Can employment subsidies save jobs? Evidence from a shipbuilding city in South Korea

Can employment subsidies save jobs? Evidence from a shipbuilding city in South Korea

Labour Economics 61 (2019) 101763 Contents lists available at ScienceDirect Labour Economics journal homepage: www.elsevier.com/locate/labeco Can e...

2MB Sizes 0 Downloads 24 Views

Labour Economics 61 (2019) 101763

Contents lists available at ScienceDirect

Labour Economics journal homepage: www.elsevier.com/locate/labeco

Can employment subsidies save jobs? Evidence from a shipbuilding city in South Korea☆ Hyejin Kim a, Jungmin Lee a,b,∗ a b

Department of Economics, Seoul National University, South Korea Institute for the Study of Labor (IZA), Germany

a r t i c l e

i n f o

JEL codes: J21 J65 R23 Keywords: Employment subsidies Counter-cyclical policy Local economic crisis Synthetic control method Program evaluation

a b s t r a c t We evaluate the effectiveness of employment subsidies when used as a countermeasure to severe recessions. In 2013 and 2014, the government of South Korea implemented an emergency policy called the Special Employment Promotion Zone program in a medium-sized city suffering a prolonged slump after the 2008 global crisis affected the city’s major industry, shipbuilding. Under the program, employers could receive subsidies to cover a significant part of the wages for retaining their employees or to create new jobs for local residents. With the synthetic control method, we found that the program was not fully utilized in its first year and had little impact on employment. In the second year, the program increased the employment rate, mainly in the non-manufacturing sector. The magnitude of the effect was small, but the effect persisted after the end of the program.

1. Introduction In many countries, subsidies are commonly used to promote employment among disadvantaged groups, for example, among the youth, elderly, or disabled. Such policies are often implemented as temporary countermeasures in response to unanticipated severe recessions. Economic theory shows that subsidies can increase employment by reducing the effective wages paid by employers, but the empirical question of the cost-effectiveness of the policy then comes into play. It is difficult to estimate the effect of subsidies because doing so requires one to estimate the number of jobs that would have been created otherwise. This problem becomes even more challenging when the policy is implemented only briefly and is targeted at a specific local area. In such cases, it is difficult to construct a reasonable counterfactual trend of employment without the policy in place. In this study, we evaluate a temporary and place-based employment subsidy program in South Korea (Korea hereinafter) that was intended

☆ We thank seminar participants at the KEA-APEA 2017 Conference, AASLE Conference 2017, and 30th EALE Conference 2018 for helpful comments and discussion. Especially, we appreciate Yoon-Gyu Yoon and Sangho Lee for their kind help when we started this research. Kim acknowledges funding from the BK21Plus Program of the Ministry of Education and National Research Foundation of Korea (NRF-21B20130000013). Lee’s work was supported by the Ministry of Education of the Republic of Korea and the National Research Foundation of Korea (NRF-2017S1A3A2066494). All remaining errors are our own. ∗ Corresponding author. E-mail addresses: [email protected] (H. Kim), [email protected] (J. Lee).

to promote employment retention and spur local job creation in a specific city Tong-Yeong (TY, hereinafter), which was experiencing a severe recession. A prolonged slump in oil prices and the Great Recession of 2008 decreased the volume of new shipbuilding orders. Therefore, all major shipbuilding companies suffered financially, leading to considerable job losses, especially in the municipalities whose economies relied substantially on the shipbuilding industry. Among the eight largest shipbuilding municipalities in Korea, TY, a medium-sized coastal city in the southern region of Korea, was affected the most. To prevent the local industry crisis from spiraling into a recession in the overall local labor market, the Ministry of Employment and Labor (MOEL) designated the city a Special Employment Promotion Zone (SEPZ) in January 2013, offering employment subsidies to local firms if they retained existing jobs or created new jobs for local residents. The program offered considerable benefits. It covered up to 90% of wage costs over six months if employers retained their employees instead of downsizing. Assuming a company received the maximum amount of subsidy per day, that is, 50,000 KRW, the program could reduce labor costs by 55% considering the average wage in TY in 2013 and 2014. Moreover, the program subsidized the hiring of local residents at a rate ranging between 33% and 50% of the wages for one year. In this study, we aim to estimate the effect of this policy on employment in the labor market of TY. Our study is mainly related to two strands of literature. First, because the subsidy program considered herein was geographically targeted at a specific city, our study relates to the strand of literature that evaluates place-based policies. Governments often implement several active labor market measures to create local jobs in economically poor

https://doi.org/10.1016/j.labeco.2019.101763 Received 15 May 2018; Received in revised form 26 June 2019; Accepted 9 September 2019 Available online 17 September 2019 0927-5371/© 2019 Elsevier B.V. All rights reserved.

H. Kim and J. Lee

areas or in areas where job prospects are weak (Neumark and Simpson, 2015). Many studies have evaluated the effectiveness of such policies, but their results have been mixed. The studies closest to our study are those that explore the Enterprise Zone (EZ). Among such studies, a few studies found little or no effect of place-based policies on employment (Wilder and Rubin, 1996; Bondonio and Greenbaum, 2007; Neumark and Kolko, 2010), while others found that place-based policies positively impact local employment (O’Keefe, 2004; Ham et al., 2011; Busso et al., 2013; Freedman, 2013). Second, our study is broadly related to studies evaluating the impact of employment subsidies that are not necessarily limited geographically. Even though it is theoretically obvious that subsidies increase employment, at least in the short run, regardless of the assumption of perfect competition or friction in labor markets, the empirical results have surprisingly been mixed. Most studies have focused on wage subsidies targeted at disadvantaged groups, such as the disabled, youth, or female workers. This might explain why a few empirical results are inconsistent with theoretical predictions. Katz (1998) argued that targeted wage subsidies for disadvantaged groups might stigmatize the said groups, making these policies somewhat less effective than more broadly targeted subsidies. This means that temporary and non-targeted subsidies might have the potential to stimulate employment growth. Only a few studies on more broadly targeted subsidies (Neumark and Grijalva, 2017; Cahuc et al., 2019), especially those on subsidies implemented temporarily during recessions, have been reported that are relevant to the subsidy program examined herein.1 The subsidy program considered in our study was implemented temporarily in one city for two years. Because the municipalities surrounding the city could potentially be influenced by the program through the spillover effect, it is difficult to apply the standard differencein-differences (DID) method to analyze the effectiveness of the program. Therefore, we use the synthetic control method developed by Abadie et al. (2010) and Abadie et al. (2015). Specifically, we construct a synthetic control unit that provides the counterfactual trend of employment growth rate, against which we can compare the actual evolution of the employment growth rate in the target city in the posttreatment period. We obtain this counterfactual trend by using data from other cities and counties in Korea, with specific weights assigned to the data to ensure that the resulting synthetic control unit best reproduces the adjusted employment growth rates in the target city before implementation of the employment subsidy program. We use administrative records from the National Unemployment Insurance (UI) system. An advantage of using UI data is that only the workers covered under UI are eligible for the subsidy program, so our employment measure from the UI data should include all of those beneficiaries. However, because the UI coverage rate among wage workers is not 100%, our results should be interpreted as the effect of the subsidy on UI-covered wage workers rather than on all workers, including non-UI-covered workers and self-employed individuals. To summarize our main findings, first, we find that the SEPZ program had little impact on employment in the local labor market in its first year. In fact, less than one-third of the government budget allocated to the program in the first year was used. This result is consistent with the previous finding that employment subsidy programs are usually underutilized initially (Katz, 1998). However, in the second year (the program was extended for a year), the program started generating a positive effect on employment. The magnitude of the effect is relatively small, especially given the possibility that our estimates are biased upward. Moreover, the effect is concentrated in the last two quarters of 1 A few studies have conducted randomized controlled experiments at the individual level, where those in the treatment group receive a subsidy (Galasso, 2004; Levinsohn et al., 2014; Groh et al., 2016). These studies provide clean evidence on the causal effect of a subsidy but are limited in that they cannot capture any spillover or general equilibrium effects that might occur among non-eligible individuals.

