key: cord-0858609-88l5h434
authors: Ando, M.; Furuichi, M.
title: The impact of COVID-19 employment shocks on suicide and poverty alleviation programs: An early-stage investigation
date: 2020-11-18
journal: nan
DOI: 10.1101/2020.11.16.20232850
sha: ac4301d51454c58b8835aa192a0273f17e8e0293
doc_id: 858609
cord_uid: 88l5h434

This paper examines whether the COVID-19-induced employment shock has increased suicides and the utilization of means-tested poverty alleviation programs in the first eight months of the COVID-19 crisis. We exploit plausibly exogenous regional variation in the magnitude of the employment shock in Japan and adopt a difference-in-differences (DID) research design to identify the employment-shock impact. Our preferred point estimates suggest that a one-percentage-point decrease in the employment rate in the second quarter of 2020 resulted in an additional 0.14-0.44 male suicides per one hundred thousand male population in June 2020 and an additional 5.3-7.9 Public Assistance recipients per one hundred thousand population in August 2020. The impacts on female suicides and other poverty alleviation programs are not precisely estimated, but there is suggestive evidence that the same employment shock has also increased the caseloads of the other poverty alleviation programs. We also examine the impact of the increase in the unemployment rate and find a positive impact on the poverty alleviation programs but not on suicides.

The COVID-19 pandemic has caused serious infections and deaths around the world, and it is becoming clearer that economic and social consequences are also devastating. Being faced with the COVID-19 pandemic and subsequent sharp economic contraction, national governments have been trying to mitigate social and economic damages by introducing new emergency cash benefits and expanding existing social programs. Nonetheless, there are daily news reports of job loss, poverty, mental disorders, and even suicides that have been directly or indirectly induced by the COVID-19 economic crisis.

Japan has also suffered from these economic and social difficulties despite its relative success in mitigating COVID-19 infection spread without imposing a strict lockdown policy. 1 As Figure 1 shows, in Japan, the COVID-19 employment shock emerges as sharp drops in the employment rate and labor force participation (LFP) rate since April 2020, while the indices of the unemployment rate and the jobs-toapplicants ratio (i.e., the ratio of job offers to job applicants) have also gradually deteriorated. Figure 1 also suggests that suicide rates have been increasing since July 2020. It is not clear from this graph to what degree the COVID-19 crisis has led to the increase in suicides, but some specialists and news media have reported that the sharp deterioration in employment in the first few months of the COVID-19 crisis may have resulted in the slightly lagged increase in suicides. 2 ".

This paper therefore provides an early-stage analysis on whether the COVID-19 employment shock affects suicide, using the latest monthly suicide data. We exploit regional (i.e. prefecture-level) variations in the magnitude of the COVID-19-induced employment deterioration and adopt a DID research design to mitigate the estimation bias arising from possible confounding factors.

We also investigate whether the employment shock has affected means-tested 1 As of November 2020, in Japan the infection and death rates of COVID-19 have been relatively low in comparison to other developed countries at least until. The cumulative numbers of COVID-19 confirmed cases and deaths per one million population on November 11th 2020 are about 871 and 15 respectively, compared with 18,911 and 453 in the EU countries and 30,990 and 724 in the U.S, according to the website Our World in Data (https://ourworldindata.org/coronavirus, accessed on 12th November 2020). In addition, Japan has not introduced a strict lockdown policy, and according to the Oxford COVID-19 Government Response Tracker Japan's government response stringency index has been the lowest among the G7 countries since April 2020, lower than that in South Korea since May 2020, and mostly lower than or roughly equal to that of Sweden since April 2020 (https://ourworldindata.org/policy-responses-covid, accessed on 12th November 2020).

2 See, among others, articles in the Japan Times such as "Suicide spike in Japan shows mental health toll of COVID-19" (9th October 2020, https://www.japantimes.co.jp/ news/2020/10/09/national/social-issues/suicide-mental-health-coronavirus) and "Japan suicides rise as economic impact of coronavirus hits home" (11th November 2020, https://www.japantimes.co.jp/news/2020/11/11/national/ japan-suicide-rise-coronavirus). The number of accepted applications for programs Sources: Employment rate, labor force participation rates, and the unemployment rate are from the Labor Force Survey and the jobs-to-applicants ratio is from the General Employment Placement Status. Suicide rates are provided as police statistics and aggregated by the Ministry of Health, Labour, and Welfare. The numbers of accepted applications for poverty-alleviation programs are provided by the central government. From April to September 2020, the monthly data aggregation is irregular as follows: April data is based on the numbers from 25th March to 2nd May, May data is from 3rd May to 30th May, June data is from 31st May to 27th June, and July data is from 28th June to 1st August.

