Labor market impact of COVID‐19 on migrants in South Korea: Evidence from local outbreaks

Abstract

Despite the growing importance of migrants and their role in the South Korean economy, how much and in which ways COVID‐19, as an adverse labor market shock, has affected them has received too little attention, with no single study published to date yet. Motivated by such a paucity, this paper investigates the impact of the COVID‐19 pandemic crisis on employment for migrants in South Korea, with special emphasis on quantifying the magnitude of its causal effect. In doing so, this study exploits the unique fact that only one specific region in South Korea had a substantial number of COVID‐19 infections in the early stage of the pandemic so that estimations can be made using a difference‐in‐differences (DD) model. The DD estimates suggest that COVID‐19 lowered migrants’ overall employment probability by 2.5 to 3.2 percent points. However, strong heterogeneity between the genders is apparent: the pandemic severely hurts female migrants’ employment, with male migrants weathering it relatively unscathed. Furthermore, female migrants seem considerably harder hit than female host populations. Heterogeneity analyses reveal that (i) a duration of stay exceeding 5 years and (ii) fluency in Korean (as a local language) protect migrants from being heavily affected by the COVID‐driven employment shock.

1. INTRODUCTION

By mid‐February 2020, South Korea, a densely populated nation, faced the highest count of new COVID‐19 infections globally after China, where the virus was first reported. Despite a resolute response by the South Korean Government, including a comprehensive border quarantine system and a 270 trillion KRW (approximately 228 billion USD) immediate fiscal stimulus, the pandemic has taken a huge toll on its economy and labor market. The outbreak forced the government to close the nation's borders and scared consumers inside and away from stores. Businesses shut down, overseas sales dried up and inbound travelers stopped coming. As of May 2020, exports from the South Korean economy, which is heavily reliant on international trade, fell by nearly half year‐on‐year as its important trade partners, such as China and the United States, were hard hit by the pandemic.

Many empirical studies on COVID‐induced employment losses, such as Cortes and Forsythe (2020) and Couch et al. (2020), reveal the unequal labor market impacts for various socio‐demographic groups. Migrants are among the most vulnerable of all workers, as they have limited means to respond. Compared to their host populations, they face several disadvantages in the labor markets of their destinations from the outset, including a lack of proficiency in the local language, cultural differences, career interruptions, undervalued previous work experience and unappreciated educational qualifications (Shin, 2021). They are also more likely to have non‐standard or informal contracts, shorter job tenure and lower seniority on the job: moreover, they usually work in less‐skilled occupations in comparison to measurably similar host populations (OECD, 2020). As such, Dustmann et al. (2010) find that migrants experience significantly larger unemployment responses to economic shocks relative to natives within the same skill group, even before the COVID‐19 outbreak. Thus, it is a logical inference that migrant workers are facing manifold setbacks amid the pandemic and are highly vulnerable to the initial economic shock induced by COVID‐19. Based on the Current Population Survey Basic Monthly Files, Borjas and Cassidy (2020) find empirical evidence that the employment decline in the United States stemming from COVID‐19 is particularly severe for migrants. They argue that such pronounced vulnerability should be attributed to the fact that migrants are less likely to work in jobs that can be performed remotely; as a result, they suffer outsized employment consequences as the lockdown measures permit those with skills more conducive to remote work to continue working from home. Fasani and Mazza (2020) also note that many migrant workers hold occupations not amenable to telework. To make matters worse, studies suggest that discrimination (against marginalized populations) tends to increase considerably in times of a slack labor market (OECD, 2020).

Despite abundant evidence suggesting that hardships multiply for migrants, their dire situation and the impact of the COVID‐19 pandemic crisis on their employment in local labor markets has received far too little attention up until now. This is especially true in the case of South Korea, where no study, to the best of the author's knowledge, has covered this topic. According to the OECD (2019), the share of migrants among the nation's economically active population has increased sharply over the past decade, and their labor market contribution is further emphasized by the fact that they are primarily employed by small and mid‐size enterprises (SMEs), for which the labor shortage has been a chronic, pivotal problem.1 In South Korea, approximately one in 10 employers with more than five employees rely on filling at least some vacancies with foreign workers (OECD, 2019). Furthermore, migrants comprise approximately 9 percent of the ‘root industry’ workforce in the nation (OECD, 2019).2 Considering the growing importance of migrants and their roles in South Korean society, as well as the aforementioned research gap, this paper explores the impact of the COVID‐19 pandemic crisis on employment for migrants in South Korea, with special emphasis on quantifying the magnitude of the impact.3 Understanding how much and in which way the pandemic‐induced employment shock has affected them is a critical element of any informed strategy for coping with the pandemic and devising proper recovery plans.

In addition to the fact that this study, to the best of the author's knowledge, is the first empirical investigation of the COVID‐19 pandemic's labor market impact on migrants in South Korea, the current paper contributes to the existing literature by also estimating the causal effect of the initial COVID‐19 shock. To this end, we exploit the unique fact that only one specific region in South Korea—Daegu‐Gyeongbuk—had a significant number of COVID‐19 infections in the early stage of the pandemic and use a difference‐in‐differences (DD) approach in this study. Such an ideal setup for causal estimations is uniquely feasible in the South Korean case, also in part because the relevant data were collected nationwide immediately following the ‘local’ outbreaks in Daegu‐Gyeongbuk. Exogenous regional variations (i.e., COVID‐affected vs unaffected) remained distinct when the data were collected in May 2020. Moreover, unlike most other countries, the government did not mandate nationwide lockdown measures in South Korea.4 In this respect, the regions unaffected by COVID‐19 (at the time of data collection) were free from both the pandemic per se and any strict containment measures, except for a voluntary social distancing campaign. As the main objective of this study is to estimate the causal effect of the COVID‐19 pandemic crisis on employment, the absence of mandatory lockdown measures is beneficial because the estimates are more likely to capture the sheer effect of the pandemic exclusively, with contamination from other factors (e.g., policy interventions) that coincide with the pandemic being less likely.

The remainder of the current paper is organized as follows. Section II presents the data used for this investigation and provides background information on the empirical setup, while Section III discusses the econometric methods used for the research question. Section IV presents the main estimation results of this study, and Section V explores whether migrants or host populations are more vulnerable to COVID‐19 as a labor market shock. Section VI discusses whether the effect of COVID‐19 on migrants’ employment is heterogeneous, and if so, what drives effect heterogeneity. Section VII concludes the paper.

