1 Online job search and unemployment insurance during the Great Recession Ioana Marinescu, University of Chicago [PRELIMINARY; DO NOT QUOTE WITHOUT AUTHOR S PERMISSION.] Abstract The recession in the U.S. was accompanied by large increases in the maximum duration of unemployment benefits. I use instrumental variables to identify the impact of weeks of unemployment benefits on applications received by jobs on Careerbuilder.com, the largest American employment website. A one week increase in maximum benefit duration decreases state level applications by 0.65%, which implies that the increase in benefit duration during led to a 39% decrease in applications in the average state.
2 1. Introduction Unemployment benefits can affect the job finding rate through two main channels: job search effort and the reservation wage. More generous unemployment benefits are predicted to decrease search effort and increase the reservation wage, i.e. the minimum wage that induces an unemployed worker to take a job. Recent empirical evidence has shown that reservation wages are generally not affected by unemployment benefits (Card, Chetty, and Weber 2007; Krueger and Mueller 2011). Therefore, the impact of unemployment benefits on unemployment must be explained by job search efforts. While a large literature has documented the negative impact of unemployment benefits on job finding (a good recent reference is Schmieder, Wachter, and Bender 2010), there is currently little to no investigation of the impact of unemployment benefits on job search effort. Krueger and Mueller (2010) are a recent exception. They examine the impact of unemployment benefits on time spent on job search activities using cross state variation in the generosity of unemployment benefits. While they do find that higher unemployment benefits decrease the time devoted to job search, their identification strategy is limited by omitted state level covariates and endogeneity, since high unemployment states may choose to enact more generous benefits. This paper investigates the impact of unemployment benefits on job search using plausibly exogenous variation in unemployment benefit duration during I measure search efforts at the statemonth level by the number of applications received by jobs on CarreerBuilder.com, the largest American employment website. Krueger and Mueller (2011) show that unemployment benefits recipients allocate between 24% and 38% of the time spent on job search activities to answering ads, sending applications and resumes, with an additional 27% of the time allocated to looking at job ads. This suggests that applications on a job search website are a good measure of job search effort. To identify the impact of unemployment insurance on applications, I use variation in unemployment benefits duration induced by the Emergency Unemployment Compensation 2008 (EUC) and extended benefits (EB). Benefit extensions depend among other things on state level unemployment rates reaching specific thresholds. In my preferred specification, I instrument the number of weeks of benefits with program rules, and I focus on states whose unemployment rate is close to unemployment rate thresholds that determine benefit duration. I find that a one week increase in unemployment benefits leads to a 0.65% decline in applications at the state level. This effect is large since it implies that the average benefit extensions between 2008 and 2009 (26 weeks to 86 weeks) led to a 39% decline in applications for the average state.
3 I also investigate the impact of unemployment benefit extensions on the number of unemployed people at the state level, using the same instrumental variable strategy as previously. In theory, it is possible for the impact of a decline in applications on unemployment to be very small: indeed, unemployed workers may be sending too many applications that have a low probability to result in a match. In fact, I find that a one week increase in unemployment benefits leads to a 0.43% increase in unemployment at the state level, which implies that the average benefit extensions between 2008 and 2009 (26 weeks to 86 weeks) led to a 26% increase in unemployment for the average state. Therefore, it seems plausible that the application decrease was instrumental in generating a higher level of unemployment. Finally, I show that the increase in unemployment generated by benefit extensions is large enough to fully account for the outward shift in the Beveridge curve during the Great Recession. In other terms, in the absence of benefit extensions, the increase in unemployment during the Great Recession would have probably been in line with the drop off in vacancies. This paper is closely related to the literature on the impact of unemployment benefits on job finding. A recent working paper (Rothstein 2011) examines the impact of EUC 2008 and EB on job finding rates using the Current Population Survey and finds that extensions indeed significantly decrease job finding rates. My paper s innovation is to examine the impact of benefit extensions on job search effort directly, while controlling for the supply of jobs, i.e. job vacancies. Using applications as an outcome instead of job finding has important advantages. Indeed, the impact of benefit extensions on job search effort is direct while the impact on job finding is mediated by job search effort and the state of the labor market. This direct relationship between unemployment benefits and job search allows me to focus on thresholds in unemployment rate that determine large changes in benefit duration, which strengthens my identification strategy. As mentioned above, there is a sparse and recent literature examining the impact of unemployment benefits on job search. Krueger and Mueller (2010) use cross state variation to investigate the impact of benefit levels on time spent on job search. Krueger and Mueller (2011) show results that are consistent with a negative impact of unemployment benefit extensions on time spent on job search during the Great Recession. However, given that their data comes from a single state (New Jersey), the identification is based only on time variation. In comparison, this paper is able to identify the impact of unemployment insurance on job search by using data on multiple states that saw different extensions in benefits at different times. This paper is thus the first to my knowledge to credibly identify the impact of unemployment benefit duration on job search.