poverty alleviation programs such as a public assistance program, a temporary housing benefit program, and two emergency loan programs. Because an increase in the recipients and caseloads of means-tested programs can be seen as the signs of both economic distress and relief, it is hard to identify the mitigating effects of poverty alleviation programs on suicide. But some insights may nevertheless be obtained about the consequences of the COVID-19 employment shock and policy responses to them. The contributions of our paper are at least three-fold. First, we provide earlystage plausible evidence of the impact of the COVID-19-induced employment shock on suicide and means-tested poverty alleviation programs, which has rarely been studied so far. Second, from a more general viewpoint, our study contributes to the literature on the impact of the world-wide economic distress on suicide and public policies, exploiting plausibly exogenous change in employment environments. Third, because we investigate ongoing economic and social incidents under the COVID-19 pandemic using the latest monthly panel data, we also contribute to the emerging literature of economic and policy evaluation using real-time data (Chetty et al. 2020) .

In previous studies, employment shocks have been considered to be important determinants of suicide. While the impacts of the COVID-19 pandemic on socioeconomic and public-policy outcomes such as unemployment and unemployment insurance benefits have been studied (Adams-Prassl et al. 2020; Bitler et al. 2020) , little research has yet examined the impact of the COVID-19-induced employment shock on socio-economic outcomes such as suicide and poverty alleviation programs. 3 When looking at broader contexts before the COVID-19 crisis, the effect of recessions on mortality has been well studied and many studies have shown that increases in unemployment are associated with a rise in the suicide rate (Ruhm 2000; Chen et al. 2009; Ruhm 2015; Matsubayashi et al. 2020) . See also Arthi and Parman (2020) for a recent review of the possible long-run impacts of COVID-19 economic consequences on various socio-economic outcomes from a historical perspective.

Another related article suggests that counter-cyclical economic policies are effective for the prevention of suicide (Christian et al. 2019; Stuckler et al. 2009; Fishback et al. 2007 ). Although we examine poverty alleviation programs only as outcomes and do not investigate whether such programs have mitigated suicides, our study may also provide some insight on this research topic.

The rest of the paper consists of the following sections. Section 2 describes our research design and the data that we use for analysis. Section 3 explains our empirical model. In Section 4 we provide estimation results and in section 5 we provide additional estimation results using an alternative employment-shock variable. Section 6 concludes.

3 Epidemiologists indicate that the rise in unemployment caused by the COVID-19 pandemic is expected to lead to an increase in the suicide rate (Kawohl and Nordt 2020; Gunnell et al. 2020) . So far, there are only a few studies that investigate suicide under the COVID-19 economic shocks. When it comes to suicide in Japan, Tanaka and Okamoto (2020) shows that the suicide rate declined after the COVID-19 crisis, comparing the citymonth-level suicide mortality rate in Japan before and after the COVID-19 crisis, Ueda et al. (2020) also found that the number of suicide deaths decreased from 2017-2019 to the early phase of the COVID crisis but increased from the June 2020. Isumi et al. (2020) shows that the suicide rate among children and adolescents did not increase during the school closure but increased in May. Using online survey data, Sueki and Ueda (2020) found suicidal ideation in April 2020 decreased compared with January 2020. They also found suicidal ideation levels among young and economically vulnerable people in Japan have been exacerbated during the COVID-19 crisis.

2 Research design and data

Exploiting the sudden and plausibly exogenous variation in the COVID-19 employment shock at the prefecture level, we adopt a difference-in-differences (DID) strategy to investigate the effect of the COVID-19 employment shock on suicides and poverty alleviation programs.

More concretely, we investigate how the regional variation in the COVID-19induced employment shock is correlated to the regional trends of suicides and povertyalleviation programs and examine to what extent we can interpret this correlation as the causal effect of the employment shock on the outcomes of interest.

We use prefecture-level monthly data from various data sources. Japanese prefectures, such as Tokyo, Kyoto, and Hokkaido, are comparable to states in the U.S. and provinces in Canada and some other European countries, and the total number of prefectures is 47. Prefectures are the second-tier local governments in Japan; the third-tier local governments are municipalities, the number of which is 1,724 (as of October 2020.) The duration of our data depends on the availability of each variable, but all the outcome variables contain data from before and after the COVID-19 crisis.