Labour Economics 61 (2019) 101763

the second year. According to the estimates obtained with the synthetic control method, the program increased the employment rate by approximately 0.5 percentage points. We conduct various robustness checks and find that our main results are robust to different outcomes, specifications, and donor pool restrictions. Especially, we find that our results are robust when we restrict the sample to those areas where shipbuilding is the major industry. This suggests that our results are not specific to the context of a particular industry but are more generalizable. Moreover, the positive effect that appeared in the last two quarters of the program seemed to persist for several months after the end of the program. Finally, we find that the positive effect is stronger in the nonmanufacturing sector and appears for both male and female workers. The remainder of this paper is organized as follows. Section 2 presents the historical background of the program and the city of TY, as well as institutional details about the SEPZ program in Korea. Section 3 introduces the data and constructs the sample for applying the synthetic control method. Section 4 presents a simple theoretical framework and our main findings. Section 5 presents the results of various robustness checks and subsample analyses. Section 6 presents our concluding remarks and discusses a few policy implications. 2. Institutional background The financial crisis of 2008 and the subsequent global economic slowdown caused a major downturn in the shipbuilding industry. The city of TY is one of several major shipbuilding municipalities in Korea, and it was adversely affected by the decline in the industry. Furthermore, all major shipbuilders in TY were small- and medium-sized shipyards, focusing on building merchant ships. Thus, they were particularly vulnerable to the negative demand shock in the global market. Although large-scale shipyards in other municipalities were able to diversify their production to better adapt to the recession, the shipbuilders in TY were hit hard, and some of them eventually went bankrupt and ceased operations. According to the 2010 Economic Census data, the shipbuilding industry accounted for only 1.4% of the total number of enterprises in TY but accounted for 18.7% of the total number of employees and 28.2% of the city’s total revenue. Therefore, the shipbuilding industry was regarded the local flagship industry in terms of the number of workers and revenue (Shim and Lee, 2014). The boom in the shipbuilding industry in the early 2000s brought not only windfall profits to shipbuilders but also facilitated the expansion of restaurants, real estate, and various rental businesses in TY. However, after the 2008 recession, the condition of the local labor market deteriorated quickly. The economic shock was observable, especially in the hollowing out of shopping streets. Fig. 1 shows the trend of the gross regional domestic product (GRDP) of TY from 1998 to 2014. It increased sharply from the mid-2000s, when the shipbuilding boom began. After the peak in 2009, it dropped sharply in 2010, mainly because of the recession in the shipbuilding industry. It has increased slightly since then, but it remains considerably lower than the peak level of 2009. Notably, the GRDP increased slightly in 2014, which was the second year of the implemented employment subsidy program, although an upward trend seemed to have been initiated in 2012.2 When the local economic situation did not recover from the crisis, civic groups and labor unions demanded government support, and the city informally asked the MOEL if TY could be designated a SEPZ in June 2012.3 At first, the government rejected the proposal, given that TY did not meet the eligibility criteria for the program. Thereafter, civic groups

2 As shown later, using the synthetic control method, we find that the subsidy program has a small but significant positive effect in the second year, 2014. It seems that the GRDP trend here is consistent with this result. 3 For detailed information on the SEPZ program in general and the program’s implementation in TY, please refer to Yoon et al. (2014).

H. Kim and J. Lee

Fig. 1. Yearly trend of gross regional domestic product of TY, 1998–2014. Data Source: Provincial Economic Statistics of the Southeast Region of Korea. Notes: The vertical line represents the year 2009. The nominal value of GRDP is converted to the real value in 2010 KRW by using consumer price index (CPI).

and unions continually urged the government to implement the program. A presidential election took place on December 19. Civic groups urged the presidential candidates to provide a possible solution for the city and the crisis of the shipbuilding industry, and the demand culminated at a press conference with the appointment of the city’s congressman to the National Assembly on November 19. Finally, on December 7, the MOEL modified a few eligibility criteria and designated TY a SEPZ.4 It is important to note that the final decision was made less than one month before the program commenced (January 1, 2013). This is an important point in our empirical analysis because to the extent that the SEPZ was unanticipated, the pretreatment trends should not be contaminated. Two specific criteria must be met for a region to be designated a SEPZ: (1) There should exist an industry accounting on average for more than 15% of all UI-covered workers in the region in the three months before being designated a SEPZ. In addition, one of the following two conditions must be satisfied: (1-a) the three-month average of the Business Survey Index (BSI), an industry-specific business cycle index constructed by the central bank, of the industry must have dropped by more than 30% compared with the same period in the previous year; or (1-b) there must be a high likelihood of massive layoffs from a major firm or establishment in the industry. (2) Additionally, at least one of the following three conditions must be met: (2-a) the three-month average number of involuntary separations must be higher than 3% of the monthly average number of UI-covered employees in the previous year; (2-b) the three-month average of the number of UI-covered employees must have decreased by more than 5% compared with the monthly average in the previous year; or (2-c) when a major firm in the region goes bankrupt or must perform severe restructuring, the ratio of layoffs to the monthly average number of UI-covered employees in the previous year must be greater than 3%.5 4 According to the original criteria, TY was not eligible for the program. Therefore, MOEL rejected the initial application made by the city. Later, MOEL modified the criteria to make TY eligible. In fact, the program’s official name was changed from Employment Development & Promotion Zone (EDPZ) to Special Employment Promotion Zone. Before TY, a city in the mid-western region of Korea Pyoungtek was designated a EDPZ. In 2009, Pyoungtek city saw massive layoffs from SsangYong Motors, an automobile maker that was the largest company in the city. About 10 billion KRW of government budget was provided to the city for one year. 5 One may believe that we can exploit the eligibility conditions and apply the difference-in-differences method or regression discontinuity design with multi-

Labour Economics 61 (2019) 101763

In the case of TY, criteria (1-a) and (2-b) were satisfied. During the three months of September to November 2012, the BSI of the shipbuilding industry, which accounted for 33.6% of the total number of employees in TY, decreased by an average of 36.3% compared with the same period in the previous year. Moreover, the average number of employees from September to November in 2012 decreased by more than 5% compared with the monthly average number of employees in 2011. The program provided a variety of employment subsidies to local firms in TY to encourage the hiring of local residents and create jobs in the region. The content of the employment subsidies, according to the SEPZ designation, can be divided into three categories: employee retention, local employment creation, and extension of the payment schedule for UI benefits and industrial accident compensation insurance premiums. The details are as follows: First, under the program, the MOEL provided up to 90% of the expenses incurred on paid leave. In addition, the expenses for training incumbent workers could be supported. Workers on unpaid leaves were granted some financial support for covering living expenses after a screening procedure. Second, to create local jobs, the government provided employment subsidies to employers who relocated, newly established, or expanded their workplaces in the region and hired local residents. In addition, some financial support was provided for local job creation projects, such as yacht schools and marine specialist schools in the case of TY.6 In addition, the scope of support for the existing employment promotion programs was expanded. Finally, the time limit for UI benefits and industrial accident insurance premiums was extended for shipbuilders in TY. Moreover, two details are important from the viewpoint of understanding our empirical analyses below. First, although the program was applied to TY because of the slump in the shipbuilding industry, there was no restriction by industry. Second, the SEPZ program was supposed to be a one-year program. However, in the case of TY, again, perhaps partly because of political pressure, the program was extended to a second year. According to a government report, the support amount totaled 17.1 billion KRW, which is equivalent to 15 million USD, and the total number of jobs subsidized was 8429. After the one-year renewal, the program was terminated in January 2015. 3. Data and synthetic control group 3.1. Data and descriptive statistics We used the administrative UI database, which encompasses all UIcovered employees. We constructed a quarterly municipality-level panel dataset for the period from 2008q1 (the first quarter of 2008), the first quarter for which data are publicly available, to 2016q2 (the second quarter of 2016). By using information about the population size of each municipality from the Population Census, we constructed our main outcome variable, the quarterly employment rate, which we defined as the number of UI-covered employees divided by the population aged 15 years and older. Notably, the employment rate computed by considering only UIcovered employees is lower than the employment rate computed by including all employees. This can be ascribed to various reasons. First, the self-employment rate is high in Korea, and self-employed individuals are ple cutoffs. However, because the program is designed as an emergency intervention policy, making the conditions intentionally stringent, few regions are close enough to satisfy the conditions. In fact, the program was implemented in only one city before TY. 6 Among the various components of the program, the employee retention program is a major component. According to the program budget report by the MOEL, this component alone accounts for 62% of all beneficiaries and 57% of the total program expense. The next largest component is the local specific job creation program in terms of the number of beneficiaries (26.2%) and the financial support for small-sized firms in terms of the amount of spending (28.4%).

H. Kim and J. Lee

Labour Economics 61 (2019) 101763

Table 1 Comparison between Tong-Yeong and the rest of Korea.

