iii .1 i 11-; ’5 a Human? a . .l. ‘ .. a. . . filing 6.9:. Amwnvfl... 3C .4 ,. m . :w........n,... .. first , a .31... 3%. . a 3 .nflfifiasn . Euwb. z. 1. us; by t u... 21.13.!!! . a. 3.3.3,}... . a B .Y......sr :21: . 1.5:... .x .3 4 ...:;Ha a . Xblu; :5: 9.." :- NI /‘\ ,.‘ _)~} LIBRARY Michin W. State University This is to certify that the dissertation entitled THREE ESSAYS IN LABOR ECONOMICS presented by MICHAEL ALLGRUNN has been accepted towards fulfillment of the requirements for the PhD. degree in Economics 3%wafi Major Professor’s Signature] /(4m"1 '3 l 7' a a a! / r Date MSU is an Affirmative Action/Equal Opportunity Employer _A_._.-._.-,—.-.-._.-.-.-.-.-.-.-.-.- -—-.—.— _ PLACE IN RETURN BOX to remove this checkout from your record. TO AVOID FINES return on or before date due. MAY BE RECALLED with earlier due date if requested. DAIEDUE DATEDUE DAIEDUE 5/08 K:lProi/Accapres/CIRC/Dateoue,indd THREE ESSAYS IN LABOR ECONOMICS By Michael Allgrunn A DISSERTATION Submitted to Michigan State University in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY Economics 2009 ABSTRACT THREE ESSAYS IN LABOR ECONOMICS By Michael Allgrunn The first of the three essays examines immigrant wage gaps from 1960 to 2000. Previous research has suggested that the U.S.-specific labor market skills of successive immigrant cohorts from declined from 1970 to 1990; that is, compared with earlier cohorts, recent cohorts started off with lower wages relative to natives and assimilated at Slower rates. We argue that the decline in immigrant skills within country-of-origin groups is not supported by the long-term evidence. Analysis of US Census data from 1970 to 2000 suggests that the unexplained wage gap after ten years contradicts the hypothesis of declining skills for cohorts after 1965. More broadly, the unexplained wage gap should not be treated solely as an indicator of a change in immigrant cohort skills. The second essay examines how benefit levels for unemployment insurance (U1) affect the duration of unemployment. Most research on the effects of U1 on unemployment duration has been limited by the use of a censored measure of unemployment spells. This essay reexamines the impact of U1 benefit levels on unemployment duration using a dataset that allows examination of actual unemployment spells. We find that while censoring concerns are legitimate, the main problem in estimating the impact of U1 benefit generosity on the duration of U1 benefit receipt and jobless duration is finding exogenous variation in UI weekly benefit amounts. Using a quasi-experimental difference-in-difference approach, we find that the effect of benefit generosity on unemployment duration may be smaller than previously estimated. The third essay considers how an increase in the potential duration of unemployment benefits affects the duration of unemployment. We examine the extent to which increasing the potential duration of unemployment benefits increases the length of unemployment spells using a national sample of workers who were laid off and claimed unemployment insurance (UI) benefits during the recession of the early 19903. The research design takes advantage of changes in the potential duration of benefits that occurred due to the Emergency Unemployment Compensation Act of 1991. Our attempts to reconcile the disparate findings of existing research suggest that different econometric estimators can produce substantially different inferences about the effects of increased potential benefit duration. We also find that estimates of the effect of potential benefit duration on weeks of benefit receipt often bear little relation to the estimates of the effect of potential duration on weeks of joblessness. Cepyright by MICHAEL ALLGRUNN 2009 ACKNOWLEDGMENTS I would like to thank Pedro Alrnoguera, Tom DeLeire, Alfonso Flores-Lagunes, Thomas J eitschko, Jack Meyer, Lucie Schmidt, and Peter Schmidt for helpful suggestions. I would also like to acknowledge the many contributions from my dissertation chair, Steve Woodbury, who is co-author of Chapters 2 and 3 and from my dissertation committee: Dale Belman, Jeff Biddle, and J eff Wooldridge. Any errors are the sole responsibility of the author. TABLE OF CONTENTS LIST OF TABLES ............................................................................ vii LIST OF FIGURES ............................................................................ x CHAPTER 1 EXAMINING IMMIGRANT WAGE GAPS, 1960-2000 ................................ 1 Introduction ............................................................................. 1 Theory ................................................................................... 2 Model ...................................................................................... 4 Data ...................................................................................... 8 Findings ................................................................................ 10 Conclusions ........................................................................... 15 Bibliography ........................................................................... 33 CHAPTER 2 UNEMPLOYMENT INSURANCE AND UNEMPLOYMENT DURATION: A REEXAMINATION ........................................................................... 34 Introduction ........................................................................... 34 Theory .................................................................................. 36 Data ..................................................................................... 37 Methodology ........................................................................... 4O Empirical Findings .................................................................. 45 Conclusions .......................................................................... 48 Bibliography .......................................................................... 66 CHAPTER 3 EMERGENCY UNEMPLOYMENT COMPENSATION AND THE DURATION OF UNEMPLOYMENT .................................................................... 69 Introduction .......................................................................... 69 Policy Background ................................................................. 71 Previous Research ................................................................ 74 Data ................................................................................... 75 Econometric Models .............................................................. 78 Parametric and Semi-parametric Models ........................... 78 Non -parametric Hazard Models ...................................... 81 Empirical Findings ................................................................. 85 Summary and Conclusions .................................................... 93 Bibliography ......................................................................... 1 12 APPENDIX .................................................................................. 116 APPENDIX A ...................................................................... 117 APPENDIX B ...................................................................... 125 vi LIST OF TABLES Table 1 Number of Observations by Country-of—Origin Groups and Census Year ............................................................................................ 26 Table 2 Coefficient Estimates of Years-since-migration Variables — 2000 Cross Section ......................................................................................... 27 Table 3 Decomposition of Relative Earnings, 1990 to 2000 comparison .............. 28 Table 4 Decomposition of Relative Earnings, 1980 to 1990 comparison ............... 30 Table 5 Decomposition of Relative Earnings, 1970 to 1980 comparison .............. 31 Table 6 Comparison of Unexplained Earnings Gap Estimates ........................... 32 Table 7 Sample descriptive statistics and changes in maximum weekly benefit amounts ....................................................................................... 55 Table 8 OLS estimates of insured unemployment duration and jobless duration 56 Table 9 Censored regression estimates of insured unemployment duration and jobless duration ............................................................................. 57 Table 10 Weibull (AFT) regression estimates of insured unemployment duration and jobless duration ........................................................................ 58 Table 11 DDD estimates of UI benefit duration and jobless duration ...................... 59 Table 12 DDD estimates of Ul benefit duration and jobless duration, omitting claims within 30 days of the increases ................................................ 61 vii Table 13 DDD estimates of Ul benefit duration and jobless duration using weeks of benefits received 63 Table 14 DDD estimates of Ul benefit duration and jobless duration using only administrative data .................................................................... 65 Table 15 Summary of Extended Unemployment Benefit Programs, 1958 to present ....................................................................................... 99 Table 16 Potential Durations of Extended UI Benefits in Pennsylvania Under Emergency Unemployment Compensation ......................................... 104 Table 17 Estimated Effects of the Impact of Increased Potential Duration of UI Benefits for the United States and Canada ...................................... 105 Table 18 Brief Variable Definitions and Sample Summary Statistics ...................... 107 Table 19 Estimated Effects of Increased Potential Benefit Duration from Parametric and Semi-Parametric Models ............................................. 109 Table 20 Estimated Effects of Increased Potential Benefit Duration from Nonparametric Models ..................................................................... 111 Table A1 Benefit Duration Estimates ............................................................... 117 Table A2 Jobless Duration Estimates .............................................................. 119 Table A3 Benefit Exit and Reemployment Hazard Estimates — Parametric and Semi-Parametric ....................................................................... 121 Table A4 Benefit Exit and Reemployment Hazard Estimates - Non-Parametric ....... 123 Table B1 Conditional Benefit Exit Probabilities (Discrete Hazards) ........................ 125 viii Table 82 Conditional Reemployment Probabilities (Discrete Hazards) ................ 127 LIST OF FIGURES Figure 1 Decomposition of the earnings gap between two immigrant cohorts A. A positive unexplained earnings gap .................................... 17 B. A negative unexplained earnings gap ................................... 18 Figure 2 Relative Earnings: White Cohorts .................................................... 19 Figure 3 Relative Earnings: Black Cohorts .................................................... 21 Figure 4 Relative Earnings: Asian Cohorts .................................................... 22 Figure 5 Relative Earnings: Mexican Cohorts ................................................ 23 Figure 6 Relative Earnings: Cuban Cohorts .................................................. 24 Figure 7 Relative Earnings: Other Hispanic Cohorts ........................................ 25 Figure 8 Budget constraint with unemployment benefits and an increase in benefits ....................................................................................... 50 Figure 9 Illustrative Ul benefit schedule with an increase in maximum benefit Amount ....................................................................................... 51 Figure 10 Distribution of Benefit Duration ......................................................... 52 Figure 11 Distribution of Jobless Duration (including censored observations) . ............. 53 Figure 12 Distribution of Jobless Duration (excluding censored observations) . ............ 54 Figure 13 Distribution of Benefit Duration ........................................................ 96 Figure 14 Distribution of Jobless Duration (including censored observations) ............ 97 Figure 15 Distribution of Jobless Duration (excluding censored observations) ........... 98 xi CHAPTER 1 EXAMINING IMMIGRANT WAGE GAPS, 1960-2000 Introduction The ongoing immigration debate in the United States has highlighted again the importance of the skills and capabilities that immigrants bring with them. Chiswick's pioneering work on the earnings of immigrants to the United States indicated a low entry wage relative to natives, but rapid ‘assimilation’ as immigrant earnings grew quickly and eventually exceeded natives’ earnings (Chiswick 1978). Borjas, by contrast, accounted for cohort effects, and revealed that Chiswick’s cross-sectional estimates overestimated the rate at which immigrants reach parity with their native counterparts (Borjas 1985). Borjas has also suggested that immigrant cohorts from 1970 to 1990 have been decreasing in ‘quality’; that is, they started off with lower wages relative to natives than previous cohorts, and have assimilated at slower rates than previous cohorts (Borjas 1985, 1995). These conclusions were drawn by examining the unexplained wage gap between cohorts. That is, any differences that Were not explained by differences in age, education, or length of time in the United States Were attributed to differences in the ‘quality’ or US-specific skills of recent l”11"1igl‘élnt cohorts.1 1\z (1 Ch} SVvick and Lalonde and Topel have suggested that the decline stems from both oeFlfning U.S. specific-job-market skills within groups and from a shift in country of l‘lgin, with more immigration coming from less-skilled countries (Chiswick 1986, slfilonde and Topel 1991). Yuengert (1994) agreed with the latter, but found that the “S of Mexican immigrants actually increased after 1964. F unkhouser and Trejo ( 1 995) and Barrett (1996) both accept the idea of declining skill since 1965, but find at the decline ended in the 19803. If immigrant quality or skill has not been decreasing since the immigration reforms of 1985, we might well rethink our opinions about immigration policy. I argue that the unexplained earnings gap does not support the idea of declining quality or skill. Using Borjas’s methodology and US. Census data from 1970 to 2000, I find no evidence of an increasing unexplained earnings gap for more recent cohorts. This is consistent with other recent work suggesting that the low wages of recent immigrants are not due to declining skill.2 I also find that the evidence of an increasing unexplained earnings gap from 1960 to 1970 is contradicted by looking at the same immigrant cohorts later in their immigration experience. Theory In a cross-section, it is typical to see recent immigrants earning lower wages than ilnmigrants who arrived earlier. Some of this wage gap can be explained by differences in age, experience, and educational levels, but part is also explained by differences in the length of time each cohort has spent in the United States. Any wage difference not explained by differences in observable socio-economic characteristics or by length of time in the United States is the unexplained wage gap. Figure 1 illustrates the wage gaps to be estimated below. In panel A, the lower line ShOws the average predicted earnings of the 1980 cohort at various times since rnrnlgl‘ation,hold1ng soc10—econom1c characteristics constant. The upper line does the \ 2 . Butcher and DiNardo (2002) find that changes in the wage structure are responsible (Eggower relative wages of more recent immigrants, not a decline in skills. Smith ent' 6) finds that the lower relative wages of more recent immigrants are explained (I lrely by changes in the wage distribution and the returns to skill, rather than a ecrease in immigrant skill by cohort. same for the 1970 cohort. The cross-sectional gap between the 1970 and 1980 cohorts observed in the year 2000 is the difference in predicted earnings for the two cohorts afier controlling for socio-economic characteristics (3720001970 — yzoomggo. where the first subscript denotes the year in which earnings are observed, and the second subscript denotes the immigrant cohort). This cross-sectional gap can be broken into two parts. The first is the growth of earnings for the 1970 cohort in during the 19903 (5)2000’1970 — 91990'1970). In 2000, the 1970 cohort had been in the United States ten years longer than the 1980 cohort, and the earnings growth they experienced in those extra ten years explains part of the cross-sectional gap. The unexplained wage gap is the difference between the predicted earnings of the cohorts at the same point in time since immigration (571990,”70 " 5720004980) SPeeifically, the unexplained wage gap shown in panel A is the difference in average PrediCted earnings of the 1970 cohort after 20 years in the United States (observed in 1990) and the average predicted earnings of the 1980 cohort at after 20 years in the United States (observed in 2000). The predicted wage lines in panel A indicate a positive unexplained earnings gap bemeen the 1970 and 1980 cohorts, indicating that something unobserved allOWed the 1970 cohort to have higher predicted earnings than the 1980 cohort. What does this mean? Borjas (1985, 1995) found a positive unexplained earnings gap bet"Veen the 1970 and 1980 cohorts, and attributed this to the higher relative quality or skill of the 1970 cohort. An alternative explanation could be that changes in the wage s . . . . 1m-‘Ctnre are responsrble for lower relative wages of more recent 1mm1grants, (Butcher and DiNardo (2002), Smith (2006)). A positive gap could also be due to factors on the demand side of the US labor market. For example, if demand for recent immigrant labor were slack or if US employers were less willing to hire recent immigrants in 1980 than in 1970, the unexplained earnings gap would again be positive. Panel B shows a negative unexplained earnings gap, which would indicate that some unobserved factor led the 1970 cohort to have lower earnings than the 1980 cohort. A negative gap could be explained by an increase in quality or skills, by an increasing willingness of US employers to hire new immigrants. It would also be possible that recent immigrants are benefiting from having and easier path to . . . . 3 aSSImllation, thanks to the efforts of and help from prevrous cohorts . Model We now look at how to estimate the unexplained wage gap for various lImmigrant cohorts following Borjas (1985). We start by estimating a cross-sectional earmngs function for each census year. Using a sample of immigrants observed in 1990: We have:4 ln(earningslggo) = X19901’199o + assDas + a80D80 + a75D75 + a7oD7o (1) + 51651365 + a60D60 + “50050 + “40040 + 51990 3\ 4 See Chiswick and Miller (forthcoming). For Simplicity, I confine the explanation in this section to the use of 1990 and 2000 ata, bUt the idea is the same for any two census years. 4 and using a sample of immigrants in 2000, we have: ln(earningszooo) = Xzoool’zooo + 395095 + 1890090 + flBSDBS + fisoDso + 575075 'I' 1870070 + 1865065 (7) + .860D60 + 550050 "I” 540040 + 81990 where ln(earnings) denotes the natural logarithm of the yearly earnings in tens of dollars, X denotes the socio-economic characteristics of the individual (no constant term included), and the Dk are dummy variables denoting the immigrant cohort; that IS, the time period during wh1ch 1mm1grat1on occurred. In order to decompose the cross-sectional earnings gap, we need to construct three Sets of predicted earnings. The first two are the predicted earnings of an average 1mmi grant for each of the cohorts in each of the two census years. 3/199o,k = X2000,kl’199o + “k (3) 3’2ooo,k = Xzooo,kl’2ooo + .31: (4) 5\ DEth Dk indicates immigration in a particular time period: D95 indicates 1995- 99, 1 90 Indicates 1990— 94, D85 1nd1cates 1985- 89, 080 indicates 1980- 84, D75 indicates 975 79, D70 indicates 1970- 74, D65 indicates 1965- 69, D60 indicates 1960- 64, D50 i tgditcates 1950- 59, and D40 indicates the individual immigrated before 1950. Notice ina the intervals are not all of equal length. These differences arise due to the manner which year of immigration information is coded in the Census, not from any them etical concerns. where k indexes the cohort (or five-year period during which immigration occurred), "' . . . 6 and X 2 000.111 denotes the average values of X for immigrant cohort k 1n 2000. Hence, J71 990k and 372000,k give the predicted 1990 and 2000 earnings for the average immigrant in cohort k. The third equation needed for the decomposition is the predicted earnings in 2000 for the average immigrant from cohort k if they had immigrated ten years later: 3’2ooo,k+1o = Xzooo,kl’2ooo + 18k+10 (5) The year 2000 cross-sectional earnings gap between cohort k immigrants and cohort k+10 immigrants is simply the difference in earnings for an individual who immigrated in period k and an observationally similar individual who immigrated ten Years later. This gap is illustrated in Figure 1 for the 1970 (k) and 1980 (k+10) cohorts _ This cross-sectional gap can be broken into two parts: y2000,k — 3’2ooo,k+1o = 072000,]: " Y1990,k) (6) + (371990.11: — 372000,k+10) he firSt term on the right is cohort k’s earnings growth during the 19903, or the ‘ W'th ~ . . . 1 11l-cohort growth.’ The second term IS the unexplamed eammgs gap, or the 6 Uélng the average values of X from the latter census year (2000 in this case) is an 3:113 ltral‘y decision. The average values of 1990 would be no less (or more) & pro.IDI‘iate. All that is required for our purposes is that the predicted earnings in quatlOns (3) and (4) be based on the same values of the vector X. 6 difference in earnings between immigrants from different cohorts but with the same length of time in the US (observed in two different census years). A positive unexplained earnings gap suggests higher wages for the earlier cohort. The decomposition in equation (6) does not take into account changes in economic conditions that may have occurred over the decade. For example, if wages generally increased during the decade, the estimated within-cohort growth will overstate the wage growth that would otherwise have occurred, and the unexplained earnings gap will understate the differences in human capital between the two immigrant cohorts. Borjas (1985) suggests handling this problem by examining the earnings of immigrants in relation to those of natives — that is, by normalizing with respect to native earnings. Again, we begin with two cross-sectional regressions. The first uses a sample of natives in 1990: ln(earnmg$1990,n) = X199O,n61990 + an (7) + 771990 and a Second using a sample of natives in 2000: 1n(earningszooo,n) = Xzooo,n5zooo + IBn (8) + 772000 Whe ’ I u a u re ’7 Indicates a nat1ve of the United States, and 17 IS the error term. We then Sub ' . . . . . . . stltute the average characteristics of 1mm1grant cohort k Into these estimated nat1ve e . llngs structures: 3’1990.k,n = X2000,k61990 + an (9) Y1990,k,n = X2000,k52000 + An (10) Equations (9) and (10) simulate the 1990 and 2000 earnings of a native who had the average characteristics of immigrants in cohort k. We can now decompose the cross- section growth (5) in a new way: 3720mm — 372000,k+1o = [(yzoook - y2000,k,n) “ (371990.1c " 371990,k,n)] (11) + [071990,]: — 9199mm) — (5)2000,k+10 _ y2000,k,n)] The first term on the right is the growth of earnings relative to natives for immigrant Cohort k over the 19905. The second term is the unexplained earnings gap, or the predicted difference in immigrant earnings, relative to natives, in 1990 and 2000, hOIding constant years since immigration. The data come from the US. Census of Population, Public Use Micro Sample 1 9 70 (Fl Metro), 1980 (5% State), 1990 (5% State sample), and 2000 (1% sample). In each set of comparisons, the earlier data contain only males in the age range 18-54, arid the later data contain only males aged 28-64. So there will be within/across- Q0Iiort effects for 1970 to 1980, 1980 to 1990, and 1990 to 2000. (The estimates for 1 970 to 1 9 8 0 have already been published in Borj as 1985.) 7 Individuals living in group quarters, part of the US armed forces, or self- employed at the time of the Census are excluded from the analysis. As in Borjas ( 1 9 8 5 ), the data are divided into six mutually exclusive groups: Cuban, Mexican, Other Hispanic, Asian, with the remainder classified as either White or Black. (In particular, the Census asks a question about ‘Hispanic origin’ and a separate question about ‘race’. If a respondent indicates Other Hispanic to the first question and White to the second, they are designated as Other Hispanic in this study.) Each group is SUbdivided into two categories: immigrant or native. Table 1 contains the number of Observations for each of these groups in each of the four census years. A few of the gI‘OUps are quite small, particularly for the pre-l950 cohort. The number of Cuban immigrants in the 1980-85 cohort is very large relative to other Cuban cohorts, cOrresponding to the Mariel Boatlift when Castro allowed Cubans to freely leave during the summer of 1980. The socio-economic variables are defined as follows: Education (edu) = Years of completed schooling. Experience (exper) = Age minus years of education minus 6. Marital Status (marr) = 1 if married with spouse present, 0 otherwise. SMSA (SMSA) = 1 if resides in a metropolitan area, 0 otherwise. Work disabilities (hlwth) = 1 if disability limits work, 0 otherwise. .7\ Going back further is not possible, as earlier censuses do not ask for year of 111:11‘nigration. However, by combining census questions on nativity and migration \Jx’lthin the last five years, it is possible to repeat this analysis for recent, or ‘short-time‘ llTln'iigrants for 1940, 1960, 1970, 1980, 1990 and 2000. The data for these additional years would come from the 1940 General, and 1960 General P.U.M.S.. 9 Findings Table 2 presents the estimated coefficients of the years-since-migration variables obtained from the 2000 cross-section for each of the six groups of immigrants (the comparison group is the 1995-99 cohort). At the 5% level, all coefficients are significant. Each coefficient in the table shows the log difference in earnings between an immigrant cohort and the most recent cohort in the same group. For example, White immigrants who came in the early 19903 have 46.93% percent hi gher earnings than White immigrants who came in the late 19903. This differential increases up to 90.5% for White immigrants who came in the early 19703. Table 3 Shows estimates of the within-cohort growth and the unexplained earnings gap from the decomposition in equation 11. The figures in the cross-section cOlumn are the difference in regression-adjusted year 2000 earnings between each immigrant cohort and the same-ethnicity immigrant cohort arriving ten years later (3’2 000,k — yzoowfi.” , where k is the cohort listed in the first column). The VVithin-cohort column shows the estimated earnings growth of cohort k over the 19903, again relative to the same-ethnicity immigrant cohort that arrived ten years later -- A A A A . . (3’ 2 000,k — y2000,k,n)—(Y1990,k - y1990,k,n)- The unexplamed earmngs gap column shows the estimated difference in relative earnings between each cohort Lt 1d the same-ethnicity cohort arriving ten years later, holding constant years since i o . A A A A Infillgmtlon ( [01990.1( ‘ y1990,k,n) _ (Y2ooo,k+1o " y2000,k,n)] ). These unexplained earnings gaps are generally negative, indicating higher predicted §3Inings for more recent cohorts. For example, the Asian 1975-79 cohort has a cross-sectional earnings gap of O - 2 '77, meaning that earnings in 2000 for the 1975-79 cohort were 27.7% higher than the earnings of the Asian cohort arriving 10 years later. This gap has two sources: First, the 64.3% increase in earnings for the 1975-79 cohort over the 19903, (as seen in the within-cohort estimate), and second, the unexplained earnings gap of -0.367, meaning the 1975-79 cohort of Asian immigrants did 36.7% worse than the cohort arriving ten years later, holding years since immigration constant. If we interpret the mieXplained earnings gap as reflecting differences in skill, the skills of Asian immigrants increased between 1975—1979 and 1985-1989. In general, the findings reported in Table 3 suggest that the skills of immigrants have increased since the 19703. While the positive unexplained earnings gaps for the earliest cohorts (1950-59 as well as 1960-64 for some groups) are Consistent with falling skills during the 19603 and early 19703, the large negative uIlextnlained earnings gaps for all other cohorts contradict such an interpretation. Tables 4 and 5 Show the results for the 1980-to-1990 and 1970-to-l980 COrtnparisons. Note that in the 1970-to-1980 comparison (Table 5), the unexplained ea-I‘llings gaps are positive or close to zero. This is consistent with Borjas’ findings for the same census years, and could be interpreted as an indication of lower skill for cohorts \ 8 The positive unexplained earnings gap for the 1970-74 Cuban cohort appears to be all artifact of the aforementioned Mariel boatlifi and its effect on the 1980-84 Cuban Q0hort. s i nce the late 19603.9 The compositions in Tables 3 and 4, however, tend to overturn this finding. In fact, the unexplained earnings gap for most of the cohorts is not constant as years since immigration increase. Consider the unexplained earnings gap for the Black 1965-69 cohort. In Table 5 their unexplained earnings gap is .289, suggesting they earned 28.9% more in their first 5 years in the US. than Black immigrants who arrived ten years later. In Table 4, however, the Black 1965-69 cohort earned 36.0% less than the Black cohort that arrived ten years later did 10-to-15 I years into theirs. Finally, in Table 3, the Black 1965—69 cohort earned 21.4% less in years 20-25 than the cohort that arrived ten years later, but is not significant at the 5% level. The changing Sign and magnitude of the unexplained earnings gap contradicts the idea that the unexplained wage gap is measuring changes in cohort quality or skill. In each case, we are comparing the 1965-69 cohort with the 1975-79 cohort, yet the LII‘leXplained wage gap is changing. To maintain that these gaps all indicate differences in the innate quality or US. specific skills of these cohorts seems nonsensical. Clearly, the unexplained earnings gap is changing over the immigration experience. Comparing two cohorts 5 and 15 years after immigration may yield di fferent results than a comparison of the same cohorts 25 and 35 years after 11'ITll‘l'ligration. This can be seen in Figures 2-7, which show the earnings (relative to 9\~ Table 7 offers a direct comparison with Borjas’ l970-to-l980 results. Of the 18 sets of estimates being compared, 4 are positive and significant in both studies, 3 are figsitive in my findings but insignificant in Borjas’, 3 are positive in Borjas’ findings, éut insignificant in mine, and 8 are insignificant in both studies, The magnitudes of Stirnates are not entirely similar, though the story of declining immigrant skill is Sup I ported by both. Unfortunately, the programming code from the original study by OI‘jas is no longer available. natives) of the various cohorts of immigrants at the same points in their immigration experience. Consider, for example, Figure 2, which shows the relative earnings for White cohorts. Each curve represents the trend of immigrant cohort earnings implied by the unexplained earnings gaps estimated above (Tables 3-5). For example, the “0- to— 5 ” curve shows the earnings of various cohorts during their first five years in the United States. The unexplained earnings gap from the 1970-1980 comparison (Table 5 ) suggests the White cohort that arrived in 1975-79 earned 60% less in their first five years than the 1965-69 cohort did in their first five years. Thus, earnings in the first five years since immigration decreases from 1 (normalized) for the 1965-69 cohort to a nadir of 0.40 for the 1975-79 cohort. Continuing in the same fashion, the uneXplained earnings gap for the 1975-79 cohort in the 1980-1990 comparison (Table 4) suggests 33% higher earnings for the 1985-89 cohort, or 0.53. Likewise, the uITleXplained earnings gap for the 1985-89 cohort in the 1990-2000 comparison (Table 3 ) Shows 75% higher earnings for the 1995-99 cohort, or 0.93. Each of the curves is Constructed in a similar manner, and allows us to see how different cohorts compare to each other holding years since immigration constant. In comparing the cohorts in their first five years after immigration, we see eviClience consistent with a skill decline for the late 19703 cohort. And while their earnings relative to natives is increasing, cohorts in the late 19805 and 19903 still have I owe! relative earnings than the late 19605 cohort in their first five years. Looking at t . . . . he Same cohorts 10-to-20 years after 1mm1grat1on, however, shows no ewdence of a 13st 19605 skill decline. Figure 2 also shows where the information available to Borjas ended. For eamings in the first five years after immigration, the last information reported in Borj as (1985) pertains to the 1975-79 cohort, indicated by the letter B by that profile. S imilarly, Borjas’ data about earnings 5-to-10 after immigration ended with the 1970- 74 cohort, and data about earnings after lO-to-20 years ended with the 1960-64 cohort. 0 Looking only at the portions of the curves available to Borjas (1985), it is easy to see evidence of a skill decline since the late 19605, with significant declines in relative earnings both 0- 5 and 5-10 years after immigration. When looking at data through the year 2000, however, the story changes. The declines for the 19705 cohorts during their first ten years are erased after that first decade. Indeed, looking at relative earnings after ten years would be consistent with higher immigrant skills. Additionally, subsequent cohorts see increases over the 19705 cohorts in all years Since immigration. The same findings hold for the other immigrant groups, with the exception of the Cuban 1965-69 cohort, which appears to do better than the 1975-79 cohort at all points in their experience (Figure 6). I suspect this is the effect of the Mariel boatlift, Wi th a large decrease in relative earnings for the group arriving just before, the 1980 census. The estimates for these Cuban cohorts actually result in a negative relative eamings; this cannot be taken literally, of course, but it does indicate a substantial CleCrease in relative earnings for this cohort. F - Borjas (1985) contains no information on earnings 20 or more years after igration. Conclum opp! —4\ \..w. " ‘1' \ u. _.‘..s t ‘ v 'I: I! -. ,‘; Q n.“ ‘ ‘l "‘ 'n‘ . I“ .h :2" - I ‘.~. .( (.‘ff‘f‘t‘ V . lg t ‘ ‘1,er . >1 liars C on clusions The unexplained wage gap between two cohorts generally varies over time. In particular, more recent cohorts have lower relative wages in their first 5-10 years, but hi gher relative wages afterward. The higher relative wages after 10 years contradict the idea of a skill decline for cohorts after 1965. The broader point, however, is that the changes in unexplained wage gaps indicate that these gaps should not be considered purely as indications of cohort skill. Using the method developed by Borj as (1985) and Census data from 1970 to 2000, I reconfirm that the relative earnings of new immigrants (that is, those within five years of arriving in the United States) were lower than those of earlier immigrants in the 19705 and 19805. . The relative earnings after five years, however, are roughly equal to earlier cohorts, and even higher for some groups. After ten years, immigrant Who arrived during the 19905 show higher relative earnings than earlier immigrant cOhorts at the same point in their immigration experience. Immigrants arriving in the 1 9905 had higher relative earnings than any other immigrant cohort in the sample, eVen 0-to-5 years into their immigration experience. If we are to interpret the unexplained earnings gap as an indicator of skill, the l Ong~term evidence shows that after ten years in the country, each cohort has higher U ‘ S . ~specific job-market skill than the one that preceded it. Overall, the evidence does not support the conclusion that the skills of successive immigrant cohorts have been declining. Rather, although the 19705 cohorts do seem to have had lower skills QQIIlpared to earlier cohorts, but this alleged skill gap was eliminated in less than ten yQaI‘s, The results could be affected by patterns of return migration that differed by skill for different cohorts. That is, if higher skilled immigrants from earlier cohorts left the United States more frequently than higher skilled immigrants from more recent cohorts, earlier cohorts would appear less skilled than they were at entry, especially late into their immigration experience. Without data on return migration rates by skill level, it is not possible to determine if there is such variation by skill. More broadly, the unexplained earnings gap should not be considered synonymous with differences U.S.-specific job-market skills. It is theoretically possible that demand for immigrant labor has increased over the decade, increasing the size of the unexplained wage gaps of more recent immigrants. This would require an increase in the demand specifically for immigrant labor that exceeded any increase in demand for native labor. Other explanations include changes in the acceptance of recent immigrants by employers, unmeasured differences in educational quality, assimilation assistance by earlier cohorts, or a combination of these factors. In fact, even after seeing the estimates of negative unexplained wage gaps, it is still possible that immigrant skills actually are in decline, but these other factors have offset the skill decline. It simply is not possible to determine the true trend in immigrant skills using the unexplained wage gap alone. Figure 1. Decomposition of the earnings gap between two immigrant cohorts A. A positive unexplained earnings gap Earnings 1970 cohon /‘ _____________________________ A . . y 2000,1970 Within cohon—< growth cross g 1980 sectional ___________ cohort gap 1990,1970 unexplained < earnings gap 3’1980,197o A 3’2000,1980 3’199o,1980 Years since 10 20 30 Immigration Notes: Illustration of the decomposition in equation (6). Each flak is the earnings in year t for cohort k. For example, 571990,1970 is the 1990 earnings of the 1970 immigrant cohort. l7 B. A negative unexplained earnings gap. Earnings 1980 cohon cross sectional _______________ A 19?“ gap 3’2ooo,1980 °° ° unexplained l 3’2000,1970 earnings gap : A i within 3’1990,1980 : cohon _______________ . Wth " | 9") l 3’1990,197o l i i i i i i A I l 3’1980,197o i l i i l l l Years since immigration 10 20 30 Notes: Illustration of the decomposition in equation (6). Each ytk is the earnings in year t for cohort k. For example, 5‘719904970 is the 1990 earnings of the 1970 immigrant cohort. 18 tozoo mafia? emiommr mmbwg vwémmr whimsmw when? @0602 $009 among 0mm Fma . _ — . — P F P \I‘./ l\\\\\.\\‘ // .liii // 805m ..\\ / \ / 829 - t. ..\\ /,/ / o F as --.9--. \.\. ............ 9&7 mos l o\ ---...o ............... /../.-.-.. ll. ...... . W XL. - - tip 0... a -- a. d 35:00 32>) 69:53 0323. .N 2:9". we we sfiuiume anireiea l9 Notes to Figures 2-7: Each curve represents the trend of immigrant cohort skill measured at the various points in their immigration experience. Each curve has been normalized to begin with a relative earnings of 1. The unexplained wage gaps from Tables 3-5 are then used to calculate the relative earnings for later cohorts. The 8 represents the last data point observable using only the 1970 and 1980 Census data for any given series. (It is in bold for the “0 to 5” curve and in light gray for the “10 to 15” curve.) 20 eocoo mafia? vmémm— mmémmw vmiowmr msémmw 3.659. mmimomv vwémmw mmémg owe-ma l p p p l b . b p O m - . - 3.1.9.2. 9 .......... - -- m) - . r . . DmOU—GW um \\\\’ / If... i\. m.— 39/. ONE? a \ /-).. .1h- 4 W a H F D Er. 2.9.... \- 85 ll .\\ \\ q a me \\ Q I d m 33:00 :35 "mafia-=3 gem—em .n 959.... sfiuiwee eniieiea eocoo mmimmmr vmiommr meme? vwiommr asimsmr 2-0an mwimmmr #9002 mmiommr 0mm 3.5 — p h p _ — — b omouom r 9.\ 2...... 822 --L. - 2% 2.5.9 moS li9|i HP «to-.00 .834. 59:53 3.3.3". .v 9.39". we we sfiuiume enireiea a}. 7. :28 n .r F _ l? p p - mmimmmw vmiommr owimwmv vmémmw whims?- vsénmw moimwmw v0.00? mmiommr 0mm Ea » 0 iii omBom e owes: 3... 92m -:9:: moS l9l «fiasco 53:85. ”3553 «>323. .m 952”. m0 ‘9. ‘— sfiuiwee eAireiea 23 tozoo mmimmmw vmémg owimwmw vmémmw whims?- gionmv moimomr vmioomw mmémmw one 35 omoaw a. one“? -_ «a a . 2m -------9 mDuD linl p h L P P b h - n \ Q» . \ .\ x /, \ \ // I \ / .. / th- V .\ / llliiilo /. Till / In!- - 9 i i -i i. i \ I / - i- .i \r.-.i.-::./ ....... t - 4 it.” _ L m» .9 ‘ I i . . - . i a” . m I \ 3.9.00 53:0 "mags-fie $323. .o 0.59". Fl 3. o a P m .3 m 9 m a m. U 5 S .3 m .3 24 :28 mmémmw vmommw mwbmg “$-0me whims? #050an meme? 006me among. com—.05 m) Wlliiillwviiwl. ....... XII/ll iW ....... iii a :1... DWOHDN an. \.....\ ...-..... ”@1122- \.\ ...X. m\l— 822 -7... ...... , . .9. 99m -..--9--- . . . .0 mos lol .. 9... 4 09.0500 03000:; .050 "005500 02.0.01 .5 0.59... mo 1.!) ‘— mm sfiuiwee eniieiea Table 1 Number of Observations by Country-of-Origin Groups and Census Year Census ear 2000 1990 1980 1970 White: native 393,444 1,810,908 1,843,576 303,926 immigrant 17,262 71,104 60,248 10,753 Black: native 59,462 213,315 262,768 38,121 immigrant 7,746 14,455 11,913 730 Asian: native 3,015 12,247 12,830 1,777 immigrant 16,772 49,913 27,338 1,205 Mexican: native 13,144 55,934 62,136 6,508 immigrant 23,372 54,514 37,117 2,158 Cuban: native 423 882 1 .041 75 . immigrant 2,133 9,415 7,943 1,141 Other Hispanic: native 7,645 17,815 37,557 6,496 immigrant 14,277 41,021 18,730 1,463 Total: native 477,133 2,111,101 2,219,908 356,903 immigrant 81,562 240,422 163,289 17,450 Source: US. Census of Population, Public Use Micro Sample 1970 (F1 Metro), 1980 (5% State), 1990 (5% State sample), and 2000 (1% sample). Table 2 Coefficient Estimates of Years-since-migration Variables - 2000 Cross Section Country of Origin Group Other White Black Asian Mexican Cuban Hispanic D40 0.578 1.144 0.654 -0.152 0.724 0270 D50 0.800 0.942 0.741 -0.061 0.620 0.359 D60 0.814 0.853 0.963 0.439 0.985 0.231 065 0.849 0.935 0.986 0.479 0.765 0.196 D70 0.905 1.119 0.997 0.664 1.001 0.583 D75 0.880 1.155 0.969 0.630 0.810 0.574 080 0.813 0.898 0.816 0.588 0.324 0.646 D85 0.638 0.976 0.692 0.474 0.889 0.620 D90 0.469 0.812 0.632 0.330 0.517 0.504 Notes: 1. Estimates of the Bk from estimation of equation (2) for immigrants from the 2000 Census of Population. Coefficients represent the difference in predicted ln(earnings) associated with being in a particular cohort, relative to the 2000 cohort of the same country-of—origin group. All coefficients are significant at the 5% level. Additional controls included in the estimated equations are education, experience, marital status, SMSA, and work-limiting disability. 27 Table 3 Decomposition of Relative Earnings, 1990 to 2000 comparison Cross-Section Within-Cohort Unexplained Gap coef. t coef. t coef. t White 1950-59 -0.014 -0.17 -0.304 -5.17 0.290 4.15 1960-64 -0.091 -0.92 -0.262 -3.75 0.171 2.14 1965-69 0032 -0.39 -0.066 -1 . 12 0.034 0.50 1970-74 0.092 0.92 0.429 5.48 -0.337 -4.74 1975-79 0.243 2.79 0.390 5.78 -0.148 -2.26 1980-84 0.344 4.04 0.472 6.68 -0.129 -2. 16 1985-89 0.638 8.20 1.389 21.76 -0.751 —13.83 Black 1950-59 0.089 0.31 0.082 0.32 0.007 0.03 1960-64 0266 -1.24 -0.253 -1.21 -0.013 -0.08 1965-69 -0.220 -1.59 -0.006 -0.05 -0.214 -1.83 1970-74 0.221 1.64 0.538 4.27 -0.317 -3.35 1975-79 0.179 1.50 0.515 4.81 -0.337 -3.51 1980-84 0.086 0.79 0.419 4.72 -0.333 -3.59 1985-89 0.976 9.12 1.577 17. 37 -0.601 -6.60 Asian 1950-59 0222 -0.90 -0.411 -1.87 0.189 1.04 1960-64 0034 -0.20 -0.038 -0.22 0.004 0.03 1965-69 0.017 0.17 0.172 1.61 -0.155 -1.73 1970-74 0.180 1.99 0.417 4.16 -0.237 -3.00 1975-79 0.277 3.79 0.644 7.88 -0.367 -4.69 1980-84 0.184 2.82 0.815 10.93 -0.630 -8.33 1985-89 0.692 10.12 1.700 22.31 -1 .008 -13.21 28 Table 3 (continued) Cross-Section Within-Cohort Unexplained Gap coef. t coef. t coef. t Mexican 1950-59 -0.500 -3.29 -0.627 —4.82 0.127 1.13 1960-64 -0.225 -2.00 -0.069 -0.63 -0.156 -2.03 1965-69 -0.151 -2.07 0.193 2.73 -0.344 -5.54 1970-74 0.076 1.10 0.514 7.82 -0.438 -8.12 1975-79 0.156 2.77 0.479 8.73 -0.323 -6.49 1980-84 0.258 4.34 0.577 1 1.29 -0.319 -5.66 1985-89 0.474 8.69 1.099 23.30 -0.625 -11.48 Cuban 1950-59 -0.365 -1.09 -0.636 -1.81 0.270 1.11 1960-64 -0.016 -0.06 -0.174 -0.79 0.158 0.55 1965-69 0045 -0.12 0.238 1.14 -0.283 -0.74 1970-74 0.678 2.51 0.445 1.58 0.233 1.08 1975-79 0079 -0.19 0.875 2.22 -0.954 -3.06 1980-84 0193 -0.85 0.357 1.71 -0.550 -2.28 1985-89 0.889 3.12 1.336 4.45 -0.447 -1.98 Other Hispanic 1950-59 0.128 0.81 -0.099 -0.52 0.227 1.17 1960-64 -0.351 -2.37 -0.150 -0.78 0201 -1.10 1965-69 -0.378 -3.63 -0.120 -0.71 -0.258 -1.51 1970-74 0064 -0.58 0.361 2.02 -0.425 -2.61 1975-79 0046 -0.51 0.517 3.06 -0.563 -3.49 1980-84 0.143 1.75 0.628 3.91 -0.485 -2.99 1985-89 0.620 7.51 1.414 8.90 -0.794 -4.84 Notes: 1. Decompositions are based on equation (11). Bold indicates significance at the 5% level. As in equation 11, a positive unexplained earnings gap indicates higher relative earnings than the cohort arriving ten years later, holding years since immigration constant. Positive within-cohort terms simply indicate higher earnings for the cohort over the decade. Table 4 Decomposition of Relative Earnings, 1980 to 1990 comparison Cross-Section Within-Cohort Unexplained Gap coef. t coef. t coef. t White 1950-59 0021 -0.74 0.085 3.30 -0.106 -3.49 1960-64 0.058 1.57 0.271 8.26 -0.213 -5.76 1965-69 0.268 7.23 0.370 10.34 -0.102 -2.83 1970-74 0.278 7.05 0.410 10.75 -0.132 -3.54 1975-79 1.176 33.44 1.510 45.82 -0.334 -10.72 Black 1950-59 0123 -0.96 0.216 1.56 -0.339 -2.66 1960-64 0029 -0.30 0.500 4.15 -0.529 -5.24 1965-69 0.082 1.06 0.443 5.03 -0.360 -4.40 1970-74 0.293 4.63 0.512 ‘ 7.13 -0.219 -3.29 1975-79 1.128 16.91 1.556 22.74 -0.428 -6. 