As a primary indicator of the employment shock caused by the COVID-19 crisis, we exploit a significant decrease in the employment rate, which is the rate of the number of employees to the population age 15 and over, that began in April 2020. That is, we define the COVID-19-induced employment shock as the following de-trended decrease in the employment rate from the first quarter (i.e. January to March) to the second quarter (i.e. April to June) in 2020 at the prefecture level:

where i indicates prefecture, X is the employment rate, and 1Q and 2Q are the first and second quarters respectively. The first term in equation (1) is a change in the employment rate from the first quarter to the second quarter of 2020 which includes a COVID-19-induced employment shock and the second term is a counterpart change in 2019 which reflects a prefecture-specific seasonal trend from the first quarter to the second quarter. The difference between these two terms is thus expected to capture the COVID-19-induced employment shock and the negative of this change represents the magnitude of the "decrease" in the employment rate caused by the COVID-19 crisis. We interpret this COVID-19-induced decrease in the employment rate as the employment shock caused by the COVID-19 economic crisis.

. CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ;

We use the employment rate, not the unemployment rate and other indicators of labor outcomes, as the primary employment indicator for the COVID-19 employment shock because the employment rate exhibits a clearer and sharper deterioration when the COVID-19 crisis started than the three other employment variables in Figure 1 . For additional employment-shock variables, we also use prefecture-level variations in two employment indicators: the labor force participation (LFP) rate and the unemployment rate, which we will discuss further in Section 5.

Finally, what we use is the prefecture-level estimates of the average employment rate, LFP rate, and unemployment rate provided in the Labour Force Survey. The prefecture-level estimates are estimated by the Statistics Bureau. The Bureau notes that some imprecision is expected for the estimates, but we treat the estimated values as true values and do not consider their statistical uncertainty in our statistical analysis due to data limitation.

For outcome variables, we focus on suicide rates and four existing means-tested poverty alleviation programs. Here we briefly explain these variables. See Ando et al. (2020) for more details about Japan's fiscal measures against COVID-19, including the four poverty alleviation programs, in the first six months of the COVID-19 crisis.

Suicide rates Our primary outcome of interest is the suicide rate, which is the number of suicides per one hundred thousand population age 15 and over. We investigate the monthly panel data of three suicide-rate variables: total suicide rate, female suicide rate, and male suicide rate. We use the prefecture-level data between January 2019 and September 2020 and the number of suicides is aggregated based on the dates and places of suicide. 4 For our DID analysis, we use both the suicide rate and the year-on-year suicide rate. The year-on-year suicide rate at month t is calculated as a difference between the suicide rate at month t and the one-year-lag suicide rates at month t − 12. This procedure is meant to control for monthly fixed effects.

Public Assistance The first variable of a poverty alleviation program is the percapita number of Public Assistance recipients. For this variable, we use both the original per-capita number of recipients and the counterpart year-on-year value as we do for the suicide rates. Note that the Public Assistance program in Japan is considered as the "final safety net" or "last resort," and as such carries a high social 4 Suicide statistics are provided as police statistics and aggregated by the Ministry of Health, Labour, and Welfare(Ministry of Health, Labour and Welfare 2020). Prefecturelevel aggregated suicide statistics based on residential addresses are also available, but we do not use these data due to the large number of missing values.

. CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ; stigma and the prerequisites for application are in general strict, such as having no savings and no assets.

Housing Security Benefit The second poverty alleviation program variable is the number of accepted applications per one hundred thousand population for the Housing Security Benefit, which is an existing scheme of short-term (i.e. three to nine months) housing allowance under the Act on Self-reliance Support for Needy Persons. This benefit scheme, originally only for those who have lost their jobs, was rarely used before the COVID crisis as shown in Figure 1 . But since April 2020, its eligibility has been extended to those who have not lost their jobs but have still experienced a large income reduction, resulting in a considerable increase in the number of households receiving the benefits. Because of the limited data availability, we can use only January, February, and March 2019 in the pre-COVID-19 period and April, May, and June 2020 in the post-COVID-19 period.