2. DATA AND EMPIRICAL CONTEXT

2.1. Data

This study uses a unique micro‐level dataset for its empirical investigation: the survey(s) of migrants’ Living Conditions and Labor Force (hereafter ‘the LCLF data’). The governmental organization Statistics Korea, in collaboration with the Ministry of Justice, has collected the cross‐sectional data on an annual basis since 2017. According to Statistics Korea, the LCLF data, as a national statistical survey mandated by Articles 17 and 18 of the Statistics Act, aim to provide basic information needed for making migrant‐related policies by identifying the socioeconomic status of non‐native residents (Statistics Korea, n.d.). The LCLF collects data on foreigners and naturalized citizens who are staying in South Korea for longer than 90 days and who are 15 years old or older. As of the year 2020, the size of the target population is approximately 25 000 persons. Under the guidance of the Statistics Korea Employment Statistics Division, micro‐level data are collected by in‐person, face‐to‐face interviews with individuals as the survey unit. Interviewers are trained specifically for the LCLF survey based on clearly stipulated data collection protocols. The data contain high‐dimensional individual variables of migrants, such as basic demographic information (e.g., age, gender), household size, labor market outcomes, place of residence, education level, Korean language proficiency, length of residence in South Korea, home country and religion. Table 8 in Appendix S1, Section A.4 summarizes important sample statistics of the data.

2.2. Empirical context

The first confirmed case of COVID‐19 in a South Korean territory was reported in a Chinese citizen on January 20, 2020.5 With the country's geographic proximity to Wuhan, the first epicenter of COVID‐19, the South Korean Government responded immediately and resolutely to the imminent risk of a pandemic by implementing diagnostic testing, contact tracing, isolating confirmed and suspected cases, providing treatment and encouraging social distancing (World Bank, 2020). International organizations, such as the World Bank (2020) and the World Health Organization (WHO, 2020), found the measures taken by South Korea very effective and thus regarded the country as an exemplary case from which other countries could learn. For an academic discussion of this topic, see Lee et al. (2020). Nevertheless, by February–March 2020, South Korea faced the second‐highest count of new infections worldwide, after China. Due to a unique set of circumstances, however, the rising caseload was ‘local’, centered in Daegu, a city of approximately 2.5 million residents in the country's southeast, and the area of its neighboring province Gyeongbuk (hereafter abbreviated as ‘DG’). Figure 1 illustrates the extraordinary ‘local outbreak’ of COVID‐19 in South Korea.

FIGURE 1.

Local outbreaks in South Korea. The dashed lines refer to the Living Conditions and Labor Force (LCLF) data collection period in the year 2020. Source: Statistics Korea, European Center for Disease Prevention and Control.

The Korea Disease Control and Prevention Agency's investigation traced the sudden spread of COVID‐19 in the DG area to the Shincheonji Church, where a super spreader congregant, known as ‘Patient No. 31’, is believed to have infected many other worshipers during religious services. Shincheonji is a religious sect known for its distinctive practices such as secrecy, the banning of face masks and praying loudly in close proximity to one another: customs that are believed to have promoted the spread of the virus among its congregants.

By February 29, the number of confirmed cases in South Korea had exploded to 3150, of which 2724 cases were in the DG area alone (Aum et al., 2021).6 On March 7, the Korea Disease Control and Prevention Agency officially reported that 63.5 percent of the country's confirmed COVID‐19 cases could be traced to the Shincheonji Church. Soon after, the number of confirmed cases surged past 6000, most of which were reported in the DG area, making it the country's single largest cluster of COVID‐19 infections. As a result, the South Korean government designated the DG area as a special quarantine management zone, where restrictions on public gatherings were stricter. In sharp contrast, the number of confirmed cases remained very low in regions other than the DG and Seoul capital areas. This facet makes it possible to use distinct regional variations in terms of COVID‐19 outbreaks for causal analyses.

This paper exploits the unique fact that only one specific region in South Korea had a significant number of COVID‐19 infections in the early stage of the pandemic so that the research question can be analyzed using a DD model. To use the DD setup in a reliable, persuasive way, local transmission should be an exogenous shock. In other words, regional variations in COVID‐19 cases should have nothing to do with socioeconomic conditions affecting employment, which is an outcome variable in this study. In this regard, recall that (i) the super spreader ‘Patient No. 31’, a Shincheonji congregant, initiated the local outburst of COVID‐19 confirmed cases and that (ii) Shincheonji's distinctive religious practices (e.g., praying and singing loudly in close proximity to one another) provided a fertile basis for the virus to spread rapidly. Whereas the former factor can be considered random, the latter factor requires careful consideration of Shincheonji as a religious sect. For detailed explanations, see Appendix S1, Section A.1.

3. ECONOMETRIC METHODS

As mentioned in Section I, the primary aim of this study is to estimate the causal effect of the COVID‐19 pandemic crisis on migrants’ employment. In doing so, it exploits the unique fact that only one specific region in South Korea had a significant number of COVID‐19 infections by the first half of the year 2020. Such extraordinary ‘local transmission’ with distinct regional variations makes it possible to use a DD model for inferring causal effects. In addition, as already pointed out in Section I, there were no government‐mandated lockdown measures in South Korea. Thus, the regions unaffected by COVID‐19 were free from both the pandemic outbreak per se and any mandatory containment measures. If the government had imposed strict lockdown measures throughout the country regardless of regional differences in confirmed COVID‐19 cases, then regions with few or no COVID‐19 cases could not easily function as a well‐controlled comparison group. In this regard, the absence of nationwide mandatory lockdown measures is beneficial for drawing causal inferences.7

3.1. Difference‐in‐differences with repeated cross‐sections

The DD estimator is a popular tool in applied economic research for evaluating the (causal) effects of policy interventions or other treatments of interest on relevant outcome variables (Abadie, 2005). However, despite its enduring popularity, the DD estimator is sometimes misused and overtrusted based on scant understanding. One of the most common pitfalls concerns how to use DD with repeated cross‐sections, which requires additional assumptions—on top of basic ones needed for panel cases. As the present study draws on the repeated cross‐sectional data, as discussed in Section II.1, this section briefly explains those additional assumptions that are important to note in applying DD to repeated cross‐sections.

Suppose there are two regions: Region 1 (treatment group) is exposed to a treatment (i.e., COVID‐19), whereas Region 0 (control group) is never treated at all (during the same study period). In the context of this study, the DG area is coded ‘Region 1: Treatment group’ because it was heavily affected by COVID‐19 in the first half of 2020, while the control area is coded ‘Region 0: Control group’ (elaborated further in Section III.2). Suppose that $a$ and $b$ (with $a<b$) denote two time periods: before and after the COVID‐19 outbreak, respectively. If migrant $i$ lives in Region 1 at time $b$, they are exposed to the economic shock of COVID‐19 and affected by the treatment:

$$d_{i,t,\text{region}}^{\text{COVID}-19}=\mathbf{1}\left[\mathbf{1}\left[t=b\right]\times \mathbf{1}\left[i\text{lives in the}\mathrm{DG}\text{area}\left(\mathrm{at}\text{time}b\right)\right]=1\right].$$ (1)

Based on this setup, three potential responses can be assumed: namely, $y_{0\mathit{ia}}$, $y_{0\mathit{ib}}$ and $y_{1\mathit{ib}}$. The subscript (i.e., 0 or 1) denotes whether migrant $i$ is treated ($d=1$) or not ($d=0$). Furthermore, each $i$ has $\left(r_{\mathit{ia}}\in \left\{0,1\right\},r_{\mathit{ib}}\in \left\{0,1\right\}\right)$, which indicates whether $i$ lives in Region 1 or not at each time $\left\{a,b\right\}$. ${\tau }_i=\mathbf{1}\left[i\text{is observed}\mathrm{at}b\right]$ is also supposed.

