Extended Unemployment Insurance and Job Search: Evidence from Time Use Data

1 Introduction

The U.S. government extended the duration of unemployment insurance (UI) benefits in several phases from mid-2008 to late 2009, raising the duration up to 99 weeks from the regular 26 weeks. The extensions, billed as the most generous in U.S. history, were targeted at financially helping job losers smooth their consumption in a period of severe unemployment resulting from the Great Recession of 2007–2009. However, the government’s move to extend benefits stirred a debate over the disincentive effect on recipients’ labor supply, which could contribute to a higher unemployment rate. For instance, in an op-ed, Barro (2010) argues that in the absence of extended UI coverage, the U.S. unemployment rate would have been 6.8% in June 2010, instead of the actual rate of 9.5%. On the other hand, the administration has argued that UI extensions have had no impact on the persistently high unemployment rate, which is viewed as caused by the severe downturn and sluggish recovery associated with the Great Recession.

Establishing the causal link between the UI benefits extension and the unemployment rate is inherently difficult because of the endogenous nature of the insurance program. In theory, the disincentive effect of an expansion in UI benefits impacts the unemployment rate through two direct channels: discouraging job search behavior of unemployed individuals and increasing their reservation wage (i.e., decreasing their job-acceptance rate). Indirectly, there are at least four additional channels through which extended UI benefits could affect the unemployment rate both in the short run and in the long run. First, the coverage can increase consumption of goods and services, thus edging up overall output and the employment rate in the short run. The Congressional Budget Office (CBO 2012) estimates that one dollar spent on extended UI benefits can create a multiplier effect of $1.10 in the economy. Second, it entails “crowd-out” effect in the short run, which is the effect of the receipt of UI benefits by a family member on the labor supply by other family members. Cullen and Gruber (2000) find that wives’ hours of labor supply could go up by about 30% during the period of their husbands’ unemployment if the latter did not have access to UI benefits. Likewise, in the long run, unemployment insurance could help better match workers and firms, thus increasing output and, consequently, employment in the economy (Acemoglu and Shimer 2000). With benefits, an unemployed individual could afford to search and wait for a better-suited, higher-paid and higher-productivity job. Finally, the disincentive effect of UI benefits may lengthen the duration of unemployment, deteriorating their skills for future jobs, thus decreasing productivity in the economy and increasing unemployment in the long run.

The direct and indirect disincentive effects require a general equilibrium analysis to quantify the link between the UI program and unemployment rate as well as unemployment spells. Therefore, in an empirical investigation of the unemployment insurance program, evaluating its effect on the key variables such as job search and reservation wage, seems to be more reasonable strategy. Yet there is a paucity of research studying the effects of recent UI extensions on the two key variables. Krueger and Mueller (2010) study the effect of regular UI benefits on job search intensity of unemployed individuals in the United States from 2003 to 2007. One of their findings is that the generosity of unemployment insurance negatively impacts job search activity among the unemployed (estimated elasticities range from –1.6 to –2.2). Econometric methodology and research question in this paper differ somewhat from Krueger and Mueller (2010). Particularly, in this paper I attempt to estimate the causal effect of the expansions in UI coverage during the Great Recession of 2007–2009 and its aftermath, on job search behavior of the unemployed.

In line with the labor supply literature, I focus on the effects on women and men separately. Women’s labor supply is considered to be more elastic than men’s. For example, Alesina, Ichino, and Karabarbounis (2011) argue for gender-based taxation because men are less responsive to market wages and compensation. They argue that because of historical, social and cultural reasons, men derive more utility from participating in the labor market even while market compensation is low. However, it is socially acceptable for women to fully engage in home production. The intuition of Alesina, Ichino, and Karabarbounis (2011) is equally valid in the UI system. Unemployment benefits could induce women to abandon the labor market in favor of home production.

Furthermore, the maximum amount of unemployment benefits, which is not tied to potential earnings, can replace only a small fraction of lost income or potential earnings for those who possess skills to command a higher salary. (An average beneficiary nationwide received about $300 a week in 2011). Hence, the UI program may be expected to create heterogeneous disincentive effects across groups. For example, educated individuals may find the UI benefits insufficient. In addition, the opportunity cost of not looking for a job and foregoing possible earning opportunities would be high. Thus we attempt to measure the disincentive effects across groups, including those with a college degree and those without it.