34 Asian 1950-59 0.037 0.47 0.140 1.50 -0.102 -1.14 1960-64 0.146 2.39 0.010 0.12 0.137 1.83 1965-69 0.261 5.78 -0.014 -0.21 0.274 4.83 1970-74 0.636 16.97 0.105 1.94 0.531 10.68 1975-79 1.339 42.05 1.073 23.24 0.266 5.91 Mexican 1950-59 -0.096 -1.97 0.026 0.47 -0.122 -2.20 1960-64 0.057 1.40 0.322 6.11 -0.265 -5.81 1965-69 0.096 2.69 0.327 7.16 -0.231 -5.68 1970-74 0.263 8.49 0.259 7.36 0.004 0.11 1975-79 0.857 27.98 0.806 24.63 0.051 1.41 Cuban 1950-59 0037 -0.41 -0.079 -0.48 0.043 0.28 1960-64 0.183 2.45 0.015 0.11 0.169 1.16 1965-69 0.100 0.78 0.034 0.24 0.066 0.37 1970-74 0.370 4.82 0.302 2.00 0.068 0.46 1975-79 0.766 5.15 1.975 10.16 -1.209 -6.80 Other Hispanic 1950-59 0041 -0.80 -0.039 -0.48 -0.002 -0.03 1960-64 0.023 0.44 -0.028 -0.40 0.051 0.76 1965-69 0.079 1.63 0.172 2.86 -0.093 -1.57 1970-74 0.106 2.43 0.037 0.66 0.069 1.31 1975-79 0.844 19.75 1.001 18.65 -0.156 -3.09 Notes: See Table 3. Table 5 Decomposition of Relative Earnings, 1970 to 1980 comparison Cross-Section Within-Cohort Unexplained Gap coef. t coef. t coef. t White 1950-59 0.063 2.22 0.105 3.05 -0.042 -1.08 1960-64 0.239 6.70 0.054 1.12 0.185 3.73 1965—69 1.070 32.16 0.466 10.78 0.604 13.94 Black 1950-59 0057 -0.47 -0.079 -0.36 0.021 0.10 1960-64 0.196 1.96 -0.400 -1.77 0.596 2.78 1965-69 1.050 13.93 0.760 5.20 0.289 1.98 Asian 1950-59 0109 -1.38 -0.026 -0.16 -0.083 -0.51 1960—64 0.306 4.78 0.141 0.89 0.166 1.10 1965-69 1.279 30.32 0.689 5.85 0.590 5.13 Mexican 1950-59 0.134 2.79 0.070 0.76 0.064 0.69 1960-64 0.092 2.03 -0.041 -0.41 0.133 1.36 1965—69 0.677 15.58 0.464 4.97 0.213 2.29 Cuban 1950-59 -0.092 -1.24 0.180 0.49 -0.272 -0.74 1960-64 0.304 4.37 0.135 0.38 0.169 0.47 1965-69 2.296 22.92 0.351 0.97 1.945 5.27 Other Hispanic 1950-59 0.034 0.46 0.294 2.01 -0.260 -1.83 1960-64 0.152 2.52 0.210 1.72 -0.057 -0.48 1965-69 0.929 17.20 0.798 7.63 0.131 1.25 Notes: See Table 3. Table 6 Comparison of Unexplained Earnings Gap Estimates Allgrunn Borjas coef. t coef. t White 1950-59 0042 -1.08 -0.022 -0.37 1960-64 0.185 3.73 0.097 2.21 1965-69 0.604 13.94 -0.012 084 Black 1950-59 0.021 0.10 0.183 1.87 1960-64 0.596 2.78 0.184 1.83 1965-69 0.289 1.98 0.297 3.71 Asian 1950-59 0083 -0.51 0.003 0.22 1960-64 0.166 1.10 0.049 0.91 1965-69 0.59 5.13 0.07 1.55 Mexican 1950-59 0.064 0.69 0.003 0.24 1960—64 0.133 1.36 0.093 2.00 1965-69 0.213 2.29 0.122 2.58 Cuban 1950-59 0272 -0.74 0.17 0.85 1960-64 0.169 0.47 0.196 1.12 1965-69 1.945 5.27 0.61 2.75 Other Hispanic 1950-59 -0.26 -1.83 0.036 0.71 1960-64 0057 -0.48 0.1 2.10 1965-69 0.131 1.25 0.147 3.31 Notes: 1. Decompostition of the unexplained earnings gap from equation (11). A positive unexplained earnings gap indicates higher relative earnings than the cohort arriving ten years later, holding years since immigration constant. 2. Borjas coefficients come from Borjas 1985, Table 5, in the column labeled “Across-Cohort Growth”. Bold indicates significance at the 5% level. For the sample sizes used in compiling these estimates, see Table 1, or Borjas 1985, Table 1. BIBLIOGRAPHY Barrett, Alan. “Did the Decline Continue? Comparing the Labor-Market Quality of United States Immigrants from the late 19705 and late 19805” Journal of Population Economics 9 (1996): 55-63. Borjas, George J. “Assimilation, Changes in Cohort Quality, and the Earnings of Immigrants” Journal of Labor Economics 3 (July 1985): 463-89 Borjas, George J. “Assimilation and Changes in Cohort Quality Revisited: What happened to Immigrant Earnings in the 19805?” Journal of Labor Economics 13 (April 1995): 201-45. Butcher, Kristin and DiNardo, John. “The Immigrant and Native-Bom Wage Distributions: Evidence from United States Censuses” Industrial and Labor Relations Review 56 (October 2002): 97-121. Chiswick, Barry R. “The Effect of Americanization of the Earnings of F oreign-Bom Men” Journal of Political Economy 86 (October 1978): 897-921. Chiswick, Barry R. “15 New Immigration Less Skilled Than the Old?” Journal of Labor Economics 4 (April 1986): 168-192. Chiswick, Barry R. and Miller, Paul W. "Immigrant Enclaves, Ethnic Goods, and the Adjustment Process," in Elliott R. Barkan, Hasia Diner and Alan Kraut, eds, Borders, Boundaries and Bonds: America and its Immigrants in Eras of Globalization, New York: New York, University Press , 2007. Pp. 80-93. Funkhouser, Edward and Trejo, Stephen J. “The Labor Market Skills of Recent Male Immigrants: Evidence from the Current Population Survey” Industrial and Labor Relations Review 48 (July1995): 792-811. LaLonde, Robert J. and Topel, Robert H. “Immigrants in the American Labor Market: Quality, Assimilation, and Distrubutional Effects” American Economic Review 81 (May 1991): 297-302. Smith, James P. “Immigrants and the Labor Market” Journal of Labor Economics 24 (May 2006): 203-233- Yuengert, A.M. “Immigrant Earnings, Relative to What? The Importance of Earnings Function Specification and Comparison Points” Journal of Applied Econometrics 9 (1994): 71-90. 33 CHAPTER 2 UNEMPLOYMENT INSURANCE AND UNEMPLOYMENT DURATION: A REEXAMINATION Introduction The effects of unemployment insurance (UI) on unemployment duration and earnings after reemployment have generated a large empirical literature and the continued interest of economists and policy makers. Curiously, though, labor economists and policy analysts have never reached a consensus on the quantitative effects of U1 on key outcomes. These effects are of more than academic interest: In the late 19905, the Clinton Administration introduced UI profiling (Black, Smith, Berger, and Noel 2003), and more recently the Bush Administration has proposed private reemployment accounts. Such reforms have the goal of speeding reemployment and improving the quality of the subsequent job match; however, whether they do so in fact depends on underlying worker behavior and market responses that are not wholly understood. There are two main difficulties in estimating the effect of unemployment benefits on the duration of unemployment spells. The first is that researchers rarely have data that includes accurate information on both unemployment benefits and the duration of unemployment spells. UI administrative data is the most widely used in this literature, and with good reason; it has the actual data used to determine a worker’s eligibility, benefit amounts, and potential duration, as well as the actual amount of benefits dispersed. It also contains data on the length of insured unemployment (the duration of benefit receipt), rather than the actual variable of 34 interest, the total length of unemployment.11 For example, a worker whose maximum duration of benefits is 16 weeks cannot be observed as unemployed for longer than 16 weeks in the administrative data. Many recent efforts in this literature have treated insured duration data as a censored measure of weeks jobless, and attempt to account for censoring using various econometric techniques. The second problem is that each worker’s weekly benefits and potential duration are correlated with his or her past earnings history through a formula that varies from state to state. Specifically, UI benefit generosity (both weekly benefits and potential duration) is greater for workers with stable earnings histories and higher earnings. This creates an endogeneity problem that has been neglected in most of the methods that have been used to date. In this paper, we discuss how the use of U1 administrative data has masked the severity of the endogeneity problem and use a difference-in-difference-in difference approach that can generate consistent estimates of the effect of benefit increases on unemployment duration. In the next section, we discuss the theoretical effect of a benefit increase on unemployment and why treating insured unemployment as a censored measure of actual unemployment is misleading. In the data section we describe the dataset and why it is particularly useful for this application. In the methodology section we describe the endogeneity problem in traditional estimation techniques, provide illustrations of the problem using our data, and introduce the difference-in-difference- . in-difference approach which allows us to avoid the same pitfall. Finally, we discuss the results and their robustness to changes in specification. 11 The effect of benefit generosity on the weeks of benefit collection itself may also be . an interesting consideration, and is discussed as part of the empirical findings section of this paper. 35 Theory Job-search models are the most common theoretical framework for considering the implications of unemployment insurance. Unemployed individuals receive wage offers from a known wage distribution with variable search intensity, and accept wage offers above a reservation wage. Job search models predict that higher weekly benefit amounts decrease the probability of becoming reemployed during the period of benefit receipt, but have no effect after benefit exhaustion. 12 If this theoretical prediction were correct, the censoring problems from using insured unemployment durations instead of actual unemployment durations would be a non-issue, as all of the effect of benefit increases would occur dining the insured spell. A simple labor-leisure model, however, provides a rationale for why a benefit increase may increase the uninsured portion of unemployment spells. The model presented here is similar to Moffitt and Nicholson (1982), and begins with the assumption that individuals faced with job 1055 make decisions about how to allocate their work and leisure over a relatively long horizon. The individual may take a job right away or choose to remain unemployed for some period. Time unemployed may be seen as utility-enhancing because it allows a period of leisure, or because it allows more time to be spent on job search. The budget constraint facing such an individual in a world with no unemployment benefits is represented as the straight line AB in Figure 8. The kinked line ACD represents the budget constraint when unemployment benefits are available for up to Umax weeks. The slope of segment CD represents the weekly wage rate of the new job, while the slope of segment AC represents the difference between the wage and the weekly benefit amount. 12 See Mortensen (1977), Moffitt (1985), Katz and Meyer (1990) for some relatively straightforward examples. 36 The effect of an increase in the weekly benefit amount is also shown in Figure 8, where the kinked line AEF is the new budget constraint. For individuals who would not have exhausted benefits under the lower benefit level (line segment AC ). the increase in benefits has both an income effect and a substitution effect. Both effects lead the individual to choose more leisure (less work). For individuals who would have chosen to exhaust under the lower benefit level (line segment CD), there is also an income effect, with individuals choosing more leisure. Thus, higher benefits will have a positive effect on both the insured portion of an unemployment spell and the uninsured portion. The labor-leisure model also provides insight as to whether insured durations should be thought of as a censored measure of total weeks jobless. For those who exhaust benefits, the end of benefits is not simply a censoring point; it marks the transition to a separate portion of the budget constraint where the effect of higher benefits is determined in a different fashion. Duration models that deal with censored data are designed to deal with censoring that occurs due to data collection techniques, typically the censoring of durations that have not ended at the time of data collection. The end of benefits, however, is not an artifact of data collection techniques. It is a policy-imposed limit that has real implications in the labor-leisure framework. The effect of higher benefits on the insured spell of unemployment is not the same as the effect on the uninsured spell, and attempts to estimate the effect of higher benefits on total weeks jobless using solely insured duration miss this difference. Data In what follows, we examine a nationally representative sample of 3,907 workers who made initial claims for U1 during calendar year 1998. The data were collected by Mathematica Policy Research, Inc., for a study of U1 exhaustees contracted by the US Department of Labor (Needels, Corson, and Nicholson 2002). 37 Mathematica used a two-stage clustered sampling design in which roughly 27,500 were randomly drawn from 25 states chosen from geographic strata. Each of the 25 states provided selected UI administrative data on the 27,500 claimants. These data include standard administrative variables such as the benefit year beginning date, first and last payments dates, base period earnings, weekly benefit amount, and the balance of U1 benefits remaining for each worker at the end of his or her benefit year. (Notably, the administrative records do not include base period earnings broken down by quarter.) Mathematica then selected random subsamples of U1 exhaustees and nonexhaustees from the 27,500 claimants and administered a follow-up survey between mid-July 2000 and mid-February 2001. Mathematica’s goal was to complete 2,000 surveys of exhaustees and another 2,000 surveys of nonexhaustees. They completed interviews with 1,864 exhaustees and 2,043 nonexhaustees. Given the timing of the survey, the interview covered labor market and other outcomes of the workers over a 2.2-year period on average, starting with the job loss that led to the UI claim.l3 From our standpoint, the most important data obtained from the survey is information on whether and when each worker became reemployed following the UI claim. This allows us to observe the length of the spell during which each worker was actually jobless. In contrast, UI administrative records track only the ntunber of weeks of U1 benefits a worker receives (up to 26 weeks of benefits in most states); whether and when a worker becomes reemployed are rarely recorded. As a result, administrative data yield information only about the duration of insured unemployment, which is a censored estimate of the duration of the jobless spell. The follow-up survey also includes information on up to five jobs held after the U1 spell, 13 For details of the sample and survey designs, see Needels. Corson, and Nicholson (2002)- 38 earnings in each of those jobs, and the industry and occupational classification of each of those five jobs. Finally, the follow-up survey includes information on characteristics of workers not typically included in administrative records, such as age, race, education, and marital status. Table 7 displays some descriptive statistics of the sample. The table lists the 25 states included in the survey and the number of observations drawn from each. It also shows the maximum weekly benefit amount in each state both before and after July 1, 1998.14 In 13 of the 25 states, the maximum weekly benefit amount increased at midyear, a fact we use below to obtain difference-in-difference-in-difference estimates of the impact of U1 benefits on insured and jobless durations. Finally, Table 7 shows the median weekly benefits amounts, median number of weeks of benefits received, and median weeks jobless for workers in the sample. Comparing median weeks of benefits (or insured duration) with median weeks jobless makes it clear that weeks of benefits is a poor proxy for the actual time a worker spends without a job. The median worker in the sample received fewer than 17 weeks of U1 benefits, but spent more than 30 weeks without a job. Figures 10-12 provide additional information about our duration variables. Figure 10 shows the distribution of benefit duration for our sample. There is a spike in the density at 26 weeks, the most common exhaustion point. Figure 11 shows the distribution of jobless duration for our sample, including workers who did not obtain reemployment before the survey date. Here we see a general decline in density until about 75 weeks. The bump in density from 75-150 weeks comes from the censored durations of individuals who did not obtain reemployment before the survey date. 14 Several states gives workers with dependents an additional “dependency allowance.” For states where a range is shown for the maximum weekly benefit amount, the lower figure is the maximum for a worker with no dependents, whereas the higher figure is the maximum for a worker with dependents. 39 Figure 12 omits these censored observations, showing the distribution of jobless duration for the portion of our sample that found reemployment. Methodology To motivate the primary methodology used in this paper, we begin with a demonstration of the problems with traditional techniques. The pioneering studies of how Ul affects unemployment duration (Ehrenberg and Oaxaca 1976, Classen 1977, Holen 1977) used Ul administrative data on Ul recipients to estimate OLS models of the form: t = a0 + aluiben + azpotdur + X6 + u (l) where t denotes either the number of weeks of U1 benefits received by the worker (durui) or the number of weeks the worker went jobless (jobless), and uiben denotes a measure of U1 benefits for which the worker was eligible — either the weekly benefit amount (wba) or the UI replacement rate (rr, defined as the weekly benefit amount divided by usual weekly earnings on the pre-UI job). Models in the literature vary in their specification, but they usually include a set of controls for the earnings history and characteristics of the worker (X), and often include the worker’s maximum potential duration of U1 (potdur). Table 8 displays results of estimating equation (1) with the Mathematica data described above; the dependent variable is the log of durui in columns 1 and 2, and the log of jobless in columns 3 and 4. Consistent with the existing literature, the coefficient on wba in column 1 suggests that a higher weekly benefit amount has a small, positive effect on the duration of unemployment; the point estimate suggests that a $100 increase in the weekly benefit amount would increase insured duration by about 10%. The inclusion of potential duration (column 2) increases this estimated 40 effect of wba to about 12%. It is tempting to think that these results give us some measure of the duration effect of benefit generosity. The problem with these estimates is twofold. First, these estimates, and other OLS estimates in this literature, are using a measure of insured duration rather than total duration: a worker whose potential duration of benefits is 16 weeks cannot be observed as unemployed for longer than 16 weeks in the administrative data. Because workers with longer potential duration can be observed longer in UI administrative records, the finding that higher benefits are associated with longer unemployment duration is hardly surprising. The second problem is that each worker’s weekly benefit amount and potential duration are correlated with his or her past earnings history through a formula that varies from state to state. Specifically, UI benefit generosity (both weekly benefits and potential duration) is greater for workers who have shown stronger labor force attachment. Although the estimates in Table 8 control for several correlates of labor force attachment —— education, pre-UI job tenure, base period earnings — it is likely that wba and potdur are correlated with u. If workers with strong labor force attachment are more likely to have shorter spells of unemployment, OLS will impart a negative bias to estimates of the effect of wba and potdur. Notice that this second problem, endogeneity bias, might be masked by the first problem, the positive bias from the use of insured durations. Using data on total weeks jobless allows us to focus solely on the endogeneity issue. The estimates in columns 3 and 4 of Table 8 show to what degree the use of insured weeks masked the endogeneity problem. The coefficients on wba (columns 3 and 4) now suggest that a higher weekly benefit amount has no significant effect on the duration of unemployment. The hypothesized positive effect of wba appears to be swamped by the fact that workers with stronger labor force attachment (as proxied by a higher wba) have shorter spells of unemployment. 41 One possible way around the first problem is to treat insured duration as a censored version of weeks jobless. Censored regression, parametric hazard models, and semi-parametric hazard models are possible econometric remedies for dealing with censored data. Censored regression deals with censoring by recognizing that censored observations have a different density than uncensored observations and adjusting the log-likelihood for each observation accordingly. 1 5 The hazard model is a tool from duration (or survival) analysis that accounts for non-normality of the error distribution.16 We estimate equation (1) using censored regression, and again using a Weibull hazard model. For simplicity, we report the accelerated failure-time analog of the Weibull hazard model, which allows the results to be interpreted, as in our other models, as the percentage effect of an increase on spell duration. (We have also estimated semiparametric Cox proportional hazard models and compared these with the Weibull hazard model. These latter two models give essentially similar results, suggesting that the distributional assumption imposed in the Weibull models is reasonable and that little is gained, in this case, by relaxing the assumption and moving to a semiparametric model.) Censored and Weibull regression estimates of equation (1) are shown in Tables 9 and 10. Columns 1 and 2 show that $100 increase in weekly benefit amount is associated with a statistically significant increase in unemployment duration of 22- 27% using the censored regression estimates, and a 14-17% increase using the Weibull estimates. If weeks of benefits were simply a censored version of weeks jobless, we 15 For a basic treatment of censored regression, see Wooldridge 2003, Chapter 17. For a more advanced discussion, see Wooldridge 2002, Chapter 16- 16 Lancaster (1979), Moffitt (1985a,b), and Solon (1985) were among the first to use parametric and semiparametric hazard models to estimate the impact of U1 benefits on unemployment duration. For a detailed discussion of survival analysis, see Cleves, Gould, and Gutierrez,2004, Allison, 1984, Cameron and Trevedi , 2005, Chapter 17, or Wooldridge 2002, Chapter 20. 