Population size 2008 2009 2010 2011 2012 Percentage of population aged 25–54 2008 2009 2010 2011 2012 Percentage of male population 2008 2009 2010 2011 2012 Employment rate 2008 2009 2010 2011 2012

Tong-Yeong

Rest of Korea

135,570 136,769 138,396 140,303 139,800

307,489 309,025 311,474 315,088 316,307

49.70 49.08 48.48 48.07 48.11

50.24 50.11 49.70 49.11 48.90

50.93 50.99 50.99 50.88 50.73

50.11 50.09 50.09 50.09 50.06

14.32 16.31 16.25 15.38 14.67

22.48 23.20 23.79 24.54 25.44

of an average municipality in Korea, and the ratio of the prime workingage population is slightly lower than the average. The share of males is slightly higher than the average. Moreover, the employment rate is 7– 11 percentage points lower than the average. The differences between the treated city and the rest of Korea indicate that it is important to find an appropriate control group to estimate the causal effect of the SEPZ program. 3.2. Synthetic control construction

Notes: Rest of Korea includes all 160 cities and counties. Except for population size, we present the population-weighted average of each variable. The employment rate of each municipality is defined as the ratio of the number of UI-covered employees to the population size.

not eligible for UI. By contrast, participation in the UI system is mandatory for wage and salary workers. Second, the UI coverage rate is not 100% for wage and salary workers. According to statistics from the Regional Employment Survey (2012), the employment rate of TY is 58%, share of wage workers among all employees is approximately 60%, and UI coverage rate among wage workers is 69%. Therefore, the UI-covered employment rate, which is the definition used in this study, is 24% (= 0.58×0.6 × 0.69). Note that this is higher than the employment rate in our data. One possible reason for this discrepancy is that we counted UI-covered workers based on the locations of establishments, while the survey data counted them based on the locations of worker residences. To qualify for the employment subsidies that we have examined herein, workers should be covered by UI, and their establishments should be located in TY. The lower employment rate in our sample implies that some workers are employed by establishments located or registered outside of TY. Because our outcome variable captures only a fraction of the labor force, our estimates presented below may be biased upwardly. First, workers might be reallocated from contracts that are not eligible for UI and wage subsidies to contracts that are eligible for UI and wage subsidies. Second, similarly, workers might be reallocated from regions outside TY to TY so that they can qualify for subsidies (Hanson and Rohlin, 2013). In either case, we would overestimate the effect of the program on total employment in the local labor market.7 In addition to the outcome variable, by using the Population Census, we included the predictors of regional labor market outcomes, namely, gender composition of the population, percentage of the population aged 25–54 years (prime age range for a working population), and population size. The yearly averages of these variables were included as the predictors. All pretreatment outcome lags were included as well.8 Table 1 presents descriptive statistics corresponding to the pretreatment period. The population of the treated city of TY is smaller than that 7 Our estimates indicate a small effect, but even this small effect could be the upper bound of the true effect. 8 For robustness, we attempted to use different sets of pretreatment outcome lags as predictors. Please see Section 5.2.

To prepare the data for the synthetic control method, following a standard time-series approach used in the literature, we took both first and seasonal differences to the log-transformed values of the outcome variables for stationarity. That is, for each regional unit (municipality), we constructed a series of Δ2 ln𝑦𝑖𝑡 = (ln𝑦𝑖𝑡 − ln𝑦𝑖𝑡−4 ) − (ln𝑦𝑖𝑡−1 − ln𝑦𝑖𝑡−5 ), where lnyit is the natural logarithm of the employment rate of municipality i in year-quarter t.9 Δ2 lnyit represents the quarterly growth rate of the seasonally adjusted employment rate. We decided to transform the data in this way because the goodness of fit for the pretreatment period was the highest with the transformed data. We compared the quality of fit in terms of the pretreatment root-mean-squared prediction error (RMSPE, the square root of the average of the squared discrepancies between the employment growth rates in TY and in its synthetic counterpart during the pre-intervention period) by using three alternative specifications10 : log employment rate, first difference of log employment rate, and seasonal difference of log employment rate. The pretreatment RMSPE was the lowest for our baseline specification (0.0035), followed by seasonal difference of log employment rate (0.0124), first difference of log employment rate (0.0127), and log employment rate (0.0219). In this regard, we follow the argument of Abadie et al. (2010) that a good pre-trend fit is a necessary condition for applying the synthetic control method.11 A few recent studies have investigated the problem of nonstationarity in the context of the synthetic control method (Harvey and Thiele, 2017; Carvalho et al., 2018; Ferman and Pinto, 2018). For this reason, using daily exchange rate data, Chamon et al. (2017) took the first difference to secure stationarity of the series. However, there is no consensus on whether it is suitable to secure data stationarity before applying the synthetic control method. Because the data were first differenced and seasonally differenced, we excluded the first five observations (quarters) for each municipality. Thus, the sample period begins in 2009q2. Because the subsidy program began in January 2013, we set the start of the treatment period to 2013q1, resulting in a pretreatment period of 15 quarters. We restricted our sample period until 2016q2 because another policy intervention was made in July 2016, where the government designated the shipbuilding industry as a sector for receiving special employment promotion support in view of the continued slump in the industry. This makes it difficult to distinguish whether the estimated effects that appear from 2016q3 can be attributed to the persistent effect of the SEPZ program or the effect of the new policy. 9 We used quarterly data because as shown later, the policy effect appears significant only in the last few months of the program. If we were to use annual data, we would not be able to capture these minute effects. 10 Please refer to Appendix A for the mathematical definition of the pretreatment RMSPE. 11 Abadie et al. (2010) stated that they would not recommend using a synthetic control method when the fit is poor. Kaul et al. (2018) argued that by optimizing the pretreatment fit of the outcome with all pretreatment outcome lags renders the covariates irrelevant, leading to a potentially biased estimator. To the best of our knowledge, whether a pre-trend fit is good or bad is theoretically ambiguous. For example, according to Abadie et al. (2010), the definition is that the characteristics of the treated unit are sufficiently matched by the synthetic control unit over the pretreatment period. The goodness of fit is usually evaluated by either the RMSPE or the “eyeball” test.

H. Kim and J. Lee

Labour Economics 61 (2019) 101763

We constructed the synthetic control as the weighted average of a comparison of municipalities in the donor pool. For our main analysis, we did not restrict the donor pool and included all 160 cities and counties in Korea.12 For robustness, we attempted to restrict the donor pool in various ways to check the sensitivity of our estimates according to the choice of donor pool. Fig. 2 shows the locations of the chosen regions on a map of Korea, and in this figure, it can be seen that they are scattered throughout the country. The counterfactual trend is the weighted average of the employment growth rates of these 12 cities and counties.13 To check how well the synthetic control unit fits the pretreatment period regarding the employment growth rate of the treated unit, we compared the trends of the employment growth rates between the treated unit and the synthetic control unit or between the treated unit and one of the two alternative comparison groups, namely, population-weighted average of all 160 regions (“Rest of Korea”) and the average of two neighboring municipalities, Geoje and Goseong (Fig. 2 show their locations relative to TY). We were especially interested in the comparison with the neighboring regions because these regions would most likely be selected as the comparison group in studies using the DID estimation method. Fig. 3 presents the gaps in the trends of employment growth rates between the treated unit and each of the three comparison groups during the pretreatment period. Here, the trend of the synthetic control unit fits that of the treated unit much better than those of the other two alternative comparison groups. In fact, the gaps between the treated unit and its synthetic counterpart are close to zero throughout the pretreatment period, reflecting that the synthetic control unit is the best control group for this study.14 4. Empirical results 4.1. Theoretical framework Before showing our empirical findings, we present a simple theoretical framework that helps us interpret the results presented below. A canonical model of labor supply and demand predicts that a wage subsidy shifts the labor demand curve outward and expands the employment of targeted workers (Katz, 1998; Neumark, 2013). The size of the employment effect depends on the wage elasticities of labor supply and demand. In a static and partial-equilibrium model with a proportional subsidy (s), the marginal effect of the subsidy on employment is given as follows: 𝜂𝜖 𝑑 ln𝐸 = 𝑑𝑠 𝜂+𝜖 where 𝜂 is the absolute value of the labor demand elasticity, and 𝜖 is the labor supply elasticity. The simple formula implies that the impact of a wage subsidy may differ based on the characteristics of target group 12 To construct this sample, three cities, Changwon, Masan, and Jinhae, which were administratively integrated in 2010, were added separately since 2008 to obtain a balanced panel. Two additional municipalities, namely, Cheongju city and Cheongwon County, were integrated in 2015, and we treated them separately in the sample for the same reason as above. Finally, we dropped one city, Sejong, which was newly created in 2012, because we could not construct a balanced panel with this city. 13 The synthetic control unit consists of 12 municipalities. Please refer to Fig. 2 for the locations of the municipalities and the weights assigned to them. In one city, the SEPZ program was implemented in 2009, but this city was not included in the synthetic control unit. 14 Recall that we used other regional characteristics to construct the synthetic control unit. For those variables, we did not produce matches that are as good when compared with the lagged outcome variables. However, this is not a serious problem because the sum of the weights assigned to those other covariates in the optimal weighting matrix is near zero. Moreover, refer to Botosaru and Ferman (2019), who argued that a perfect match of the covariates is not required for the synthetic control method so long as there is a perfect match on a long set of the pretreatment outcomes.