This paper’s objective is to focus on just one “channel,” where UI extension can be treated as plausibly exogenous. In my research design, I assume that eligibility for UI is not affected by search activity. I use the American Time Use Survey (ATUS) data, as it provides information about job search activities of unemployed individuals. This paper’s identification strategy uses unemployed persons eligible for UI benefits as a treatment group and those ineligible as a control group (e.g., Farber and Valletta 2013; Rothstein 2011; Krueger and Mueller 2010).

I employ the difference-in-differences (DD) method, since the extensions in 2008–2009 could be treated as a natural experiment or pseudo-experiment. Before proceeding to the actual estimation, I test if the control group could serve as a true counterfactual, so that the proposed DD methodology is a statistically valid approach. The post-treatment sample consists of the period 2010–2011. Even though the UI extension began in June 2008, Congress successively revised and introduced new tiers of extensions with different features and benefits until November 2009 (see Section 2 for details). More importantly, the extensions were still intact through 2010 and 2011 (and continued into 2013, albeit with some changes made in 2012). At the time of writing, the latest year of data available from the ATUS is 2011. I choose the period before June 2008, specifically from April 2004 to May 2008, as the pre-treatment period. Prior to March 2004, another moderately extended UI coverage – Temporary Extended Unemployment Compensation (TUEC) – was in effect. As UI coverage was extended during recessions in the past, the unemployed might have anticipated an extension in UI coverage, especially after the beginning of recession (December 2007) and might subsequently have reduced job search efforts. To overcome this anticipation effect, I also estimate models using pre-treatment period only until December 2007. It is worth mentioning that even though the potential duration of benefits has been extended, the replacement rate (ratio of the UI benefits relative to previous earnings) has been left unchanged. Therefore, the recipient received the weekly dollar amount of the extended benefits (EBs) equal to the regular benefit amount.

To preview the results, I find women are more responsive to the extended UI benefits (the effect for women is negative and statistically significant) than men. The average effect of the extensions of UI coverage on job search behavior of unemployed women is over a 10 percentage point decline in the probability of job search. I find that the probability of unemployed men searching for a job decreases by one percentage point; however, it is not statistically significant. I apply a placebo test to check whether the model is capturing actual treatment effects, not the pre-existing differential trends between treatment and comparison units.

This paper is related to the literature empirically investigating the disincentive effect of UI benefits. The majority of the literature that is focused on the effect of UI benefits on unemployment durations or exits (e.g., Katz and Meyer 1990; Meyer 1990; Card and Levine 2000; Jurajda and Tannery 2003) shows a wide range of disincentive effects. Furthermore, this paper complements a strand of the literature that looks at the disincentive effect of EBs associated with the Great Recession on unemployment rates. Rothstein (2011), along with examining the effect of UI extensions on unemployment exits and durations, investigates the effect on the unemployment rate, and finds a small effect of about a 0.1 to 0.5 percentage point increase in the unemployment rate. Farber and Valletta (2013) apply a similar approach to Rothstein (2011), and their estimates show a statistically significant and small increase in unemployment duration, and around a 0.4 percentage point rise in the unemployment rate. Fujita (2011) and Mazumder (2011) employ a simulation approach, based on the estimated parameters in the previous research, to measure the disincentive effect of UI on unemployment rates. They find that extended UI benefits contributed to around a 0.8 to 1.2 percentage point increase in the unemployment rate. Using a structural calibrated model, Nakajima (2012) finds that UI extensions of 2008–2009 caused the unemployment rate to increase by 1.4 percentage points, accounting for about 29% of the total unemployment rate increase in the United States between 2005–2007 and 2009–2011. This paper contributes to this literature by looking at the effect of extended UI benefits on job search behavior, a key variable affected by the UI program.