4 The remainder of the paper is organized as follows. Section 2 discusses how unemployment benefit extensions were decided, describes the data and the identification strategy. Section 3 presents the key results. Section 4 discusses the results, and in particular their relevance to the Beveridge curve. Finally, Section 5 concludes. 2. Policy background, data and identification strategy Policy background Standard unemployment benefit duration in the United States is 26 weeks. However, during times of high unemployment, this duration can be extended based on state level determinants. The extended benefits (EB) program activates in a state under one of two conditions: (1) if the state's 13 week average insured unemployment rate (IUR) in the most recent 13 weeks is at least 5.0 percent and at least 120 percent of the average of its 13 week IURs in the last 2 years for the same 13 week calendar period; or (2) at state option, if its current 13 week average IUR is at least 6.0 percent, and regardless of the experience in previous years. I will say that the IUR option is in place if the state chose this option. States have the option of electing an alternative trigger authorized by the Unemployment Compensation Amendments of 1992 (Public Law ). I will say that the TUR option exists if the state chose this option. This trigger is based on a 3 month average total unemployment rate (TUR) using seasonally adjusted data. EB is turned on: If this TUR average exceeds 6.5 percent and is at least 110 percent of the same measure in either of the prior 2 years, a State can offer 13 weeks of EB. If the average TUR exceeds 8 percent and meets the same 110 percent test, 20 weeks of EB can be offered. Normally, extended benefits are financed 50% by states and 50% by the federal government. Under the American Recovery and Reinvestment Act of 2009 (ARRA) passed on Feb. 17, 2009, the benefits are financed entirely by the federal government. This provided many states with an incentive to choose the TUR option (the IUR option is mostly irrelevant because few states reach 6% IUR without already having
5 EB under the regular IUR condition). Federal funding of EB is currently set to expire on December 31, The Tax Relief, Unemployment Insurance Reauthorization, and Job Creation Act of 2010 (P.L , passed on Dec and signed on Dec. 17, 2010) temporarily changed the look back to three years, as unemployment indicators in most states have been consistently high for the past two years and would have resulted in many states being unable to meet the 120% IUR or 110% TUR conditions. This three year look back exception is set to end on December 31, As of Feb , Arkansas, Iowa, Louisiana, Maryland, Mississippi, Montana, Oklahoma, Utah, and Wyoming could qualify for extended benefits under TUR but chose not to use that option. Another reason why unemployment benefits were extended during the Great Recession is the federal Emergency Unemployment Compensation (EUC) EUC08 is an emergency federal benefits program that is payable to individuals who have exhausted all rights to regular compensation with respect to a benefit year that ended on or after May 1, 2007, and have no rights to regular compensation or extended benefits (EB). Therefore, EUC benefits come into play after any EB benefits expire. The EUC08 program, signed into law on June 30, 2008, provides up to 13 weeks of 100 percent federally financed compensation to eligible individuals in all states. Public Law (P.L.) expanded the EUC08 program on November 21, 2008 to provide up to 20 weeks of 100 percent federally funded unemployment compensation to eligible individuals in all states. This constitutes tier 1 or EUC1. Tier 2 of EUC (EUC2) was created by Public Law (P.L.) It provides 13 weeks of benefits to eligible individuals in states where TUR (defined as for EB) is above 6% or IUR (defined as for EB) is above 4%. Public Law No , enacted on November 6, 2009, expanded the EUC08 program, in the following ways: o It increased the maximum EUC2 entitlement from 13 weeks to 14 weeks of benefits in all states, and this Tier is no longer triggered on by a state reaching a specified rate of unemployment; o It created EUC3 providing up to 13 additional weeks of benefits in states with IUR above 4 percent or TUR above 6 percent;
6 o It created EUC4 providing up to 6 additional weeks of benefits in states with IUR above 6 percent or TUR above 8.5 percent. Just as in the case of EB, the TUR conditions are much more likely to be satisfied than the IUR conditions. For example, EUC2 and EUC3 require that TUR be above 6% or IUR above 4%. In my sample, when TUR is above 6%, IUR is above 4% in 97% of the cases. On the other hand, when IUR is below 4%, TUR is nonetheless above 6% in 54% of the cases. This overview of the policies suggests that one can use sharp changes in benefit duration at 6.5% and 8% TUR (for EB) and 6% and 8.5% TUR (for EUC) to identify the impact of benefit duration on applications. I will discuss below how I make use of these in practice. Data The data on applications and job vacancies comes from proprietary data provided to me by CareerBuilder.com. Job vacancies are the total number of job ads posted in a given state during a given month. The data spans September 2007 to July An application is defined as a person clicking on the Apply Now button in a job ad. Applications are the number of applications received by all jobs in a given state and month. These applications can therefore partially come from out of state jobseekers. However, jobseekers overwhelmingly apply in state, so in first approximation we can consider the applications to come from jobseekers in the state where the job is posted. Additionally, when an individual state experiences an increase in benefit duration, there is no strong reason to expect that applications from out of state will also decline. One should also note that the recorded applications are not coming from unemployment benefit recipients only. According to CarrerBuilder s applicant survey, about half of the applicants are employed 1. As such, applications measure job search effort for individuals in a given state, whether or not they are employed. Employed individuals may also react to increases in the duration of unemployment benefits by searching less; this would particularly apply to employed individuals who engage in job search because they think they may lose their job soon. However, since the end of unemployment benefits is further away for a currently employed worker who envisions becoming unemployed than for a currently unemployed worker, discounting predicts that the impact of benefit duration on employed workers search effort should be lower (see Card, Chetty and Weber, 2007, for a discussion of the role of discounting in estimating the impact of benefit duration on unemployment duration). Therefore, the estimated impact of unemployment benefits on applications is 1 This statistic is interesting but cannot be taken at face value since it is based on the selected sample of those applicants who were willing to answer the survey