Fig. 2. Municipalities included in synthetic control, and seven major shipbuilding municipalities. Notes: Panel A shows the 12 cities and counties included in synthetic control with the weights assigned to them. The locations of the seven major shipbuilding municipalities in Korea are shown in Panel B.

H. Kim and J. Lee

Labour Economics 61 (2019) 101763

Fig. 3. Pretreatment period trend gaps between treated unit and each of the comparison units. Notes: The graph shows the difference in the adjusted employment growth rate for each quarter between TY and each possible control group: synthetic control, rest of Korea, and neighboring municipalities (Geoje and Goseong).

individuals across local labor markets as well as across industries. For example, the labor demand elasticity tends to be larger in the manufacturing sector (Lichter et al., 2015), so the impact of wage subsidy on manufacturing employment is likely to be larger. Moreover, because the labor supply elasticity is larger for females than it is for males, it is expected that the impact is larger for females. Furthermore, according to a behavioral model, apart from the wage elasticities of labor demand and supply, the effectiveness of a policy may depend upon institutional factors, such as regulatory burden, existence of a stigma effect, and level of policy awareness (Perloff and Wachter, 1979).15 Therefore, often, the utilization rate or take-up rate is pretty low (Hamersma, 2005), especially when the program is first enacted. While a static model clearly predicts a positive effect on employment in the short run, a wage subsidy program, even if only temporary, may have a few dynamic effects. The lasting impact occurs through several mechanisms (Hamersma, 2008; Mckenzie, 2017; de Mel et al., 2019). First, wage subsidies may enable firms to overcome the matching frictions. Firms may not know workers’ types, which may make them reluctant to hire workers, especially during recessions. This problem could be exacerbated when firing costs are high. A wage subsidy can reduce the cost of hiring and induce firms to consider hiring new workers, leading to an increase in employment. If firms retain their workers who have turned out to be good matches, the program will have a persistent impact. A second possibility stems from combination of firm-specific human capital and minimum wages. Newly hired workers may be less productive because of low firm-specific human capital, but their productivity can be enhanced through on-the-job training, which is also subsidized by the program. In a competitive model, firms pay workers their value of marginal product, and wages increase as worker productivity increases. However, the minimum wage laws set a lower bound on wages, and this may prevent firms from hiring workers with productivity lower than that corresponding to the minimum wage. Wage subsidies could compensate firms for hiring such low-skilled and low-productivity workers. However, if workers could increase their productivity to a level above the minimum wage during the subsidy program, firms will retain them after the end of the program.

15 Chetty et al. (2013) showed that the level of knowledge about the EITC, which is a type of worker subsidy, differs significantly across areas in the U.S., and, accordingly, the impact of EITC on labor supply differs.

Fig. 4. Treatment effects over time: Seasonally adjusted employment growth rates. Notes: The first dashed vertical line corresponds to the time (2013q1) when TY was designated a SEPZ, and the second dashed vertical line refers to the time (2014q4) when the SEPZ program ended.

4.2. Main findings Fig. 4 displays the trends in the adjusted employment growth rates of the treated unit and the synthetic control unit from 2009q2 to 2016q2. The employment growth rates of the synthetic control unit closely track the trajectory of those in TY in the pretreatment period. The close fit with the pretreatment outcomes in Fig. 4 demonstrates that it is possible to construct a combination of other regions in the donor pool that could well reproduce the pretreatment characteristics of the treated unit.16 Therefore, the synthetic control unit can seemingly provide a good approximation of the adjusted employment growth rate that would have

16 To test whether we over-fitted the pretreatment trend, we divided the pretreatment period into a training period and a testing period, and checked the synthetic control fit for the testing period. For robustness, we divided the pretreatment period by four different points of time: 2009q4, 2010q3, 2011q2, and 2012q1. The lengths of the training period were 3, 6, 9, and 12 quarters, respectively. The synthetic control unit for TY fits well the trend in the testing period. Of course, the fit improves as the length of the testing period increases. Additionally, we used the cross-validation technique, as recommended in Abadie et al. (2015). The same definitions of training and testing periods were used herein, and we obtained results similar to our baseline results.

H. Kim and J. Lee

evolved in the treated unit during the post-treatment period in the absence of the SEPZ program. Our estimates of the effects of the employment subsidy program in the treated unit are shown by the gaps between the adjusted employment growth rates of the treated unit and the synthetic control unit after the start of the program, as illustrated in Fig. 4. In its first year, the program had a negligible effect on the local labor market. This result is not surprising because the program was not implemented properly in 2013. In 2013, only approximately 28% of the total budget for employee retention was utilized. Of the 7 billion KRW in the budget for employment creation, approximately 4 billion KRW was not spent. However, the data for the second year look different. The magnitude of the estimated impact of the SEPZ was greater in the second year. The gap between the treated and the synthetic control units started to increase notably immediately after extension of the program. Immediately after completion of the program, the treated unit experienced a decline in employment growth. However, the treatment effect strengthened again, and the program may seemingly have had some lasting effects on the adjusted employment growth rates.17 To demonstrate the effect of the program more intuitively, we accumulated the adjusted employment growth rates for the treated unit and the synthetic control unit and plotted the gap in the level of employment rate, as shown in Fig. 5.18 Consistent with our findings pertaining to the employment growth rate, we found that the program had little or no impact on the level of employment in 2013. However, the effect seemed to emerge in the second year. The gap between the treated unit and the synthetic control unit grew notably in the second year. Specifically, the employment rate in TY was estimated to be approximately 1.83 percentage points higher than that in the synthetic control unit in 2014q4, which corresponds to the period just before the end of program. During the implementation of the policy from 2013q1 to 2014q4, the employment rate increased by approximately 0.52 percentage points. Our findings are consistent with those of a previous research that found some positive effects of place-based policies or wage subsidies. Regarding the impact of the place-based policy, Freedman (2013) examined the effect of the EZ program and found that the EZ designation increases resident employment by 1–2% per year. O’Keefe (2004) found that the EZ program raises employment growth by approximately 3% per year during the first six years after designation. Busso and Kline (2008) used rejected applicants as the control group and found EZ designation to be associated with an increase of approximately 4 percentage points in the local employment rate. Regarding the effect of the wage subsidy program, Bruhn (2016) found a positive but not statistically significant effect on employment over the duration of the program, ranging from 5.7% to 13.2%. Cahuc et al. (2019) found that the hiring credit in France significantly increased the employment growth rate of targeted firms by 0.8 percentage points. Because these policies are different in many regards, it is not possible to compare the effect sizes directly, but it seems that our estimates are within the range of the estimates in the previous studies or slightly lower. The growth in the employment rate slowed down immediately after the program ended, but soon after, the employment rate increased again. It is interesting to find that the policy effect did not dissipate after the subsidy ended and for several months thereafter. The reason

17 One valid concern in this study is that the designation of TY as the SEPZ could have affected the labor market of the cities, especially Geoje city and Goseong County, which are located near TY, creating a spillover effect. In our judgement, this potential problem does not appear to be serious because Geoje and Goseong are not included in the synthetic control unit. Moreover, we could obtain almost similar results, showing that the synthetic control estimates were not affected when conducting the analysis after excluding Geoje and Goseong from the donor pool. Moreover, we considered the alternative outcome variable, that is, the number of jobs, and obtained similar results. 18 In Appendix C, we have explained in detail how we recovered the level of employment rates from our estimates of the employment growth rates.