This paper is also related to a growing literature studying the job search effort of the unemployed. Aguiar, Hurst, and Karabarbounis (2013), who study how individuals utilize their lost work hours during the recent recession in the United States, find that around 2 to 6% of the lost market work hours were devoted to job search. Likewise, DeLoach and Kurt (2013), Gomme and Lkhagvasuren (2013), and Mukoyama, Patterson, and Sahin (2014) study the job search behavior over the business cycles using the ATUS data. Though not a primary objective of this paper, its results have implications for understanding the prediction of the standard search and matching model. For instance, Shimer (2005) finds that the textbook search and matching model is unable to account for the observed volatility of the labor market tightness, defined as the vacancy to unemployment ratio. The volatility might be affected by fluctuations in job search behavior over the business cycle due to changes in generosity of unemployment benefits. Note that the U.S. government has historically extended the unemployment insurance during periods of economic downturn.

The remainder of this paper proceeds as follows. Section 2 offers a brief description of the unemployment insurance system in the United States. Section 3 describes data and presents basic estimates. I design a difference-in-differences specification in Section 4. Section 5 presents results, and Section 6 includes placebo tests and further robustness checks. Section 7 presents its conclusions.

2 Background on unemployment insurance extensions

Established under the Social Security Act of 1935 as a joint federal-state program, the unemployment insurance (UI) system provides temporary financial support to workers who lose lobs. The UI aims to ease financial hardships of the unemployed and facilitate their job search behavior. Normally, states set the parameters of the regular UI benefits, including the maximum and minimum amount of the benefits, the potential duration of the benefits, and eligibility criteria. The duration of availability is typically 26 weeks in most states ^[1] during normal economic periods. The weekly benefit amount varies from state to sate. In 2011, an average worker received around $300 a week nationwide, replacing almost half of his or her previous earnings. To receive benefits, workers must lose their job through no fault of their own, must have a certain amount of previous earnings, and must actively search for work. Generally, new entrants and re-entrants into the labor force as well as those who voluntarily leave jobs or are fired for a cause like misconduct are not eligible for the benefits. According to Krueger and Meyer (2002), about 40% of all unemployed individuals (both eligible and ineligible) qualify for and receive UI benefits.

One important aspect of the U.S. unemployment insurance system is that the government has historically extended the duration of UI benefits during periods of economic downturn through two channels. Since 1970, EB, a joint federal-state program, has provided an additional 13 and 20 weeks of benefits in states exceeding the unemployment rate of 6.5% and 8%, respectively. On top of that, the federal government has extended the duration of UI coverage on a temporary basis to the unemployed during recessions or periods of severe unemployment. So far Congress has activated such temporary programs eight times – 1958, 1961, 1971, 1974, 1982, 1991, 2002, and 2008 (see Whittaker and Isaacs 2013).

As part of its response to severe unemployment brought about by the Great Recession, Congress enacted the Emergency Unemployment Compensation (EUC) program in June 2008 (see Appendix Table 13). In the beginning, the EUC offered an additional 13 weeks of benefits to the unemployed who exhausted their regular benefits. Congress expanded the duration of EUC benefits in successive phases until November 2009, increasing the duration of the benefits under the EUC up to 53 weeks. There are four tiers of EUC. EUC Tiers I and II offer an additional 20-weeks and 14-weeks of compensation for the unemployed in all states irrespective of unemployment rates. However, the remaining two tiers are based on unemployment rates. EUC Tier III provides benefits for a further 13 weeks in states where the unemployment rate exceeds 6%. EUC Tier IV is implemented in states exceeding 8.5% unemployment and offers an additional 6 weeks of benefits. In total, an unemployed individual in 2010 and 2011 was able to receive UI benefits as long as 99 weeks (26 weeks under the standard UI, 20 weeks under the EB, and 53 weeks under the EUC). These unprecedented extensions have been called the most generous in U.S. history (e.g., Barro 2010).

3 Data and basic estimates

This paper draws its main data from the ATUS as it is the only nationally representative survey that contains information about job search activity. This information enables me to directly investigate the effect of unemployment insurance on job search behavior. The U.S. Census Bureau conducts the ATUS, which is sponsored by the Bureau of Labor Statistics. The ATUS selects respondents from the outgoing sample of the Current Population Survey (CPS). Individuals aged 16 or above are surveyed in the ATUS, 2–5 months after their final CPS interview; about 71% after 3 months. Individuals are asked to report their detailed activities on the basis of their previous day’s time use diary. One of the several activities reported in the ATUS is the amount of time spent on job search. Job search includes making contacts to employers, looking at and responding to job advertisements, filling out applications, and traveling to job interviews. A complete list of activities used as job search in this paper is given in Table 14 in Appendix B.