While what we can conceptually consider (and what we can observe from panel data) is

$$\left\{y_{0\mathit{ia}},y_{0\mathit{ib}},y_{1\mathit{ib}},r_{\mathit{ia}},r_{\mathit{ib}},{\tau }_i\right\},$$ (2)

what we can observe from repeated cross‐sections is

$$\left\{y_i,r_i,{\tau }_i\right\}.$$ (3)

This is the point at which it becomes necessary to impose additional assumptions. In Equation (3), where $i$ lives (i.e., $r_i$) is observed as the result of

$$r_i=\left(1-{\tau }_i\right)r_{\mathit{ia}}+{\tau }_i{r}_{\mathit{ib}}.$$ (4)

Likewise, the outcome $y_i$ is observed as the result of

$$y_i=\left(1-{\tau }_i\right)y_{0\mathrm{i}a}+\left\{{\tau }_i\left(1-r_i\right)y_{0\mathit{ib}}+{\tau }_i{r}_i{y}_{1\mathit{ib}}\right\}.$$ (5)

The subscript $i$ is suppressed hereafter for the sake of notational simplicity, and, in an attempt to control for compositional changes, the inclusion of observable covariates $\mathbf{x}$ is assumed.8 Applying DD to repeated cross‐sections requires two (additional) conditions as follows. First,

$$\tau \perp y_{0a},y_{0b},y_{1b}\mid r_t,\mathbf{x},t\in \left\{a,b\right\}$$ (6)

is needed. Equation (6) requires that, as long as $r_t$ (for $t\in \left\{a,b\right\}$) and $\mathbf{x}$ are conditioned on, $i$'s potential responses do not affect whether $i$ is observed at $b$ in such a way that $\mathbb{E}\left(\tau |y_{0a},y_{0b},y_{1b}\right)=\mathbb{E}\left(\tau \right)$ holds (i.e., mean independence). Second,

$$\mathbb{E}\left(y_{0a}\mid r_a=c,\mathbf{x}\right)=\mathbb{E}\left(y_{0a}\mid r_b=c,\mathbf{x}\right),c\in \left\{0,1\right\}$$ (7)

should be satisfied. In words, Equation (7) means that the average of untreated, baseline $y_{0a}$ should be the same as long as a region is specified with $\mathbf{x}$ being conditioned on. Applied to the context of this study, Equation (7) requires that those living in the DG area before the COVID‐19 outbreak are similar to those residing in the DG area after the COVID‐19 outbreak in terms of their untreated, baseline employment status. If Equation (7) holds, the possibility of any endogenous, systematic moves across regions can be ruled out, and the composition of the treated and untreated groups is assumed to be stable across $a$ and $b$ (Cameron & Trivedi, 2005).

In the given context, Equation (7) is thought to hold for the following reasons. First, there were no crucial events during the study period, other than the COVID‐19 outbreak, that were big enough to change the composition of $i$ in each region. Second, as mentioned above and portrayed in Figure 1, the local transmission of COVID‐19 in the DG area occurred explosively (over a short period of time) in (late) February–March, and the LCLF data were collected in May shortly after the localized spread. The period between March and May can be considered too short for substantial interregional moves. Third, the sample individuals of the LCLF data are migrants to South Korea, for whom it is even harder to make interregional moves than the host population, especially amid the pandemic.9 Notwithstanding, a caveat is needed in that the plausibility of Equations (6) and (7) is not directly testable.

3.2. Defining treatment and control groups

In using the DD model, defining the treatment and control groups is critically important. In this study, the massive localized outbreak in the DG area makes it the clear choice for the treatment group. In contrast, a control group must be chosen with great care from among the many other regions not seriously affected by COVID‐19.

This study uses the Gwangju‐Jeonnam‐Jeonbuk area (hereafter abbreviated as ‘GJ’) as a control region for several reasons. First, the number of COVID‐19 confirmed cases in the GJ area, as of the LCLF data collection period in May 2020, was extremely low: only 63 cumulative confirmed cases out of 5 121 575 inhabitants.10 The GJ area was considered a ‘COVID‐free’ region until the second wave of COVID‐19 infections hit the country in June; see Figures 1 and 2. Second, the DG and GJ areas are not directly adjacent, and the borderline region they share is only sparsely populated, as Figure 3 shows. Hence, the two regions can be deemed geographically well‐separated.

FIGURE 2.

The cumulative number of COVID‐19 confirmed cases, as of 1 May 2020

FIGURE 3.

Population density (treatment vs. control regions), as of 1 May 2020

Third, according to the LCLF data, the DG area and the GJ area are similar in terms of industries in which migrants are employed. In both regions, the mining and manufacturing sector is the largest employer of local migrant workers; the wholesale–retail and food–accommodation sector hires the second‐largest share. Such similar industrial composition is beneficial for the DD setup because it makes the expectation more plausible that, in the absence of COVID‐19, the employment dynamics of migrant workers in the DG area would have been similar to those of migrant workers in the GJ area.

More formally, recall that, in the conventional DD setup, observable covariates $\mathbf{x}$ and unobservable confounder $\epsilon$, as well as a treatment variable $d$, can vary over time. In estimating the effect of $d$, ‘before–after’ changes in $\mathbf{x}$ do not seriously matter because observable variables can be (linearly) controlled for. However, the possibility of changes in unobservable $\epsilon$, which cannot be directly tested, should be taken into careful consideration. If treatment and control groups are different in terms of their temporal changes in $\epsilon$, what is captured by the coefficient estimate of $d$ cannot be solely attributed to the treatment itself. Thus, the treatment and control groups must experience the same temporal changes in $\epsilon$, and two groups’ similar sectoral composition is advantageous in this regard so that unobservable $\epsilon$ can be removed by differencing steps.

Based on the aforementioned aspects, the treatment and control groups are defined at a regional level as follows:

$$\left\{\begin{array}{ll}\text{Treatmentgroup}:& \text{Daegu}-\text{Gyeongbuk}\left(\text{the}\mathrm{DG}\text{area}\right)\\ \text{Controlgroup}:& \text{Gwangju}-\text{Jeonnam}-\text{Jeonbuk}\left(\text{the}\mathrm{GJ}\text{area}\right)\end{array}\right..$$ (8)

3.3. Model specification

In the DD framework, a ‘treatment‐on’ indicator $d$, defined in Equation (1), appears in such a way that the treatment group is treated at some point in time, whereas the control group is never treated (Lee, 2016). This is distinguished from the simpler ‘before–after’ approach in the sense that DD uses the difference of the two group‐wise ‘before–after’ differences (Lee, 2016). As a result, any changes stemming from temporal effects that are common to both groups can be thinned out.