Temporary loan programs The third and fourth variables of the poverty-alleviation programs are the numbers of accepted applications per one hundred thousand population for two types of interest-free and guarantor-free emergency temporary loan programs: Emergency Small Amount Funds for those who urgently need cash and General Support Funds for those who need cash for a certain period. They are both existing means-tested emergency loan programs, but they were also rarely utilized before the COVID-19 crisis (see Figure 1 ). However, their requirements have also been relaxed since March 2020 and the number of loan recipients has increased dramatically. We use the monthly data between January 2019 and July 2020, but due to the limited data availability, we do not have the data for February and March 2020. 5 In addition, the monthly-level data aggregation from April to July 2020 is slightly irregular, although it should not cause serious estimation bias. 6 Table 1 shows the descriptive statistics of our treatment and outcome variables. Because we use the panel data of 47 prefectures from January 2019 to September 2020 (21 months), the full sample size is 987. Because the employment-shock variables defined in equation (1) are time-invariant and cross-sectional variables, we have 47 observations. The variables of suicide rates are 21-months data while the variables of public assistance recipients lack the data for September 2020. The variables of 5 Statistics in February and March 2020 were missing from the original data provided by the government. 6 The number of loan decisions between April and July 2020 in the statistics provided by the central government is aggregated as follows: April data is based on the numbers from 25th March to 2nd May, May data is from 3rd May to 30th May, June data is from 31st May to 27th June, and July data is 28th June to 1st August.

. CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ;

Housing Security Benefit, Emergency Small Amount Funds, and General Support Funds are more limited due to data limitation as already explained. Figure 2 shows the correlations of our main treatment variable (i.e. the employment shock) and changes in our two main outcome variables from January 2020 (just before the COVID-19 pandemic) to September or August 2020 (in the midst of the COVID-19 pandemic). The changes in the outcome variables are used in the left graphs and the de-trended year-on-year changes that additionally subtract the counterpart changes in 2019 like equation (1) are used in the right graphs. The observed positive correlations indicate that the larger COVID-19-induced employment shock may be related to more suicides and Public Assistance recipients, although the correlations are not crystal clear.

The model specification takes the following form:

where Y it is the outcome variable such as a suicide rate for prefecture i at time t, π i and θ t are prefecture and month fixed effects, respectively, Empshock i is a continuous treatment variable of the COVID-19 employment shock defined as in 8 . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ; Notes: In all four graphs, the X axis is the size of an employment shock defined as in (1). The Y axis in the left panels is a change in the suicide rate (from January to September 2020) and a change in the number of Public Assistance recipients per one hundred thousand people (from January to August in 2020). For the Y axis in the right graphs, a counterpart outcome change in 2019 is further subtracted from the original outcome change. The size of each circle is based on the population size of each prefecture. The dashed line is the fitted linear regression line based on the Ordinary Least Square (OLS) method.

(1), 1[t = τ ] is a dummy variable that takes the value of one if t = τ and zero otherwise, and ε it is an error term. The coefficients of interest are the time-varying coefficients β τ , which capture the correlation between EmpShock i and the outcome trend from January 2020 to time τ . When time τ is before January 2020, β τ can be interpreted as a placebo estimate that is expected to be zero if no confounding trends exist before the COVID-19 crisis. When month τ is after January 2020, β τ can be seen as the average impact of EmpShock i given that the placebo estimates are around zero and the assumption of no differential trends in the COVID-19 period is plausible.

In the regression analysis, we use both the ordinary least square (OLS) estimation method and the weighted least square (WLS) estimation method in which prefecturelevel population sizes are used as weights. With this WLS model, the heterogeneous effect of the employment shock on an outcome for a larger prefecture is more highly weighted in a DID estimator. We use both estimation methods because, as Solon et al. (2015) point out, it is not straightforward whether a WLS estimator is more appropriate than an OLS estimator as an estimator of a population average effect when effect heterogeneity exists.

Furthermore, we also provide estimation results with prefecture-level linear trends 9 . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ;  in the model (2) and examine how DID estimates change by following this procedure. Controlling for the prefecture-level linear trends may reduce the bias from confounding differential trends. Figure 3 shows the estimation results for the suicide rates based on four different model specifications. First, DID estimates in the pre-COVID-19 period are more or less zero, implying that there is no clear statistically significant correlation between the employment shock and all the outcome trends before January 2020. While some fluctuations in the pre-COVID-19 estimates suggest that there may be some idiosyncratic monthly correlations that exist between the shock and the outcome trends before the COVID-19 period, they should not be sources of confounding trends in the post-COVID-19 estimates.