7 a lower bound on the impact of unemployment benefits on the job search effort of unemployed workers. This estimate is interesting in its own right because it constitutes a measure of the impact of unemployment benefits on overall labor market tightness at the state level. I use data from the Department of Labor EB and EUC trigger notices to determine when the conditions for each extension are realized. This data contains all the relevant variables by state and week: TUR, IUR and applicable look back criteria. Using statutory regulations, I can thus determine the maximum weeks of benefits available in a given state or month based on rules only. I calculate two versions of the statutory weeks of benefits. One assumes that all states use both optional IUR and TUR triggers for EB. The other version takes out the TUR trigger conditions and will only be used in specifications restricted to states that never had a TUR option in place during my sample frame. Since this data is at weekly frequency, I average the weeks of benefit at the monthly level to merge it with the data on applications and vacancies. At the same time, I define the maximum of IUR and TUR in each month, and merge these values in with the data on applications and vacancies. I take the maximum of IUR and TUR and not the average so that I can clearly determine whether a IUR or TUR threshold was ever crossed during a given month. To determine the number of weeks of unemployment benefits that are actually available, I use data from the Department of Labor detailing monthly first payments made by states for each of the EUC and EB 2. I compute weeks available by using policy rules, and requiring that the state made some first payments in that month. There are differences between this variable and the number of weeks based on rules because payments tend to start later than when conditions are met (on average two months later), and because the variable based on rules ignores states specific choice of TUR and IUR options. The choice of TUR and IUR options appears to be indeed highly endogeneous: in particular, after ARRA is passed (making EB fully federally funded), and when states reach the optional TUR threshold, they are much more likely to elect the TUR option. Finally, data on the total number of unemployed people by state and month comes from the Bureau of Labor Statistics. 2 I have checked the correspondence between this dataset on payments and the states status based on program rules. I eliminated the only obvious inconsistency: Louisiana erratically reports some extended benefits payments, even though according to the rules the state never had extended benefits during my sample frame. I therefore chose to assume that indeed Louisiana never had extended benefits, so that actual weeks of extended benefits are always 0.
8 Table 1 shows summary statistics for key variables. Notice in particular the large number of applications per state. The number of applications is about twice as high as the number of unemployed individuals, and about 30 times as high as vacancies, implying that each vacancy receives about 30 applications on average. Other statistics on the unemployment rate are familiar. Identification strategy The outcome of interest is applications at the state level, and the key explanatory variable is the number of weeks of unemployment benefit available. I will start with using a state panel. Applications, vacancies and the number of unemployed people are all in logs. I regress applications on the maximum weeks of benefits available, controls, and state, year and quarter fixed effects. I cluster standard errors at the state level. I always control for a quadratic in vacancies because the number of applications should positively depend on the number of vacancies: mechanically, applications are to these specific vacancies. When only controlling for vacancies, the panel specification identifies the impact of maximum weeks of unemployment benefits on applications using within state variation in weeks of benefit available. However, both applications and benefits depend on unemployment. Applications should increase with the number of unemployed people. Additionally, benefit extensions are conditional on states reaching certain thresholds of the unemployment rate, and controlling for flexible functions of IUR and TUR allows the estimates to mostly come from discontinuities in weeks of benefits induced by IUR and TUR thresholds. This suggests that one should control for unemployment. Concretely, when controlling for unemployment, I use a quadratic in the number of unemployed people, a cubic in TUR and a cubic in IUR. On the other hand, unemployment is endogenous in this specification: indeed, lower job search effort in terms of applications should lead to higher unemployment. To limit the scope of this problem, when controlling for unemployment, I always use lagged values 3. In regressions that do not control for unemployment, some of the applications response to increased benefit duration may be due to changes in state specific economic conditions that are correlated with unemployment but not explained by job vacancies. It is important to note that benefit duration and unemployment are positively correlated due to regulations: for this reason, not controlling for unemployment could lead to an underestimate of the impact of benefit duration on applications. A second specification uses the same panel regressions, but instruments the maximum weeks available with the weeks available according to rules. As mentioned above, the maximum weeks available differs 3 This does not fully address the endogeneity issue because of autocorrelation over time in the unemployment variables.