Based on what is discussed in Sections III.1 and III.2, the most basic specification that this study estimates takes the following form, with a migrant worker's employment status as a binary outcome variable $y\in \left\{0,1\right\}$:

$$\begin{array}{ll}& y_{i,t\in \left\{2017,2018,2019,2020\right\}}^{\text{Region}}\\ & ={\beta }_1+{{\mathbf{x}}^{\prime }}_{i,t}{\mathit{\beta }}_{\mathbf{x}}+\mathbf{1}\left[t\ge 2020\right]·{\beta }_{\tau }+\mathbf{1}\left[\text{Region}=\mathrm{DG}\right]·{\beta }_r\\ & +\mathbf{1}\left[\mathbf{1}\left[t\ge 2020\right]\times \mathbf{1}\left[\text{Region}=\mathrm{DG}\right]=1\right]·{\beta }_d+{\epsilon }_{i,t}\end{array}.$$ (9)

In Equation (9), ${\beta }_{\tau }$ captures the effect of $\mathbf{1}\left[t\ge 2020\right]$ (i.e., the ‘after’ period, $b$), common across regions, for all observations. If changes (other than the COVID‐19 outbreak) occur in the year 2020 that commonly affect all, ${\beta }_{\tau }$, as a single term, picks up their effects. In contrast, ${\beta }_r$ captures the intercept shift in $y$ for those who live in the DG area. Most importantly, ${\beta }_d$ is the key parameter of interest: it picks up the interregion difference in region‐specific ‘before–after’ differences in $y$. For ${\beta }_d$ to be reliably informative as a causality parameter,

$$\mathbb{E}\left(y_{d=0,b}-y_{d=0,a}\mid \text{Region}=\mathrm{DG},\mathbf{x}\right)=\mathbb{E}\left(y_{d=0,b}-y_{d=0,a}\mid \text{Region}=\mathrm{GJ},\mathbf{x}\right)$$ (10)

should be satisfied, which is often called the ‘same time‐effect condition’ or the ‘parallel trend condition.’ In plain terms, Equation (10) means that, in the absence of treatment $d$, $y$ would have followed the same trend in both the treated (DG area) and untreated (GJ area) regions. The plausibility of Equation (10) can be tested by using more than two time periods; see Section IV.2 for the test results in the given context.

3.4. Nonlinear dependent variable

The dependent variable of the current study is not continuous but binary. This aspect, in the context of using the DD approach, necessitates further econometric considerations because the conventional DD model (and its differencing logic) works under the additivity of different effects (Lee, 2016). For the sake of brevity, detailed explanations are provided in Appendix S1, Section A.2. A robustness check in Appendix S1, Section A.3 addresses this facet.

4. ESTIMATION RESULTS

4.1. Descriptive evidence

Before proceeding to regression‐based analyses, we compare temporal changes in unemployment rates by region based on simple, descriptive statistics: see Figures 4 and 5 that serve as initial evidence. While descriptive statistics for migrants (Figure 4) are calculated directly from the LCLF data, those for host populations (Figure 5) are computed from the Economically Active Population Survey(s), which targets general populations in South Korea and collects sampled individuals’ labor market outcomes on a monthly basis. For the sake of comparison, the corresponding statistics for the OECD average are presented in parallel,11 and each group's unemployment rate in the year 2019 is set to zero as a baseline value. As portrayed in Figure 4, the treatment region (denoted by a solid black line) experienced a sharp, unprecedented increase in migrants’ unemployment in 2020, in contrast to the control region (denoted by a gray dashed line) encountering only a moderate change. In contrast, in the case of host populations, Figure 5 suggests that both regions remained relatively unscathed, with the post‐pandemic increase in unemployment rates therein being far below the OECD average. This descriptive evidence forms the ground for hypothesizing that the impact of COVID‐19's regional outbreaks was indeed ‘local’ in two respects: (i) being specific to the treatment region and (ii) selectively hurting migrants, a distinct group of high vulnerability.

FIGURE 4.

Changes in unemployment rates—migrants

FIGURE 5.

Changes in unemployment rates—host populations

4.2. Testing the same time‐effect condition

Before discussing this study's main estimation results, it should be checked whether the same time‐effect condition (10) holds, which is a necessary condition for using DD. To check the plausibility of Equation (10), the following model is applied to the pre‐pandemic period that spans from 2017 to 2019:

$$\begin{array}{ll}& y_{i,t\in \left\{2017,2018,2019\right\}}^{\text{Region}}\\ & ={\beta }_1+{{\mathbf{x}}^{\prime }}_{i,t}{\mathit{\beta }}_{\mathbf{x}}+\sum _{j=2017}^{2018}{\beta }_{\tau }^j·\mathbf{1}\left[t=j\right]+\mathbf{1}\left[\text{Region}=\mathrm{DG}\right]·{\beta }_r\\ & +\sum _{j=2017}^{2018}{\beta }_q^j·\mathbf{1}\left[\mathbf{1}\left[t=j\right]\times \mathbf{1}\left[\text{Region}=\mathrm{DG}\right]=1\right]+{\epsilon }_{i,t}.\end{array}$$ (11)

Observable covariates $\mathbf{x}$ include a wide range of individual characteristics, such as age, gender, education level, the country in which $i$ obtained their final education degree, how long $i$ has lived in South Korea, religion, home country, naturalization, parents’ nationality, and household size. The inclusion of $\mathbf{x}$ is to control compositional changes in terms of measurable characteristics.

Note that pre‐pandemic temporal changes are captured by year‐specific dummies in Equation (11). All of the other notations are the same as defined above. While researchers often run Equation (11) based on all available time periods (i.e., including a post‐treatment period[s]) to check the parallel trend condition, it should be noted that we cannot rule out the possibility of COVID‐19 affecting covariates $\mathbf{x}$. Thus, a more accurate and rigorous approach for testing the necessary condition (10) is to use time periods strictly free from the pandemic; for this reason, the last year completely unaffected by COVID‐19 (i.e., 2019) is used as a baseline for Equation (11). Based on this, ${\widehat{\beta }}_q^j$ for $j\in \left\{2017,2018\right\}$ are both marginally and jointly tested.

The results are presented in the form of confidence intervals in Figure 13 in Appendix S1, Section A.4. The marginal null hypothesis of ${\beta }_q^j=0$ cannot be rejected for $j\in \left\{2017,2018\right\}$. In addition, the joint test fails to reject the null hypothesis of ${\beta }_q^{2017}={\beta }_q^{2018}=0$: its $F$‐test $p$‐value is $0.4079$. Based on these results, it may be said that, in the absence of COVID‐19, the employment dynamics in the DG area would have been similar to those in the GJ area, and interregional differences in time‐driven $y$ differences would have been zero, which is comforting for further analyses.12 The same time‐effect condition holds when male and female subgroups are analyzed separately.