I limit the data to unemployed individuals between ages 25 and 64 years in the main estimations since this group has a strong participation in, and inclination toward, the labor market. ^[2] This paper’s choice of a lower-bound age restriction follows Blau and Kahn (2007) and Cullen and Gruber (2000), in order to avoid complications related to choices of schooling versus market work by young people. According to Mulligans (2012), more than 75% of unemployed youths between ages of 16 and 24 years do not receive UI benefits. Upon reaching 65, an individual’s labor market behavior may change due to eligibility for social security and Medicare benefits. I use data from April 2004 to 2011. The ATUS began collecting data in 2003. Its 2003 survey consists of around 20,000 respondents. Since 2004, the number of respondents has decreased to approximately 13,000 each year. The sample (which excludes individuals who are employed, are out of the labor force or aged below 25 and above 64) includes 3,259 unemployed persons. The data are repeated cross-sections.

The definition of an unemployed person in the ATUS is the same as in the CPS. The definition includes those aged 16 years and older who are unemployed, but available for work in the reference week and who have actively been searching for work in the past 4 weeks. Search rule is exempted for those individuals who are expecting to be recalled to a job. (These individuals are still considered unemployed even if they have not looked for work in the past 4 weeks prior to the survey.)

The ATUS does not contain any explicit information about eligibility and ineligibility for UI benefits, nor does it have any information on whether eligible individuals indeed claim the benefits. However, the data contain information about individuals’ reasons for unemployment. Using these reasons for unemployment, I follow the classification strategy of Krueger and Mueller (2010) to infer an individuals’ eligibility for UI benefits. In recent years, this strategy of classification has become popular in the literature e.g., Farber and Valletta (2013), Rothstein (2011). I first divide unemployed persons into four categories: (i) job losers, (ii) those expecting a recall from their previous employer, (iii) voluntary job leavers, and (iv) new entrants and re-entrants into the labor force. Then, I categorize: (i) job losers as well as (ii) those expecting a recall from the previous employer as eligible for UI benefits, and (iii) voluntary job leavers as well as (iv) re- or new entrants into the labor force as the ineligible group. More specifically, each category is defined in the following way:

1. Job losers: This category includes persons who reported in the CPS interview that they were on layoff, or that they had become unemployed due to the ending of a temporary job, and are still unemployed in the ATUS interview. It also includes those who reported in the CPS that they are employed, but are unemployed in the ATUS interview. It could be recalled that CPS data were collected 2–5 months prior to the ATUS.

2. Expecting a recall: This category consists of unemployed persons who indicate in the ATUS interview that they are expecting a recall for work from their previous employer.

3. Voluntary job leavers: This category includes those who reported in the CPS that they quit their job and are unemployed in the ATUS. It is worth mentioning that the ATUS questionnaire does not include any information about job quitters, so I turn to the CPS data to identify them.

4. Re- and new entrants: This category includes those persons who reported in the CPS that they are out of the labor force, but who are unemployed at the time of the ATUS interview. It also includes those who reported in the CPS interview that they are re-entrants or new entrants in the labor force and are still unemployed at the time of the ATUS.

4 Difference-in-differences estimates for job search probability

My goal in this section is to assess the effect of UI extensions on job search behavior during the Great Recession and its aftermath. As the extensions could be seen as a pseudo-experiment, I apply a difference-in-differences (DD) regression, which is very popular in the literature to measure the policy effect or intervention. I use the period from March 2004 to May 2008 as the pre-treatment period and 2010–2011 as the post-treatment period. One of the major issues in DD analysis is whether the control units in a study can truly serve as the counterfactual of the treatment units. If there are systematic differences between the comparison and treatment units, the estimates from the DD approach may reflect results of their differences, not the policy effect (treatment). This is true even if the unconfoundedness assumption still holds in the regression. One indirect way of testing such a systematic difference is to test if there is a sufficient overlap in the covariates between treatment and control units. For this purpose, one may think of an conventional approach which is to use a t-test:

However, the major drawback with this test is that it is very sensitive to sample size, and it tends to over-reject the null hypothesis when the sample size increases. To address this limitation, Imbens and Wooldridge (2009) suggest using a normalized difference (which is also scale-invariant), where the difference is:

The authors also suggest that researchers can use the threshold of one-quarter normalized difference as a rule of thumb. Exceeding this specified threshold implies that covariates are different between control and treatment units. In Tables 5 and 6, I report normalized differences for each individual-specific covariates used in the DD model for women and men respectively. The differences are calculated for the control and treatment groups during pre- and post-treatment periods, separately. Almost all of the differences are less than one-quarter. (For women, only one variable – a dummy variable for the white – is slightly higher than one-quarter, and for men, age variables slightly exceed this threshold.) Overall, this suggests that individual characteristics are not systematically different by the treatment status. Therefore, the difference-in-differences estimator could be seen as a plausible empirical strategy.

As many respondents in the sample are not participating in job search (i.e., they have zero-search) on a given diary day, I am interested in looking at the impact of the UI benefits on the probability of their decision to participate in the job search activity. (In an alternative specification and robustness check later, I actually look at the effect on job search intensity using the DD method in the form of eq. [3].) It is important to note that the probability of job search in this paper could be viewed as the probability of sampling an individual who spends a positive amount of time in job search activities. It is true that to be counted as unemployed in the CPS data, an individual should have at least one active job search in the last 4 weeks preceding the surveyed date. Note that respondents in the ATUS data are drawn from the CPS’s respondents, and the ATUS follows the CPS’s definition of unemployment. In the CPS (and the ATUS), active job search refers to an activity which could land an individual on a job without any additional work. Otherwise, an activity is called passive, which is not enough for an individual to be categorized as an unemployed. For example, looking at an ad on newspaper is a passive activity. However, in the definition of job search in this paper, both active and passive activities are included. I have listed a complete list of job search activities in Appendix A.

One of the shortcomings of time diary data is that we can observe a respondent’s activities only on a particular day. Hence, it is possible that the respondent might have spent time looking for a job on other days. Nonetheless, sampling an individual with positive amount of time devoted to job search could be a close proxy to the job search probability of the individual. The intensity with which a job seeker pursues a position should be positively correlated with the probability to have a positive amount of job search on a random day of the survey. For instance, if an individual has only one active job search in the past 4 weeks and is categorized as unemployed, it is very likely that we observe zero job search probability on the surveyed day. On a given surveyed day, around 18.38% unemployed women have greater than zero job search activities (i.e., around 81.62% of them have zero job search). Likewise, around 26.94% unemployed men have greater than zero job search activities. Table 3 contains further details on how many respondents have job search activities greater than zero on a diary day, by the type of unemployment before and after the treatment. Note that the probability of the participation represents the share of unemployed individuals in the sample who have a greater than zero job search on a given diary day.

One advantage of using the binary dependent variable (probability of job search instead of job search intensity) is that it is less prone to measurement error. The reason is that it is easier for individuals to remember whether they looked for a job on a given diary day, rather than recalling the amount of time they spent looking for a job. Specifically, I estimate the logit DD model below on repeated cross-sectional data to measure the causal effects of UI extensions on job search decision.

where i = 1, 2, 3,…N denotes individuals and s represents state. The outcome variable, yi(t,s), is an indicator variable equal to 1 if an individual participates in job search; otherwise 0. Λ(.) is the logistic cumulative distribution function. The variable Postt is a dichotomous variable which equals 1 if it is the post-extension period. The variable Treati is a dummy variable, equal to 1 if an individual is eligible for UI benefits. The variable Postt∗Treati is the variable of interest (its coefficient is the treatment effect). The treatment and control groups have the same linear time trend. (In a later robustness check, instead of time trend I use a year fixed-effect model, and also use a quadratic time trend.) xi(t,s) is a vector of individual characteristics, including age, age-squared, level of education, race, sex, presence of child in house, and marital status. ^[3] The variable AWBs is average weekly UI benefits (which represents the average dollar amount of benefits received by a recipient per week, calculated by month and state). ^[4] I also include other control variables described below.