9 from what is determined by rules because of delays in the payment of benefits, and because some states choose not to enact TUR and IUR options for EB. However, the choice of states to enact TUR and IUR options is likely endogenous. In particular, states that expect to experience high unemployment rates have a strong incentive to enact TUR after ARRA was passed in February Indeed, ARRA makes EB fully federally funded. Therefore, it makes sense to use the instrument in order to circumvent this source of endogeneity. These instrumental variables regressions thus estimate the impact of weeks of unemployment benefit on applications using only plausibly exogenous within state variation in weeks of benefits available. A third specification makes use of TUR thresholds that determine benefit extensions for both EUC and EB. These thresholds are 6.5 % and 8% TUR for EB under the TUR option, and they are 6% and 8.5% TUR for EUC. As mentioned above, the IUR thresholds are less likely to be determinant since they are almost always reached after the TUR thresholds are reached. Still, reaching the TUR threshold is not enough: it must also be that unemployment is strongly increasing as determined by the look back conditions. Finally, delays in benefit payment intervene even when all thresholds are reached. For these reasons, this is not exactly a regression discontinuity design, though the identification strategy I use here is similar to what was done in Angrist and Lavy (1999). Instead of having a sharp discontinuity in benefit duration around TUR thresholds, we expect a strong increase in benefit duration around these thresholds. In practice I use three different sample to focus on various TUR thresholds, and in each case I report an OLS specification, and an IV specification where actual weeks of benefit available are instrumented with weeks available based on rules. All these specifications include state fixed effects, quarter fixed effects, and a quadratic in vacancies. I do not use year fixed effects because higher TURs occur systematically at later dates during my sample frame, and I therefore lack power when using year fixed effects. I experiment with including a quadratic in the lagged number of unemployed people, with the caveat that this variable may be endogenous. I use three different samples. The first one restricts TUR to be between 5.5 and 7%, and includes the 6% (EUC) and 6.5% (EB) discontinuities. The second one restricts TUR to be between 7.5 and 9%, and includes the 8% (EB) and 8.5% (EUC) discontinuities. Finally, the third sample restricts TUR to be between 5 and 7%, and states to those that never had a TUR option during my sample frame. As such, this third sample only includes the 6% TUR discontinuity coming from EUC rules. The instrument I use for weeks of benefit does not take into account the rules from the TUR option since these states never have a TUR option. These specifications estimate the impact of weeks of unemployment benefits on applications using the discontinuities in weeks of benefit based on TUR.
10 3. Results Panel regressions By examining the across state average of key variables over time (not shown), one can see that the maximum weeks of benefits available tracks pretty closely the number of unemployed people. One can notice a downward trend in applications during the recession (Dec to June 2009), despite the fact that at the same time the number of unemployed people increases sharply. In particular, applications drop by about 50% after February 2009 (when ARRA was passed) and through July 2009, which suggests that unemployment benefit extensions, and in particular EB, may play a role in the application decline. We can also note that the number of applications is positively correlated with the number of vacancies. This explains why, in , applications increase again as the number of vacancies increases. Table 2 presents panel regression results. In the first column, I find that maximum weeks of benefit available do not have a significant impact on applications; the point estimate is small and positive. Note that vacancies and the number of unemployed people both have a significant impact on applications. At sample means, the impact of both the number of vacancies and the number of unemployed people on applications is positive. In column 2, I instrument maximum weeks available by the number of weeks available based on program rules (see section 2). The coefficient on weeks of benefits becomes negative and significant at the 10% level, implying that a one week increase in unemployment benefits decreases applications by 0.4%. In columns 3 and 4, I drop controls for unemployment, since unemployment may be endogenous. In this case, both OLS and IV specifications show a negative and significant impact of weeks of benefit on the number of applications. In the IV specification (col. 4), I find that a one week increase in benefit duration is associated with a 0.5% decline in applications. This effect is similar in magnitude to what was found in column 2, when controlling for unemployment, which suggests that controlling for unemployment does not introduce a large bias in the estimate. The estimated effect of benefit duration on applications is substantial since it implies that the average benefit extensions between 2008 and 2009 (26 weeks to 86 weeks) were associated with a 30% decline in applications. Overall, I find that during the recession the increase in maximum benefit duration was associated with a significant decline in job search effort, with applications declining by 30%. The estimated decline in applications due to increases in benefit duration is only slightly smaller than the decline in applications observed in the raw application data. This suggests that increases in maximum benefit duration can account for most of the decline in applications observed in the raw data in However, one may question the causal interpretation of these estimates based on panel data. I have controlled for
11 flexible functions of IUR and TUR, which should have allowed me to identify the impact of maximum weeks of benefit mostly from discontinuities in the maximum weeks of benefit around TUR and IUR thresholds. At the same time, I have used some data that is far away from the thresholds and therefore the identification strategy is not as strong as what one could get by focusing on discontinuities in weeks of benefit introduced by program rules. To strengthen our confidence in these results, I will now exploit discontinuities in maximum weeks of benefits by restricting the sample to observations close to TUR thresholds. Focusing on discontinuities in program rules As mentioned in section 2 above, the maximum weeks of benefits available depend on reaching certain thresholds in IUR and TUR, and this is true both for EB and EUC. However, the TUR thresholds are generally reached before the IUR thresholds are reached. As such, to exploit discontinuities in maximum weeks of benefit available, one should focus on TUR. EUC2 between November 2008 and October 2009 and EUC3 from November 2009 both depend on TUR being above 6%. EUC4 depends on TUR being above 8.5%. These discontinuities due to EUC are made somewhat fuzzy by existence of the IUR condition, as well as the delays in implementation. EB, when the TUR option exists, depends on TUR being above 6.5% to deliver 13 weeks of benefits and TUR above 8% to deliver 20 weeks of benefits. However, for EB, it is additionally necessary that current TUR is high relative to TUR one to three years ago (see section 2 above). EB discontinuities are therefore fuzzier than EUC discontinuities because TUR and IUR also need to satisfy a trend condition. Figure 1 plots a lowess smooth of maximal weeks of UI available (solid line) and maximum weeks of UI available based on program rules (dashed line). Solid vertical lines represent the thresholds for EUC, and dashed vertical lines represent the thresholds for EB. There is a strong increase in weeks of benefits available for both the actual measure and the measure based on rules at 6%. At 6.5%, the maximum weeks available based on rules increase more than weeks available. This is because not all states have the TUR option for EB. There is an increase in weeks available according to both measures at 8%. However, at 8.5% the increase is barely noticeable for EUC4. Generally, while increases in weeks of benefit available around the threshold are very large, especially for the 6 and 6.5% thresholds, there is no actual discontinuity. This is for the following reasons: program rules are not based on the level of TUR only, there were delays in implementation, the data is aggregated to the monthly level from the weekly