4.3. Two‐period difference‐in‐differences

This section, as a point of departure, starts with the estimates from a two‐period DD model that spans from 2019 (i.e., the ‘before’ period denoted by $a$) to 2020 (i.e., the ‘after’ period denoted by $b$). Recall that the LCLF data were collected in May of each year, which means that the $a$‐period is completely unaffected by COVID‐19, while the $b$‐period, only in the case of the treatment region, is in the middle of the pandemic. This two‐period model can be regarded as the most conservative approach in that a shorter time window reduces the likelihood that additional confounding shocks other than COVID‐19 happen between $a$ and $b$. The following specification is applied to the period that spans from 2019 to 2020:

$$\begin{array}{ll}& y_{i,t\in \left\{2019,2020\right\}}^{\text{Region}}\\ & ={\beta }_1+{{\mathbf{x}}^{\prime }}_{i,t}{\mathit{\beta }}_{\mathbf{x}}+\mathbf{1}\left[t\ge 2020\right]·{\beta }_{\tau }+\mathbf{1}\left[\text{Region}=\mathrm{DG}\right]·{\beta }_r\\ & +\mathbf{1}\left[\mathbf{1}\left[t\ge 2020\right]\times \mathbf{1}\left[\text{Region}=\mathrm{DG}\right]=1\right]·{\beta }_d+{\epsilon }_{i,t}.\end{array}$$ (12)

Its results are summarized in Table 1.

TABLE 1.

Difference‐in‐differences (DD) estimates: Two periods

Based on specification (12)	[1] Both genders	[2] Males	[3] Females
βd (DD)	−0.0252** [0.0120]	−0.0113 [0.0125]	−0.0488** [0.0242]
Periods	2019–2020	2019–2020	2019–2020
Number of observations	10 204	6135	4069

Note: Robust standard errors are in brackets.

^*** p < 0.01;

^** p < 0.05;

^* p < 0.1.

As can be seen in Column 1 in Table 1, the parameter of interest ${\beta }_d$ is estimated to be negative with statistical significance. ${\widehat{\beta }}_d=-0.0252$ means that COVID‐19 lowers migrants’ employment probability by 2.52 percentage points. This result is in accordance with the (non‐causal) finding of Borjas and Cassidy (2020) that the COVID‐19 pandemic has a substantial effect on migrant workers’ employment in the United States.13

Meanwhile, in most countries, the COVID‐19 pandemic has affected women's employment much more negatively than a typical recession by striking a severe blow to sectors in which women are highly represented; South Korea is no exception in this regard. Many empirical studies, such Alon et al. (2020) and Blundell et al. (2020), point to the existence of the gendered impacts of the pandemic crisis and corroborate that COVID‐19 has hit females harder than males in terms of employment. In light of this aspect, the two‐period DD specification (Equation 12) is applied separately to each gender subgroup as in Borjas and Cassidy (2020), and Columns 2 and 3 in Table 1 present the results for males and females, respectively. In the case of male migrants, the null hypothesis of ${\beta }_d=0$ cannot be rejected, which implies that the negative labor market impact stemming from COVID‐19 lacks statistical evidence. In stark contrast, in the case of female migrants, ${\beta }_d$ is estimated to be substantially negative with statistical significance. Thus, it can be argued that the pandemic indeed hit female migrants’ employment harder and that a migrant's gender is a critical driver of the differing levels of vulnerability to the initial COVID‐19 shock to the labor market.

To compare ${\beta }_d^{\text{Male}}$ and ${\beta }_d^{\text{Female}}$ in a more authentic way, a female dummy is interacted with other terms in Equation (12)—in the spirit of the triple difference estimator—without splitting the sample into gender‐wise subgroups. By doing so, ${\beta }_d^{\text{Male}}={\beta }_d^{\text{Female}}$ can be directly tested in a single model. The results hardly differ from those presented in Table 1. For males, the labor market impact of COVID‐19 is not statistically significant (coefficient estimate: $-0.0125$, $p$‐value: $0.317$). In sharp contrast, the negative impact of COVID‐19 on females is evident with substantial magnitude (coefficient estimate: $-0.0501$, $p$‐value: $0.042$).

4.4. Multiple‐period DD

As a next step, all available periods, spanning from 2017 to 2020, are exploited for estimating the causal effect of COVID‐19 on migrants’ employment. The most basic all‐period specification for DD takes the following form:

$$\begin{array}{ll}& y_{i,t\in \left\{2017,2018,2019,2020\right\}}^{\text{Region}}\\ & ={\beta }_1+{{\mathbf{x}}^{\prime }}_{i,t}{\mathit{\beta }}_{\mathbf{x}}+\sum _{j=2018}^{2020}{\beta }_{\tau }^j·\mathbf{1}\left[t=j\right]+\mathbf{1}\left[\text{Region}=\mathrm{DG}\right]·{\beta }_r\\ & +\mathbf{1}\left[\mathbf{1}\left[t\ge 2020\right]\times \mathbf{1}\left[\text{Region}=\mathrm{DG}\right]=1\right]·{\beta }_d+{\epsilon }_{i,t},\end{array}$$ (13)

where the key parameter of interest is ${\beta }_d$. Note that in Equation (13), the year 2017 is used as a baseline reference. The results, summarized in Table 2, are in accordance with those presented earlier in Table 1. Comparing the results in Tables 1 and 2 reveals the clear advantage of using multiple periods: the larger number of observations results in higher efficiency and stronger statistical significance. For the whole sample (i.e., Column 1) and the female subgroup (i.e., Column 3), ${\beta }_d$ is estimated to be negative with statistical significance even at the 1‐percent level. In contrast, as discussed in Table 1 already, ${\widehat{\beta }}_d$ in Column 2 suggests that male migrants seem to have weathered the storm of the COVID‐19 pandemic relatively unscathed. The gendered impacts of COVID‐19 are also present when tested in a single model without splitting the sample into male and female subgroups. For males, the labor market impact of COVID‐19 is not statistically significant (coefficient estimate: $-0.0127$, $p$‐value: $0.310$). The impact of COVID‐19 is clearly negative for females, however (coefficient estimate: $-0.0506$, $p$‐value: $0.039$).

TABLE 2.

Difference‐in‐differences (DD) estimates: Multiple periods

Based on specification (13)	[1] Both genders	[2] Males	[3] Females
βd (DD)	−0.0264*** [0.0101]	−0.0080 [0.0105]	−0.0591*** [0.0205]
Periods	2017–2020	2017–2020	2017–2020
Number of observations	17 441	10 149	7292

Note: Robust standard errors are in brackets.

^*** p < 0.01;

^** p < 0.05;

^* p < 0.1.

