5.7 Caribean (26) 1983 Uruguay 15.4 9.1 8.5 -6.3 -6.9 1984 Jamaica 25.6 18.9 15.7 -6.7 -9.9 1985 Colombia 14.0 8.9 9.5 -5.1 -4.5 1985 El Salvador 16.9 8.4 7.9 -8.5 -9.0 1987 Bolivia 20.5 5.9 3.1 -14.6 -17.4 1993 Barbados 25.6 14.6 9.3 -11.0 -16.3 1994 Bahamas, The 13.4 7.6 6.9 -5.8 -6.5 1995 Mexico 6.9 2.5 2.9 -4.4 -4.0 1995 Nicaragua 16.9 10.9 12.2 -6.0 -4.7 1996 Trinidad and 16.3 12.1 10.5 -4.2 -5.8 1997 Cuba 7.1 4.1 1.9 -3.0 -5.2 1999 Jamaica 15.7 11.7 9.6 -4.0 -6.1 2000 Colombia 20.5 13.7 10.9 -6.8 -9.6 Middle East 3.0% (2/66) 1995 Morocco 22.9 13.9 11.6 -9.0 -11.3 and Northern 2000 Algeria 29.8 17.7 13.8 -12.1 -16.0 Africa South Asia (2) 4% (1/24) 1994 Sri Lanka 13.0 9.1 7.9 -3.9 -5.1 Sub Saharan (5) 0% (0/25) Africa Notes: the number of observations refers to the number of country-year observations characterized by unemployment in excess of 3% that have not experienced the start of an episode within the last 7 years. ∆Unemployment refers to the 4 year change in unemployment from the onset of the event onwards. * indicates unemployment rates were imputed. 24 Table 3: List of Unemployment Reduction Episodes by Income Group and Initial Unemployment Income Group Unemployment Lower Middle Upper High: High: Total Quartile Middle Non-OECD OECD Q1 No. 1 0 0 0 1 (3.0