The unemployment rate (un_rates): One of the major factors that might influence an individual’s job search behavior is a perception about the macroeconomic environment. Unemployed individuals may not look for jobs during a period of severe unemployment if they perceive that jobs may not be available and that the marginal cost of job search will be higher than the expected payoff. Conversely, it could be argued that an individual needs to put extra time and effort to find a job in a period of mass unemployment due to a decline in jobs arrival rate. And, the literature is not conclusive about the cyclical property of job search behavior. DeLoach and Kurt (2013) and Gomme and Lkhagvasuren (2013) find job search behavior to be pro-cyclical (unemployed decrease search efforts during the labor market slack), while Mukoyama, Patterson, and Sahin (2014) show that the aggregate job search effort is countercyclical. To capture the potential change in an individual’s job search behavior, I control for the monthly unemployment rate by state. ^[5] In the ATUS data, I can observe the month of interview of the respondents, so I am able to match the interview month with the monthly unemployment rate.

The unemployment duration (un_duri): The canonical model of Mortensen (1977) predicts that unemployed individuals ineligible for UI benefits put constant effort toward job search irrespective of their unemployment duration, while eligible ones speed up their search effort when their benefits are about to expire. Therefore, controlling for unemployment duration is essential. However, the ATUS data do not include the unemployment duration of individuals. As in Krueger and Mueller (2010), I use the duration reported by unemployed individuals in the CPS interviews. I then add the number of weeks between their CPS and ATUS interview to the duration reported in the CPS. However, for the respondents who are not unemployed in the CPS interview but are in the ATUS, I am not able to observe their duration. Therefore, I impute a proxy duration for them, which is half of the weeks between the CPS and ATUS interviews.

Potential wage (ln(w)^i(t,s)): Potential wage is believed to influence an unemployed individual’s behavior. For those unemployed persons who expect a higher wage in the labor market, unemployment benefits might have minimal effects as their opportunity cost of staying unemployed would be high. For the low earnings group, the weekly benefits can replace a significant portion of their potential earnings, encouraging them to substitute their labor with leisure. However, the data do not contain information about previous earnings of the unemployed, even if they have work history. To deal with this wage issue, I estimate a predicted wage for each individual. First I use the Current Population Survey’s (CPS) Outgoing Rotation Group data and estimate the following state-fixed effect model separately for each year ^[6] for log-wage:

where xis is a vector of individual characteristics, which include the level of education, age, age-squared, gender, marital status, and presence of children. I also use state dummies. Then, on the basis of estimated parameters from eq. [4] for a particular year, I predict expected log-wage for each respondent in the ATUS for that year. ^[7]

Distribution of the potential wage offer (stateresids): Another determinant of job search is the potential wage offer distribution. The rationale behind this is that the more dispersed the distribution is, the more time individuals spend looking for a job as they try to get the best possible wage. As a proxy for the potential wage-offer distribution, I use residual wage dispersion calculated as the standard deviation of the residual (εi,s) from wage eq. [4] for each state and year (see Krueger and Mueller 2010).

For the non-linear DD model, the interpretation of the policy parameter, δ, is different from the case of a linear regression. The average treatment effect on the treated (ATT) equals the differences in the two cross differences (e.g., Eissa and Liebman 1996; Puhani 2012). Let τ be the effect:

where Zi(t,s) is a vector of all control variables used in regression [3]. Bertrand, Duflo, and Mullainathan (2004) show that the conventional DD standard errors are severely understated due to serial correlation. As suggested by the authors to correct for possible serial correlation, I calculate the variance–covariance matrix of errors, clustering at the state level, from the regression of eq. [3]. ^[8] Then, I apply the Delta Method to calculate standard errors for estimates.

5 Results

In this section, I present results from a variety of specifications of difference-in-differences (DD) estimation based on eq. [3] to explore the causal effect of the recent UI extensions on job search behavior across gender. The focus of this paper on gender differences in the impact of the extensions in UI benefits is in line with the labor supply literature.