4.5. Multiple‐period difference‐in‐difference: With region‐specific time trends

If there are multiple periods available, DD can be implemented through diverse model specifications with varying degrees of robustness. Seminal empirical studies, such as Wolfers (2006), emphasize the importance of carefully modeling group‐specific time trends in DD analyses. In the context of this study, such analyses can be executed in the following manner:

$$\begin{array}{ll}& y_{i,t\in \left\{2017,2018,2019,2020\right\}}^{\text{Region}}\\ & ={\beta }_1+{{\mathbf{x}}^{\prime }}_{i,t}{\mathit{\beta }}_{\mathbf{x}}+\sum _{j=2018}^{2020}{\beta }_{\tau }^j·\mathbf{1}\left[t=j\right]+\mathbf{1}\left[\text{Region}=\mathrm{DG}\right]·{\beta }_r\\ & +\mathbf{1}\left[\text{Region}=\mathrm{DG}\right]·\lambda \left(t\right)·{\beta }_{\lambda }+\mathbf{1}\left[\mathbf{1}\left[t\ge 2020\right]\times \mathbf{1}\left[\text{Region}=\mathrm{DG}\right]=1\right]·{\beta }_d+{\epsilon }_{i,t}.\end{array}$$ (14)

Note the inclusion of a new interaction term $\mathbf{1}\left[\text{Region}=\mathrm{DG}\right]·\lambda \left(t\right)$, where $\lambda \left(t\right)$ refers to a linear time trend. In Equation (14), while year‐specific fixed effects that are common to all regions are captured by ${\beta }_{\tau }^j$ for $j\in \left\{\mathrm{2018,2019,2020}\right\}$, ${\beta }_{\lambda }$ allows for the possibility that treatment and control regions follow different trends over time in a limited but potentially reasonable way (Angrist & Pischke, 2008).14 The results, presented in Table 3, lead to the same conclusion, which is reassuring. COVID‐19 seems to have lowered the employment probability of migrants by 3.16 percentage points (Column 1). As above, strong gender‐wise heterogeneity is apparent: the crisis severely hurts female migrants’ employment (Column 3), whereas the negative labor market impact stemming from COVID‐19 does not clearly register for male migrants (Column 2).15

TABLE 3.

Difference‐in‐differences (DD) estimates: Multiple periods

Based on specification (14)	[1] Both genders	[2] Males	[3] Females
βd (DD)	−0.0316** [0.0153]	−0.0123 [0.0161]	−0.0612** [0.0303]
Periods	2017–2020	2017–2020	2017–2020
Number of observations	17 441	10 149	7292

Note: Robust standard errors are in brackets.

^*** p < 0.01;

^** p < 0.05;

^* p < 0.1.

As a more robust approach, region‐specific time trends can be captured by region‐year dummies in place of $\lambda \left(t\right)$. Accordingly, the following specification is applied:

$$y_{i,t\in \left\{2017,2018,2019,2020\right\}}^{\text{Region}}={\beta }_1+{\mathbf{x}}_{i,t}{\mathit{\beta }}_{\mathbf{x}}+\sum _{\begin{array}{l}j=2017\\ j\ne 2019\end{array}}^{2020}{\beta }_{\tau }^j\cdot \mathbf{1}\left[t=j\right]+\mathbf{1}\left[\text{Region}=\mathrm{DG}\right]\cdot {\beta }_r+\sum _{\begin{array}{l}j=2017\\ j\ne 2019\end{array}}^{2020}{\beta }_d^j\cdot \mathbf{1}\left[\mathbf{1}\left[t=j\right]\times \mathbf{1}\left[\text{Region}=\mathrm{DG}\right]=1\right]+{\epsilon }_{i,t},$$ (15)

with the year 2019 as a baseline reference.16 Note that, in Equation (15), the key parameter of interest is ${\beta }_d^{2020}$ because $d$ is switched on only in 2020, as defined in Equation (1). The results from Equation (15) are presented in Figure 6 and Table 4. In particular, Figure 6 provides strong visual evidence of treatment and control regions with a common (pre‐pandemic) underlying trend and of the pandemic effect that induces a sharp deviation from such a (pre‐pandemic) parallel trend. Despite including region‐year dummies for a more robust estimation, the results lead to the same conclusion as follows, which is comforting. It seems that COVID‐19 lowered the employment probability of migrants by 2.55 percentage points (Column 1). Clear gender‐wise heterogeneity is found: female migrants’ employment is seriously impaired by the pandemic crisis (Column 3), whereas there is no clear negative effect related to COVID‐19 for male migrants (Column 2). Gendered impacts of COVID‐19 are also found when tested in a single model without splitting the sample into male versus female subgroups. The labor market impact of COVID‐19 lacks statistical significance for males (coefficient estimate: $-0.0127$, $p$‐value: $0.310$), yet females experience a substantial negative impact of COVID‐19 (coefficient estimate: $-0.0506$, $p$‐value: $0.039$).

FIGURE 6.

Multiple periods difference‐in‐differences estimates. Source: Surveys on Immigrants’ Living Conditions and Labor Force, South Korea. Estimation by the author.

TABLE 4.

Difference‐in‐differences (DD) estimates: Multiple periods

Based on specification (15)	[1] Both genders	[2] Males	[3] Females
β2020 (DD)	−0.0255** [0.0120]	−0.0110 [0.0125]	−0.0506** [0.0242]
Periods	2017–2020	2017–2020	2017–2020
Number of observations	17 441	10 149	7292

Note: Robust standard errors are in brackets.

^*** p < 0.01;

^** p < 0.05;

^* p < 0.1.

4.6. Discussion of the main findings

A critical question should be raised here: what makes migrants sensitive to the pandemic‐induced labor market shock? Most importantly, they are predominantly concentrated in low‐skilled, low‐paid professions (Fasani & Mazza, 2020). While different skill distributions are considered a crucial factor, Dustmann et al. (2010) find significantly larger unemployment responses to economic shocks for migrants relative to natives, even within the same skill group. In addition, migrant workers are much more likely to have temporary employment contracts than host populations, leading to their precarious, fragile job security (Fasani & Mazza, 2020). Another important reason is that they are far less likely to work in jobs that can be performed remotely (Borjas & Cassidy, 2020).

Since they exist at the intersection of many unfavorable aspects, it is inevitable that migrant workers are highly prone to pandemic‐driven lay‐offs. Unfortunately, they, nevertheless, are placed in a blind spot with manifold setbacks in South Korea. Non‐native workers usually do not enjoy the same social benefits as host populations. As a stark example, most migrant workers are not enrolled in employment insurance, unless they are granted permanent residency. As another striking example, unlike host populations, migrant workers cannot apply for government‐supported job retention schemes.

One more important question concerns gender‐based heterogeneity: have all migrants felt the economic consequences of COVID‐19 equally severely, regardless of gender? The estimation results discussed in Sections 4.3, 4.4 and IV.5 coherently reveal a clear, definite answer to this question. Female migrants are far more vulnerable to the initial COVID‐19 shock to the labor market, with male migrants weathering it relatively unimpaired. Therefore, not all migrants are evenly affected, and a migrant's gender represents a critical driver that deepens or mitigates the pandemic‐driven employment shock.