42 might expect to see similar results using actual weeks jobless as the dependent variable. Columns 3 and 4 of Tables 9 and 10 show that this is not the case, with small, statistically insignificant coefficients for wba. Neither censored nor Weibull regression attempt to address endogeneity bias, so these estimates likely understate the effect of higher benefits on jobless duration- Given these findings it seems clear that econometric techniques that account for censoring in the dependent variable are unconvincing approaches to estimating the effect of unemployment benefits on jobless duration. Rather, the primary consideration in estimating the effect on duration is finding exogenous variation in UI weekly benefit amounts. While three social experiments have been devoted to determining whether a reemployment bonus would shorten unemployment spells of U1 recipients (Robins and Spiegelman 2001), the reemployment bonus experiments did not examine whether changes in weekly benefit amounts would affect unemployment duration. Indeed, it is difficult to imagine an experiment in which workers were assigned randomly to different weekly benefit amounts or potential durations of benefits. Accordingly, we need to find a quasi-experiment — an event that gives rise to (arguably) exogenous changes in UI benefits. Meyer (1989) did just this by taking advantage of 15 increases in the maximum weekly benefit amount that occurred in 5 states during 1979 through 1983. Using a difference-in-differences approach, he examined the effects of these benefit increases on the duration of unemployment using data from the Continuous Wage Benefit History, which comes from UI administrative records and contains information on the length of insured unemployment spells. Meyer’s difference-in-differences approach is 43 attractive because it avoids the problems caused by the correlation of past earnings histories and weekly benefit amounts. 17 Here we follow a similar approach, using a difference-in-difference-in- difference model and data on both insured duration (from the administrative records component of the Mathematica data) and jobless duration (from the survey component). Thirteen of the twenty-five states in our data raised the maximum weekly benefit amount on July 1, 1998 (see Table 7). These increases were regularly scheduled increases and were not in response to changing labor market conditions. Treating these increases as a quasi-experiment, we first divide workers in the sample into two groups based on whether they began their benefit year before or after July 1, 1998. We then divide workers into high-eamings and low-earnings groups. The high- earnings group consists of workers whose earnings histories would (or did) put them at the maximum weekly benefit amount before July 1, 1998 — these are the workers who would be affected by an increase in the maximum weekly benefit amount. The low-earnings group consists of workers whose earnings histories place them below the pre-July maximum weekly benefit amount, so they would not be affected by an increase in the maximum. Figure 9 illustrates these groups and the effect of an increase in the U1 benefit schedule. Finally, we divide workers by whether or not they were in one of the 13 states that increased the maximum weekly benefit amount on July 1, 1998. The basic difference-in-difference-in-differences (DDD) regression model is: 17 Meyer, Viscusi, and Durbin (1995) also use this approach to examine the effects of a similarly structured increase in Workers’ Compensation benefits. 44 In t = 20 + Alofter - high - increase + Azafter - high + 113after - increase + A4high - increase + Asafter + Aéhilgh (3) + Mincrease + 2 (pistate, i=1 + XB + 8 where t denotes either duration of insured unemployment (durui) or duration of joblessness (jobless), after is an indicator equal to 1 for a U1 claim starting after July 1, 1998, high is an indicator equal to l for a claimant whose earnings history places him or her at the maximum weekly benefit amount, and increase is an indicator equal to 1 for claimants residing in states that increased the maximum weekly benefit amount on July 1,‘ 1998. The basic specification also includes state indicators for states where an increase took place. Augmented specifications add various additional controls (X ), as will be seen in Table 11. The main coefficient of interest is it] , the percentage increase in duration resulting from an increase in the weekly benefit amount. Empirical Findings The estimates obtained using the DDD method explained above can be seen in Table 11, with results from using weeks of insured unemployment seen in the six specifications in Panel A. The point estimates suggest that an increase in the maximum benefit amount increases weeks of benefits collected by about 4 to 9 percent; however, none of these estimates is statistically significant. In the data we are examining, the average increase in the maximum benefit amount was about 5 percent. 45 Accordingly, the point estimates suggest that a 5-percent increase in the weekly benefit amount increases the duration of insured unemployment by 4 to 9 percent — an increase of roughly 0.7 to 1.5 weeks of insured unemployment for those at the median weeks of benefits (16.5). Meyer’s estimates suggest that an increase in the maximum benefit amount increases insured duration by between 3 and 8 percent. Notably, four of Meyer’s six estimates are statistically significant at the 5% level, perhaps because his samples are in the range of 12,500 to 16,000, or because the increases in the maximum weekly benefit amount for that time period are somewhat larger— an average increase of about 9 percent. Meyer's estimates correspond to an increase of 0.6 to 1.8 weeks. Meyer interprets his estimates as the effect of the increases on actual weeks jobless, but it is important to recall that they are based on insured unemployment spells. I 8 For reasons discussed above, it seems more reasonable to interpret the estimates as effects of a benefit increase on insured duration. Estimates displayed in Panel B use actual jobless duration. The point estimates suggest that an increase in the maximum weekly benefit amount increases the duration of jobless spells by 6 to 12 percent. At the median weeks jobless in our sample, 30.6, this corresponds to an increase of 1.8 to 3 .7 roughly double that of the effect on insured weeks alone. While the effects are relatively large, and much larger than any of the estimates obtained using equation (1), none are statistically significant. The estimates using weeks jobless are generally higher than those using weeks of benefits received, offering some support for the idea that insured duration and total duration respond differently to changes in the weekly benefit amount. 18 Meyer includes a brief discussion of why data on weeks insured can be used to estimate weeks jobless (p. 21-22), but appears to have assumed that benefits only affect the insured portion of the unemployment spell. 46 It should be noted that the effects estimated by our approach would not necessarily mirror the effect we would see in the entire population. Since the changes involved are with the maximum benefit amount, and not the underlying benefit formula, the effect is coming entirely from high-wage earners. It is possible that the treatment effect for this group differs from the effect we would see from increasing benefits for lower wage workers. 19 One possible concern with the approach used in Table 11 is that Ul claimants may have known about the upcoming benefit increase, and timed their claim to come just after the increase rather than just before.20 (For example, a claimant in Tennessee who would be eligible for the maximum weekly benefit of $255 if she filed on July 1 would only get $220 if she filed on June 30.) If individuals planning on longer unemployment spells are more likely to wait until after the change to file for benefits, the estimated effect would be biased upward. Table 12 shows the results of the same difference-in-difference approach, but excluding from the sample all claims within 30 days of the July 1 change. This exclusion shrinks all estimates of the effect of the weekly benefit amount, which may indicate that individuals expecting longer unemployment spells may have been more likely to time their claims to take advantage of the increase. As in Table 1 1, estimates using weeks jobless are generally higher than those using weeks of benefits received. Given the relative difficulty of obtaining data on the full duration of jobless spells, it is worth considering whether it is possible to obtain estimates of the effect on actual weeks jobless using only data on weeks of insured unemployment. Table 13 19 There are periodic changes in the minimum weekly benefit amount in some states, so it would theoretically be possible to measure the effect of benefits for low-wage workers. The minimum levels, however, are low enough that very few individuals are affected by such increases. 20 Rogers (1998) finds that individuals have “significant foresight” about changes in UI provisions. 47 shows the results of using a difference-in-difference technique with censored regression (Panel A) or a representative hazard model technique (Panel B) on insured weeks. As in Table 10, we report estimates from an accelerated failure-time Weibull model; this allows the results to be interpreted, as in our other models, as the percentage effect of a weekly benefit increase on spell duration.21 The point estimates from both techniques are generally larger than the estimates found in the other estimations of equation 2. This is consistent with the idea that treating insured weeks as a censored measure of total weeks leads to overestimates of the effect of higher benefits, but the estimates are still statistically insignificant. Finally, we expand our data to the full sample of administrative data. In table 14, we report estimates of equation (2) for the full sample (comparable to the estimates in Table 11, Panel A), the full sample omitting those filing within 30 days of the increase (comparable to the estimates in Table 12, Panel A), and using censored regression (comparable to the estimates in Table 13, Panel A). Since the administrative data does not include all the variables necessary for specifications 2-6, we only estimate the first specification. In each case, the point estimate is smaller than the corresponding estimates using the survey sample, and each is still statistically insignificant. Conclusions The endogeneity of weekly benefit amounts is a problem that cannot be ignored. Both the potential duration and level of weekly benefits depend intentionally on variables associated with past employment patterns. It stands to reason that 21 Again, we have also estimated Weibull hazard models [the hazard model analog to the accelerated failure-time model reported here] and semiparametric Cox proportional hazard models. These latter two models again give similar results, suggesting that little is gained by relaxing the distributional assumption imposed in the Weibull. 43 unobservables that affect unemployment duration are correlated with both the weekly benefit amount and potential duration of benefits. Estimates of the effect of benefits based on insured duration tend to mask this problem due to the disproportionate right- censoring of the spells of those who have short benefit durations. Using data on actual unemployment duration reveals the limitations of methods that do not account for the relationship between earnings history and unemployment duration. Our quasi- experimental approach finds that increased benefits do not have a statistically significant effect on either insured duration or weeks jobless. 49 Figure 8 Budget constraint with unemployment benefits and an increase in benefits Income wage B Leisure Umax Notes: 1. AB is the budget constraint absent unemployment benefits. ACD is the budget constraint With unemployment benefits. AEF is the budget constraint with an increase in the weekly benefit amount from wga to wga’. Go 00:00 0002 .6 Lothuémf 0550 09:58 20.003 000.020. 09» oov 0mm oom 0mm CON om.- ooF om mm on 0300200 50:09 /. k b... l -. mm 02 me om? mt. oom 000000 mNN c0e0 omv 30.5 36: 0:20 209500: 0320 50:00:80 Awe EnoEm 50:00 582, «£3056 «50:00 ESP-glue a: 0&005093 :0 £3; 0_350£0m “$0209 _: O>muflhum3=_ 0 032.... 51 Figure 10 Distribution of Benefit Duration N _ m .— E U) C 1— _. 0.) O LO Q _ O — l l i 0 10 20 30 4O 50 Weeks of benefits received Notes: 1. Spike occurs at the most common exhaustion point of 26 weeks- 2. Sample size = 3,158 52 Figure 11 Distribution of Jobless Duration (including censored observations) Density 0 50 1 00 1 50 200 Weeks jobless (censored) Notes: 1. The increase in density after 75 weeks is almost entirely due to censored observations. (Workers who fail to find reemployment before the survey date.) See Figure 12 for the distribution without censored observations. 2. Sample size = 3,158 Figure 12 Distribution of Jobless Duration (excluding censored observations) Density .02 l .01 1 ol 1, l 0 T 1 0 150 Weeks jobless Notes: 1. Only includes workers who found reemployment before the survey date. 2. Sample size = 3,158 54 Table 7 Sample descriptive statistics and changes in maximum weekly benefit amounts maximum weekly benefit amount (6) median median median weekly weeks of weeks Stgte N before after benefits 13) benefits Jobless CA 429 230 same 150 21.0 40.6 FL 127 275 same 229 16.9 39.0 GA 71 224 244 200 14.3 33.3 HI 103 356 same 330 24.0 40.7 ID 94 265 273 198 14.0 23.6 IL 110 269-355 same 259 19.1 35.1 IA 111 239-293 251-307 248 12.3 16.6 KY 100 256 268 238 8.8 22.6 ME 101 216-324 227-340 202 18.0 31.3 MI 156 300 same 287 15.0 19.0 MN 116 , 331 same 306 15.0 19.8 MS 99 180 190 162 15.9 31.7 MT 93 237 246 197 14.7 19.7 NJ 132 390 same 255 20.1 41.8 NY 169 300 same 268 26.0 53.3 NC 94 322 same 221 12.1 45.6 OH 102 267-358 same 255 7 17.9 OK 114 255 262 254 14.0 24.6 PA 210 375-383 same 262 19.9 37.9 RI 89 347-433 364-455 255 13.9 25.7 TN 102 220 255 197 16.1 36.6 TX 167 280 same 235 17.1 41.3 VA 89 226 228 220 11.0 26.0 WA 76 384 410 221 19.2 24.2 WI 104 282 290 257 9.2 24.1 Total 3158 na na 224 16.5 30.6 Sources: Authors' tabulations of data from Needels, Corson, and Nicholson (2002) and US Department of Labor (various years). 55 Table 8 OLS estimates of insured unemployment duration and jobless duration (1) (2) (3) (4) Dependent variables Independent variables In (Weeks of benefits) ln(Weeks jobless) weekly benefit amount 0.1023 0.1213 0.0714 0.0661 ($1005) (0.0358) ** (0.0366) ** (0.0467) (0.0477) potential duration -- 0.0122 -- 00034 of Ul benefits (0.0049) * (0.0064) age 0.0118 0.0118 0.0166 0.0166 (0.0017) '* (0.0017) ** (0.0022) ** (0.0022) *" years of education 0.0267 0.0272 -0.0179 -0.0180 (0.0060) ** (0.0060) ** (0.0078) * (0.0079) * pre-UI job tenure -0.0115 -0.0116 -0.0009 -0.0009 (years) (0.0027) ** (0.0027) ** (0.0035) (0.0035) white -0.1322 -0. 1332 -0.0953 -0.0950 (0.0410) ** (0.0410) ** (0.0534) (0.0534) female 0.0275 0.0191 0.0884 0.0907 (0.0375) (0.0376) (0.0488) (0.0490) married -0.0639 -0.0623 -0.0499 -0.0503 (0.0376) (0.0376) (0.0491) (0.0491) dependents < 18 0.0366 0.0367 0.0308 0.0308 (0.0157) * (0.0157) * (0.0205) (0.0205) log of base period -0.1662 -0.2206 -0.2665 -0.2514 earnings (0.0408) *" (0.0463) ** (0.0532) ** (0.0604) ** state unemployment rate 0.1477 0.1410 0.1410 0.1429 (0.0190) ** (0.0192) ** (0.0248) ** (0.0250) ** constant 2.4744 2.6958 4.6234 4.5620 (0.3483) ** (0.3593) ** (0.4538) ** (0.4686) ** Observations 3158 3158 3158 3158 Adjusted R-squared 0.0537 0.0553 0.0526 0.0523 Notes: 1. OLS estimates of equation (1 ). Coefficients are the percentage change in weeks of benefit duration orjobless duration with respect to a 1 unit increase in each independent variable. 2. ** significantly different from 0 at the 1% level; * significantly different from 0 at the 5% level. Standard errors in parentheses. 3. Each specification includes a set of industry indicators. 56 Table 9 Censored regression estimates of insured unemployment duration and jobless duration (1) (2) (3) Dependent variables (4) Independent variables In (Weeks of benefits) ln(Weeks jobless) weekly benefit amount 0.2694 0.2229 0.0816 0.0812 ($1005) (0.0647) ** (0.0660) ** (0.0585) (0.0597) potential duration - -0-0362 -- -0.0003 ofUl benefits (0.0093) ** (0.0081) age 0.0218 0.0216 0.0238 0.0238 (0.0031) ** (0.0031) ** (0.0028) ** (0.0028) ** years of education 0.0467 0.0455 -0.0396 -0.0396 (0.0109) ** (0.0109) ** (0.0102) ** (0.0102) ** pre-Ul job tenure -0.0154 -0.0153 0.0027 0.0027 (years) (0.0047) ** (0.0047) ** (0.0044) (0.0044) white -0.2634 —0.2593 -0.1195 -0.1195 (0.0742) ** (0.0744) ** (0.0671) (0.0671) female 0.0440 0.0639 0.1043 0.1045 (0.0675) (0.0680) (0.0613) (0.0615) married -0.0960 -0.0990 -0.0463 -0.0464 (0.0676) (0.0678) (0.0616) (0.0616) dependents < 18 0.0649 0.0645 0.0388 0.0388 (0.0283) * (0.0283) * (0.0257) (0.0257) log of base period -0.5494 -0.4096 -0.3330 -0.3318 earnings (0.0758) ** (0.0840) ** (0.0669) ** (0.0758) ** state unemployment rate 0.2326 0.2505 0.1636 0.1637 (0.0339) ** (0.0343) ** (0.0310) ** (0.0313) ** constant 5.5249 5.0642 5.2965 5.2915 (0.6461) ** (0.6603) ** (0.5698) ** (0.5875) ** Observations 3158 3158 3158 3158 censored 1508 1508 1508 1508 LR chi—squared 215 230.31 213.48 213.48 Pseudo R-squared 0.0252 0.027 0.0198 0.0198 Notes: 1. Censored regression estimates of equation (1). Coefficients are the percentage change in weeks of benefit duration or jobless duration with respect to a 1 unit increase in each independent variable. 2. ** significantly different from 0 at the 1% level; * significantly different from 0 at the 5% level. Standard errors in parentheses. 3. Each specification includes a set of industry indicators. 57 Table 10 Weibull (AFT) regression estimates of insured unemployment duration and jobless duration (l) (2) (3) l4) Dependent variables Independent variables Weeks of benefits Weeks jobless weekly benefit amount 0.1713 0.1443 0.0380 0.0472 ($1005) (0.0544) ** (0.0557) * (0.0523) (0.0534) potential duration -- -0.0187 -- 0.0062 of U! benefits (0.0074) * (0.0071) age 0.0181 0.0179 0.0219 0.0220 (0.0028) ** (0.0028) ** (0.0025) ** (0.0025) ** years of education 0.0358 0.0358 -0.0523 -0.0519 (0.0094) ** (0.0095) ** (0.0089) ** (0.0089) ** pre-Ul job tenure -0.0084 -0.0078 0.0143 0.0142 (years) (0.0041) * (0.0041) (0.0041) " (0.0041) ** white -0.2305 -0.2297 -0.1208 -0.1211 (0.0673) ** (0.0678) ** (0.0601) * (0.0601) * female 0.1146 0.1252 0.0921 0.0882 (0.0594) (0.0598) ‘ (0.0546) (0.0547) married -0.0914 -0.091 1 -0.0249 «0.0248 (0.0598) (0.0601 ) (0.0549) (0.0549) dependents < 18 0.0520 0.0524 0.0495 0.0503 (0.0251) (0.0252) * (0.0228) * (0.0228) * log of base period -0.3711 -0.3031 -0.2529 -0.2797 earnings (0.0640) ** (0.0706) ** (0.0595) '* (0.0670) ** state unemployment rate 0.2122 0.2187 0.1096 0.1061 (0.0294) ** (0.0296) ** (0.0279) ** (0.0282) ** constant 4.7260 4.5469 5.6318 5.7350 (0.5367) ** (0.5547) ** (0.5011) ** (0.5130) ** shape parameter (p) 0.884566 0.879708 0.780283 0.780388 Observations 3158 3158 3158 3158 LR chi-squared 180.71 187.07 262.89 263.63 Notes: 1. Weibull (AFT) regression estimates of equation (1 )- Coefficients are the percentage change in weeks of benefit duration or jobless duration with respect to a 1 unit increase in each independent variable. 2. ** significantly different from 0 at the 1% level; * significantly different from 0 at the 5% level. Standard errors in parentheses. 3. Each specification includes a set of industry indicators. 58 Table 11 DDD estimates of UI benefit duration and jobless duration Panel A Weeks of benefits Specifications received (1 ) (2) (3) (4) (5) (6) afier°high-increase 0.0704 0.0697 0.0525 0.0392 0.0883 0.0681 (0.1545) (0.1533) (0.1530) (0.1510) (0.1434) (0.1423) log of base period -- -- -0.1709 -0.1313 -0.1301 -0.1522 earnings (0.0379) ** (0.0382) ** (0.0363) ** (0.0366) ** potential duration of -- -- 0.0127 0.0106 0.0087 0.0107 Ul benefits (0.0052) * (0.0051) * (0.0049) * (0.0049) * age, race, gender, -- yes yes yes yes yes educafion industry indicators -- -- -- yes yes yes recall indicators -- -- -- —- yes yes occupation, married, -- -- -- -- -- yes dependents Observations 3158 3158 3158 3158 3158 3158 Adjusted R-squared 0.0295 0.0441 0.0497 0.0756 0.1665 0.1825 59 Table 11 (continued) Panel 8 Weeks jobless Specifications (1) (2) (3) (4) (5) (6) after-high-increase 0.0667 0.0764 0.0628 0.061 1 0.1 197 0.1064 (0.2017) (0.1991) (0.1984) (0.1977) (0.1915) (0.1914) log of base period -- -- -0.2127 -O.2037 -0.2114 -0.2300 earnings (0.0492) ** (0.0500) ** (0.0485) ** (0.0492) ** potential duration of -- -- -0.0028 -0.0050 -0.0070 -0.0055 Ul benefits (0.0067) (0.0067) (0.0065) (0.0065) age, race, gender, -- yes yes yes yes yes educafion industry indicators -- -- -- yes yes yes recall indicators -- -- -- -- yes yes occupation, married, -- -- -- -- -- yes dependents Observations 3158 3158 3158 3158 3158 3158 Adjusted R-squared 0.0246 0.0495 0.0573 0.0649 0.1234 0.1270 Notes: 1. Estimates of equation (2). Coefficients give the percentage change in weeks of benefit duration (Panel A) or jobless duration (Panel B) with respect to a 1 unit increase in each independent variable. the percentage change in weeks of benefits received or weeks jobless for those affected by the July 1 increase in benefits. 2. ** significantly different from 0 at the 1% level; * significantly different from 0 at the 5% level. Standard errors in parentheses. 