When it comes to the post-COVID-19 estimates after January 2020, the estimates for total and male suicide rates are consistently positive from May onward regardless of model specifications and outcome specifications and the lower bounds of the 90% confidence intervals are often above or near zero. In particular, the estimates for male suicide rate are robustly above zero from May 2020. The sizes of the estimates differ between OLS and WLS: the OLS estimates for the male suicide rate in June, the last month of the second quarter, suggest that the one-percentagepoint reduction in the employment rate led to around a 0.25-0.44 point increase in the suicide rate whereas the counterpart WLS estimates indicate around a 0.14-0.18 point increase.

On the other hand, we do not find noticeable changes in DID estimates for the female suicide rate in the COVID-19 period. Although the estimate is significantly different from zero in May 2020 when the simple, not year-on-year, female suicide rate is used as an outcome, the other post-COVID-19 estimates for the female suicide rate are mostly not significantly different from zero.

Overall, the DID estimates suggest that the negative employment shock caused by the COVID-19 economic crisis has led to the increase in male suicides. However, we cannot find robust evidence of a similar effect for female suicides.

The magnitude of the estimated impact is not small. The minimum estimate for the year-on-year male suicide rate in June 2020 is 0.14 and the maximum is 0.44. Because the average monthly male suicide rate (i.e. the average number of suicides per one hundred thousand population in a month) during the sample period is 1.85, the increase by 0.14-0.44 is not negligible; as of October 2020 the male population in Japan is about 61,260,000, so a back-of-the-envelope calculation based on the estimates suggests that the one-percentage point decrease in the employment rate under the COVID-19 crisis has led to an additional 86-270 male suicides per 10 . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ; https://doi.org/10.1101/2020.11.16.20232850 doi: medRxiv preprint 

Notes: Each plot indicates a point estimate and a vertical line indicates a 90% confidence interval. The treatment variable is the COVID-19-induced employment shock, which is calculated as a decrease in the prefecture-level employment rate from the second quarter of 2019 to the second quarter of 2020. A year-on-year outcome variable is calculated as a difference between an outcome value at month t and an outcome value at month t − 12. WLS estimation is weighted by prefecture population size. Estimation "with trends" incorporates prefecture-level (i.e. individual) linear trends and time-varying estimates can be obtained from February 2019 with this model specification.

. CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ; https://doi.org/10.1101 https://doi.org/10. /2020 month at the national level. Figure 4 provides estimation results for the means-tested poverty alleviation programs. To begin with, DID estimates for the number of Public Assistance recipients (the upper two figures) are significantly different from zero and increasing in the COVID-19 period (i.e. after January 2019). While the non-zero estimates in the pre-COVID-19 period based on the models without prefecture linear trends suggest there exist some differential trends that are correlated with the employment shock, the introduction of prefecture linear trends successfully mitigates such confounding trends and suggests that the employment shock has caused the increase in Public Assistance recipients.

The DID estimates in August 2020 are around 5.3 and 7.9 according to the OLS and WLS estimation methods with prefecture linear trends respectively. This implies that the one-percentage point decrease in the employment rate resulted in about 5 to 8 Public Assistance recipients per one hundred thousand population. Because the average number of Public Assistance recipients in the sample period is about 1,400 people per one hundred thousand population (Table 1) , the margin of 5 to 8 can be considered relatively small.

Estimates for the other three poverty alleviation programs are mostly not significantly different from zero (the lower three figures), but the estimates are increasing and sometimes significantly different from zero, suggesting the possibility of some positive effects on these programs. Note that the numbers of accepted applications for these programs are discontinuously smaller before the COVID-19 crisis than after the crisis, the estimates before the COVID-19 period are negligible and we cannot access the pre-COVID-19 trends in these outcomes.

The results suggest that, if the causal interpretation of the DID estimates is plausible, the one-percentage-point increase in the employment rate leads to around 19-33 accepted applications per one hundred thousand population for Emergency Small Amount Funds in June 2020 and roughly 14-26 accepted applications per one hundred thousand population for General Support Funds in July 2020. 7 These numbers are much higher than the estimated marginal increase in Public Assistance recipients.