The most important reason for such unequal impacts stems from the fact that, among migrant workers, male and female employment are concentrated in different sectors in South Korea. The mining and manufacturing sector employs the largest share of male migrants: in contrast, the share of female migrants in the mining and manufacturing sector is approximately 70 percent that of males. Female migrant workers are instead active in the wholesale–retail and food–accommodation sector, with the shares hovering around 25 percent, compared with just 5 to 7 percent of employed male migrants. COVID‐19 has understandably dealt an acute blow to the service sector, primarily because of the strict measures enacted to combat the spread of the virus, such as social distancing and curfews, that have rocked the foundation of the service industry by curbing face‐to‐face contact and activities. Since female migrants disproportionately work in the service sector, which essentially depends on customer–provider interactions and large groups congregating, they clearly faced a greater risk of becoming unemployed due to the pandemic.

5. COMPARISONS WITH HOST POPULATIONS

The LCLF data only cover non‐native residents and provide no information about native citizens and their labor market outcomes. Nevertheless, it is important to address which group (i.e., migrants versus host populations) is more vulnerable to the initial COVID‐19 shock to the labor market. Thus, this section examines the question by additionally using the Local Area Labor Force (LALF) survey, which is a nationally approved dataset that Statistics Korea collects about the host populations’ employment.

To estimate the differentiated impact of the pandemic on migrant versus native populations, however, some practical limitations exist. First, the two datasets use different sampling strategies and include different variables. Second, some important variables, such as education levels, are recorded in different, non‐unifiable ways based on different definitions. Third, the two datasets were not collected at exactly the same time: the annual cross‐sections of the LALF data are collected every April, whereas those of the LCLF data are collected in May. In light of such limitations, the two datasets are not combined into one, and estimations are made separately for each group. The results in this section, therefore, should be interpreted with care and can only be considered as interim evidence. The same estimation strategy explained in Section III is applied to the LALF data.

Table 5 presents the estimation results for the sake of comparison, and some noteworthy implications are as follows.17 First, for both host populations and migrants, the pandemic hit women substantially harder than men. This finding corroborates the fact that the labor market impact of COVID‐19 is by no means gender‐neutral. Second, the gendered impacts are more pronounced in the case of migrants, among whom males weathered the shock more unscathed: it is only ${\widehat{\beta }}_d$ for male migrants (in Column 2), which lacks statistical significance.

TABLE 5.

Difference‐in‐differences (DD) estimates: Host populations versus migrants

Multiple‐period DD	[1] Host populations	[2] Migrants	[3] Difference
Data	LALF	LCLF
β d (males)	−0.0144** [0.0068]	−0.0046 [0.0105]	0.0098 (0.0123)
Confidence interval (95%)	[−0.0278, −0.0011]	[−0.0246, 0.0154]
β d (females)	−0.0219*** [0.0087]	−0.0657*** [0.0205]	−0.0439** (0.0206)
Confidence interval (95%)	[−0.0390, −0.0049]	[−0.1023, −0.0292]
Number of observations	258 725	17 441

Note: Robust standard errors are in brackets. The statistical test in Column (3) is based on the method of Clogg et al. (1995).

^*** p < 0.01;

^** p < 0.05;

^* p < 0.1.

Most importantly, the core question concerns which group is more heavily affected by the initial COVID‐19 shock to the labor market. An answer to this question requires the difference between the estimates in Columns 1 and 2 in Table 5 (for each gender group) to be tested statistically. Because the two estimates, as mentioned above, are made separately based on different data, this study adopts the test method suggested by Clogg et al. (1995) and Paternoster et al. (1998). Column 3 in Table 5 shows these results. For males, the null hypothesis of an equal impact on host and migrant populations cannot be rejected. The reason may have to do with the fact that many male migrant workers in South Korea are employed in industries with little face‐to‐face interaction, such as agriculture, forestry and fisheries, which have been relatively less affected by COVID‐19. Moreover, those industries suffer from chronic labor shortages as most native males are less willing to work in these sectors. As a result, native and migrant males are unlikely to compete with one another in the labor market due to clear sectoral segregation; hence, the pandemic was unlikely to intensify competition. In stark contrast, it seems that female migrants are considerably harder hit than host populations, with the difference being statistically significant. In the case of females, the sectoral segregation between host populations and migrants is not as distinct as that of males, with service‐related industries employing disproportionately large shares in both groups. Thus, this ‘blurry’ distinction in the female labor market means that some migrants likely compete with natives, and this competition may have intensified when shock‐driven labor market slack exists. In this case, if employers are inclined to lay off migrant workers first and retain native workers while waiting out the pandemic upheaval, the result in Column 3 in Table 5 is understandable. Furthermore, most migrant workers, for whom taking out employment insurance is not mandatory under the current system in South Korea, are not eligible for government‐supported job retention schemes. Hence, when all other conditions are equal, it is more costly for employers to continue employing migrant workers than native workers. It seems that the double whammy of both competing against host populations and lacking social safety nets amid the pandemic has made female migrants the most vulnerable group and triggered a substantial drop in their employment.

6. EFFECT HETEROGENEITY

6.1. Duration of stay in South Korea

The driver expected to have the greatest effect on heterogeneity is the duration of a migrant's stay in South Korea. In migrant‐related studies, how long a migrant stays in his or her host country is critical information that, in general, is positively correlated with labor market outcomes, such as employment and wages. Many channels may cause such a positive association, such as acquiring the local language, adapting to the local culture and institutions, accumulating country‐, industry‐, and firm‐specific human capital, and establishing social networks, which tend to increase with time spent in the destination country (Isphording, 2015).18

In this study, the question is whether migrants, who have resided in South Korea longer, are less affected by the negative employment shock stemming from COVID‐19. For empirical evidence, this aspect is analyzed in the manner of the triple difference estimator so that duration‐specific treatment effects can be estimated. Figures 7 and 8 portray the estimation results in the form of 95‐percent confidence intervals for males and females, respectively. The same inference can be made in both figures: the longer a migrant has lived in South Korea, the less affected he or she is by COVID‐19 as a labor market shock, which is intuitively expectable. The lack of statistical significance for those residing in South Korea for less than 1 year, despite sizable coefficient estimates, could be attributed to the small number of observations falling within such categories. Even in the case of female migrants, who overall are much harder hit than male migrants, the negative effect seems to vanish as one's stay duration exceeds 5 years. Such an important ‘buffer’ role of stay duration against the COVID‐19 employment shock is largely in line with Niimi et al. (2008): they underscore the positive link between the duration of residence and various labor‐related outcomes, such as earnings and remittances.

FIGURE 7.

Effect heterogeneity (by stay duration)—male migrants. Source: Surveys on Immigrants’ Living Conditions and Labor Force, South Korea. Estimation by the author.

FIGURE 8.

Effect heterogeneity (by stay duration)—female migrants. Source: Surveys on Immigrants’ Living Conditions and Labor Force, South Korea. Estimation by the author.