60 For after-high-increase, the coefficients give Table 1 2 DDD estimates of Ul benefit duration and jobless duration, omitting claims within 30 days of the increases Panel A Weeks of benefits Specifications received (1 ) (2) (3) (4) (5) (6) after-high-increase 0.0468 0.0470 0.0354 0.0215 0.0513 0.0107 (0.1697) (0.1683) (0.1680) (0.1667) (0.1592) (0.1577) log of base period -- -- -0.1316 -0.1164 -0.1253 -0.1455 earnings (0.0413) ** (0.0417) ** (0.0400) ** (0.0402) ** potential duration of -- -- 0.0109 0.0098 0.0093 0.0104 Ul benefits (0.0057) * (0.0057) (0.0054) (0.0054) * age, race, gender, -- yes yes yes yes yes educafion industry indicators -- -- -- yes yes yes recall indicators -- -- -- -- yes yes occupation, married, -- -- - -- -- yes dependents Observations 2601 2601 2601 2601 2601 2601 Adjusted R-squared 0.0260 0.0426 0.0456 0.0626 0.1455 0.1643 61 Table 12 (continued) Panel 8 Weeks jobless Specifications (1) (2) (3) (4) (5) (6) after-high-increase 0.0535 0.0474 0.0344 0.0275 0.0665 0.0325 (0.2155) (0.2125) (0.2120) (0.2117) (0.2056) (0.2052) log of base period -- -- -0.1529 -0.1658 -0.1848 —0.2039 earnings (0.0521) ** (0.0530) ** (0.0516) ** (0.0523) ** potential duration of -- - -0.0058 -0.0068 -0.0073 -0.0062 Ul benefits (0.0072) (0.0072) (0.0070) (0.0070) age, race, gender, -- yes yes yes yes yes educafion industry indicators -- -- -- yes yes yes recall indicators -- - -- -- yes yes occupation, married, -- -- —- -- -- yes dependents Observations 2601 2601 2601 2601 2601 2601 Adjusted R-squared 0.0257 0.0526 0.0577 0.0630 0.1162 0.1226 Notes: 1. Estimates of equation (2). The sample is limited to those filing 30 days before or after the July 1 increases. Coefficients give the percentage change in weeks of benefit duration (Panel A) or jobless duration (Panel B) with respect to a 1 unit increase in each For afterohigh-increase, the coefficients give the percentage change in weeks of benefits received or weeks jobless for those affected by the July 1 independent variable. increase in benefits- 2. ** significantly different from 0 at the 1% level; * significantly different from 0 at the 5% level. Standard errors in parentheses. Table 13 DDD estimates of Ul benefit duration and jobless duration using weeks of benefits received Panel A Censored Specifications regression (1 ) (2) (3) (4) (5) (6) after-high-increase 0.1494 0.1390 0.1451 0.1125 0.2248 0.2079 (0.2735) (0.2711) (0.2709) (0.2667) (0.2516) (0.2486) log of base period -- -— -0.2789 -0.2204 -0.2224 -0.2738 earnings (0.0688) *" (0.0691) ** (0.0655) ** (0.0659) ** potential duration -- -- -0.0432 -0.0461 -0.0459 -0.0418 ofUl benefits (0.0099) ** (0.0098) ** (0.0092) ** (0.0091) ** age, race, gender, -- yes yes yes yes yes educafion industry indicators -- -- -- yes yes yes recall indicators -- -- - - yes yes occupation, married, -- -- - -- -- yes dependents Observations 3158 3158 3158 3158 3158 3158 censored 1508 1508 1508 1508 1508 1508 LR chi-squared 110.53 167.74 235.24 313.01 631.56 712.42 Pseudo R-squared 0.013 0.0197 0.0276 0.0367 0.0741 0.0836 63 Table 13 (continued) Panel B Weibull (AFT) Specifications regression (1 ) (2) (3) (4) (5) (6) afier °high°increase 0.2026 0.1966 0.1675 0.1464 0.1227 0.0359 (0.2327) (0.2307) (0.2332) (0.2322) (0.2178) (0.2159) log of base period -- -- -0.2030 —0.1844 -0.2330 -0.0349 earnings (0.0609) ** (0.0616) ** (0.0583) ** (0.0076) ** potential duration of -- -- -0.0278 -0.0289 -0.0385 -0.0349 UI benefits (0.0082) ** (0.0079) ** (0.0075) ** (0.0076) ** age, race, gender, -- yes yes yes yes yes educaflon industry indicators -- -- .. yes yes yes recall indicators -- -- -- -- yes yes occupation, married, -- - -- -- -- yes dependents Shape parameter (p) 0.8824 0.8902 0.8814 0.8866 0.9471 0.9571 Observations 3158 3158 3158 3158 3158 3158 LR chi-squared 103.06 167.94 208.29 267.37 677.73 755.95 Notes: 1. Estimates of equation (2). Coefficients give the percentage change in weeks jobless with respect to a 1 unit increase in each independent variable. For afler-high-increase, the coefficients give the percentage change in weeks jobless for those affected by the July 1 increase in benefits. 2. ** significantly different from 0 at the 1% level; * significantly different from 0 at the 5% level. Standard errors in parentheses. 64 Table 14 DDD estimates of UI benefit duration and jobless duration using only administrative data Censored DDD DDD (Omit) regression after-high-increase -0.0059 -0.0186 -0.0002 (0.0563) (0.0603) (0.0765) log of base period -- -- -- earnings potential duration of -— -- -- Ul benefits age, race, gender, -- -- __ educafion industry indicators -— -- -- recall indicators -- -- -- occupation, married, -- -- -- dependents Observations 25,572 21,112 25,572 Adjusted R-squared 0.0094 0.0077 -- Pseudo R-squared -- -- 0.0043 LR chi-squared -- -- 306.62 Notes: 1. Estimates of equation (2). Coefficients give the percentage change in weeks of benefit duration (Columns 1 and 2) or jobless duration (Column 3) for those affected by the July 1 increase in benefits. 2. ** significantly different from 0 at the 1% level; * significantly different from 0 at the 5% level. Standard errors in parentheses. 65 BIBLIOGRAPHY Allison, Paul D. Event History Analysis.“ Regressionfor Longitudinal Event Data. Beverly Hills and London: Sage Publications, 1984. Black, Dan A., Jeffrey Smith, Mark C. Berger, and Bret Noel. “Is the Threat of Reemployment Services More Effective than the Services Themselves? Evidence from Random Assignment in the UI System.” American Economic Review 93 (September 2003): 1313-1327. Cameron, A.Colin and Trivedi, Pravin K. Microeconometrics: Methods and Applications. Cambridge University Press, New York, May 2005. Classen, Kathleen. “The Effect of Unemployment Insurance on the Duration of Unemployment and Subsequent Earnings.” Industrial and Labor Relations Review 30 (July 1977): 438-444. Classen, Kathleen P. "Unemployment Insurance and Job Search." In Studies in the Economics of Search, edited by S.A. Lippman and J.J. McCall. Amsterdam: North— Holland, 1979. Pp. 191—219. Cleves, Mario A., Gould, William W., and Gutierrez, Roberto G. An Introduction to Survival Analysis Using STA T A.. College Station, TX: Stata Press, 2004. Ehrenberg, Ronald G. and Ronald L. Oaxaca. "Unemployment Insurance, Duration of Unemployment, and Subsequent Wage Gain." American Economic Review 66 (December 1976): 754-766. Folks, Laura, Melynda Ihrig, and Karen Needels. “Study of Unemployment Insurance Exhaustees: Public Use File Documentation.” Princeton, NJ: Mathematica Policy Research, Inc., October 2001. Holen, Arlene. "Effects of Unemployment Insurance Entitlement on Duration and Job Search Outcome." Industrial and Labor Relations Review 30 (July 1977): 445—450. Katz, Lawrence F. and Bruce D. Meyer. "The Impact of the Potential Duration of Unemployment Benefits on the Duration of Unemployment." Journal of Public Economics 41 (February 1990): 45-72. Kiefer, Nicholas M. "Economic Duration Data and Hazard F unctions." Journal of Economic Literature 26 (June 1988): 646—679. Lancaster, Tony. "Econometric Methods for the Duration of Unemployment." Econometrica 47 (July 1979): 939—956. 66 Mathematica Policy Research. "A Longitudinal Study of Unemployment Insurance Exhaustees: Waves I and II." Report prepared for the Employment and Training Administration, US. Department of Labor, January 1976. Moffitt, Robert. "Unemployment Insurance and the Distribution of Unemployment Spells." Journal of Econometrics 28 (April l985a): 85—101. Moffltt, Robert. "The Effect of the Duration of Unemployment Benefits on Work Incentives: An Analysis of Four Data Sets." Unemployment Insurance Service Occasional Paper 85—4. Washington, DC: US. Department of Labor, Employment and Training Administration, Unemployment Insurance Service, 1985b. Meyer, Bruce D. “A Quasi-Experimental Approach to the Effects of Unemployment Insurance.” National Bureau of Economic Research Working Paper No. 3159, November 1989. Meyer, Bruce D., Viscusi, W. Kip, and Durbin, David L. “Workers’ Compensation and Injury Duration: Evidence from a Natural Experiment.” American Economic Review 85 (June 1995): 322-340 Needels, Karen, Walter Corson, and Walter Nicholson. “Left Out of the Boom Economy: UI Recipients in the Late 19903.” ETA Occasional Paper 2002—03. Washington, DC: US Department of Labor, Employment and Training Administration, May 2002. Robins, Philip, and Robert G. Spiegelman, editors. Reemployment Bonuses in the Unemployment Insurance System: Evidence from Three Field Experiments. Kalamazoo, MI: W.E. Upjohn Institute, 2001. Rogers, Cynthia L. “Expectations of Unemployment Insurance and Unemployment Duration.” Journal of Labor Economics 16 (July 1998): 630-666. Solon, Gary. "Work Incentive Effects of Taxing Unemployment Benefits." Econometrica 53 (March 1985): 295—306. Sueyoshi, Glenn T. "Semiparametric Proportional Hazards Estimation of Competing Risks Models with Time-Varying Covariates." Journal of Econometrics 51 (January/February 1992): 25-58. Wooldridge, Jeffrey M. Introduction to Econometrics: A Modern Approach, second edition. Cincinatti, OH: Thomson South-Western, 2003. Wooldridge, Jeffery M. Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: MIT Press, 2002. 67 US Department of Labor. Comparison of State Unemployment Insurance Laws. Prepared by the Employment and Training Administration, Office of Workforce Security, various years. 68 CHAPTER 3 EMERGENCY UNEMPLOYMENT COMPENSATION AND THE DURATION OF UNEMPLOYMENT Introduction In each economic downturn since 1958, the US. Congress has extended by fiat the potential duration of unemployment benefits available to laid-off workers. Roughly a dozen empirical studies have examined the extent to which extended unemployment benefits lead to longer spells of unemployment in the United States. The studies have used a variety of data sources covering different time periods, different states, and different extended benefit programs. Also, they have used a variety of the econometric methods to model duration and the timing of reemployment. It is perhaps not surprising that the estimated effect of an additional week of unemployment insurance (UI) on the duration of unemployment varies dramatically over these studies — from 0 to 0.9 week. By implication, a typical 13- week emergency extension of benefits could add anywhere from 0 to nearly 12 weeks to a typical worker’s spell of unemployment. This paper seeks to understand the extent to which differing econometric models and assumptions can explain the wide range of estimates of the impact of extended benefits. To do so, we use a single detailed dataset collected specifically to analyze the Emergency Unemployment Compensation Act of 1991 (EUC), which extended benefits between November 1991 and February 1994. The data come from the administrative records of 16 states, supplemented by an extensive follow-up 69 survey. They include information on 2,304 workers who began collecting UI benefits between January 1991 and September 1993. Our empirical work addresses two main questions. First, when Congress increases the potential duration of U1 by one week, by how much does the average number of weeks of U1 paid to claimants increase? This first question is important because policymakers need to know the financial implications of extending UI benefits. Second, when the potential duration of U1 increases by one week, by how much does the average jobless spell increase? This is the question usually addressed in empirical research on extended benefits because search and matching models imply that increased availability of benefits will lead workers to reduce their job search effort and extend their spells of joblessness. To address these questions, we systematically apply a range of estimators and assumptions to the data with the goal of sorting out how these differences affect statistical inferences. Our results suggest that much of the variation in existing estimates can be attributed to variation in econometric methods and assumptions. We describe the relationship between methods and inferences and offer some speculation on which methods (and hence which inferences) are most convincing. The next section gives background on unemployment benefit duration and extended benefits, with a focus on EUC. We then briefly review previous studies of the effect of UI benefit duration on the duration of unemployment. After describing the data we use, we present estimates and discuss the findings from different econometric specifications. 70 Policy Background ’7’) The UI system in the United States can be thought of as having three “tiers.”“ The first tier consists of regular state UI benefits, which have a maximum potential duration of 26 weeks in all but two states. (The exceptions are Massachusetts and Washington, where the maximum potential duration is 30 weeks.) Regular state benefits vary in amount and duration from state to state, are funded by experience- rated payroll taxes collected from employers. and have the stated goal of offering a limited number of weeks of benefits during a temporary spell of unemployment. In 8 states, the potential duration of benefits is 26 weeks for all claimants who qualify for any benefits:23 In all other states, potential duration varies with a claimant's work experience in the base period — roughly the year preceding the claim for benefits — according to formulas that vary from state to state. The second tier of the UI system is the “standby” Extended Benefits program (EB), a federal program enacted by Congress in 1970 that is intended to activate automatically and increase the potential duration of U1 benefits in an economic downturn. EB extends benefits to claimants who exhaust their regular state benefits by one—half of regular benefit duration, up to 13 weeks. It is financed equally by federal and state revenues and provides the same weekly benefit amount to a claimant as the regular state program. 22 For detailed descriptions of U1 benefit duration in the United States, see Woodbury 3nd Rubin (1995). Connecticut, Hawaii, Illinois, Maryland. New Hampshire New York. Vermont, and West Virginia. 71 Vroman and Woodbury (2004, chapter 3) and Wenger and Walters (2006) show that during its first 13 years (1971—1983) EB activated frequently, but since then it has activated rarely and is no longer an important program for the unemployed. This is at least partly by design: In 1981 Congress revised EB to make it more difficult for EB to activate. Also, the indicator that activates EB — the insured unemployment rate in a state — has trended downward since the early 19803, making it difficult for EB to activate. Finally, state governors have frequently suspended EB when federal emergency benefits are available because EB is funded half by state UI trust funds, whereas emergency benefits are wholly federally funded. The third tier of the UI system consists of federal “emergency” benefit extensions. Unlike the EB program, emergency extensions do not activate automatically. Rather, Congress has enacted them by fiat in every recession starting with the recession of 1958 — a total of seven to date. Table 15 gives a brief history of the three most recent emergency extensions, which have varied significantly in the number of additional weeks they provide, sources of fimding, and eligibility criteria. In this paper, we examine the Emergency Unemployment Compensation program (EUC), the sixth emergency benefit extension, which was enacted in November 1991 in response to the downturn of the early 19905. EUC was the most complicated emergency benefit extension to date: It went through five “phases” as Congress amended it, provided different potential benefit durations among different states at a given time (depending on whether a state was classified as “high unemployment” or low unemployment”), and changed potential durations within a state over time (both by Congressional fiat and when a state changed between the “high” and “low” unemployment classification) — see again Table 15. As an example, Table 16 shows how the potential duration of benefits varied under EUC in Pennsylvania. Between November 1991, when EUC became effective, and February 1994, when Phase V of EUC ended, the potential duration of benefits in Pennsylvania changed nine times. Six of these changes resulted from enactment of EUC or legislative changes during the life of EUC, and three resulted because Pennsylvania was reclassified as high- or low-unemployment. The empirical work below takes advantage of the variation over time and among states in the potential duration of unemployment; however, we do not claim that EUC represents a quasi-experiment because it was implemented during a recession and provided longer extensions in states designated as high-unemployment than in those designated as low-unemployment. Because longer spells of unemployment are expected in bad labor market conditions, and EUC coincided with bad labor market conditions, potential benefit duration under EUC must be considered endogenous. We try to c0pe with this endogeneity by including state-level indicators and indicators of the labor market conditions facing unemployed workers in an attempt to control for workers’ reemployment opportunities. This amounts to a “plug- in” approach to controlling for otherwise omitted variables. Also, the link between labor market conditions in a state and potential benefit duration under EUC and was quite loose. Specifically, although the recession of the early 19905 was concentrated in the northeastern states and on the west coast, all states were covered by EUC and had 73 benefit extensions, including many (like Illinois and North Carolina) that experienced only mild downturns. Previous Research Efforts to estimate the impact of extended UI benefits date to the mid 19703, not long after the standby EB program came into being. Holen (1976) used administrative records on UI claim spells and regressed the number of weeks of U1 benefits received by a claimant on appropriate explanatory variables, including measures of the UI replacement rate and potential duration of benefits. Her estimates suggested that a one-week increase in the potential duration of benefits led to an additional 0.8 weeks of benefits received. (Table 17 summarizes Holen’s and others’ findings.) Solon (1979) found that a one-week increase in potential duration led to an additional 0.16 weeks of unemployment for a sample of New York workers. Classen (1979) and Newton and Rosen (1979) both examined UI claim spells but estimated Tobit models to handle truncation of the dependent variable (weeks of U1 claimed). Classen found that a one-week increase in potential duration led to at most an additional 0.12 week of benefits received, whereas Newton and Rosen's estimates suggested an additional 0.6 week. Katz and Ochs (1980), Moffitt (1985b), and Solon (1985) estimated parametric models of unemployment duration, along the lines of those described by Lancaster (1979 ). Katz and Ochs estimated that an additional week of potential duration increased unemployment duration by 0.17—0.23 weeks. Moffitt (1985b) obtained a range of estimates between 0.17 and 0.45 weeks when he examined a lS-state sample 74 and data from Georgia data. Solon (1985) examined Georgia data and estimated that an additional week of potential benefit duration leads to 0.36 additional weeks of unemployment. Moffitt (l985a, I985b), Ham and Rea (1987), Grossman (1989), and Katz and Meyer (1990) estimated semi-parametric hazard models along the lines of those developed by Cox (1972). Moffitt (1985a, 1985b) and Katz and Meyer (1990) suggest an additional week of potential duration leads to an increase in unemployment duration of 0.15 to 0.2 week. Ham and Rea’s (1987) estimate of 026-035 week is somewhat higher and comes from data on Canadian men. Grossman's (1989) estimate of 0.9 week, from Phase IV of the Federal Supplemental Compensation program (the emergency extension program of the early 19803), is the highest of all; however, her estimate comes from a sample of claimants who hadcxhausted their regular state UI benefits, so it is an estimate of the impact of an additional week of U1 benefits on the expected duration of unemployment conditional on exhausting regular UI benefits. Jurajda and Tannery(2003) use a flexible hazard model and find that potential duration decreases both the new job and recall hazard in individuals with more than 1 weeks of remaining benefit eligibility. Data We examine a nationally representative sample of 2,304 workers from 16 states who made initial claims for U1 between January 1992 and September 1993 and/or received EUC between July 1992 and May 1994. The data were collected by Mathematica Policy Research, Inc., under contract to the US Department of Labor. 75 They are unusual because they include two matched components — UI administrative records and survey data from interviews conducted between April 1996 and April 1997 (roughly three and a half years after the workers’ first benefit payments). The following brief description is based on Cederbaum (1997) and Corson, Needels, and Nicholson (1999), who fully describe the sample design and public use data. Information available in UI administrative data vary from state to state, but variables central to administering UI are always included — base period earnings. weekly benefit amount, the initial claim date, benefit year beginning date, benefit payment dates, and the balance of U1 benefits remaining for each worker at the end of his or her benefit year. Many analyses of potential benefit duration have relied solely on UI administrative, which are extremely rich: For example, they often include basic demographic data on claimants, in addition to information on the amount and timing of benefits received. UI administrative records are rich in that they offer relatively complete information about UI benefits paid to workers; however, they generally exclude information on reemployment and claimants’ subsequent earnings. Hence, although they indicate the duration of U1 claim spells, they do not provide data on spells of unemployment or joblessness. The follow-up survey conducted by Mathematica was designed to remedy this problem by collecting information on whether the claimant became reemployed, the timing of reemployment, and post-reemployment earnings. The follow-up survey also collected a consistent set of data on various worker characteristics that are not always available in UI administrative data. The combination of U1 administrative records with the follow-up survey results in a dataset 76 in which each component complements the other. Specifically, the follow-up survey includes information on reemployment, jobless duration, marital status, educational attainment, race, previous earnings, and recall variables, while the administrative data includes information about benefit duration, potential benefit duration. benefit exhaustion, weekly benefit amount, gender, age, the state in which the UI claim was filed, and whether the claimant’s previous job was in manufacturing. After dropping observations with missing values, we have a sample of 1,821 unique individuals. Table 18 defines the variables and displays summary statistics of this sample. The average duration of insured unemployment is 25.09 weeks, while the average potential duration is 41.58 weeks. 40% of workers exhausted their UI and/or EUC benefits. Jobless duration shows that the average time to a stable job (one lasting more than 3 weeks) is 153.23 weeks, though this is censored for 63.7% of our sample (that is, the interview date occurs before the individual finds stable employment). 