Labour Economics 61 (2019) 101763

Fig. 5. Treatment effects on the level of employment rates. Notes: The first dashed vertical line corresponds to the time (2013q1) when TY was designated a SEPZ, and the second dashed vertical line refers to the time (2014q4) when the SEPZ program ended. The figure plots the level of employment rate of treated unit and counterfactual unit, which are the cumulative changes in the log of the seasonally differenced employment rate for treated unit and synthetic control unit.

for the prolonged effect after the end of the program is unknown. As we explained in the previous subsection, the effect may persist through various channels, for example, via reduced matching frictions or firmspecific human capital accumulation of the newly employed during the program period. In addition, the persistent effect can be ascribed to institutional factors. Especially, in the program examined herein, only UIcovered workers were eligible for subsidies. Because those workers are protected by unions and labor laws, this restriction may make it difficult for firms to re-designate those jobs to non-UI-covered jobs. 5. Robustness checks 5.1. Donor pool restrictions As a robustness check, we analyzed whether our results are adequately robust to restrict the donor pool in different ways and include alternative control groups. First, we restricted the donor pool to the municipalities in the South Kyeongsang (Southeastern) Province, to which TY belongs, and South Jeolla (Southwestern) Province. The two provinces are adjacent to each other. Second, we restricted the donor

H. Kim and J. Lee

Fig. 6. Donor pool restrictions. Notes: The first dashed vertical line corresponds to the time (2013q1) when TY was designated a SEPZ, and the second dashed vertical line refers to the time (2014q4) when the SEPZ program ended. For panel B, see the notes to Fig. 5.

pool to seven major shipbuilding municipalities, as shown in the panel B of Fig. 2. It is reasonable to assume that they should constitute a good control group because they share many common factors given their similar industrial structures. However, a disadvantage of this restriction is that there are only seven municipalities in this category. Finally, we selected and included in the donor pool the municipalities with similar levels of employment rates as those in the treated unit. Specifically, we selected the municipalities whose trends of employment rates were within one standard deviation above or below the mean of the treated unit during the pretreatment period. The results of all three robustness checks are presented in Fig. 6. Not surprisingly, the goodness of the pre-trend fit is not as good as before when the entire sample is used as the donor pool without any restrictions (the upper left graph in Fig. 6A). However, the results are qualitatively similar, regardless of the donor pool restrictions. All three results show that the program has little impact in its first year, but it does have some positive impact in its second year. 5.2. Pretreatment outcome usage There is an ongoing debate over the choice of predictors in the application of the synthetic control method. In our baseline specification, we employ all outcome lags as predictors. However, Kaul et al. (2018) showed that using all pretreatment outcomes results

Labour Economics 61 (2019) 101763

Fig. 7. Pretreatment outcome usage. Notes: The first dashed vertical line corresponds to the time (2013q1) when TY was designated a SEPZ, and the second dashed vertical line refers to the time (2014q4) when the SEPZ program ended. For panel B, see the notes to Fig. 5.

in biased and less precise estimates. They compared the results across three cases: when all lags are used, when the average lag is used, and when only the last pretreatment outcome is used. Considering six different specifications, Ferman et al. (2018) showed that limiting the number of pretreatment outcome lags causes misallocation of weights. Given this concern in the literature, we checked the robustness of our results with different specifications of the pretreatment outcome usage. We employed four alternative specifications: (1) using the annual averages instead of all quarterly outcomes, (2) using only the outcomes of each year’s last quarter, (3) using only the average of the entire pretreatment period, and (4) using only the outcome in the quarter immediately before the start of the program. Fig. 7 shows our results (see Table B1 in Appendix B for predictor weights). Panel A shows the trends of the treated and synthetic control units over the entire sample period. In panel B, as before, we convert the growth rate results into the level of employment rate. The results show that except for the specification of using the pretreatment period average, the results are qualitatively similar to each other and to those obtained using our baseline specification. This shows that the program had no significant effect in its first year but a positive effect in its second year, especially in the last two quarters. The positive effect in the second year seems to be stronger with the specifications using either the

H. Kim and J. Lee

outcomes of each year’s last quarter or the outcome of the last quarter of the pretreatment period. Surprisingly, the effect is negative over the entire program period in panel B when we use the average of the entire pretreatment period. This is most likely because of an abnormal jump in the employment growth rate for the synthetic control unit in the first quarter of the program period. We do not know the underlying reason, but we guess that the use of only the pretreatment average is inadequate for predicting the counterfactual outcome owing to substantial variations in the outcomes of the treated unit over the pretreatment period. 5.3. Placebo tests and statistical inference We conducted a placebo test in which the treatment was reassigned to a different region which was selected randomly from the donor pool. We iteratively applied the synthetic control method to every other re-

Fig. 8. True and placebo treatment effects. Notes: The first dashed vertical line corresponds to the time (2013q1) when TY was designated a SEPZ, and the second dashed vertical line refers to the time (2014q4) when the SEPZ program ended. The graphs depict the gap between the adjusted employment growth rates in the treated and the synthetic control units. The thick dark line represents the gap for TY. The gray lines indicate the gap for the 160 placebo regions in panel A and for the 130 placebo regions after discarding cities with pretreatment MSPE five times higher than that of TY.

Labour Economics 61 (2019) 101763

gion. That is, we proceeded as if one of the regions in the donor pool would have been designated the SEPZ in 2013q1 instead of TY. We then computed the estimated effect for every placebo test. This procedure yielded the distribution of the estimated gaps for the treated unit and for the 160 placebo regions. We will consider the effect for the actual treated unit as significant if the estimated gap for the treated unit is unusually large compared with the distribution of the placebo effects. Fig. 8 displays the results of the placebo test. The gray lines represent the estimated gaps associated with each of the 160 runs of the placebo test, that is, the gray lines show the difference in the adjusted employment growth rate between each placebo unit and its synthetic counterpart. The superimposed black solid line denotes the estimated gap for the actual treated unit, TY. The estimated gap for the treated unit during the treatment period is positive, but the magnitude of the effect is modest compared with the gaps for the placebo units. However, in 2014q2, the estimated gap for the treated unit during the treatment period seems unusually large relative to the distribution of gaps for the 160 placebo regions. The graph in panel A of Fig. 8 shows that in the cases of a few placebo units, the pretreatment trends are not well fitted. This means it is difficult to find a convex combination of other municipalities in the donor pool that can best reproduce the adjusted employment growth

Fig. 9. Treatment effects over time for males. Notes: The first dashed vertical line corresponds to the time (2013q1) when TY was designated a SEPZ, and the second dashed vertical line refers to the time (2014q4) when the SEPZ program ended. For panel B, see the notes to Fig. 5.

H. Kim and J. Lee

Labour Economics 61 (2019) 101763

Table 2 Estimated treatment effects and empirical p-values.

2013q1 2013q2 2013q3 2013q4 2014q1 2014q2 2014q3 2014q4 2015q1 2015q2 2015q3 2015q4 2016q1 2016q2 Average (2013q1-2014q4)

(1) Treated unit (TY)

(2) Synthetic control unit

(3) Gap

(4) 160 placebo municipalities

(5) 130 placebo municipalities

0.1423 0.0017 0.0026 0.0126 0.0108 −0.0363 0.0096 −0.0197 −0.0154 0.0385 −0.0135 0.0041 0.0029 0.0071 0.0156

0.0814 0.0000 0.0028 0.0025 −0.0083 −0.0007 −0.0154 −0.0111 0.0090 0.0008 0.0142 0.0101 −0.0129 −0.0396 0.0067

−0.0001 0.0160 0.0037 −0.0183 0.0064 0.0100 0.0456 0.0616 −0.0002 −0.0424 −0.0372 −0.0132 0.0159 0.0467 0.0139

0.2625 0.2500 0.4250 0.9188 0.5625 0.2688 0.0250 0.0188 0.5438 0.9375 0.9625 0.7938 0.2875 0.1000 0.0438

0.2385 0.2154 0.4308 0.9615 0.5462 0.2385 0.0231 0.0231 0.5615 0.9615 0.9846 0.7923 0.2692 0.0769 0.0308

Notes: Empirical p-values in columns (4) and (5) have been determined by estimating the treatment effect for each placebo unit in the donor pool and then calculating the fraction of such placebo effects greater than or equal to the effect for the treated unit. In column (4), we have used all 160 municipalities in the donor pool. In column (5), we have restricted the donor pool to 130 municipalities after excluding 30 outliers whose pretreatment fit was found to be bad.

Table 3 Difference-in-differences estimation results. Control group Treatment After Observations R-squared Control group Treatment After Observations R-squared

(1) (2) Rest of Korea

(3) (4) Southern provinces

0.0341∗ (0.0182)

0.0300∗∗∗ (0.0102)

0.0338∗∗ (0.0171) 0.0014 (0.0188) 3703 4669 0.1501 0.1359 (5) (6) Seven major shipbuilding municipalities 0.0283∗∗∗ 0.0286∗∗∗ (0.0104) (0.0099) 0.0084 (0.0108) 184 232 0.6347 0.6234

0.0291∗∗∗ (0.0112) 0.0044 (0.0123) 966 1218 0.3171 0.2488 (7) (8) Municipalities with comparable pretreatment employment rates 0.0311∗∗∗ 0.0312∗∗∗ (0.0097) (0.0104) −0.0028 (0.0114) 483 609 0.5717 0.4888