12 level, the graph uses all the data from late 2007 to 2011 while EUC TUR discontinuities only existed from November 2008 on 4. Figure 2 plots the number of weeks of UI available according to rules against the log number of applications. We can see that applications decline across the 6% and 6.5% thresholds, but continue to increase afterwards. However, this data comes from different samples of states at different levels of the TUR. Therefore, it is useful to calculate residuals from a regression of applications on state fixed effects. Figure 3 shows a lowess plot of these residuals against TUR. The impact of TUR thresholds on applications is now much more visible: application residuals decline both across the 6% and 6.5% thresholds and across the 8 and 8.5% thresholds. One issue is that the TUR thresholds 6 and 6.5%, and 8 and 8.5% are close together. Therefore, one cannot hope to separately identify the impact of each of the discontinuities on applications. In order to focus on a sharper discontinuity, I restrict the sample to states that never had a TUR option during my sample. These states should then only be affected by the EUC thresholds at 6% and 8.5%. Figure 4 shows a lowess of the application residuals for these states on TUR. Maximum weeks of benefits available according to program rules are calculated in such a way that they do not use the optional TUR rules. For these states, there is essentially no data around 8.5%. So I focus on the 6% discontinuity. The figure clearly shows a large decline in application residuals across the 6% threshold. In Table 3, I estimate the impact of benefit extensions on application using samples close to discontinuities in TUR. In columns 1 3, I use the sample for which TUR>=5.5 and TUR<=7, which contains the 6% and the 6.5% discontinuities. OLS and IV estimates in columns 1 2 both show a negative and significant impact of weeks of benefits on applications. The IV estimate in column 2 implies that a oneweek increase in the maximum duration of benefits leads to a 0.57% decline in applications. In column 3, I repeat the IV estimate for column 2, but excluding potentially endogenous unemployment controls. The estimate remains negative and significant, but is smaller. In columns 4 6, I repeat the exercise in columns 1 3, but for the sample with TUR>=7.5% and TUR<=9%. I find negative and significant impacts of weeks of benefits on applications in all specifications. Interestingly, in this case IV estimates with or without unemployment controls are almost identical. The IV specification in column 5 shows that a one 4 This choice was made to maximize the amount of data that can be used for the plot, especially at low levels of unemployment. TUR discontinuities for EB existed also before November 2008 for those states that had the TUR option. If we restrict to the period after November 2008, we have less data at low unemployment levels, and there is still no sharp discontinuity in benefit duration around the thresholds because of all the other reasons listed above.
13 week increase in the maximal duration of benefits decreases applications by 0.65%. Finally, in columns 7 9, I repeat the exercise for the sample of states that never had the TUR option for EB and where TUR>=5% and TUR<=7%. This focuses on the EUC discontinuity at 6% TUR. OLS and IV estimates both yield a negative and significant impact of maximum weeks of UI on applications of about 1%. When not controlling for unemployment in column 9, the estimate size drops slightly and the coefficient becomes marginally insignificant, but the point estimate is not statistically significantly different from the estimate in column 8. Overall, I conclude that the estimates from the three discontinuity samples yield consistent results, with very similar magnitudes. I consider the estimate in column 5 to be the most credible because it is not sensitive to excluding the potentially endogenous unemployment control. Additionally, this estimate also falls in the middle between the three IV estimates that control for the number of unemployed people (cols. 2, 5, and 8). The impact of one week of benefit extension on applications, estimated to be a negative 0.65% in column 5, is very close to the estimate from Table 2, column 4, i.e. the IV estimate without unemployment controls in the panel sample. The estimate from column 5 of Table 3 implies that the average increase in maximum weeks of unemployment benefits between 2008 and 2009 (26 weeks to 86 weeks) led to a 39% decline in applications. The increase in maximum weeks of unemployment benefits during the Great Recession led to a substantial decline in applications. My preferred estimate indicates that a one week increase in the duration of benefits yields a 0.65% decline in job search effort as measured by applications. The impact of unemployment benefit duration on unemployment The focus of this paper is the impact of unemployment benefit extensions on job applications. However, the same identification strategy can be used to investigate the impact of benefit extensions on the number of unemployed people at the state level. This exercise is somewhat more fragile econometrically than the estimation of the impact of benefit duration on applications. Indeed, we expect benefit duration to increase unemployment, but at the same time, due to program rules, benefit duration depends on past unemployment. This makes it difficult to clearly identify the impact of benefit duration on unemployment. By contrast, we expect applications to decline with benefit duration, but benefit duration and applications are both positively correlated with unemployment. Therefore, to the extent that one does not properly account for the role of unemployment in determining both applications and benefit duration, bias should yield a positive impact of benefit duration on applications. In that sense, finding a significant and negative impact of benefit duration on applications is a strong