To sum, the estimation results show the COVID-19-induced employment shock measured by the decrease in the employment rate has had a clear but small impact on the increase in the number of Public Assistance recipients. The estimates for the other three programs are not precise, but the results suggest that the magnitude of the impact on the utilization of the temporary loan programs may be larger than that on Public Assistance. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ; https://doi.org/10.1101 doi: medRxiv preprint The treatment variable is the COVID-19-induced employment shock, which is calculated as a decrease in the prefecture-level employment rate from the second quarter of 2019 to the second quarter of 2020. A year-on-year outcome variable is calculated as a difference between an outcome value at month t and an outcome value at month t − 12. WLS estimation is weighted by prefecture population size. Estimation "with trends" for the outcome of Public Assistance incorporates prefecture-level (i.e. individual) linear trends and time-varying estimates can be obtained from February 2019 with this model specification. For the outcome of the accepted applications for the Housing Security Benefit, the reference year is March 2019 instead of January 2020 due to the lack of data. But this should not affect estimation results because the number of accepted applications for the Housing Security Benefit is negligible before the COVID-19 crisis.

This subsection provides DID estimation results when we use the unemployment rate, instead of the employment rate, as an indicator of the COVID-19-induced employment shock.

As Figure 1 shows, the employment rate and the LFP rate has dropped sharply since April 2020, whereas the unemployment rate has gradually increased during the COVID-19 pandemic. This is the primary reason we use the decrease in the employment rate as a baseline treatment variable for the employment shock.

We think this is a reasonable choice, but the fact that the employment rate and the LFP rate has dropped in a similar manner in Figure 1 suggests that a large part of the prefecture-level employment shock adopted in the baseline analysis is explained by the significant reduction in the labor force (i.e. the number of people who are either working or actively looking for work) under the COVID-19 pandemic, 13 . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ; not the increase in unemployment. Figure 5 shows this is the case. In this figure, the "employment-rate shock" is defined as in equation (1), the "LFP-rate shock" is defined in the same way but the LFP rate is used for X and the "unemployment-rate shock" is defined in a similar way to equation (1) but the unemployment rate is used for X and the negative sign at the beginning of equation (1) is dropped for easier comparison with the employment-rate shock.

Notes: This correlation matrix shows the correlation coefficients and scatter plots among the three employment shock variables. The employment shocks are defined based on equation (1) where the estimated quarter averages of the employment rate, the LFP rate, and the unemployment rate are used for X. When the unemployment rate is used for X, the negative sign at the beginning of equation (1) is not used to ensure better comparability.

This figure shows that there is a very strong correlation between the employmentrate shock and the LFP-rate shock (the correlation coefficient is 0.958) and a weaker correlation between the employment-rate shock and the unemployment-rate shock (the correlation coefficient is 0.346). It implies that the COVID-19 employment shock examined in the baseline analysis can be interpreted as the impact of the reduction in the labor force. In fact, Figure A .1 and Figure A .2 in the Appendix provide similar estimation results for the outcomes of the suicide rates and the poverty alleviation programs when the LFP rate is used for the employment-shock variable instead of the employment rate.

On the other hand, the weaker correlation between the employment-rate shock and the unemployment-rate shock indicates that the two variables reflect different 14 . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ;

features of the COVID-19 employment shock. More concretely, because the unemployment rate is defined as the ratio of the number of unemployed people to the labor force, the unemployment-rate shock should directly reflect the COVID-19-induced increase in the unemployed population rather than the increase in the non-working population.

We thus estimate with the same empirical model (2) but using the unemploymentrate shock as EmpShock i . 8 Figure 6 shows estimation results for suicides, using the same four estimation models as in the baseline analysis. Unlike the baseline analysis in Figure 3 , the positive post-COVID-19 estimates for the male suicide rate are not observed in this case. Post-COVID-19 estimates for the total and female suicide rates are also noisy and the confidence intervals tend to be large while some estimates are positive in May and June.

In Figure 7 , we present counterpart DID estimates for the poverty alleviation programs. First, we observe similar positive post-COVID-19 estimates for Public Assistance recipients to the baseline estimates with the employment-rate shock when the year-on-year variable is used (the upper right graph), while results are noisier when the simple, not year-on-year, outcome is used (the upper left graph). In particular, when the prefecture linear trends are incorporated in the model for the year-on-year outcome, the estimates clearly suggest that the increase in the unemployment rate has resulted in the increase in the number of Public Assistance recipients.

The lower three graphs of Figure 7 provide counterpart DID estimations for the outcomes of the other poverty alleviation programs. Interestingly, there seem to be some positive correlations between the COVID-19-induced unemployment increase and the utilization of all of these programs, although the confidence intervals tend to be relatively large. We speculate that these positive correlations, which are not precisely observed when we use the employment-rate shock as the treatment variable, can be explained by the fact that these poverty alleviation programs are primarily for those who have lost their income by losing or stopping their jobs and need urgent financial support. The increase in the unemployed population, rather than the decrease in the labor force, should be more strongly related to the increasing caseloads of these poverty alleviation programs. 