Notes: All specifications have been controlled for municipality demographic characteristics, year-quarter time fixed effects, and municipality-specific fixed effects. Standard errors are given in parentheses. ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

rate in the pretreatment period. If the synthetic control unit would have failed to fit the trend of the outcome variable before the intervention, we would have interpreted that the post-treatment gap between the actual and the synthetic units was created artificially by the lack of fit rather than the effect of intervention. Thus, when the pretreatment fit is poor for placebo units, we cannot obtain the information required to measure the “relative rarity” of estimating a large gap for the treated unit (Abadie et al., 2010). For this reason, we conducted another set of placebo tests after excluding thirty municipalities with a pretreatment MSPE that is more than five times higher than the MSPE of the treated unit. The results are presented in panel B in Fig. 8. Now, fewer lines deviate from the zero-gap line in the pretreatment period. Evaluated against the distribution of gaps for these 130 remaining placebo units, the magnitude of the treatment effect is more pronounced, especially in the last two quarters of the program. In this context, an empirical p-value can be constructed by calculating the fraction of placebo effects greater than or equal to the effect estimated for the treated unit (Abadie et al., 2015). Table 2 reports the p-value for the dynamic treatment effect in each quarter in the treatment period and for the average treatment effect during the policy implementation period. The estimated treatment effect from 2013q1 to 2014q2 is not significant, but the estimated treatment effect for 2014q3 and 2014q4 is significant. Moreover, the average treatment ef-

fect during the intervention is significant at the conventional 5% level. If one were to pick a city randomly and assign it to the intervention, the chance of getting a treatment effect is as large as the one for TY is 7/160 ≅ 0.0438. With 130 placebo municipalities that have a preMSPE of less than five times the MSPE of TY, the p-value becomes smaller. 5.4. Difference-in-differences estimates As another robustness check, we employed the standard DID method. We obtained estimates with the following two standard models: Δ2 lny𝑖𝑡 = 𝛼 + 𝜆𝑖 + 𝜏𝑡 + 𝛽1 𝑇 𝑌𝑖 ⋅ 𝑃 𝑜𝑙𝑖𝑐 𝑦𝑡 + 𝑋𝑖𝑡′ 𝛾 + 𝜀𝑖𝑡 Δ lny𝑖𝑡 = 𝛼 + 𝜆𝑖 + 𝜏𝑡 + 𝛽1 𝑇 𝑌𝑖 ⋅ 𝑃 𝑜𝑙𝑖𝑐 𝑦𝑡 + 𝛽2 𝑇 𝑟𝑒𝑎𝑡𝑖 ⋅ 𝐴𝑓 𝑡𝑒𝑟𝑡 + 2

(1) 𝑋𝑖𝑡′ 𝛾

+ 𝜀𝑖𝑡 (2)

where i indicates the municipality, and t indicates the year-quarter (2009q2–2014q4 for Eq. (1) and 2009q2–2016q2 for Eq. (2), including the period after the end of the program). The dependent variable Δ2 lnyit was defined in the same way as before. 𝜆i and 𝜏 t are the municipalityand quarter-specific fixed effects, respectively. TYi was defined as 1 if the municipality is the treated unit (TY) and as zero otherwise. In Eq. (1), the binary variable Policyt was set to 1 for quarters between 2013q1 and 2014q4 (policy period) and to zero for quarters before 2013q1. Thus, 𝛽 1

H. Kim and J. Lee

Labour Economics 61 (2019) 101763

captures the effect of the policy intervention between TY and the control group before and after implementation of the SEPZ policy. With Eq. (2), we examined whether the effect persists after the end of the program. We used two time dummies to separate the treatment effects during the policy period and after the end of the program. Policyt was defined in the same way as in Eq. (1). We constructed Aftert , a dummy variable set to 1 for quarters after 2014q4, and added the interaction term with Treati . Aftert was subsumed to Tt . Thus, 𝛽 1 captures the treatment effect during the policy period and 𝛽 2 the effect after the end of the policy. Xit denotes the vector of the covariates, such as the percentage of male population, the percentage of prime-aged workers, and population size. All covariates were included in the adjusted growth rate form. Table 3 reports the results. Given the arbitrariness surrounding the selection of the correct control group, we tried four different groups, namely, rest of Korea (all 160 municipalities), the municipalities in two southern provinces, seven major shipbuilding municipalities, and municipalities with employment rates comparable to that of the treated unit. These groups were selected because we restricted the donor pool in Section 5.1. The results show that the adjusted employment growth rate of the treated unit is approximately 3 percentage points higher than that of the control group. The estimate is close to the result of a synthetic control analysis we performed, which is 2.95 percentage

points. If we convert the estimate to the effect in terms of the level of the employment rate, it corresponds to an increase of 0.5 percentage points. Moreover, we found the DID estimates to be robust to the definition of the control group. The results consistently show that the policy intervention has a positive effect on the adjusted employment growth rate, without any immediate reverting effect after the end of the intervention.

Fig. 10. Treatment effects over time for females. Notes: The first dashed vertical line corresponds to the time (2013q1) when TY was designated a SEPZ, and the second dashed vertical line refers to the time (2014q4) when the SEPZ program ended. For panel B, see the notes to Fig. 5.

Fig. 11. Treatment effects over time for manufacturing workers. Notes: The first dashed vertical line corresponds to the time (2013q1) when TY was designated a SEPZ, and the second dashed vertical line refers to the time (2014q4) when the SEPZ program ended. For panel B, see the notes to Fig. 5.

5.5. Subsample analyses In this subsection, we conduct subsample analyses by gender and industry. Given that the policy was initiated because of a crisis in the shipbuilding industry, in which most workers are male, it is reasonable to hypothesize that the program would have heterogeneous effects across genders and industries. Moreover, as mentioned in Section 4.1, the impact of employment subsidy may differ because of various factors, for example, economic environments, including labor demand and supply elasticities, and institutional factors, such as program stigma or policy awareness. First, when examining males and females separately, we can see in Figs. 9 and 10 that the results for both genders are qualitatively similar to our results for the entire sample. We found no effect on the employment growth rate until 2014q3. However, for females, some positive

H. Kim and J. Lee

effects emerged slightly earlier from 2014q1. The employment rate increased by approximately 0.45 percentage points for males and by 0.34 percentage points for females. The positive effect was persistent for both males and females after the end of the policy. Since female labor supply is more elastic than male supply, we expect that the impact is greater for females. However, our finding is inconsistent with this theoretical prediction. This suggests the possible existence of other offsetting factors. For example, labor demand might be more elastic for male workers, or some institutional or behavioral factors might equalize the effects between males and females. Moreover, note that the trend of female employment was increasing before the program, while that of male employment was declining. Thus, it might be difficult for firms to accelerate the hiring of females, perhaps because of a shortage of female labor supply or temporary capacity constraints in female-dominated occupations. Next, we compare the results by industry. For the manufacturing sector, in panel B of Fig. 11, we find an intriguing result: the employment rate of the treated unit is significantly lower than that of the synthetic control unit over most of the program period. The negative effect is the strongest in 2014q2, when the employment rate of the treated unit is 1.1 percentage points lower than that of the synthetic control unit. On the contrary, Fig. 12 shows that for the non-manufacturing

Fig. 12. Treatment effects over time for non-manufacturing workers. Notes: The first dashed vertical line corresponds to the time (2013q1) when TY was designated a SEPZ, and the second dashed vertical line refers to the time (2014q4) when the SEPZ program ended. For panel B, see the notes to Fig. 5.

Labour Economics 61 (2019) 101763

sector, the treatment effect is positive throughout the policy period. This is intriguing given that workers in all industries are eligible for subsidies so long as they are covered by the UI. The growth rate decelerates after the end of the policy, but the positive effects do not dissipate. The findings by industry are inconsistent with the theoretical prediction based on a simple framework considering labor demand and supply elasticities. We expect that because the labor demand elasticity is larger, the subsidy impact is likely to be greater in the manufacturing sector. Owing to data limitations, we cannot investigate the mechanisms explaining why the impact is larger for the non-manufacturing sector. This topic should be researched in the future.