14 result that goes against the potential positive bias and confirms theoretical expectations. By contrast, the expected effect and the expected bias go in the same direction when estimating the impact of benefit duration on unemployment. Nonetheless, we can still use discontinuities in program rules to identify the impact of benefit duration on unemployment. Table 4 and Table 5 repeat the analysis in Table 2 and Table 3 respectively, but using log unemployment on the left hand side instead of log applications. The only difference is that these specifications do not control for lagged log unemployment: indeed, doing so would yield biased estimates unless the dynamic nature of the panel model is fully taken into account. Results in Table 4 are almost the mirror image of the results in Table 2: benefit duration has a positive impact on the number of unemployed people at the state level that is roughly as large in absolute value as the negative impact of benefit duration on applications. For example, in col. 2 of Table 2, I find that a one week increase in benefit duration decreases applications by 0.41%, while col. 2 of Table 4 shows that a one week increase in benefit duration increases unemployment by 0.43%. Given the concern that unemployment benefit duration is positively correlated with past unemployment due to program rules, it is particularly important to control for a cubic in unemployment rate and the insured unemployment rate. This is why the IV estimate in column 2 is the most credible. Still, it is interesting to note that the IV estimate in the absence of controls for past unemployment (col. 4) is only slightly larger. Table 5 focuses on sub samples close to TUR thresholds. The impact of benefit duration on unemployment is consistently positive in all samples and all specifications. Interestingly, the impact of benefit duration on unemployment is smaller when the unemployment rate is higher. A one week increase in benefit duration increases unemployment by 0.73% when the TUR is between 5.5 and 7% (col. 2), but only by 0.3% when TUR is between 7.5 and 9% (col. 4). Interestingly, the impact of benefit extensions on applications does not appear to be smaller when unemployment rates are higher (Table 3). This suggests that the same proportional drop in applications is associated with a smaller increase in unemployment when the unemployment rate is higher. This may indicate that some jobseekers are sending too many applications when unemployment rates are high (I will discuss this issue in more detail in the discussion section below). The overall estimate from the panel IV specification in col. 2 of Table 2 falls in between the estimates for low and high TUR, at 0.41%, but it is closer to the high unemployment value, presumably because TUR is closer to the high range than to the low range for a large share of the sample. For this reason, the estimate in col. 2 of Table 2 is the one that I consider to be most credible. It
15 implies that the average benefit extensions between 2008 and 2009 (26 weeks to 86 weeks) led to a 26% increase in unemployment for the average state. Overall, I find that unemployment benefit duration has a negative and significant impact on job search effort as measured by job applications. At the same time, unemployment benefit duration has a positive impact on the level of unemployment, consistent with previous literature. This suggests that job search effort is an important channel explaining the positive impact of unemployment benefits on unemployment. 4. Discussion Robustness tests One may wonder whether applications on CareerBuilder.com are a good measure of search effort. As mentioned in the introduction, unemployment benefit recipients during the Great Recession spent between 24 and 38% of their job search time sending applications and answering job ads (Krueger and Mueller, 2011). Additionally, 27% of job search time is spent browsing job ads. Nowadays, the overwhelming majority of vacancies are posted on the Internet (Regis Barnichon 2010). Therefore, it is plausible to think that more than two thirds of job search time is spent on the Internet. Another way to investigate whether applications capture job search effort is to gauge whether more applications lead to more hires. The idea is that, if job search effort increases, this should all other things equal lead to more hires, and therefore applications should have a positive effect on hires. To test this, I merge monthly national data on hires and vacancies from JOLTS with data on monthly national applications in CareerBuilder, and data on the total number of unemployed people from the BLS. In Table 6, column 1, I regress the log number of hires on the log number of vacancies and the log number of unemployed people (restricting data to September 2007 and later). I find that more vacancies are associated with more hires, and that more unemployment is associated with fewer hires. However, surprisingly, none of the coefficients is significant. In column 2, I add log applications from CareerBuilder to this specification, and I find that applications have a positive and significant impact on hires. Remarkably, this is the only variable that has a significant impact on hires, as the number of vacancies and the number of unemployed people remain insignificant. In as much as higher job search effort should lead to more hires, this exercise suggests that applications are a good measure of search effort. While I have shown that the increase in unemployment benefits duration led to a substantial decrease in applications, one may wonder whether this decline was due to a decrease in applications on the
16 CareerBuilder website only. If so, it is possible that jobseekers applied through other channels and so the overall applications in the state may not have decreased as much as it seems. This scenario is very unlikely. First, I have just shown that the number of applications on CareerBuilder is significantly and positively related to hires in national monthly data. If applications on CareerBuilder were substituted by applications through other channels, this relationship would likely not be significant. Additionally, when graphing hires and applications between 2007 and 2011, one does not see that applications fall behind hires in relative terms as time goes by, as would be the case if applications in CareerBuilder had represented a lower and lower share of total applications (not shown). Second, the number of active jobseekers on CareerBuilder has been increasing essentially linearly over (not shown). An active jobseeker is defined as someone who either applied to at least one job during the month or modified their resume. Third, I present independent evidence showing that jobseekers are unlikely to have moved away from Internet job search during the Great Recession. Indeed, while there is no representative survey about online job search in the US during the Great Recession, a study using quarterly labor force survey data from the United Kingdom (Green et al. 2011) shows that there was an increased use of Internet for job search purposes among jobseekers during In April to June 2009, over 4 in 5 British jobseekers use the Internet to look for jobs, and this proportion is even higher among the recipients of unemployment benefits. Therefore, there is no reason to assume that jobseekers moved away from online job search during the recession. I conclude that my results cannot be explained by jobseekers moving away from CareerBuilder.com in order to apply elsewhere. Interpretation of the results I have shown that the increase in the generosity of unemployment benefits during the Great Recession led to a substantial drop in job search effort. Does this make the extension of benefits a wrong headed policy? Was the large unemployment rise during the Great Recession caused by overly generous unemployment benefits? I offer two partial answers to this question. First, I will argue that, from a theoretical perspective, the decrease in job search effort may have been socially beneficial, preventing too many workers from chasing too few jobs. Second, I will show that, empirically, the decline in applications can only account for the increase in unemployment above and beyond what could have been predicted from the drop in vacancies. That is, even without the benefit extensions, unemployment would probably still have been high during the Great Recession, but its level would have been in line with the low level of vacancies.