I estimate four-baseline nonlinear DD models for men and women separately; the results are presented in Tables 7 and 8. The first three models include same control group, but slightly different treatment groups. The first model consists of the full sample (all individuals eligible for UI benefits as treatment group). In the second and third models I attempt to exclude those eligible unemployed persons, who are perceived to have exhausted their UI coverage. In the pre-treatment period, UI benefits were usually available for 26 weeks. Therefore, I exclude those eligible individuals having the duration of unemployment more than 26 weeks in the pre-treatment time (from the treatment group). In the second model, I exclude all individuals who are unemployed more than 99 weeks in 2010 and 2011 (from the treatment group), as they were eligible for UI benefits up to 99 weeks. In the third model, I exclude all eligible individuals who are unemployed more than 60 weeks in the post-treatment period (for the reasons explained in the earlier Section).

In model 4, I exclude those unemployed persons who are on temporary layoff and expecting a recall from their previous employer (from both the control and treatment group), and those who voluntarily left the job (from the control group). It is expected that these people behave quite differently from the average labor market participant. The sample then consists of individuals who are on layoff through no fault of their own as eligible (treatment) group, and re- or new entrants in the labor force as ineligible (control) group.

In all these models, I find consistently close results (Table 7) for women. I get positive and significant effect of the eligibility status (i.e., the variable eligible) on job search behavior. One of the potential reasons for having higher job search probability for eligible group could be the requirement of job search to collect UI benefits. And, it is possible that eligible individuals might have kept up looking for a job; however, they might have refused to take up job offers, increasing reservation wage and extending the duration of unemployment (another disincentive effect of UI benefits which I do not explore here). Another possibility for the higher probability of job search by the eligible group than the ineligible group could be the result of the initial differences in job search behavior. It could be noted that the eligible group mostly consists of those who are laid off from work, and the ineligible group consists of re- or new entrants in the labor force. Arguably, a layoff can be considered as a shock to the worker. In light of the shock, the laid-off worker could be in greater need of a job to smooth or maintain consumption than that of a re- or new entrant who can have a relatively longer period to adjust for consumption. Nonetheless, this variable does not represent the effect of UI extensions.

The interaction term between eligibility and post-treatment period (referred as Treati×Postt) is the variable of interest as its coefficient is the implied causal effect of treatment (extensions of UI benefits). In the technical language of program evaluation, the coefficient is termed as average treatment effect on the treated (ATT). The average effect of the UI extensions in 2010 and 2011 is over 10 percentage points decline in the probability that an eligible unemployed woman searches a job. It is statistically significant at the 5% level in the first and third model, and at the 1% level in the second and fourth models. This represents around a 56% of the average predicted probability of job search (i.e., the ratio of the effect to \bar \hat p). I find around a one percentage point decline in the probability of job search for men (Table 8). However, it is not statistically significant.

6 Robustness check

7 Conclusion

In this paper, I investigate the effects of unemployment insurance (UI) extensions on job search, using data from the ATUS. The government made successive decisions about the extensions of UI coverage from June 2008 to November 2009, raising the duration of benefits to as many as 99 weeks. The extensions that are termed as the most generous in history were continued into 2011. In the research design of this paper, I use unemployed persons eligible for UI benefits as the treatment group and the ineligible as the control group.

I apply a difference-in-differences estimator, as the extensions provide a pseudo-experiment. My focus is gender differences in the effect of UI extensions. I find a negative and statistically significant effect on women, but for men the estimate is not statistically different from zero. The average treatment effect for women is around a 10 percentage point decline in a job search probability.

Some caveats apply to this study. Presumably, there have been some classification errors when it comes to classifying unemployed workers into eligible (for UI benefits) and ineligible groups. On top of that, I use eligible workers as the treatment group. In practice, a moderate portion of the eligible workers do not claim benefits (which makes it inappropriate to put them into treatment group). Both of these limitations will likely cause the results to be underestimated (i.e., downward bias). If the ATUS begins collecting information about UI benefits received by individuals, more precise evaluation of the UI benefits system will be possible.