6. Conclusion We attempted to estimate the causal effect of the SEPZ program, the geographically targeted wage subsidy program, by using the synthetic control method. Our main finding is that the program had little impacts in its first year, but after extension of the program for a year, the positive effect occurred, but its magnitude was small. The estimated gain in the employment rate is approximately 0.5 percentage points. This gain is small, and even the small effect possibly reflects a positive bias. A more positive finding is that the effect did not dissipate after the end of the program. Our findings suggest that the policy achieved its goal of promoting local employment to a limited extent. However, the benefits were not uniformly distributed. Contrary to our expectation, the policy had a stronger effect in the non-manufacturing sector. In fact, we found no positive effect of the policy on employment in the manufacturing sector, including the shipbuilding industry which was struggling during the recession. This does not necessarily indicate the failure of the policy because its main goal was the promotion of local employment overall. However, the result suggests that the policy might not have solved the fundamental problem of the region’s economic vulnerability to external shocks. Employment subsidies are widely used in many countries. Indeed, such subsidies are politically popular because they are typically aimed at disadvantaged groups or local areas that have undergone severe economic recessions. It seems less politically costly to rationalize such a policy because most taxpayers are sympathetic to these groups. When a recession impacts a socio-economically disadvantaged group, politicians must act in a way that is observable and immediate, and in this regard, employment subsidies are quite attractive politically. The SEPZ program examined herein was actually adopted and implemented through the same political process described above. When the shipbuilding industry declined owing to a global depression, the government was forced to take some actions. Therefore, the government started the program by modifying the eligibility conditions. Even a month before the start of the policy, both firms and workers were uncertain about whether the policy would be enacted. This is probably why we found that the policy was unsuccessful in its first year. The policy started to yield a positive effect on employment in its second year. However, the size of the effect was not substantial. Furthermore, as we found, the actual beneficiaries of the program did not well coincide with those who were hit the hardest by the economic shock to the shipbuilding industry, which may suggest that the economic shock to the shipbuilding industry more likely reflects a longterm fundamental problem rather than a temporary one. The city and its industrial structure might need more structural adjustments rather than temporary aids. Furthermore, the temporary aid provided through the program could distort firms’ incentive to make some long-term adjustments. Indeed, the economy of TY has been struggling constantly. In April 2018, the government designated TY a SEPZ for the third time.

H. Kim and J. Lee

Labour Economics 61 (2019) 101763

Supplementary material

close to zero. Then, we can use 𝛼̂ 1𝑡 = 𝑌1𝑡 −

Supplementary material associated with this article can be found, in the online version, at doi 10.1016/j.labeco.2019.101763. Appendix A: Synthetic control method We briefly present the synthetic control approach proposed by Abadie et al. (2010) and Abadie et al. (2015). Suppose we observe 𝐽 + 1 units (municipalities in Korea in our case) indexed by j in periods 1, 2, … , 𝑇 . Assume without loss of generality that region 1 is the case of interest (treatment unit, i.e., TY in our case) and units 𝑗 = 2 to 𝐽 + 1 are comparison units. Borrowing from the literature on statistical matching, the set of comparison units is referred to as the “donor pool.” Let 𝑌𝑖𝑡𝑁 be the outcome that would be observed for region i at time t in the absence of intervention for units 𝑖 = 1, … , 𝐽 + 1 and periods 𝑡 = 1, … , 𝑇 . Let T0 be the number of pretreatment periods, where 1 ≤ T0 < T. Let 𝑌𝑖𝑡𝐼 be the outcome that would be observed for region i at time t if region i is exposed to the intervention (the SEPZ project in TY) in periods 𝑇0 + 1 to T. We assume that the intervention has no effect on the outcome before it is implemented, so for 𝑡 ∈ {1, … , 𝑇0 } and all 𝑖 ∈ {1, … , 𝐽 + 1}, we have 𝑌𝑖𝑡𝑁 = 𝑌𝑖𝑡𝐼 . Let Dit be an indicator that takes a value of one if unit i is exposed to the intervention at time t and zero otherwise. Then, the observed outcome for unit i at time t is 𝑌𝑖𝑡 = 𝑌𝑖𝑡𝑁 + 𝛼𝑖𝑡 𝐷𝑖𝑡 . Because we assume region “one” is exposed to the intervention after period T0 , we have { 1 if 𝑖 = 1 and 𝑡 > 𝑇0 𝐷𝑖𝑡 = . 0 otherwise The goal of this study is to estimate the effect of the intervention on the treated unit given by 𝛼1𝑡 = 𝑌1𝐼𝑡 − 𝑌1𝑁𝑡 = 𝑌1𝑡 − 𝑌1𝑁𝑡 from 𝑇0 + 1 to T (post-treatment period). Because 𝑌1𝐼𝑡 is observed, to estimate 𝛼 1t , we simply need to estimate 𝑌1𝑁𝑡 by using the synthetic control method. The synthetic control method is based on the premise that a combination of comparison units, which we term “synthetic control,” is often superior any individual comparison unit in reproducing the characteristics of the treated unit. Motivated by this consideration, we defined a synthetic control unit as the weighted average of the units in the donor pool that best resembles the characteristics of the case of interest. That is, a synthetic control unit can be represented with a (J × 1) vector comprising weights 𝑊 = (𝑤2 , … , 𝑤𝐽 +1 )′ , with 0 ≤ wj ≤ 1 for 𝑗 = 2, … ., 𝐽 + 1 and 𝑤2 + … + 𝑤𝐽 +1 = 1. Suppose there is an optimal weight vector W∗ , such that characteristics of the treated unit are replicated precisely by the characteristics of the synthetic control unit. 𝑗+1 ∑ ∗ Abadie et al. (2010) show that under regular conditions 𝑌1𝑁𝑡 − 𝑤𝑖 𝑌𝑖𝑡 is 𝑖=2

𝑗+1 ∑ 𝑖=2

as an estimator of 𝛼 1t . To implement the synthetic control method numerically, we must measure the distance between the synthetic control units and the treated unit. Let X1 be a (k × 1) vector containing the values of the pretreatment characteristics of the treated unit, and X0 be the (k × J) matrix collecting the values of the same variables for the units in the donor pool. The pretreatment characteristics may include pretreatment values of outcome variables. We select the synthetic control based on optimal weight vectors W∗ that minimize ‖𝑋1 − 𝑋0 𝑊 ‖. Abadie et al. (2010) choose W∗ as the value of W that minimizes √( )′ ( ) ‖𝑋 − 𝑋 𝑊 ‖ = 𝑋1 − 𝑋0 𝑊 𝑉 𝑋1 − 𝑋0 𝑊 0 ‖ 1 ‖𝑣 where V is some (k × k) symmetric and positive-semidefinite matrix. The solution to the above equation, W∗ (V), depends on the diagonal matrix V, whose diagonal elements are weights that reflect the relative importance of each variable in X1 and X0 according to their predictive power on the outcome variable. Occasionally, the choice of V∗ can be based on subjective assessments of the importance of the variables in X1 and X0 . Alternatively, the choice can be data-driven. Following a datadriven procedure proposed by Abadie et al. (2010), we choose optimal V∗ among all positive-definite and diagonal matrixes, such that the mean squared prediction error (MSPE) of the outcome variable is minimized over some set of pretreatment periods. The MSPE measures the lack of fit between the path of the outcome variable for the treated unit and its corresponding synthetic control units. The RMSPE is simply the square root of the MSPE. ( )2 𝑇0 𝐽∑ +1 1 ∑ ∗ MSPE = 𝑌1𝑡 − 𝑤𝑗 𝑌𝑗𝑡 𝑇0 𝑡=1 𝑗=2 The use of large sample inferential techniques is not suitable for comparative studies when the number of units in the comparison group and the number of periods are relatively small. Therefore, Abadie et al. (2010) propose a “placebo test,” the basic principle of which is to iteratively apply the synthetic control method by randomly assigning the intervention across units (i.e., to control units where the intervention did not occur). Subsequently, we can assess whether the effect of the actual intervention is large relative to the effect of each “placebo control” selected at random. By construction, this exercise produces exact inferences, regardless of the number of available comparison units and time periods (Abadie et al., 2011). Appendix B: Predictor weights Table B1.

Table B1 Predictor weights.

Employment rate Population size Percent of population aged 25–54 Percent of male population

𝑤∗𝑖 𝑌𝑖𝑡 for 𝑡 = 𝑇0 + 1, … , 𝑇

All

Yearly average

Fourth quarter

Average

Last quarter

1.000 0.000 0.000 0.000

0.803 0.048 0.108 0.041

0.574 0.150 0.187 0.089

0.651 0.131 0.106 0.112

0.410 0.300 0.122 0.168

Note: Columns one to five report the predictor weights v (in percent) of the lagged outcome variable and three covariates. We take both the first and the seasonal differences to the log-transformed values of the variables for stationarity. The counterfactual “All” is calculated using the three annual average covariates and all lagged outcome variables. The counterfactual “Annual average” is calculated using the three annual average covariates and the annual averages of the lagged outcome. The counterfactual “Fourth quarter” is calculated using three annual covariates and only the outcomes of the last quarter of each year. The counterfactual “Average” is calculated using the average pretreatment values of the covariates and the lagged outcome. The counterfactual “Last” is calculated using the average pretreatment value of the covariates and only the last pretreatment value of the lagged outcome.