34% of the sample expected recall, but only 9% had a definite recall date. 83% of workers are reemployed at some point before the survey date, but only 36% find stable jobs in that time. The Continuous Wage and Benefit History data used in Moffit (1985b) had an average potential duration of 33 weeks with an average duration of only 13 weeks. Of the papers with measures of actual reemployment, J urajda and Tannery have 86% of their sample becoming reemployed by their censoring date, while Katz and Meyer (1990) have 76%. Neither provide figures for stable reemployment. Figures 13-15 provide additional information about our duration variables. Figure 13 shows the distribution of benefit duration for our sample. There is a spike in 77 the density at the beginning (workers who only take a few weeks of benefits), and then a general decline with spikes at various exhaustion points: 25-26 weeks, 33-34 weeks. 45-46 weeks, and 51-52 weeks. These spikes are consistent with the exhaustion of benefits (26 weeks of U1 and either 0, 8, 20, or 26 weeks of EUC). Figure 14 shows the distribution of jobless duration for our sample, including workers who did not obtain stable reemployment before the interview date. Here we see a general decline in density until about 26 weeks, higher densities from 26 to 60 weeks, then a steady decline until 120 weeks. The large second bump (from 120 to 280) consists almost entirely of censored observations (individuals who did not obtain stable reemployment before the interview date). Figure 15 omits these censored observations, showing the distribution of jobless duration for the portion of our sample that found stable reemployment. Econometric Models Parametric and semi-parametric models We model the effect of potential duration of benefits on both benefit duration and on jobless duration. These duration models can be written as: durationi = 30 + fllpotduri + BZXZL' + + [31(th + at where duration 1' denotes either the number of weeks of benefits received by worker i, or the number of weeks worker i was jobless, potduri denotes the potential duration of 78 u. u. If; 1\“$ HA Ll. .n 315.13. 51°3qu- >¥l\-‘l.5 1950 " ‘1 bis: .!I.J1' ‘ sl sh KL- ‘w‘v‘m I bit-ll.“ \\ 'l'm .z a - hilliS \15 M' '1‘“ 1131.31.11 ‘1‘: . innit: 4.43m ‘~‘ ‘1»; mar? A)“. Works; UI and EUC benefits (combined) for which the worker was eligible at the time of the initial benefit claim, and X21: through xK, denote controls. In the case of benefit duration, each observation is a spell of benefit receipt. In the case of jobless duration, each observation is a spell ofjoblessness. In both cases, no individual contributes more than one spell to the data. Table 18 describes the control variables included in the specification. We include indicators for female, marital status (married; divorced, separated, widowed; with never married as the reference group), highest level of educational attainment (high school dropout; associate’s degree or vocational degree; college or graduate degree; some other form of education; with high school only as the reference group) , state in which the U1 claim was filed, race (African-American; Asian; Mexican; other non-White; with white as the reference group), whether the claimant’s previous job was in manufacturing, whether the claimant expected recall, and whether the claimant had a definite recall date. (Claimants who expected recall may or may not have had a definite recall date). We also include interactions of female with the two indicators for marital status listed above. To control for the labor market conditions facing the worker, we include the monthly state unemployment rate at the time the initial claim for benefits was made. Finally, we include the claimant’s average weekly earnings during the year before he or she claimed benefits, the weekly benefit amount for which the claimant was eligible, and age at first claim. We estimate four versions of equation (1) for both benefit and jobless duration. In the first we ignore censoring issues, simply estimating the model by OLS. In the second we recognize that benefit exhaustion (in the case of benefit duration) and the 79 gr. '9: mad For are same \1 Heels. 11.215111 the 11:12 Infidel i COHesp $0110.11. cox aria: Weibull paldmeu end of the survey (in the case of jobless duration) right—censor our dependent variable, and estimate using censored normal regression. For the third version we use the Weibull hazard model, which recognizes the censoring issue, but assumes that the duration distribution is distributed Weibull and that the baseline hazard is monotonically increasing, monotonically decreasing, or stable. The Weibull model can be parameterized either as an accelerated failure time model, providing duration estimates that can easily be compared with estimates from OLS and Censored regression, or as a proportional hazard model, which provides hazard estimates that can be compared with the results from other hazard models. In any proportional hazard model, covariates are assumed to have a proportional effect on the hazard rate. For example, the effect of an extra week of potential benefits is assumed to be the same whether it is an increase from 12 to 13 weeks or an increase from 51 to 52 weeks. While Figures 1—3 suggest that the duration distributions are not distributed Weibull, we include the Weibull model as a bridge between the duration models and the hazard models. Finally, we employ the Cox proportional hazard model (Cox 1972). This model is semi-parametric in that it does not assume a particular duration nor a corresponding baseline hazard, but it does assume that the conditional hazard is proportional to the baseline hazard (Cameron and Trivedi 2005), meaning that the covariates are assumed to have a proportional effect on the hazard rate just as in the Weibull hazard model. The Cox PH model imposes fewer assumptions than parametric models, and is generally preferred, a priori, to the other models (Box- Steffenmeier & Jones 2004). 80 Non-parametric hazard models To estimate the effect of EUC on the probability of reemployment, we specify a flexible discrete-time hazard model with time—varying explanatory variables. This involves transforming the data from spell-level data (in which the unit of observation is the spell of unemployment experienced by a person) to person-week data (in which each observation represents a week a person spends unemployed). Each spell of unemployment contributes T,- observations to the person-week dataset, where T, is the . . 24 number of weeks of unemployment experienced by person i. We construct two dependent variables: benefitEndiz, equal to I if person i stopped taking unemployment benefits in week I, 0 otherwise. reemployment”, equal to 1 if person i became reemployed in week t, 0 otherwise. While it is possible that an individual may stop taking benefits for a time before resuming within the same benefit year, benefitEndi, treats all spells as continuous (or compressed). As before, only stable jobs (those lasting more than three weeks) are counted as reemployment events (reemploymentn). The empirical model for either dependent variable can be written: 24 See Allison (1984) and Han and Hausman (1990). 81 Pr (eventit = 1|X) = Flflo + B1x1i+ 182x2i + + 18KxKi + 051211“ 'I’ aZZZI + "l' aKZKi 'l' ul‘ where event), is either benefitEndi, or reemployment“, x 1 ,- through x10 denote K characteristics of person i that do not vary over time, and Z 1 it through ZN), denote N time-varying influences on reemployment, including measures of the potential duration of U1 benefits (see below). Setting F = 1 implies a linear model in which the coefficients are interpreted as effects of a unit change in an independent variable on the probability of reemployment. A positive estimated coefficient suggests a higher probability of reemployment and hence a shorter spell of unemployment. Coefficient estimates from equation (2) can also be used to infer the effect of each independent variable on jobless duration following the method described in Davidson and Woodbury (1991). The time-invariant person characteristics we include in equation (2) include indicators for female, marital status (married; divorced, separated, widowed; with never married as the reference group), highest level of educational attainment (high school dropout; associate’s degree or vocational degree; college or graduate degree; some other form of education; with high school only as the reference group) , state in which the UI claim was filed, race (African-American; Asian; Mexican; other non- White, with white as the reference group), whether the claimant’s previous job was in manufacturing, whether the claimant expected recall, and whether the claimant had a 82 definite recall date. (Claimants who expected recall may or may not have had a definite recall date). We also include interactions of female with the two indicators for marital status listed above. We also include the claimant’s age at first claim, base period earnings (earnings during approximately the year before he or she claimed Ul benefits), and UI weekly benefit amount. The main time-varying influence we include in equation (2) is the potential duration of benefits. We model this in two alternative ways. First, we follow Grossman (1989) and Katz and Meyer (1990) and construct a variable indicating the expected potential duration of benefits for which a claimant believes herself eligible in the current week, based on the status of her regular state benefits and EUC in the current week — maxexp. This variable is constant during a spell of unemployment unless the potential duration of EUC changes. Specifically, for weeks when a claimant is drawing regular state benefits, we define maxexp as the potential duration of regular benefits for which she was eligible at the initial claim plus the potential duration of EUC for which she would be eligible if regular benefits were exhausted in the current week. (EUC entitlements were set at the time a claimant first claimed EUC, typically when she exhausted regular state benefits. Once set, the entitlement could increase but not decrease.) Because the EUC entitlement was not set until a worker exhausted her regular benefits, and EUC durations both rose and fell while workers were drawing regular benefits, maxexp can either rise or fall in weeks when a worker is drawing regular benefits. For weeks when a claimant is drawing EUC, we define maxexp as the potential duration of regular benefits for which she was eligible at the initial claim plus the 83 potential duration of EUC for which she is eligible in the current week. In this latter case, maxexp can change if (and only if) a claimant’s EUC entitlement increased because a state changed from low- to high-unemployment status or Congress increased the duration of EUC benefits. We also include a quadratic term to allow variation in the change in the benefit exit hazard (or reemployment hazard) associated with a one- week increase in the maximum potential duration of benefits. The second approach to specifying the potential duration of benefits is to include two time-varying independent variables in equation (2): the number of weeks until the claimant can expect to exhaust benefits [wksleft(t)] and the number of elapsed weeks since the beginning of the spell of unemployment [elapsed(t)]. Specifically, for weeks when a claimant is drawing regular state benefits, we define wksleft(t) as the number of weeks of regular benefits remaining plus the number of weeks of EUC benefits that would be provided if the worker exhausted regular benefits in the current week. For weeks when a claimant is drawing EUC, we define wksleft(t) as the number of weeks of EUC benefits currently remaining in week i. Note that wksleft(t) typically falls each week as the claimant moves closer to exhausting benefits; however, wksleft(t) may also change from one week to the next if the expected potential duration of benefits (potdur) changes. As with maxexp, we include a quadratic term to allow the effect on benefit exit (or reemployment) to vary over time. Ham and Rea (1987) also used a quadratic specification. We also specify elapsed(t) as a quadratic. To control for the labor market conditions facing the worker, we include the monthly state unemployment rate corresponding to the week being observed in all specifications. 84 Empirical Findings The top panel of Table 19 reports the results of applying three alternative estimators to the duration model we have written as equation (1). The left column shows estimated effects of a one-week increase in potential benefit duration—— potdur—on the expected duration of U1 benefit receipt. The right column shows estimated effects of a one-week increase in potdur on the duration of joblessness. (We report estimates of the full models from which all estimates in Tables 19 and 20 are derived in Appendix A, Tables A1, A2, A3, and A4.) The OLS estimates suggest that a one-week increase in potential duration increases the duration of benefit receipt by 0.41 weeks and increases the duration of joblessness by 1.64 weeks, other things equal. Both results are statistically significant at the 5% level. The former estimate is comparable to those of Holen (1977) and Moffitt (1985b), both of whom used OLS to estimate the effect on weeks of benefit receipt. Holen (1977) estimated the effect to be between 0.77 and 0.81 week, whereas Moffitt (1985b) reported estimates of 0.11 weeks for males and 0.19 for females. The OLS estimator does not account for censoring of the dependent variable; accordingly it is a biased and inconsistent estimator of the effect of potential duration. This is because OLS treats censored durations the same as durations that actually end in a true event (in this case, exit from benefit receipt or joblessness) in the same time period. As will become clear, the bias in this case will be upward—an overstatement of the effect of potential duration on both benefit receipt and on joblessness. For joblessness, the upward bias may be due to the changing size of EUC benefit 85 extensions. The largest increase in potential duration occurred in the early phases of EUC, with maximum extensions as large as 33 weeks early in the program, and as small as 7 weeks by the end of the program. Individuals filing claims near the end of EUC are closer to the interview date, which is the censoring point for our measure of joblessness. Such individuals will have both shorter potential duration and be censored earlier in their jobless experience (that is, they would appear to have shorter jobless spells due to the censoring), thus causing an upward bias in the OLS estimate. This would also explain why our estimate of the effect of a one-week increase in potential duration on jobless duration (an increase of 1.64 weeks) is substantially larger than that estimated by Solon (1979)——0.16 week for a sample of workers in New York. In Solon’s case, the increase in benefits was a single increase in benefit duration of 13 weeks, which occurred only for individuals filing later in the sample period. This would make individuals with longer potential duration appear to have shorter spells ofjoblessness, implying a downward bias to his estimate. For benefit receipt, censoring occurs at the time of benefit exhaustion. The OLS estimator would be upward-biased if workers were more likely to exhaust benefits during longer EUC extensions. This would be consistent with the idea that EUC benefit extensions increased in response to poor labor market conditions. Table 19, Panel A also shows the censored regression estimates of the effect of potential duration on benefit receipt and joblessness. Censored regression accounts for censoring of the dependent variable (weeks of benefit or weeks of joblessness) and assumes that the underlying distribution of durations is normal. A one-week increase in potential duration is associated with a 0.17 week increase in benefit receipt (column 86 1) and a 1.44 week increase in joblessness (column 2). Both results are statistically significant at the 5% level. That both estimates are smaller than the corresponding OLS estimates is consistent with the above reasoning about the bias of the OLS estimator. Using censored regression estimators, C lassen (1979) found that a one-week increase in potential duration leads to at most an additional 0.12 week of benefit receipt; Newton and Rosen (1979) obtained an estimate of 0.6 week; and Moffitt (1985b) obtained estimates of 0.085 week for males, and 0.116 for females. Our estimate for weeks of benefit receipt is similar to Classen’s and Moffitt’s. Again, the censored regression estimate we obtain is smaller than our OLS estimate (.41 weeks), a pattern also seen in Moffitt’s results. Newton and Rosen speculate that their relatively large estimate may have to do with Georgians having higher sensitivity to changes in potential duration, but ultimately conclude that it is “difficult to ascertain” the source of the differences between their results and other estimates. The censored regression estimates we obtain for both UI duration and weeks of joblessness are lower than their OLS counterparts. This is consistent with the upward bias in the OLS estimators as suggested above. The decline is more dramatic in the case of benefit exit, perhaps because the censoring occurs much sooner’than in the case of joblessness. The bottom row of Panel A reports estimates from a Weibull accelerated failure time estimator, which takes into account the censoring issue and assumes that the underlying distribution of durations can be characterized by the Weibull distribution. (This is equivalent to assuming that the baseline hazard is monotonically 87 increasing, monotonically decreasing, or constant.) The estimate of 0.001 (in Column 1) suggests that an additional week of potential benefit duration has no significant effect on the duration of benefit receipt. Similarly, the estimate on 0.014 (in Column 2) suggests that an additional week of potential benefit duration also has a negligible (and statistically insignificant) effect on weeks jobless. Both estimates are far smaller than those obtained using the OLS and censored regression estimators. Perhaps the most informative comparison is with Moffitt’s (I985b) estimates, which assume the distribution of insured spells is exponential (a special case of the Weibull). He finds no significant effect on benefit receipt for either males or females. It is worth noting that Moffitt’s results show the same pattern as ours—OLS estimates are larger than censored regression estimates (which assume a normal distribution), . . . . . . . . 25 and W eibull (exponentlal) estimates are the smallest and statistically 1n31gn1ficant. If we assume that the underlying distribution of spells is closer to a Weibull than to a normal distribution, then the Weibull estimates are preferable to the censored regression estimates given earlier. Nevertheless, the distributions shown in Figures 1, 2, and 3 do not appear consistent with either assumption, so the estimators are still likely to be inconsistent. Panel B of Table 19 again shows results of applying the Weibull estimator, but this time in hazard form. These estimates differ from the first three we report in that we are now looking at the effect of potential duration on the probability of benefit exit rather than on the duration of benefit receipt. (Similarly, we are looking at the effect 25 . . Katz and Ochs (1990) also use an estimator that assumes an exponential distribution and find that an additional week of potential duration increased jobless duration by 0.17—0.23 weeks. 88 of potential duration on the probability of becoming reemployed rather than the duration ofjoblessness.) A negative estimate implies a decrease in the probability of exit from benefit receipt (or joblessness) and an increase in the duration of benefit receipt (or joblessness). The Weibull hazard estimate of -0.001 (column 1) suggests that an additional week of potential benefit duration decreases in the likelihood of benefit exit, though the estimate is not statistically significant. The estimate of -0.009 (column 2) suggests that a one-week increase in potential duration is associated with a (marginally statistically significant) 0.009 decrease in the hazard of moving to reemployment. Panel B also reports estimates from the Cox proportional hazard model. In this case, an added week of potential benefits implies a 0.003 decrease in the hazard of benefit exit, and a 0.006 decrease in the hazard of reemployment, although neither estimate is statistically significant. Katz and Meyer (1990) estimate a comparable model and find a hazard coefficient of -0.025 for benefit exit—an order of magnitude larger than our estimate of -0.003. As mentioned above, the Cox model is less restrictive than the Weibull hazard model; however, it does impose the proportional hazards assumption. When we test this assumption using a Schoenfeld residuals test, we soundly reject the null hypothesis that the proportional hazards assumption holds in either the benefit exit and . 26 . . . . reemployment regreSSions. This implies that the estimated standard errors are inconsistent and that the coefficient on potential duration overstates the effect on the 26 The chi-square test statistic is 124.57 for the benefit exit regression, and 64.97 for the reemployment regression, each with 37 degrees of freedom. Both reject the null of proportional hazards at the 99% level. 89 hazard in both regressions (Box-Steffensmeier and Zorn 2001 ). Accordingly, it makes sense to move to a fully non-parametric with these data. Table 20 reports the results of estimating four alternative specifications of equation (2). In Panel A, potential benefit duration is specified as the maximum expected duration of all UI benefits (both regular state and EUC) the claimant could expect to receive in the observed week—maxexp—and its square. The left column reports the estimated coefficients from models in which benefit exit is the dependent variable. The estimated coefficient on maxexp (-0.002) suggests that an added week of potential benefits decreases the hazard of benefit exit by 0.2%, although the positive estimate on the quadratic term also suggests that this effect may diminish as maxexp increases. (The main effect is statistically significant, and the quadratic term is marginally so.) For an individual with 26 weeks of potential benefits, the marginal effect of an added week of potential benefits on benefit exit is -0.0018—a 0.18% decrease in the conditional probability of benefit exit. As shown in Appendix B (Table B1),27 this marginal effect corresponds to a 08-week increase in benefits received. This estimate is at the low end of estimates obtained by others by others who have used hazard models (Moffitt 1985a, 1985b; Ham and Rea 1987; and Katz and Meyer 1990), whose estimates range from .10 to .35 weeks. Although the sizes of the estimated effects vary, all suggest that increased potential duration may reduce the probability of ending a spell of U1 and lengthen the duration of benefit receipt. Because the estimates reported in Table 20 make the fewest restrictive assumptions 27 The conversion from hazard rates to expected duration follows the method outlined in Davidson and Woodbury (1991). See Tables B1 and B2 for the hazard fiinctions from which the durations are calculated. 