17 First, one can ask whether the decline in job search effort induced by higher unemployment benefits is socially harmful. There are two elements to this question: are more generous unemployment benefits socially harmful in general, and are even more generous unemployment benefits during a recession socially harmful? Independently of any business cycle consideration, there are two reasons why the duration of unemployment benefits affects job search effort: liquidity constraints and moral hazard (Card, Chetty and Weber, 2007). Assume that benefits do not affect the reservation wage, only search effort; and assume search effort is costly. Unemployment benefits decrease the income differential between employment and unemployment, and therefore reduce job search effort (liquidity effect). This liquidity effect is the unavoidable side effect of providing income support to unemployed individuals, which is the main aim of unemployment insurance. Additionally, since unemployment insurance is conditional on being unemployed, taking a job would make one loose unemployment benefits, and hence the benefits of working are diminished relative to the case of no unemployment benefits (moral hazard effect). Chetty (2008) shows that 60% of the effect of unemployment benefits duration on unemployment duration can be explained by liquidity constraints. Therefore, liquidity constraints play an important role, and given that the Great Recession was also a credit crisis, these liquidity constraints probably played an even greater role than usual in jobseekers decisions. If liquidity constraints are the key reason why extended unemployment benefits decreased search effort, then the decrease in search effort can be seen as the unavoidable side effect of unemployment insurance playing its income support role at a time where such support was particularly needed due to the situation of the credit and labor markets. Based on previous research on the reasons behind the effect of unemployment benefits on unemployment duration, I conclude that the impact of unemployment benefits on job search is not necessarily welfare decreasing. There is a newer strand of research that focuses on the idea that unemployment benefits should be higher during recessions because the impact of unemployment benefits on unemployment duration is lower during recessions (Kroft and Notowidigdo 2011; Landais, Michaillat, and Saez 2011). In particular, a working paper by Landais, Michaillat and Saez (2011) develops a model with job rationing during recessions. In their model, job search has negative externalities: if a larger number of unemployed workers search for jobs, it makes it harder for each one of them to find one of the rationed jobs. Therefore, providing a more generous UI in recessions reduces search effort, which improves welfare by addressing the negative externality. The authors also present evidence consistent with unemployment insurance effects on unemployment duration being lower during recessions. This paper also finds that indeed the impact of benefit duration on unemployment is
18 smaller when unemployment rates are higher (see Table 5), even though the impact of benefit duration on job search effort does not significantly vary with the unemployment rate. From this perspective, it is therefore possible that, by reducing the negative externality of job search effort during recessions, some of the reduction in job search effort that I document here was actually socially beneficial. To further assess whether extending unemployment benefit duration was socially harmful, one can investigate to what degree unemployment benefit extensions contributed to higher unemployment during the Great Recession. To do so, I start with predicting, for each state and month, the level of unemployment if unemployment benefit duration had stayed at 26 weeks. For each state and month, let u be the log number of unemployed people and b the number of weeks of benefits available. Using the estimated impact of one week of benefits on unemployment from my preferred specification i.e % (column 2 of Table 4), I calculate counterfactual unemployment in each state and month as exp(u (b 26)). I then add up counterfactual unemployment of all states for each given month, which yields national counterfactual unemployment. Based on this calculation, in the absence of any benefit extensions, the national unemployment rate would have been 1.6 percentage points lower in 2009, and 2.4 percentage points lower in Finally, I merge this data on counterfactual unemployment with monthly national JOLTS data on hires and vacancies. In Figure 5, I plot unemployment and counterfactual unemployment against vacancies. The hollow circles use actual unemployment and vacancies data from JOLTS In this scatter plot, we can see the strong negative relationship between unemployment and vacancies: this relationship is linear at low to medium levels of unemployment. However, there is a group of points to the far right of the graph, corresponding to the Great Recession: these points lie above the virtual line constituted by earlier points, implying that unemployment is too high relative to the number of vacancies. This phenomenon represents an outward shift of the Beveridge curve. The red triangles represent the relationship between vacancies and counterfactual unemployment between September 2007 and July Remarkably, no outward shifting of the Beveridge curve can be observed in this plot: the triangles line up nicely with the virtual line drawn by vacancies and unemployment observed between 2001 and This shows that the increase in unemployment induced by the extensions in the duration of unemployment benefits can explain the outward shift in the Beveridge curve. This exercise suggests that the extension of unemployment benefits duration during the Great Recession led to a decline in job search effort that is large enough to account for the outward shift of the Beveridge curve. Were it not for benefits extension, unemployment may have been in line with vacancies: still high given that vacancies saw a large decline during the Great Recession, but not as high as was actually observed. This