H. Kim and J. Lee

Labour Economics 61 (2019) 101763

Appendix C: Recovering employment rate levels We explain how we draw Fig. 5. To draw the graph, we must convert the estimated employment growth rate to the level of the employment rate. First, we set 𝑙𝑛𝑦𝑇𝑡 = 𝑙𝑛𝑦𝑆𝑡 for the pretreatment period. Next, we iteratively obtain the employment rate of the synthetic control unit for the treatment period. For example, for 𝑡 = 0, the first quarter of the treat̂ 𝑆 = 𝑙𝑛𝑦𝑇 − 𝛼̂ because 𝛼̂ is the estimated difference in ment period 𝑙𝑛𝑦 0

0

0

0

the seasonally adjusted growth rate {(𝑙𝑛𝑦𝑇0 − 𝑙𝑛𝑦𝑇−1 ) − (𝑙𝑛𝑦𝑇−4 − 𝑙𝑛𝑦𝑇−5 )} − {(𝑙𝑛𝑦𝑆0 − 𝑙𝑛𝑦𝑆−1 ) − (𝑙𝑛𝑦𝑆−4 − 𝑙𝑛𝑦𝑆−5 )}. Rearranging it, we obtain (𝑙𝑛𝑦𝑇0 − 𝑙𝑛𝑦𝑆0 ) − (𝑙𝑛𝑦𝑇−1 − 𝑙𝑛𝑦𝑆−1 ) − (𝑙𝑛𝑦𝑇−4 − 𝑙𝑛𝑦𝑆−4 ) + (𝑙𝑛𝑦𝑇−5 − 𝑙𝑛𝑦𝑆−5 ), which is simply (𝑙𝑛𝑦𝑇0 − 𝑙𝑛𝑦𝑆0 ) because all other terms are pretreatment differences and are thus zero. Because 𝑙𝑛𝑦𝑇0 is observed, ̂ 𝑆 . For the second treatment quarter (𝑡 = 1), we can obtain 𝑙𝑛𝑦 0

we estimate 𝛼1 = (𝑙𝑛𝑦𝑇1 − 𝑙𝑛𝑦𝑆1 ) − (𝑙𝑛𝑦𝑇0 − 𝑙𝑛𝑦𝑆0 ), where we can get ̂ ̂ 𝑆 = 𝑙 𝑛𝑦𝑇 − (𝑙 𝑛𝑦𝑇 − 𝑙̂ 𝑆 , which we 𝑙𝑛𝑦 𝑛𝑦𝑆0 ) − 𝛼̂1 . We can plug in 𝑙𝑛𝑦 1 1 0 0 obtained from the first quarter. Similarly, we can iteratively recover the employment rates of the synthetic control unit for t > 1 during the treatment period. References Abadie, A., Diamond, A., Hainmueller, J., 2010. Synthetic control methods for comparative case studies: estimating the effect of California’s tobacco control program. J. Am. Stat. Assoc. 105 (490), 493–505. Abadie, A., Diamond, A., Hainmueller, J., 2011. Synth: an R package for synthetic control methods in comparative case studies. J. Stat. Softw. 42 (13), 1–17. Abadie, A., Diamond, A., Hainmueller, J., 2015. Comparative politics and the synthetic control method. Am. J. Pol. Sci. 59 (2), 495–510. Bondonio, D., Greenbaum, R.T., 2007. Do local tax incentives affect economic growth? What mean impacts miss in the analysis of enterprise zone policies. Reg. Sci. Urban Econ. 37 (1), 121–136. Botosaru, I., Ferman, B., 2019. On the role of covariates in the synthetic control method. Econ. J. 1–14. Bruhn, M., 2016. Can wage subsidies boost employment in the wake of an economic crisis? Evidence from Mexico. Policy Research working paper. World Bank Group, Washington, D.C. WPS 7607. Busso, M., Kline, P., 2008. Do local economic development programs work? Evidence from the federal empowerment zone program. Busso, M., Gregory, J., Kline, P., 2013. Assessing the incidence and efficiency of a prominent place-based policy. Am. Econ. Rev. 103 (2), 897–947. Cahuc, P., Carcillo, S., Le Barbanchon, T., 2019. The effectiveness of hiring credits. Rev. Econ. Stud. 86 (2), 593–626. Carvalho, C., Masini, R., Medeiros, M.C., 2018. ArCo: an artificial counterfactual approach for high-dimensional panel time-series data. J. Econom. 207 (2), 352–380. Chamon, M., Garcia, M., Souza, L., 2017. FX interventions in Brazil: a synthetic control approach. J. Int. Econ. 108, 157–168. Chetty, R., Friedman, J.N., Saez, E., 2013. Using differences in knowledge across neighborhoods to uncover the impacts of the Eitc on earnings. Am. Econ. Rev. 103 (7), 2683–2721.

de Mel, S., McKenzie, D., Woodruff, C., 2019. Labor drops: experimental evidence on the return to additional labor in microenterprises. Am. Econ. J. Appl. Econ. 11 (1), 202–235. Ferman, B., & Pinto, C. (2018). Synthetic controls with imperfect pre-treatment fit, Working paper. Ferman, B., Pinto, C., & Possebom, V. (2018). Cherry picking with synthetic controls, Working paper. Freedman, M., 2013. Targeted business incentives and local labor markets. J. Hum. Resour. 48 (2), 311–344. Galasso, E., Ravallion, M., Salvia, A., 2004. Assisting the transition from workfare to work: a randomized experiment. Ind. Labor Relat. Rev. 58 (1), 128–142. Groh, M., Krishnan, N., McKenzie, D., Vishwanath, T., 2016. Do wage subsidies provide a stepping stone to employment for recent college graduates? Evidence from a randomized experiment in Jordan. Rev. Econ. Stat. 98 (3), 488–502. Ham, J.C., Swenson, C., Imrohoroğlu, A., Song, H., 2011. Government programs can improve local labor markets: evidence from state enterprise zones, federal empowerment zones and federal enterprise community. J. Public Econ. 95 (7), 779–797. Hamersma, S., 2005. The Work Opportunity and Welfare-to-Work Tax Credits. Urban Institute, Washington, D.C.. Hamersma, S., 2008. The effects of an employer subsidy on employment outcomes: a study of the work opportunity and welfare to work tax credits. J. Policy Anal. Manag. J. Assoc. Public Policy Anal. Manag. 27 (3), 498–520. Hanson, A., Rohlin, S., 2013. Do spatially targeted redevelopment programs spillover? Reg. Sci. Urban Econ. 43 (1), 86–100. Harvey, A., & Thiele, S. (2017). Co-integration and control: assessing the impact of events using time series data. University of Cambridge, Cambridge Working Paper Economics 1731. Katz, L.F., 1998. Wage subsidies for the disadvantaged. In: Generating jobs: How to Increase Demand for Less-Skilled Workers. Russell Sage Foundation, New York, NY, pp. 21–53. Kaul, A., Klößner, S., Pfeifer, G., & Schieler, M. (2018) Synthetic control methods: never use all pre-intervention outcomes together with covariates, Working paper. Levinsohn, J., Rankin, N., Roberts G., & Schöer, V. (2014). Wage subsidies and youth employment in South Africa: evidence from a randomized control trial. Stellenbosch Economic Working Paper 02/14. Lichter, A., Peichl, A., Siegloch, S., 2015. The own-wage elasticity of labor demand: a meta-regression analysis. Eur. Econ. Rev. 80, 94–119. McKenzie, D., 2017. How effective are active labor market policies in developing countries? A critical review of recent evidence. World Bank Res. Obs. 32 (2), 127– 154. Neumark, D., 2013. Spurring job creation in response to severe recessions: reconsidering hiring credits. J. Policy Anal. Manag. 1 (32), 142–171. Neumark, D., Grijalva, D., 2017. The employment effects of state hiring credits. Ind. Labor Relat. Rev. 70, 1111–1145. Neumark, D., Kolko, J., 2010. Do enterprise zones create jobs? Evidence from California’s enterprise zone program. J. Urban Econ. 68 (1), 1–19. Neumark, D., Simpson, H., 2015. Place-based policies. In: Handbook of Regional and Urban Economics, 5. Elsevier, pp. 1197–1287. O’Keefe, S., 2004. Job creation in California’s enterprise zones: a comparison using a propensity score matching model. J. Urban Econ. 55 (1), 131–150. Perloff, J.M., Wachter, M.L., 1979. The new jobs tax credit: an evaluation of the 1977-78 wage subsidy program. Am. Econ Rev. 69 (2), 173–179. Shim, S., Lee, S., 2014. Special employment rate promotion zone: the case of Tong-Yeong city. Korean J. Reg. Empl. 6 (2), 1–24. Wilder, M.G., Rubin, B.M., 1996. Rhetoric versus reality: a review of studies on state enterprise zone programs. J. Am. Plann. Assoc. 62 (4), 473–491. Yoon, Y., Shim, S., Oh, S., Lee, S., 2014. Employment Impact Assessment of Designation of Special Employment Promotion Zone: The Case of Tong-Yeong City. Korea Labor Institute.