9O about the distribution of spells, their findings should be preferred to the Weibull or Cox estimates. The right column of Panel A shows the estimated effect of maxexp on the conditional probability of reemployment. The coefficient of -0.00008 suggests that an added week of potential benefits decreases the hazard of reemployment by only 008%. and the estimate is statistically insignificant. The quadratic term again indicates that this effect may diminish at higher levels of potential duration. (Taken ’7 together, maxexp and maxexp are jointly significant at the 95% level.) As shown in Appendix Table 82, these estimates suggest that the effect of an added week of U1 benefits for an individual with 26 weeks of benefits, is a negligible .01-week increase in weeks jobless. The existing literature offers few opportunities for comparison, as past studies using hazard models have modeled the probability of exiting UI, not the probability of exiting a jobless spell. Juarajda and Tannery (2003) do have data on jobless spells and find that the new job hazard is depressed by increased potential duration, but they do not report a marginal change that would allow a clear comparison with our estimates. Panel B of Table 20 displays estimates from models in which potential benefit duration is specified as the number of weeks of all Ul benefits remaining for the claimant in the observed week—wksleft—and its square. This specification also includes a control for the number of weeks in the current spell (that is, weeks since the start of benefits). Using this specification, the estimated effects of potential benefit duration on the conditional probabilities of U1 exit and on reemployment are larger (and also statistically significant) than in those from the maxexp specifications. In 91 particular, the estimated coefficients on the main effects (wksleft) are -0.008 (for U1 exit) and -0.0002 (for reemployment). The marginal effect of an added week of benefits on the UI exit hazard at 26 weeks of benefits is -0.0047, a larger decrease than in the maxexp specification. and statistically significant. As shown in Appendix Table Bl, these hazard estimates correspond to a .28-week increase in benefits received in response to one additional week of benefit entitlement. This is close to the upper end of the range of effects estimated by other authors using hazard models (Moffitt 1985a,b; Ham and Rea 1987; Katz and Meyer 1990), which range from .10 to .35 weeks. The right column of Panel A shows the estimated effect of wksleft on the conditional probability of reemployment. The estimated marginal effect of wksleft on the conditional probability of reemployment is again larger than the estimated effect of maxexp, and corresponds to a .06-week increase joblessness for workers at 26 weeks of benefits. Although somewhat larger than the effect of potential duration on jobless duration estimated using the maxexp specification, it is still quite small. Estimates from the wksleft specifications are substantially larger than those from the maxexp specifications. Is there a reasonable way of choosing between the two? The wksleft specifications have a better fit than their maxexp counterparts, as measured by the adjusted-R2. Although it would be useful to explore differences between to the two specifications further, the better fit of the wksleft specifications may imply that workers respond more to their weeks of remaining benefits than to the maximum potential duration of benefits. If so, using maximum potential duration of benefits as the key right-hand-side variable in parametric models (like those reported in Table 19) may be yet another specification error to which those models are vulnerable. The nonparametric estimates in Table 20 suggest that an addition week of potential benefit duration increases a typical UI spell by about 08-28 weeks (and these estimates are statistically significant at conventional levels). The estimated effect of an addition week of potential benefit duration on a typical jobless spell is substantially less—only .01-.06 weeks (and only the higher end of this range is statistically significant at the 5-percent level). Why are the estimated effects of increased UI benefit duration on UI spells larger than the estimated effects on the length of jobless spells? Consider a one-week reduction in the potential benefit duration of a worker who exhausts benefits. By definition, such a worker would receive one less week of U1 benefits, but his or her jobless duration may not change at all. It makes sense, then, that estimates of the effect of potential benefit duration on the length of U1 spells should be larger than estimated effects on the length of jobless spells. Summary and Conclusions When we apply a range of econometric estimators to a high-quality dataset that includes information on both the duration of U1 spells and the duration of jobless spells, we obtain a wide range in the estimated effect of extended benefits on the duration of U1 benefit receipt and the duration of joblessness. The estimated effect of an additional week of potential duration on benefit receipt ranges from .00-.41 weeks (Tables 19, 20, and Appendix Table B1). The most convincing estimates—that is, the 93 estimates that come from models that impose the fewest restrictions—come from non- parametric hazard models (Table 19 and Appendix Table BI) and suggest a narrower range of.08-.28 weeks. The estimated effects of an additional week of potential duration on weeks jobless cover a wider range—from .01-1.64 weeks (Tables 19, 20, and Appendix Table B2). Estimates at the high end of this range come from the parametric and semi-parametric estimators. For example, OLS and censored regression suggest that an additional week of potential benefit duration increases jobless duration by about 1.5 weeks. In contrast, nonparametric estimates, which impose fewer restrictions, suggest far more modest effects in the range of .01-.06 weeks. The estimates we find most convincing are the nonparametric estimates reported in Table 20 and Appendix Tables B1 and 82. These estimates suggest that a 13-week increase in potential benefits (as in phase 1 of EUC) would increase a typical UI spell by about 1.0-3.6 weeks, and would increase the typical jobless spell by .1-.8 week. Because we have used a single dataset, we can safely attribute the variation in our estimates to differences in model specification and the estimators we apply to the data. The findings suggest that such differences substantially affect the inferences we derive from the data. To the extent possible, we have traced the connections between these differences in specification and estimators (on one hand) and the disparate results found in existing research (on the other). For the most part, the inferences we derive are consistent with the inferences past researchers have obtained using various datasets and a range of estimators. Perhaps the most consistent finding we have is that 94 estimates of the effect of potential benefit duration on weeks of benefit receipt often bear little relation to the estimates of the effect of potential duration on weeks of joblessness. 95 Figure 13 Distribution of Benefit Duration Density 20 30 Weeks of benefits received Notes: 1. Spikes occur at exhaustion points: 26, 34, 46, and 52 weeks. (26 weeks of Ul and either 0, 8, 20, or 26 weeks of EUC). 2. Sample size = 1,821 96 Figure 14 Distribution of Jobless Duration (including censored observations) oo o- 0. Density 004 .006 .002 1 00 200 300 Weeks jobless (censored) Notes: 1. The large increase in density after 120 weeks is almost entirely due to censored observations. (Workers who fail to find stable reemployment before the interview date.) See Figure 3 for the distribution without censored observations. 2. Sample size = 1,821 97 Figure DistribI Densily )1 (.115 (1') ("I l ”I .6 Notes intervit Figure 15 Distribution of Jobless Duration (excluding censored observations) 0 5‘0 100 150 200 250 Weeks jobless Notes: 1. Only includes workers who found stable reemployment (> 3 weeks) before the interview date. 2. Sample size = 656 98 zones. 3 33. .mEBmocu Eoeom EmEonEocb tobzotameo 3.355% a hp 03.2. .NmE 803.5 #5.: 99 80863895 .92 m Eoom 2 ooEEoo $me .35 9:on 9:88. 99: mm .9 a___n_m__m SE £26869 908 oomE 996.5 99m .oomumcomfi come .0 :20 on: on; 9.58.90 2 can .803 999:6 Eooom Lo xoom 2 @8629 98985.0 2 3:8 mm .59 989w 338% .585 m3. .826: m 3 meadow =m 5 99:8 238 mm outcome. 5 99m 80¢ 22.95 583009.». 8033:8800 «5863895 ooocoaxw 2 Eng mos—.99 xm... Eo8>o_a8o83 exec; 932 4.8 8-3 ....n. .818 4.8 +8-8 4n. :_ wEmE—ucmrcm BEE 5? 65-5 ....n. .89 29 .0 ~o< 8053:8800 ”Samoa 9 ohm .LommE ED: 99 Eo8>o_o895 .935“. Q. 9 a: .80:an “comma Eo8>o_o8ocD 859: cm 3 99m m E o9mszom mm 80: 22.80 99w 89.69 co 0&8 2 ohm poocowxw 88w 5 $8. do? 5 as: 93 48 Sam: xm._. Eo8>oa895 8< cozmmcoQEoO .maooou 3095 2.5863895 E mom—webs me 2 a: 80:93.6 poocoaxm 39888. 99w 5:69 ho o\oom N06 - 5? 39883 311mm ....n. 856 9695 5.4. cocoa—.8800 @5699th t 2 me 8 a: 80:98 Eo8>o_a8083 29:29 80:86:88 99m 950. 09276998. 99m 6:69 ..o £00m mmR - 8% 39888. nouoz 9:0: «Eu 629305 Eco: on «80.9.8.8 80:299.. oooc3xo “—0 one count @5393 80:93. .oacouon. 9509mm one 8959a «some... 3 wow? .mEocmocm «co-5m 8953389.: oooceuxm ho Em883m m—. «30... 99 I lit 8982803 m.» 928.5 87mm 48. R3. 8.55036. 8886 8:85.986 ocm 3......90 .908. 82.5 .03 8:56 .8..an 289.. .90. KB 0:0 003:90. 90:03 ”003:99 :90: mm... :0 0x003 0m 2 a: .m. .9... Re .38 00050.3 9.00; we 0: a: .:0:930 0:90 .9300. n.0 38 92.6.8 .8853 mem .._.n_ .vam ....n. .9106 J.n. .Nkmém J.n_ .990 E... 0:0 ohm c858 mm 05.9.9.8 .9900 EB. .80 East: 909., wk: .80“: 9.050 0005 .8 9.003 mm 095 :0:930 «coco: _0::0:00 (ODm 0: 000:900 9 0: 03 60:930 0: 000:0:x0 .08089003w .22 8 .89... ..m c. 8.9.80 Emacs. mm. 03980: 906 5.38. .o 98 0&9 - an: .908“. www-mm 203m. .583. ....n. {~98 ....n. 5:00:00800 0x095 mtm .:o< 6:00:00800 8930.: 99:00 0. 0003 2.086.088: 2 0. 03 80:93: 0: 000:0:x0 808.6388: .989: mm EB. 229.3 989.. 92.990 8823 99m .938. .0 $8 .95 - NS 58905. 00:02 959.05... 000303. 3:0:00 0:59.050 p.000—050— 000C0ux0 M.0 0:0 0000 5530.0 :0:930 .9539. 0>:00tm 0:0 89005 .8558. B 0.an 100 {r «00:22:08 m“ .3ka .00... 02000. 0. 00.03050 8 o. 00.. 60.9.60... 00 9.0 0.00:2 0.0.0 .0300. 0.00_.0>< Own. .0 0000:... :0 o. 8.0.0 92.8 0.3.9.0 mm. .0...ch 2-00. .0: .30 .0.c0:00 ....Omn. .00:00 0.300 000.030cx0 ..-00. .80 .00... 009.20 030.0 00... .0 mm. 0.0.0.0 :0:>> b9 0.30.5 :0..0.30 .0..:0.00 0020.50.00 00:0 b0) .0: 0.0 E05955 .3 ....-00. 00 0600 ”60.0.85: 2-00. 0.00; 3 .0 m 0. 03 .:0..930 0.0.0 .0500. .0 $00 ”800003. ...00. 0.00.5 or .0 w 0. 03 .:0..0.30 0.0.0 .0300. .0 $00 ”.8000... .700“. 00.002. 0.. .0 0 0. a3 .:0..930 mm}. 0:0 090 00700 ....n. €0-00 ....n. €00.00 ....n. .0030 4.0 .60.. :0..00:00E00 .>_-Owu. :. .0096 0.0000. 00.. 0.0.0000 0:0 0.0.0 .0300. .0 $00 0. 000:0.x0 .0.:0E0_003m 03.0.0 mm. 0.0.0.0 5.3 00:0) :0..0.30 .0..:0.0n. 03:90. 90:00 A00. 73.0. 700... .008 I 00.0 .9000“. 00.02 050:0:..... 00:0:00 0:50:0.xo :000350. 000:0.x0 .0 0:0 00.00 05... 0:0 :00930 30:80.. 03.005 0:0 50.00.... 6035500. 0.. 030... 10] «00......200. m ._ 0305 .030 08:08 9.000. 2 3.3.0.0 .3 .0300. :.000. 0. 00000: .:0:..0.0 0 .00.030...x0 03m. 00:0 .030 .o 0000... ..0 o. 00.300 0:0...0 3.3.0.0 mm 3.0. 000. .00.. 00:00 30: 0.0 00.0 0.... .0 0.00:00 03m. 05.590 .0 8:8 00.. 0.56.0.0 .2030 0:0 ...-030 .00:3 50.00 0.00:9. 02.00030. 0: .33. 0:20.505 2.0005000... .00:3 0.00:00 02000. 0.300 .0:. 0:0 .00 503.00 0.00.00 05.0320 0.:0E.0.0 0000.000 .0: 0.300 .:0E0....:0 003:90. .90:00 E2. 2-030 9.0 ...-030 .568... 0x003 m. .0 n .0 0.00:9. .0300. .0 0000 .o .0000. ”$00 00.0: >-ODm. 00.002, 0. .0 or .0 0.00:9. .0300. .0 0000 .0 .0000. H80.07003. 2 .03m 0.003 mm .0 om .0 .0:—0:00 .0300. .0 .000. .o .0000. ”80.0 -005 ...-03m 00.00; mm .0 mm .0 .0:—0:00 .0300. .0 0000.. www-mop ....n. 6.00.. .....n. .0500? ....n. .3muwow .....n. .Nww 03m. 030.050 0.0:. 0E0000 E9090 :0..00:09:00 .0 .0000. AN0R-N0\N. 3R .3. i..n. .00 Two. .. .o 020.0 :0... o. 26. ES. 006... 0.0.0 .. 2.08.20893 ..-030 0.00; 00 .o .80 .005. ....n. .030. .00. .6 00090:. 0.300 .:0:.0_...:0 03m. ..:0E>0.0:.0:3 000:0.xm 0P .0 .0...0:00 .0300. .00.... .N0R .0< :0..00:00E00 .50. .0 -..0... 00 :0..00...000.0 0.0.0.0 :0 000:0000 E9. >-03w .0 0009 .0 .0000. 0. 000:0.x0 .:0:.>0.0:.0:D .0:.E .0 05.. .0 00:.:..0.00 :0..930 .0..:0.0& 0:0 ....03m ...03m. x00N-.0;... ...ODm. .NOB I .0:. >0:00.0Ew 00.02 050: 0:... 00030... 0.00:2. 0:0.0:0.x0 :000350. 000:0.x0 .0 0:0 00.00 0::00c0 0000.30 .09—0.0a 0>000tw 0:0 50.00:. 6035.08. 0. 0.000 102 I, 2.0.5.38. 0. 030.. .r.. . . .. P 1.0.0.07 00.0.0 .:0E>0.0E0:: :0... :. 0.003 0. o. 0: 000.00 .078. ....n. .0300 080.0 .::000< .0300. .0 0000 .0< :0..00:00:.00 :0..00:00E00 _0:0...000 ”0.003 no.0 .:0E>0_0:.0:D 0.0. .:0:.>0_0E0:: 00.30:. 0.0.0 :0 00000 .:0E>0_0E0:D m. 0. 03 .:0..0.:0 o. 000:0.x0 00050.0 00.0.0 :. 0.00:0..0 0.00:00 000:0.x0 .0:0...00< 000:0.xm. ...0:00 .0300. .0 000... .NO.N.-N0.0 30.00:.00 00.02 05050:... 00030.0 0.00: 00 0:20:008 :000030. 000:0.x0 .0 0:0 00.00 0:..00:0 :0..0.:0 .00:300 0>000tm 0:0 E0590 602.....8. 0. 0.00. 103 Table 16 Potential Durations of Extended Ul Benefits in Pennsylvania Under Emergency Unemployment Compensation Potential EUC Unemployment duration phase Dates level (weeks) Notes I 11/17/91 — L 13 EUC I began 11/17/91 1/25/92 PL. 102-164 a. 102-182 I 1/26/92 - H 20 2/08/92 II 2/09/92 - H 33 EUC ll approved 2/07/92 6/13/92 PL 102-244, 33 or 26 weeks for claims on or before 6/13/92 ll 6/14/92 — H 26 Per PL. 102-244, 13 (L) and 20 7/04/92 (H). This was changed retroactively. See note 2. III 7/05/92 - H 26 EUC III approved 7/03/92 8/15/92 PL. 102-318 Ill 8/16/92 — L 20 3/20/93 IV 3/21/93 — H 26 EUC IV approved 3/04/93 6/ 19’ 93 PL. 103—6. Limits are the same as EUC III. IV 6/20/93 — L 20 9/11/93 IV 9/12/93 — L 10 10/02/93 V 10/03/93 — L 7 EUC V approved 11/24/93, See 2/05/94 note 3. Notes: 1. Per PL. 102-318, the reduction in the maximum EUC entitlement when the NTUR fell below 7% affected only new EUC claimants, not those already collecting benefits. 2. As noted above, according to PL. 102-244, maximums of 20 or 13 weeks would apply to new claims made after 6/13/92. However, in PL. 102-318, this was changed retroactively to 20 or 26 weeks for new claims made after 6/13/92. 3. EUC legislation lapsed from 10/03/93 to 11/27/93. When EUC V was approved on 11/24/93, the new EUC entitlements were made retroactive to 10/03/96. (Per conversations with Mike Miller at DOL/ETA/UIS.) 4. EUC trigger reports for 7/11/93 - 7/31/93 indicate a potential duration under EUC of 10 weeks. According to Mike Miller, the maximum of 20 weeks was maintained throughout the period from 6/20/93 to 9/11/93. Due to confusion over interpretation of NTUR, the 20 week maximum was made retroactive to cover the weeks in July that were reported as 10 weeks. 104 Table 17 Estimated Effects of Increased Potential Duration of Ul Benefits for the United States and Canada Study Change in weeks of unemployment from 1 added week Remarks Holen (1977) Solon (1979) Classen (1979) Newton and Rosen (1979) Katz and Ochs (1980) Moffitt and Nicholson (1982) Mof‘fitt (1985a) Moffitt (1985b) Data of potential Ul UI claimants in San Francisco, 0.77-0.81 Boston, Phoenix, Seattle, Minneapolis, 1969-70 Ul exhaustees in New York, 0.16 1972-1973 (pre- and post-EB, surveyed after exhaustion) Ul claimants in Arizona and 0-0.12 Pennsylvania, 1967-69 Ul recipients in Georgia, 1974- 0.6 76 Current Population survey, 0.17-0.23 individuals in 26 states, 1968-70 and 1973-77 Recipients of EB and FSC, 15 0.1 states, 1975-77 Continuous Wage and Benefit 0.15 History, 1978-83 Continuous Wage and Benefit History, 1978-83: White men 0.17 White women 0.10 F88 and EB recipients in 15 states, 1975-78: Men 0.45 Women 0.28 Ul recipients in Georgia, 1974—76: Men 0.17 Women 0.37 105 OLS linear duration estimates OLS linear duration estimates Tobit duration estimtates Tobit duration estimates Maximum likelihood duration estimates Labor supply model, maximum likelihood UI exit rate estimates Ul exit rate estimates Maximum likelihood duration estimates Maximum likelihood duration estimates Table 17 (continued) Study Solon (1985) Ham and Rea (1987) Grossman (1989) Katz and Meyer (1990) Jurajda and Tannery (2003) Data UI claimants in Georgia, 1978- 79 Canadian men, 1975-80 Continuous Wage and Benefit History, individuals in 3 states, 1981-1984 Continuous Wage and Benefit History, individuals in 3 states, 1981-84 Continuous wage and Benefit History, Pittsburg and Philadelphia, 1981-84 Change in weeks of unemployment from 1 added week of potential Ul 0.36 026—035 0.9 0.16—0.20 not reported Remarks Maximum likelihood duration estimates Ul exit rate estimates Ul exit rate estimates of FSC impacts UI exit rate estimates New job rate estimates - hazard of new job is decreased by increased potential duration 106 Table 18 Brief Variable Definitions and Sample Summary Statistics Variable Description Mean Std. dev. benefit duration Weeks of benefits received 25.09 16.08 jobless duration Weeks from initial claim until taking a job 153.23 84.65 lasting at least 3 weeks (censored) reemployed 1 if became reemployed by the survey 0.83 0.37 date, 0 otherwise stable reemployed 1 if became reemployed in a job lasting at 0.36 0.48 least 3 weeks by the survey date, 0 potential benefit Sum of weeks of UI and EUC the 41.58 8.79 duration individual is eligible to receive female 1 if female, 0 otherwise 0.41 0.49 age Age in years 39.06 11.89 married 1 if currently married 0.63 0.48 not married 1 if divorced, separated, or widowed; 0 0.16 0.37 otherwise female, not married 1 if female and married, 0 otherwise 0.09 0.28 female, married 1 if female and not married, 0 otherwise 0.25 0.43 dropout 1 if less than high school, 0 otherwise 0.16 0.36 post secondary 1 if associate's or vocational degree, 0 0.17 0.37 otherwise college 1 if college or graduate degree, 0 0.14 0.34 otherwise other education 1 if some other form of education, 0 0.02 0.13 otherwise african-american 1 if African-American, 0 othewvise 0.11 0.32 asian 1 if Asian, 0 otherwise 0.01 0.12 mexican 1 if Mexican, 0 otherwise 0.07 0.26 other non-white 1 if Non-White, 0 otherwise 0.05 0.21 107 Table 18 (continued) Variable Description Mean Std. dev. manufacturing 1 if previous job in manufacturing, 0 0.33 0.47 otherwise state unemployment Unemployment rate in the state at the 7.37 1.86 rate time of claiming benefits (°/o) previous weekly Average weekly earnings ($10) in the year 45.95 30.33 earnings preceeding benefit collection expect recall 1 if worker expects recall, 0 otherwise 0.34 0.47 definite recall date 1 if worker has definite recall date, 0 0.09 0.29 otherwise weekly benefit amount Ul/EUC weekly benefit amount 183.79 68.37 exhausted benefits 1 if exited Ul/EUC by exhausting 0.40 0.49 benefits, 0 otherwise sample size = 1,821 Source: Author’s tabulations of variables constructed from the Emergency Unemployment Compensation Public Use File compiled by Mathematica Policy Research, Inc. All calculations are computed using spell data. Notes: 1. The variable jobless duration denotes the time from the initial claim to the first stable reemployment (a job lasting at least 3 weeks). For workers interviewed before finding stable reemployment, jobless duration is time from the initial claim to the interview date. 2. Demographic variables (those involving gender, marital status, race, or educational attainment) were collected at the time of the survey, not at the time of the initial claim. 108 Table 19 Estimated Effects of Increased Potential Benefit Duration from Parametric and Semi-Parametric Models Estimation procedure Change in weeks of benefit duration from 1 Change in weeks jobless added week of potential from 1 added week of Panel A: Duration estimates benefits potential benefits OLS 0.413 1.639 (0.046) (0.252) adjusted R-squared 0.1777 0.0990 Censored regression 0.172 1.441 (0.074) (0.292) log likelihood -5385.7 -5009.5 Weibull accelerated 0.001 0.014 failure time (0.004) (0.008) log likelihood -2442.2 -2161.9 Change in benefit exit Change in reemployment hazard from 1 added hazard from 1 added Panel 3; Hazard estimates week of potential benefits week of potential benefits Weibull hazard -0.001 -0.009 (0.004) (0.005) log likelihood -2442.2 -2161.9 Cox proportional -0.003 -0.006 hazard (0.004) (0.005) log likelihood -7404.1 -4748.1 sample size 1,821 1,821 Notes: 1. Estimates of equation (1 ). Panel A contains estimates in which the dependent variable is either weeks of benefits received during the benefit year (column 1) or weeks jobless (column 2). See Appendix A, Tables A1 and A2 for estimates of the complete model. OLS and Censored regression estimates are the change in weeks of benefit duration or jobless duration with respect to a 1 week increase in potential benefits. Weibull accelerated failure time estimates give the proportional change in benefit duration orjobless duration with respect to a 1 week increase in potential benefits. Panel B gives estimates in which the dependent variable is either benefit exit or reemployment. See Appendix A, Table A3 for estimates of the complete 109 Table 19 (continued) model. Estimated coefficients give the proportional change in the hazard of benefit exit or jobless duration with respect to a 1 week increase in potential benefits. Standard errors in parentheses. 2. Included control variables: previous weekly earnings, weekly benefit amount, age, monthly state unemployment rate, and indicators for gender, marital status, interactions of gender and marital status indicators, educational attainment, state, race, manufacturing, expected recall, and recall with a definite date. F‘- Efil—u Jd'; . A .r‘ 53.1.1 I 1‘ 110 Table 20 Estimated Effects of Increased Potential Benefit Duration from Nonparametric Models Change in benefit exit Change in reemployment hazard from 1 added week hazard from 1 added week Panel A of potential benefits of potential benefits maxexp -0.002076 -0.000079 (0.000592) (0.000047) 2 maxexp 0.000012 0.000002 (0.000008) (0.000001) adjusted R-squared 0.0050 0.0011 sample size 33,048 254,191 Panel B wksleft -0.007675 -0.000203 (0.000446) (0.000030) 2 wksleft 0.0001 18 0.000003 (0.000008) (0.000001) adjusted R-squared 0.0443 0.0025 sample size 33,048 254,191 Notes: 1. Estimates of equation (2). See text for further discussion. See Appendix A, Table A4 for estimates of the complete model. Estimated coefficients give the change in the probability of benefit exit or reemployment with respect to a 1 week increase in potential benefits. Standard errors in parentheses. 