19 exercise suggests that we do not need an increase in mismatch to account for the outward shift in the Beveridge curve; a decrease in job search effort does the trick. This may explain why papers that investigate the role of mismatch in accounting for unemployment in the Great Recession found that mismatch played a limited role in explaining the growth in unemployment (R. Barnichon and Figura 2011; Sahin et al. 2011). 5. Conclusion This paper has used state level variation in the maximum weeks of unemployment benefit during the Great Recession to investigate the impact of unemployment benefits on job search effort. I measure job search effort by the number of applications received by jobs on CareerBuilder.com, the largest American employment website. Using instrumental variables, I show that a one week increase in the duration of unemployment benefits leads to a 0.7% decline in applications. This estimate implies that the average increase in benefit duration at the state level in led to a 39% decrease in applications. Similarly, I estimate that the average increase in benefit duration at the state level in led to a 26% increase in unemployment. If one uses my preferred estimates to predict counterfactual unemployment in the absence of benefit extensions, one finds that the recent outward shift of the Beveridge curve can be fully accounted for by unemployment benefit extensions above the regular 26 weeks. If the outward shift of the Beveridge curve is in fact due to benefit extensions and not increased mismatch, this is good news for policy makers and the economy. Indeed, as the economy recovers, benefit extensions will end and so will ultimately any excess unemployment that they have generated. By contrast, if the shift in the Beveridge curve were due to mismatch between jobs and unemployed individuals, this problem could be more persistent and lead to a higher level of equilibrium unemployment. For this reason, policy makers should not be discouraged by these results, which offer hope for a better recovery of the labor market. It is important to stress once again that these results do not suggest that benefits extensions are fully responsible for the high level of unemployment experienced during the great recession. This paper instead points to the fact that, because of benefit extensions, unemployment may have been higher than what could have been predicted based on the drop in job vacancies. Benefit extensions could not have prevented the drop in vacancies, which is a major driver of unemployment. Finally, even if benefit extensions increased unemployment by decreasing job search effort, this does not imply that this policy was socially harmful. Indeed, unemployment benefits income replacement role entails a decline in
20 search efforts in as much as jobseekers are credit constrained. Given the large tightening of credit during the Great Recession, income support was particularly important to help smooth consumption, and the decline in job search effort may have been a necessary evil to achieve enough consumption smoothing. Whether benefit extensions during the Great Recession were in fact higher than was necessitated by the consumption smoothing goal of unemployment insurance is an open question for future research. Angrist, Joshua D., and Victor Lavy Using Maimonides Rule to Estimate the Effect of Class Size on Scholastic Achievement. The Quarterly Journal of Economics 114 (2) (May 1): Barnichon, R., and A. Figura What drives matching efficiency? a tale of composition and dispersion. FRB Staff WP 10. Barnichon, Regis Building a composite Help Wanted Index. Economics Letters 109 (3). Economics Letters: Card, David, Raj Chetty, and Andrea Weber Cash on Hand and Competing Models of Intertemporal Behavior: New Evidence from the Labor Market. The Quarterly Journal of Economics 122 (4) (November 1): doi: /qjec Chetty, Raj Moral Hazard versus Liquidity and Optimal Unemployment Insurance. Journal of Political Economy 116 (2) (April): doi: / Green, A.E., M. de Hoyos, Y. Li, and D. Owen Job Search Study: Literature review and analysis of the Labour Force Survey. Department for Work and Pensions Research Report 726. Kroft, Kory, and Matthew J. Notowidigdo Should Unemployment Insurance Vary With the Unemployment Rate? Theory and Evidence. Working Paper. Krueger, Alan B., and Andreas Mueller Job search and unemployment insurance: New evidence from time use data. Journal of Public Economics 94 (3 4). Journal of Public Economics: Job Search and Job Finding in a Period of Mass Unemployment: Evidence from High Frequency Longitudinal Data. Princeton University, Department of Economics, Industrial Relations Section., January. Landais, Camille, Pascal Michaillat, and Emmanuel Saez Optimal Unemployment Insurance over the Cycle. NBER working paper (August). Rothstein, Jesse Unemployment Insurance and Job Search in the Great Recession. Working Paper (September). Sahin, A., J. Song, G. Topa, and G.L. Violante Measuring Mismatch in the US Labor Market. Manuscript, revised March. Schmieder, Johannes F., Till von Wachter, and Stefan Bender The effects of unemployment insurance on labor supply and search outcomes : regression discontinuity estimates from Germany. Institut für Arbeitsmarkt und Berufsforschung (IAB), Nürnberg [Institute for Employment Research, Nuremberg, Germany], February.