6.2. Education level

Recent empirical studies on COVID‐19 as a labor market shock, such as Che et al. (2020) and Aum et al. (2021), find that the less‐educated experienced a larger negative impact on their employment. To figure out whether the same is the case for migrants in South Korea, their level of schooling is taken into account as a moderating factor causing effect heterogeneity, and treatment effects are estimated for each subgroup with different education levels. Figures 9 and 10 present the estimation results in the form of 95‐percent confidence intervals for males and females, respectively. In the case of male migrants, while the overall pattern seems to be inverse U‐shaped, all estimates are statistically insignificant. A counterintuitive pattern is observed for female migrants, however: college‐educated female migrants seem to have been the hardest hit ($p$‐value: $0.018$), followed by high school graduates ($p$‐value: $0.056$). This finding indicates that higher levels of education do not ‘insure’ female migrants from being heavily affected by COVID‐19 and losing their jobs; rather, they seem to burden non‐native females in the pandemic era.

FIGURE 9.

Effect heterogeneity (by education level)—male migrants. Source: Surveys on Immigrants’ Living Conditions and Labor Force, South Korea. Estimation by the author.

FIGURE 10.

Effect heterogeneity (by education level)—female migrants. Source: Surveys on Immigrants’ Living Conditions and Labor Force, South Korea. Estimation by the author.

Why would COVID‐19 have a greater impact on college‐educated migrants? One of the most plausible reasons may reside in the possibility that highly educated migrants can compete with similarly skilled host populations, as mentioned by Globerman (2019). Studies such as Shirmohammadi et al. (2019) address the issue that skilled migrants often encounter the situation of competing with host populations. When the crisis puts employers in a dire position of having to cut operating costs, including labor costs, it is conceivable that they tend to lay off migrant workers first and retain native workers as long as possible while waiting out the pandemic turbulence. Moreover, as mentioned in Section V, most foreign workers in South Korea are not eligible for government‐supported job retention schemes; this facet may make employers feel more financially burdened to maintain migrants’ employment. As a result, the negative impact stemming from COVID‐19 could be more severe and direct for college‐educated migrants, who often compete with host populations, work in sectors where labor shortages are less of an issue and cannot rely on the social safety net.

Another question pertains to why, among college‐educated migrants, females are more heavily affected. A probable reason lies in the sector in which female migrants with a university degree are largely employed. Tables 9 and 10 in Appendix S1, Section A.4 summarize the relevant information, based on the LCLF data, for females and males, respectively. As these statistics reveal, the largest share of college‐educated female migrants is employed in the business, personal or public services sector, which, according to the classification rule, includes education‐related services. The largest proportion of male migrants with a university degree, by contrast, is employed in the manufacturing and mining sector. It is a known fact that COVID‐19 has taken a heavy toll on education jobs, the daily tasks of which require face‐to‐face contact and in‐person activities (Aum et al., 2021). In this regard, it is understandable that female foreign workers with a university degree, a considerable number of whom are employed in education services, are among the hardest hit. Note, however, that there is a lack of conclusive evidence; further research is needed to investigate this counterintuitive moderating role of education.

6.3. Korean language proficiency

It is axiomatic that being able to communicate in the language of the host country is one of the most vital drivers affecting migrants’ economic and non‐economic outcomes. A low level of proficiency in the local language presents high hurdles to participating in the labor market and engaging in social interactions: having adequate language skills allows migrants to increase their probability of employment, improve their access to better‐paying jobs and progress along the career ladder (Isphording, 2015). Several empirical studies corroborate a significantly positive response of labor market outcomes to local language proficiency, such as Chiswick (1991). In this regard, a reasonable hypothesis is thus, among migrants in South Korea, those having higher proficiency in Korean, the host country language, may have been less affected by COVID‐19 in terms of employment. As in the previous sections, this hypothesis is investigated in the fashion of the triple difference estimator, exploiting a relevant variable (i.e., ‘Korean proficiency’) provided by the LCLF data. Unfortunately, the information on a migrant's Korean proficiency is available only for those surveyed in 2018 and 2020, which means that the sample size in this section is substantially smaller compared to that used in Sections VI.1 and VI.2. Table 10 in Appendix S1, Section A.4 summarizes the distribution of migrants’ proficiency in Korean by gender.

The estimation results portrayed in Figures 11 and 12 for males and females, respectively, are in accordance with the aforementioned hypothesis. Those with low proficiency in Korean, regardless of gender, seem to have been most vulnerable to the COVID‐19 employment shock. In stark contrast, fluency in Korean seems to have insured migrants from being heavily affected by the crisis and losing their jobs. Such results pointing to the highly heterogeneous impacts of COVID‐19 are unambiguous in the sense that a migrant's familiarity with the host country's language makes them more productive and less replaceable. Thus, host country employers, when being in a situation of having to lay off non‐native employees, understandably tend to keep those who are more language‐proficient, which leads to the findings shown in Figures 11 and 12.

FIGURE 11.

Effect heterogeneity (by local language proficiency)—male migrants. Source: Surveys on Immigrants’ Living Conditions and Labor Force, South Korea. Estimation by the author.

FIGURE 12.

Effect heterogeneity (by local language proficiency)—female migrants. Source: Surveys on Immigrants’ Living Conditions and Labor Force, South Korea. Estimation by the author.

7. CONCLUSION

In South Korean society, which is already burdened by depopulation and labor shortage issues, the growing importance of migrant workers and their role is recognized widely. Nevertheless, despite the fact that they serve the country's economic needs, no single study has investigated how much and in which ways COVID‐19 has impaired migrants’ labor market outcomes. As underscored in many studies conducted in other countries, COVID‐19 represents an adverse labor market shock that is greatest for non‐native workers, as they have limited means to respond. This is even more so in the case of South Korea, where most foreign workers lack employment insurance and thus protection with the social safety net, motivating this study to diagnose the impact of COVID‐19 on migrants’ employment through the lens of causal estimations. Exploiting the unique regional outbreaks of COVID‐19 in South Korea, the DD estimates of this paper suggest that COVID‐19 substantially lowered migrants’ employment probability and that female migrants are far more vulnerable to its initial shock to the labor market. Male migrants are weathering the storm relatively unscathed.

Migrants’ pandemic‐driven job loss can lead to various other economic consequences through many channels. Clearly, on the supply side, labor costs will likely increase due to a reduced migrant labor supply and accompanying labor shortage issues (Laborde et al., 2020). In South Korea, damaging impacts are already manifest in the agriculture, forestry and fishery industries. As another example, harmed employment for migrants and its resulting decline in remittances can even transmit pandemic‐triggered negative impacts to their families in sending countries (Moroz et al., 2020). Considering that the decline can push countless people in sending countries into serious poverty, policy interventions are urgently needed to address the amplified vulnerability of migrants. Strengthening social security measures and extending them to include migrants as well would be of the utmost importance.

Supporting information

Appendix S1 Supporting Information

Shin, S. (2022). Labor market impact of COVID‐19 on migrants in South Korea: Evidence from local outbreaks. Asian Economic Journal, 36(3), 229–260. 10.1111/asej.12280