2. Included control variables: previous weekly earnings, weekly benefit amount, age, monthly state unemployment rate, and indicators for gender, marital status, interactions of gender and marital status indicators, educational attainment, state, race, manufacturing, expected recall, and recall with a definite date. 111 BIBLIOGRAPHY Advisory Council on Unemployment Compensation. Report and Recommendations. Washington, DC: Advisory Council on Unemployment Compensation, February 1994. Advisory Council on Unemployment Compensation. Unemployment Insurance in the United States: Benefits, Financing, Coverage. Washington, DC: Advisory Council on Unemployment Compensation, 1995. Box-Steffensmeier, Janet M. and Bradford S. Jones. Event History Modeling: A Guide for Social Scientists. NY: Cambridge University Press, 2004. Box-Steffensmeier, Janet M. and Christopher J. W. Zorn. “Duration Models and Proportional Hazards in Political Science.” American Journal of Political Science 45(2001): 9722-988. Cameron, A. Colin and Pravin K. Trivedi. Microeconometrics: Methods and Applications. New York, NY: Cambridge University Press, May 2005. Cederbaum, R. “Emergency Unemployment Compensation: Public Use File Documentation.” Princeton, NJ: Mathematica Policy Research, Inc., December 1997. Classen, Kathleen P. “Unemployment Insurance and Job Search.” In Studies in the Economics of Search, edited by S.A. Lippman and J .J . McCall. Amsterdam: North— Holland, 1979. Pp. 191—219. Corson, Walter, Jean Grossman, and Walter Nicholson. “An Evaluation of the Federal Supplemental Compensation Program.” Unemployment Insurance Occasional Paper 86—3. Washington, DC: US. Department of Labor, Employment and Training Administration, Unemployment Insurance Service, 1986. Corson, Walter, David Long, and Walter Nicholson. “An Analysis of the 1981-82 Changes in the Extended Benefit Program.” Unemployment Insurance Occasional Paper 85—1. Washington, DC: US. Department of Labor, Employment and Training Administration, Unemployment Insurance Service, 1985. Corson, W., K. Needels, and W. Nicholson. “Emergency Unemployment Compensation: The 1990’s Experience (Revised Edition)” Unemployment Insurance Occasional Paper 99-4. Washington, DC: US Department of Labor, January 1999. Corson, Walter and Walter Nicholson. The Federal Supplemental Benefits Program. Kalamazoo, Michigan: W.E. Upjohn Institute for Employment Research, 1982. 112 Corson, Walter, and Anu Rangarajan. 1994. “Extended UI Benefit Triggers.” Unemployment Insurance Occasional Paper 94-2. Washington, DC: US. Department of Labor, Employment and Training Administration, Unemployment Insurance Service. Cox, D. R. “Regression Models and Life Tables.” Journal of the Royal Statistical Society (Series B) 34 (Number 2, 1972): 187—220. Davidson, Carl and Stephen A. Woodbury. “Effects of a Reemployment Bonus under Differing Benefit Entitlements, or, Why the Illinois Experiment Worked.” Working paper, Michigan State University and WE. Upjohn Institute, March 1991. Felder, Henry E. and Richard W. West. “The Federal Supplemental Benefits Program: National Experience and the Impact of PL. 95—19.” Washington, DC: US. Department of Labor, Employment and Training Administration, Unemployment Insurance Service, 1978. Grossman, Jean Baldwin. “The Work Disincentive Effect of Extended Unemployment Compensation: Recent Evidence.” Review of Economics and Statistics 71 (February 1989): 159—164. Ham, John C. and Samuel A. Rea, Jr. “Unemployment Insurance and Male Unemployment Duration in Canada.” Journal of Labor Economics 5 (July 1987): 325—353. Han, Aaron and Jerry A. Hausman. “Flexible Parametric Estimation of Duration and Competing Risk Models.” Journal of Applied Econometrics 5 (January-March 1990): 1-28. Holen, Arlene. “Effects of Unemployment Insurance Entitlement on Duration and Job Search Outcome.” Industrial and Labor Relations Review 30 (July 1977): 445—450. J urajda, Stepan and Frederick J. Tannery. “Unemployment Durations and Extended Unemployment Benefits in Local Labor Markets.” Industrial and Labor Relations Review 56 (January 2003): 324-348. Katz, Arnold and Jack Ochs. “Implications of Potential Duration Policies.” In Unemployment Compensation: Studies and Research, Volume 1. Washington, DC: National Commission on Unemployment Compenastion, July 1980. Pp. 111—126. Katz, Lawrence F. and Bruce D. Meyer. “The Impact of the Potential Duration of Unemployment Benefits on the Durationof Unemployment.” Journal of Public Economics 41 (February 1990): 45—72. 113 Katz, Lawrence F. and Bruce D. Meyer. “Unemployment Insurance, Recall Expectations, and Unemployment Outcomes.” Quarterly Journal of Economics (November 1990): 973—1002. Kiefer, Nicholas M. “Economic Duration Data and Hazard Functions.” Journal of Economic Literature 26 (June 1988): 646—679. Lalive, Rafael. “How do Extended Benefits Affect Unemployment Duration? A Rergression Discontinuity Approach.” CESifo Working Paper No. 1765, July 2006. Lancaster, Tony. “Econometric Methods for the Duration of Unemployment.” Econometrica 47 (July 1979): 939—956. Mathematica Policy Research. “A Longitudinal Study of Unemployment Insurance Exhaustees: Waves I and 11.” Report prepared for the Employment and Training Administration, US. Department of Labor, January 1976. Moffitt, Robert. “Unemployment Insurance and the Distribution of Unemployment Spells.” Journal of Econometrics 28 (April l985a): 85-101. Moffitt, Robert. “The Effect of the Duration of Unemployment Benefits on Work Incentives: An Analysis of Four Data Sets.” Unemployment Insurance Service Occasional Paper 85—4. Washington, DC: US. Department of Labor, Employment and Training Administration, Unemployment Insurance Service, 1985b. Moffitt, Robert and Walter Nicholson. “The Effect of Unemployment Insurance on Unemployment: The Case of Federal Supplemental Benefits.” Review of Economics and Statistics 64 (February 1982): 1—11. Mortensen, Dale T. “Job Search and Labor Market Analysis.” In Handbook of Labor Economics, Volume 11, edited by Orley Ashenfelter and Richard Layard. Amsterdam: North—Holland, 1986. Pp. 849—919. Newton, Floyd C. and Harvey S. Rosen. “Unemployment Insurance, Income Taxation, and Duration of Unemployment: Evidence from Georgia.” Southern Economic Journal 45 (January 1979): 773—784. Solon, Gary. “Work Incentive Effects of Taxing Unemployment Benefits.” Econometrica 53 (March 1985): 295—306. Storey, James R. and Gene Falk. “Unemployment Compensation: The Emergency Unemployment Compensation (EUC) Program.” Washington, DC: The Library of Congress, Congressional Research Service, 93—367 EPW, March 1993. 114 Vroman, Wayne and Stephen A. Woodbury. “Trend and Cycle Analysis of Unemployment Insurance and the Employment Service.” Employment and Training Administration Occasional Paper 2005-04. Washington, DC, US. Department of Labor, 2005. Wenger, Jeffrey B. and Matthew J. Walters. “Why Triggers Fail (and What To Do About It): An Analysis of the Unemployment Insurance Extended Benefit Program.” Journal of Policy Analysis and Management 25 (Summer 2006): 553-576. Woodbury, Stephen A. and Murray Rubin. “The Duration of Benefits.” In Unemployment Insurance in the United States: Analysis of Policy Issues, edited by Christopher J. O’Leary and Stephen A. Wandner. Kalamazoo, MI: W.E. Upjohn Institute, 1997. Pp. 211-283. 115 APPENDICES 116 Table A1 Benefit Duration Estimates APPENDIX A Censored Weibull OLS regression (AFT) potential benefit 0.4129 0.1719 0.0007 duration (0.0460) (0.0742) (0.0044) female -1.3360 -2. 1761 -0.0203 (1.5941) (2.5015) (0.1454) age 0.2702 0.3748 0.0169 (0.1754) (0.2837) (0.0175) age squared -0.0001 -0.0001 -0.0000 (0.0002) (0.0003) (0.0000) married -3.7177 -5.7919 -0.2702 (1.2227) (1.9103) (0.1086) not married -1.6670 -1.3791 -0.0320 (1.6459) (2.6293) (0.1584) female, not married 1.8769 3.2371 0.0762 (2.3419) (3.7427) (0.2252) female, married 3.7301 6.9092 0.3007 (1.8360) (2.8837) (0.1681) dropout -0.3215 0.1131 0.0336 (1.0589) (1.6544) (0.0953) post secondary 2.8040 4.4008 0.3114 (0.9765) (1.5495) (0.0936) college 1.5697 3.1390 0.2163 (1.1089) (1.7731) (0.1082) other education 5.4173 5.8423 0.0916 (2.7455) (4.4528) (0.2578) african-american 3.5408 6.2799 0.3940 (1.1923) (1.9455) (0.1238) asian -4.4658 -6.2346 -0.3339 (3.0331) (4.6214) (0.2652) mexican 2.2149 3.4075 0.2354 (1.5630) (2.4798) (0.1430) other non-white 1.3566 1.8871 0.1381 (1.7050) (2.7550) (0.1694) 117 Table A1 (continued) Censored Weibull OLS Regression (AFT) manufacturing -1.0527 -0.9298 -0. 1482 (0.7633) (1.1947) (0.0683) state unemployment 0.7988 1.0521 0.0483 rate (0.5078) (0.7855) (0.0464) previous weekly 0.0101 0.0044 0.0002 earnings (0.0149) (0.0232) (0.0013) expect recall -4.0104 -6.6714 -0.3789 (0.8454) (1.3060) (0.0753) definite recall date -9.0943 -12.6815 -0.8681 (1.3362) (1.9871) (0.1039) weekly benefit 0.0085 0.0196 0.0013 amount (0.0071) (0.0112) (0.0007) constant -8.1138 -0.7546 2.2411 (5.0617) (8.0015) (0.4753) sample size 1,821 1,821 1,821 adjusted R-squared 0.1777 - -- Log likelihood -- -5385.7 -24422 Notes: 1. Estimates of equation (1) in which weeks of benefits received during the benefit year is the dependent variable. OLS and censored regression estimates independent variable. Weibull AFT estimates give the proportional change in benefit duration with respect to a unit change in each independent variable. Standard errors in parentheses. 2. All specifications also include state indicators. 118 Table A2 Jobless Duration Estimates Censored Weibull OLS regression (A FT) potential benefit 1.6435 1.4408 0.0139 duration (0.2535) (0.6604) (0.0082) female -5.0420 -17.0974 -0. 1900 (8.7863) (21.6193) (0.2389) age 3.1319 7.8451 0.0978 (0.9670) (2. 5860) (0.0327) age squared -0.0024 -0.0055 —0.0001 (0.001 1) (0.0030) (0.0000) married 13.3215 29.9840 0.3821 (6.7394) (17.1283) (0.2026) not married 20.7402 44.4206 0.5187 (9.0719) (23.7364) (0.2928) female, not married -20.1865 48.5688 -0. 5571 (12.9082) (33.0753) (0.3931) female, married 0.9796 13.8123 0.1900 (10.1199) (25.4291) (0.2945) dropout 1 1 .5686 22.9998 0.2555 (5.8365) (15.8824) (0.2058) post secondary -6.0042 -18.9077 -0.2453 (5.3825) (13.8622) (0.1652) college -1 1.1150 -28.9345 -0.3545 (6.1122) (15.7627) (0.1887) other education 4.8369 -14. 1281 -0. 1562 (15.1330) (39.2833) (0.4641) african-american 8.0855 12.0489 0.0498 (6.5717) (17.1702) (0.2026) asian 12.5858 46.3573 0.5960 (16.7182) (46.6493) (0.6051) mexican 25.3250 47.8970 0.5166 (8.6148) (22.8936) (0.2919) other non-white 2.7780 8.3174 0.1724 (9.3977) (24.8922) (0.3136) manufacturing -1.1818 -2.9079 -0.0185 (4.2073) (11.0689) (0.1370) state unemployment 12.9625 18.4753 0.1985 rate (2.7990) (7. 3581) (0.0894) previous weekly -0.0267 -0. 1705 -0.0028 earnings (0.0819) (0.2158) (0.0026) 119 Table A2 (continued) Censored Weibull OLS regression (AFT) expect recall 5.9793 15.3085 0.2241 (4.6594) (12.2895) (0.1532) definite recall date 3.5007 19.7948 0.1949 (7.3652) (20.1784) (0.2622) weekly benefit -0.0733 -0.1081 -0.0009 amount (0.0391) (0.1039) (0.0012) constant -88. 5374 -143. 9320 1.8464 (27.8994) (73.2411) (0.9013) sample size 1,821 1,821 1,821 adjusted R-squared 0.0990 — - Log likelihood - -5009.5 -2161.9 Notes: 1. Estimates of equation (1) in which weeks jobless is the dependent variable. OLS and censored regression estimates give the change in jobless duration with respect to a unit change in each independent variable. Weibull AFT estimates give the proportional change in jobless duration with respect to a unit change in each independent variable. Standard errors in parentheses. 2. All specifications also include state indicators. Table A3 Benefit Exit and Reemployment Hazard Estimates - Parametric and Semi-Parametric Weibull (hazard) Cox PH Benefit exit Reemployment Benefit exit Reemployment potential benefit -0.0007 -0.0090 -0.0028 -0.0055 duration (0.0043) (0.0053) (0.0044) (0.0055) female 0.0197 0.1226 0.0099 0.1194 (0.1409) (0.1541) (0.1389) (0.1538) age -0.0163 -0.0631 -0.0153 -0.0635 (0.0169) (0.0210) (0.0166) (0.0208) age squared 0.0000 0.0000 0.0000 0.0000 (0.0000) (0.0000) (0.0000) (0.0000) married 0.2620 -0.2466 0.2558 -0.2286 (0.1052) (0.1306) (0.0984) (0.1334) not married 0.0310 -0.3347 0.0182 -0.3140 (0.1535) (0.1888) (0.1528) (0.1873) female, not married -0.0738 0.3595 -0.0487 0.3397 (0.2183) (0.2535) (0.2160) (0.2520) female, married -0.2915 -0.1226 -0.2839 -0.1303 (0.1628) (0.1900) (0.1617) (0.1913) dropout -0.0326 -0. 1649 -0.0335 -0.1526 (0.0924) (0.1327) (0.0905) (0.1331) post secondary -0.3019 0.1583 -0.3026 0.1448 (0.0907) (0.1065) (0.0921) (0.1064) college -0.2097 0.2287 -0.2123 0.2253 (0.1048) (0.1216) (0.1055) (0.1234) other education -0.0888 0.1008 -0.0868 0.0910 (0.2499) (0.2995) (0.1881) (0.3016) african-american -0.3820 -0.0321 -0.3846 -0.0433 (0.1199) (0.1307) (0.1254) (0.1259) asian 0.3237 -0.3846 0.3293 -0.3714 (0.2571) (0.3903) (0.2476) (0.4069) mexican -0.2282 -0.3334 -0.2138 -0.3214 (0.1386) (0.1882) (0.1321) (0.1914) other non-white -0.1338 -0.1113 -0.1268 -0.1200 (0.1642) (0.2023) (0.1559) (0.2037) manufacturing 0.1436 0.0119 0.1369 0.0172 (0.0662) (0.0884) (0.0654) (0.0909) Table A3 (continued) Weibull (hazard) Cox PH _ Benefit exit Reemployment Benefit exit Reemployment state unemployment -0.0469 -0.1281 -0.0481 -0.1084 rate (0.0449) (0.0576) (0.0465) (0.0580) previous weekly -0.0002 0.0018 -0.0002 0.0018 earnings (0.0013) (0.0017) (0.0013) (0.0016) expect recall 0.3673 -0. 1446 0.3766 -0.1354 (0.0728) (0.0988) (0.0714) (0.1000) definite recall date 0.8416 -0. 1258 0.8270 -0.1158 (0.1006) (0.1692) (0.1028) (0.1773) weekly benefit -0.0012 0.0006 -0.0013 0.0005 amount (0.0006) (0.0008) (0.0006) (0.0008) constant -2. 1726 -1.1915 -- -- (0.4631) (0.5861) sample size 1,821 1,821 1,821 1,821 Log likelihood -2442.2 -2161.9 -7404.1 -4748.1 Notes: 1. Estimates of equation (1) with distributional assumption and dependent variable as noted. Estimated coefficients give the proportional change in the hazard of benefit exit or jobless duration with respect to a unit change in each independent variable. Standard errors in parentheses. _ 2. All specifications also include state indicators Table A4 Benefit Exit and Reemployment Hazard Estimates - Non-Parametric maxexg wksleft Benefit exit Reemployment Benefit exit Reemployment maxexp -0.00208 -0.00008 - -- (0.0006) (0.0000) 2 maxexp 0.00001 0.00000 -- -- (0.0000) (0.0000) wksleft -- -- -0.00767 -0.00020 (0.0004) (0.0000) wksleft -- -- 0.00012 0.00000 (0.0000) (0.0000) elapsed -- -- 0.00254 -0.00013 (0.0004) (0.0000) elapsedz -- .- 0.00002 0.00000 (0.0000) (0.0000) weekly benefit -0.00004 0.00002 -0.00005 -0.00000 amount (0.0000) (0.0000) (0.0000) (0.0000) base period 0.00000 -0.00000 0.00000 0.00000 earnings (0.0000) (0.0000) (0.0000) (0.0000) female 0.00109 0.00032 -0.00143 0.00032 (0.0058) (0.0005) (0.0056) (0.0005) age -0.00053 -0.00020 -0.00110 -0.00021 (0.0006) (0.0000) (0.0006) (0.0000) age squared 0.00000 0.00000 0.00000 0.00000 (0.0000) (0.0000) (0.0000) (0.0000) married 0.00747 -0.00113 0.01040 -0.00092 (0.0043) (0.0004) (0.0042) (0.0004) not married 0.00201 -0.00129 0.00320 -0.00104 (0.0057) (0.0005) (0.0055) (0.0005) female, not married -0.00262 0.00067 -0.00362 0.00055 (0.0082) (0.0007) (0.0080) (0.0007) female, married -0.00677 -0.00016 -0.00807 -0.00026 (0.0066) (0.0006) (0.0065) (0.0006) dropout 0.00023 -0.00035 -0.00076 -0.00031 (0.0038) (0.0003) (0.0038) (0.0003) post secondary -0.00514 0.00021 -0.00800 0.00022 (0.0034) (0.0003) (0.0033) (0.0003) college -0.00263 0.00057 0.00045 0.00058 (0.0039) (0.0003) (0.0039) (0.0003) Table A4 (continued) maxexp wksleft Benefit exit Reemployment Benefit exit Reemployment other education -0.00402 -0.00012 -0.00794 -0.00033 (0.0093) (0.0008) (0.0091) (0.0008) african-american -0.00698 -0.00033 -0.01052 -0.00027 (0.0042) (0.0003) (0.0041) (0.0003) asian 0.00869 -0.00087 0.00863 -0.00110 (0.0131) (0.0010) (0.0128) (0.0010) mexican -0.01056 -0.00056 -0.01355 -0. 00043 (0.0058) (0.0005) (0.0057) (0.0005) other non-white -0.00249 -0.00010 -0.00348 -0.00016 (0.0060) (0.0005) (0.0058) (0.0005) manufacturing 0.00669 -0.00008 0.00803 0.00000 (0.0027) (0.0002) (0.0027) (0.0002) previous weekly -0.00004 0.00000 -0.00001 0.00000 earnings (0.0001) (0.0000) (0.0001) (0.0000) expect recall 0.00461 -0.00033 0.00705 -0.00032 (0.0030) (0.0002) (0.0030) (0.0002) definite recall date 0.03490 0.00005 0.05066 -0.00001 (0.0057) (0.0004) (0.0056) (0.0004) state 0.05127 -0.08086 -0.11672 -0.07999 unemployment rate (0.7103) (0.0582) (0.6962) (0.0581) constant 0.12579 0.01000 0.14869 0.01763 (0.0198) (0.0016) (0.0179) (0.0014) sample size 33,048 254,191 33,048 254,191 adjusted R squared 0.0050 0.0011 0.0443 0.0025 Notes: 1. Estimates of equation (2) with dependent variable as noted. See text for further discussion. Estimated coefficients give the change in the probability of benefit exit or reemployment with respect to a unit change in each independent variable. Standard errors in parentheses. 2. All specifications also include state indicators APPENDIX B Table B1 Conditional Benefit Exit Probabilities (Discrete Hazards) Hazards Weeks since initial claim Risk set Unadjusted Cox maxexp potdur 0-1 1821 0.0992 0.0964 0.0974 0.0927 2-3 1652 0.0612 0.0585 0.0595 0.0548 4-5 1558 0.0493 0.0465 0.0475 0.0428 6-7 1487 0.0474 0.0447 0.0457 0.0410 8-9 1422 0.0483 0.0455 0.0465 0.0418 10-11 1359 0.0564 0.0536 0.0546 0.0499 12-13 1289 0.0631 0.0604 0.0614 0.0567 14-15 1215 0.0638 0.0610 0.0620 0.0573 16-17 1145 0.0710 0.0682 0.0692 0.0646 18-19 1072 0.0953 0.0925 0.0935 0.0889 20-21 981 0.0891 0.0863 0.0874 0.0827 22-23 904 0.1029 0.1002 0.1012 0.0965 24-25 823 0.8385 0.8358 0.8368 0.8321 26-27 231 0.0702 0.0674 0.0684 0.0637 28-29 223 0.0660 0.0633 0.0643 0.0596 30-31 216 0.1111 0.1083 0.1093 0.1047 32-33 205 0.2841 0.2813 0.2823 0.2777 34-35 180 0.0159 0.0131 0.0141 0.0094 36-37 179 0.0968 0.0940 0.0950 0.0903 38-39 173 0.1607 0.1579 0.1589 0.1543 40-41 164 0.0851 0.0823 0.0833 0.0787 42-43 160 0.1373 0.1345 0.1355 0.1308 44-45 153 0.1364 0.1336 0.1346 0.1299 46+ 147 0.7895 0.7867 0.7877 0.7830 Implied expected 19.43 19.55 19.50 19.70 benefit duration (in weeks) Change in weeks of - 0.12 0.08 0.28 benefit duration from 1 added week of potential benefits Notes: 1. The risk set is the number of workers in the sample who are still on benefits at the start of each two-week period. The unadjusted hazard is the proportion of workers in the risk set who exit benefits during the period. The final three columns are the adjusted hazards for each of the three hazard specifications. 125 Table B1 (continued) 2. For each model, the hazard of benefit exit in each period is adjusted according to the change in the benefit exit hazard rate due to a one week increase in potential benefits. 3. Implied expected durations, d calculated as: d = 22?; ftt where ft = (1-m1)(1-m2)...(1-mt_1)(mt) and where mt is the conditional probability of benefit exit in time period t. 126 Table 32 Conditional Reemployment Probabilities (Discrete Hazards) 127 Hazards Weeks since initial claim Risk set Unadjusted Cox maxexp potdur 0-1 1821 0.0927 0.0872 0.0926 0.0925 2-3 1764 0.0466 0.0411 0.0466 0.0464 4-5 1738 0.0432 0.0377 0.0432 0.0431 6-7 1715 0.0452 0.0397 0.0451 0.0450 8-9 1692 0.0432 0.0377 0.0432 0.0430 10-11 1671 0.0344 0.0289 0.0344 0.0342 12-13 1655 0.0490 0.0435 0.0490 0.0488 14-15 1633 0.0351 0.0296 0.0351 0.0350 16-17 1618 0.0291 0.0236 0.0291 0.0290 18-19 1606 0.0550 0. 0495 0.0550 0. 0548 20—21 1584 0.0185 0.0130 0.0185 0.0184 22-23 1577 0.0270 0.0215 0.0269 0.0268 24-25 1567 0.0305 0.0250 0.0304 0.0303 26-27 1556 0. 0257 0.0202 0. 0257 0.0255 28-29 1547 0.0469 0.0414 0.0469 0.0468 30-31 1531 0.0400 0.0345 0.0400 0.0398 32-33 1518 0.0417 0.0362 0.0416 0.0415 34-35 1505 0.0201 0.0146 0.0200 0.0199 36-37 1499 0.0341 0.0286 0.0341 0.0340 38-39 1489 0.0247 0.0193 0.0247 0.0246 40-41 1482 0.0326 0.0271 0.0326 0.0324 42-43 1473 0. 0637 0.0582 0.0636 0.0635 44-45 1456 0.0320 0.0265 0.0320 0.0318 46-47 1448 0.0413 0.0358 0.0413 0.0412 48-49 1438 0.0259 0.0204 0.0258 0.0257 50-51 1432 0.0442 0.0388 0.0442 0.0441 52-53 1422 0.0370 0.0316 0.0370 0.0369 54-55 1414 0.0817 0.0762 0.0817 0.0816 56-57 1397 0.0628 0.0573 0.0628 0.0627 58-59 1385 0.0503 0.0448 0.0502 0.0501 60-61 1376 0.0941 0.0886 0.0941 0.0940 62-63 1360 0. 0260 0.0205 0. 0259 0.0258 64-65 1356 0.0667 0.0612 0.0666 0.0665 66-67 1346 0.0429 0.0374 0. 0428 0.0427 68-69 1340 0.0149 0.0094 0.0149 0.0148 70-71 1338 0. 0303 0.0248 0. 0303 0.0301 72-73 1334 0.0469 0.0414 0.0468 0.0467 74-75 1328 0.0410 0.0355 0.0409 0.0408 76-77 1323 0.0342 0.0287 0.0341 0.0340 78-79 1319 0.0354 0.0299 0.0354 0.0352 Table 82 (continued) Hazards Weeks since initial claim Risk set Unadjusted Cox maxexp potdur 80-81 1315 0.0275 0.0220 0.0275 0.0274 82-83 1312 0.0377 0.0323 0.0377 0.0376 84-85 1308 0.0196 0.0141 0.0196 0.0194 86-87 1306 0.0500 0.0445 0.0500 0.0498 88-89 1301 0.0632 0.0577 0.0631 0.0630 90-91 1295 0.0787 0.0732 0.0786 0.0785 92-93 1288 0.0610 0.0555 0.0609 0.0608 94-95 1283 0. 0649 0.0595 0. 0649 0.0648 96-97 1278 0.0139 0.0084 0.0139 0.0137 98-99 1277 0.0282 0.0227 0.0281 0.0280 100-101 1275 0.0435 0.0380 0.0434 0.0433 102-103 1272 0.0455 0.0400 0.0454 0.0453 104—105 1269 0.0317 0.0263 0.0317 0.0316 106-107 1267 0.0328 0.0273 0.0327 0.0326 108-109 1265 0.0169 0.0115 0.0169 0.0168 110+ 1264 0.5043 0.4989 0.5043 0.5042 Implied expected 66.05 67.58 66.06 66.11 joblessness duration (in weeks) Change in weeks jobless — 1.53 0.01 0.06 from 1 added week of potential benefits Notes: 1. The risk set is the number of workers in the sample who have not yet found a a stable job at the start of each two-week period. The unadjusted hazard is the proportion of workers in the risk set who become reemployed during the period. The final three columns are the adjusted hazards for each of the three hazard specifications. 2. For each model, the hazard of benefit exit in each period is adjusted according to the change in the reemployment hazard rate due to a one week increase in potential benefits. 3. Implied expected durations, d calculated as: d = 228-(31 ftt where ft = (1-m1)(1-m2)...(1-mt_1)(mt) and where m, is the conditional probability of benefit exit in time period t. TYLB l RARIES 9517 mallllllllillllliImilllll 3 1293 13062