OLS without trends WLS without trends OLS with trends WLS with trends Notes: Each plot indicates a point estimate and a vertical line indicates a 90% confidence interval. The treatment variable is the COVID-19-induced employment shock, which is calculated as an increase in the prefecture-level unemployment rate from the second quarter of 2019 to the second quarter of 2020. A year-on-year outcome variable is calculated as the difference between an outcome value at month t and an outcome value at month t − 12. WLS estimation is weighted by prefecture population size. Estimation "with trends" incorporates prefecture-level (i.e. individual) linear trends and time-varying estimates can be obtained from February 2019 with this model specification.

. CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ; Notes: A blue circle indicates a point estimate and a black line indicates a 90% confidence interval. The treatment variable is the COVID-19-induced employment shock, which is calculated as an increase in the prefecture-level unemployment rate from the second quarter of 2019 to the second quarter of 2020. A year-on-year outcome variable is calculated as the difference between an outcome value at month t and an outcome value at month t − 12. Estimation "with trends" for the outcome of Public Assistance incorporates prefecture-level (i.e. individual) linear trends and time-varying estimates can be obtained from February 2019 with this model specification. For the outcome of the accepted applications for the Housing Security Benefit, the reference month is March 2019 instead of January 2020 due to the lack of data. But this should not affect estimation results because the number of accepted applications for the Housing Security Benefit is negligible before the COVID-19 crisis.

When it comes to effect sizes, the DID estimates by OLS and WLS with prefecture trends suggest that a one-percentage-point increase in the unemployment rate has led to an increase in Public Assistance recipients of around 11-15 people per one hundred thousand population. The DID estimates for the other three programs indicate that the same unemployment increase may result in increases in the accepted applications for the Housing Security Benefit, Emergency Small Amount Fund, and General Support Funds by roughly 12-30, 24-58, and 16-41 people per one hundred thousand population in June or July respectively.

To sum up, the increase in the unemployment rate in the COVID-19 crisis has not clearly resulted in an increase in suicides, but it may have led to an increase in the recipients of the means-tested poverty alleviation programs.

. CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ; https://doi.org/10.1101/2020.11.16.20232850 doi: medRxiv preprint

Using prefecture-level monthly panel data and exploiting seemingly exogenous regional variations in the employment situation caused by the COVID-19 pandemic, this paper examines whether the COVID-19-induced employment shock has affected suicide rates and poverty alleviation programs in the first several months of the COVID-19 economic crisis.

Our estimation results suggest that the COVID-19-induced decrease in the employment rate has increased male suicides and Public Assistance recipients. We, however, do not find evidence that the decrease in the employment rate has led to an increase in female suicides. The employment-shock impacts on the caseloads of three poverty alleviation programs (Housing Security Benefit and the two emergency temporary loan programs) are not clear, but the effect sizes on the two loan programs may be larger than that on the Public Assistance recipients if these effects exist.

When we use the COVID-19-induced increase in the unemployment rate as an alternative employment shock variable, DID estimates for the suicide rate in the COVID-19 period are not significantly different from zero in most cases. On the other hand, our estimates show relatively robust evidence of the positive impact on the number of Public Assistance recipients and weak but suggestive evidence of the positive impacts on the accepted applications for the other means-tested poverty alleviation programs.

There are several issues to be discussed regarding our findings. Here we briefly discuss three of them as concluding remarks.

First, why do we observe the positive impact of employment-rate shock on only male, not female, suicides? One possible explanation is that, although the COVID-19 employment shock itself is larger for female workers than male workers (Kikuchi et al. 2020 ), a persistent decrease in female employment in the COVID-19 pandemic is largely due to the job loss of part-time or married female workers who may not be breadwinners in their households. 9 The financial and psychological distress of job loss or temporary job retirement for this female cohort may thus have been milder than that of male workers who have lost their jobs in the COVID-19 crisis, at least on average. However, given the higher increase in female suicides than that in male suicides in the COVID-19 period, this interpretation should be considered as speculative and requiring further examination.

Second, why do we observe the employment-shock impact on male suicides only when the employment or LFP rate, not the unemployment rate, is used as the indica-9 Documents and statistics presented in the "Study Group on the Impacts and Challenges for Women under the COVID-19" held by the Gender Equality Bureau Cabinet Office (https://www.gender.go.jp/kaigi/kento/covid-19/index.html) provide useful information about female employment under the COVID-19 crisis.

. CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ;  tor of the employment shock? One possibility is that the decrease in the male labor force is related to an increase in the male population with a high risk of suicide, but the increase in male unemployment is not. That is, because "unemployed workers" are those who are currently not working but are willing and able to work for pay, the short-term increase in unemployment caused by the COVID-19 pandemic may not be an appropriate indicator to capture the increasing population who are at the brink of suicide. Another possible reason is that the magnitude and the variation of the unemployment-rate shock in Japan is not large enough to have a detectable impact on suicides.

Third, how can we interpret the robustly positive but relatively small impact of all three employment-shock variables on Public Assistance recipients and imprecise but suggestive impact on the other means-tested poverty alleviation programs? We consider these results an indication that the COVID-19 employment shock has unarguably increased financial distress, and the means-tested benefits and temporary loan programs have partly alleviated such distress. At the same time, the Public Assistance scheme has strict certification criteria and this may be the reason the estimated effect size is somewhat small even though the estimates are relatively precise. On the other hand, the other three benefit and loan programs are accessible to a larger number of people, but this also implies that the relation between the employment status and the utilization of these programs is relatively weak. Of course, data limitation is another non-negligible feature in our analysis of these three poverty alleviation programs.

Finally, while we provide some evidence of the impact of the COVID-19 employment shock on suicides and poverty alleviation programs, we cannot address whether the poverty alleviation programs have mitigated suicides under the COVID-19 crisis. This is a much more difficult task because of the endogenous determination of the poverty alleviation programs, but future research on this question should be undertaken.

OLS without trends WLS without trends OLS with trends WLS with trends Notes: Each plot indicates a point estimate and a vertical line indicates a 90% confidence interval. The treatment variable is the COVID-19-induced employment shock, which is calculated as an increase in the prefecture-level LFP rate from the second quarter of 2019 to the second quarter of 2020. A year-on-year outcome variable is calculated as the difference between an outcome value at month t and an outcome value at month t − 12. WLS estimation is weighted by prefecture population size. Estimation "with trends" incorporates prefecture-level (i.e. individual) linear trends and time-varying estimates can be obtained from February 2019 with this model specification.

. CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted November 18, 2020. ; https://doi.org/10. 1101 

Inequality in the impact of the coronavirus shock: Evidence from real time surveys

Fiscal Responses to the coViD-19 cRisis in Japan: the FiRst six Months

Disease, downturns, and wellbeing: Economic history and the long-run impacts of COVID-19

Cyclicality of the US Safety Net: Evidence from the 2000s and Implications for the COVID-19 Crisis

How is suicide different in Japan?

The Economic Impacts of COVID-19: Evidence from a New Public Database Built Using Private Sector Data

Income shocks and suicides: causal evidence from Indonesia

Births, deaths, and New Deal relief during the Great Depression

Suicide risk and prevention during the COVID-19 pandemic

Do suicide rates in children and adolescents change during school closure in Japan? The acute effect of the first wave of COVID-19 pandemic on child and adolescent mental health

COVID-19, unemployment, and suicide

Who Suffers from the COVID-19 Shocks? Labor Market Heterogeneity and Welfare Consequences in Japan

Government spending, recession, and suicide: evidence from Japan

Suicide statistics: Suicide statistics by region

Are recessions good for your health?

What are we weighting for?

The public health effect of economic crises and alternative policy responses in Europe: an empirical analysis

Short-term effect of the COVID-19 pandemic on suicidal ideation: A prospective cohort study

Suicide during the COVID-19 pandemic in Japan

Mental Health Status of the General Population during the COVID-19 Pandemic: A Cross-sectional National Survey in Japan

2: DID estimates for poverty alleviation programs (LFP-rate shock)

The treatment variable is the COVID-19-induced employment shock, which is calculated as an increase in the prefecture-level LFP rate from the second quarter of 2019 to the second quarter of 2020. A year-on-year outcome variable is calculated as the difference between an outcome value at month t and an outcome value at month t − 12. Estimation "with trends" for the outcome of Public Assistance incorporates prefecture-level (i.e. individual) linear trends and time-varying estimates can be obtained from February 2019 with this model specification. For the outcome of the accepted applications for the Housing Security Benefit, the reference month is

A Appendix