1. Are Training Programs More Effective When Unemployment is High? Michael Lechner and Conny Wunsch www.siaw.unisg.ch/lechner October 2006

2. © Michael Lechner, Conny Wunsch, p. 1 Idea of the paper Understand whether the effectiveness of training programs for the unemployed depends on the state of the labour market Why is that important? optimize timing and volume of ALMP over the business cycle better understanding of empirical evidence for different time periods and countries Problem: hard to find suitable data regional cross-sectional studies face a lot of regional heterogeneity macro studies cannot address relevant selection problem meta studies have a lot of individual study-specific heterogeneity experiments typically do not trace program entries long enough

3. © Michael Lechner, Conny Wunsch, p. 2 What do we know so far? Meta study by Kluve (2006) seems to indicate a positive dependence of program effectiveness on unemployment Johansson (2001) uses variation of Swedish active labour market programs over municipalities  programs prevent unemployed from leaving the labour force during a downturn  positively related to UE Raaum, Torp, Zhang (2002): - 12 cohorts of labour market training participants in Norway 1991-1996 - 4 subgroups: men/women with/without UB - outcome variable: annual earnings 1/2/3 years after program - meta analysis: (i) pool all estimates, (ii) exploit county-level variation - results: positive correlation of the ATET with unemployment rate as well as exit rate to employment at outcome measurement - do not control for changing composition of participants and program mix

4. © Michael Lechner, Conny Wunsch, p. 3 Our contribution Systematic investigation of the relationship between program effectiveness and labour market conditions using administrative data for West Germany 1980-2003 10 years of monthly program entries (1986-1995) 8 years after program start to observe outcomes at least 6 years before program start to control for selectivity institutional environment relatively stable control for changes in composition of participants and program mix

5. © Michael Lechner, Conny Wunsch, p. 4 The West German economy 1984-2003

6. © Michael Lechner, Conny Wunsch, p. 5 Training as a part of German ALMP ProgramDescription Job search assistance Learn how to write CV / application, how to locate job vacancies; practice job interviews etc. (only until end of 1992) Short training Further training: general update or adjustment of skills; planned duration  6 months. Practice firmsSimulate working in a specific profession Long trainingSame types as short training with a planned duration > 6 months. Retraining Training to obtain a new professional degree in a field other than the profession currently held. Ranked in ascending order of planned program duration: min. 1 week max. 48 months

7. © Michael Lechner, Conny Wunsch, p. 6 Training as a part of German ALMP

8. © Michael Lechner, Conny Wunsch, p. 7 Institutional changes 1994 (1-3%) cut in replacement rates of unemployment benefits (UB), unemployment assistance (UA) and benefits during participation in training (maintenance allowance, MA) eligibility: reduction in required work experience before program by 3 years  changes small and can be controlled for in the data

9. © Michael Lechner, Conny Wunsch, p. 8 The data Employment subsampleBenefit payment registerTraining participant data Source Employer supplied mandatory social insurance entries. Benefit payment register of the FEA. Questionnaires filled in by caseworker for statistical purposes ( ST35 ). Population 1% random sample of persons covered by social insurance for at least one day 1975-1997. Self- employed, civil servants, students are not included. Data 1980-2003 Recipients of UA, UB, or MA 1980-2003. Participants in further training, retraining, short programs (§41a EPA), German language courses and temporary wage subsidies 1980-1997. Available information Personal characteristics and history of employment. Information about the receipt of benefits, mainly UB, UA, MA. Personal characteristics of participants and information about training programs. Important variables Gender, age, nationality, education, profession, employment status, industrial sector, firm size, earnings, regional information Type and amount of benefits received. Type, duration and result of the program, type of income support paid during participation. StructureSpells based on daily information.Spells based on daily info.Spells based on monthly information. Note: The merged data is based on monthly information. For detailed information on the merging and recoding procedures see Bender et al. (2004). The creation of this data base is a result of a three year joint project of research groups at the Universities of Mannheim (Bergemann, Fitzenberger, Speckesser) and St. Gallen (Lechner, Miquel, Wunsch) as well as the Institute for Employment Research of the FEA (Bender).

10. © Michael Lechner, Conny Wunsch, p. 9 Sample definition For each month 1986-1995 check: participants start training in that month nonparticipants do not enter training but are unemployed with receipt of UB/UA, no program also in the 11 following months age 20-55, no homeworkers/students, no part-time workers < ½ full-time eqiv. no program in the 4 years before eligibility: receipt of UB/UA in month before  multiple appearance of a person as participant or nonparticipant possible  pool participants/nonparticipants over 6 months to obtain sufficient sample sizes  check sensitivity w.r.t. these choices

11. © Michael Lechner, Conny Wunsch, p. 10 Program starts in our sample (pooled)

12. © Michael Lechner, Conny Wunsch, p. 11 The formula that explains what we estimate 1.participants in month t (ATET at t) 2.population which has same personal characteristics as pool of all participants 1986-1995 reduced to common support 3.population which has same personal characteristics and program mix as pool of all participants 1986-1995 reduced to common support Vary population P t in an interesting way:

13. © Michael Lechner, Conny Wunsch, p. 12 Plausibility of the conditional independence assumption in our data:  eligibility: ensured by sample definition  selection of caseworkers: detailed personal, regional, employer information  selfselection of UE: initial and remaining UE benefits claim, previous earnings  6 years of monthly pre-program employment history Potentially important variables that are missing:  jail and health status histories  caseworker assessment of the UE (about motivation etc.)  unobservable factors captured to the extend to which they had impacts on pre-program employment history  focus on correlation of effects with labour market conditions: bias no problem if uncorrelated with labour market conditions Identification: selection on observables

14. © Michael Lechner, Conny Wunsch, p. 13 Estimation: modified matching estimator as in LMW `05 Matching (for each month/6-month window): for each participant find one or more nonparticipants who are as similar as possible in all characteristics that jointly influence participation and the outcome of interest similarity within a prespecified radius comparisons are weighted according to their distance in characteristics characteristics can be summarised by participation probability to overcome curse of dimensionality (Rosenbaum and Rubin, 1983) and allow for semiparametric estimation of the effect common support: effect can only be estimated for the group of people for which there are comparable participants and nonparticipants [apply – per period – matching estimator of Lechner, Miquel, Wunsch, 2005]

15. © Michael Lechner, Conny Wunsch, p. 14 Outcomes of interest employment (subject to social insurance) cumulated employment registered unemployment (receipt of UB/UA, participation in training) cumulated registered unemployment (cumulated) monthly earnings  6 months after program start (average of months 5-7) [lock-in effect]  3 years after program start (average of months 34-39)  6 years after program start (average of months 61-72)  8 years after program start (average of months 85-96) [long-run effect]

16. © Michael Lechner, Conny Wunsch, p. 15 Results: program effects Unemployment 6 months after program start Employment 8 years after program start Employment 6 months after program start Unemployment 8 years after program start

17. © Michael Lechner, Conny Wunsch, p. 16 Results: employment rates Employment 8 years after program start Employment 6 months after program start

18. © Michael Lechner, Conny Wunsch, p. 17 Results: correlation with macro indicators Outcome Unemployment rate at Quarterly GDP growth rate # of participants in training programs program start outcome measurement Unemployment 6 months program after entry -43** -33*319 3 years after entry-36*21810 6 years after entry-27*24*1521 8 years after entry2617 Employment6 months after entry25*58 3 years after entry45**-45**2-3 6 years after entry43**-33**-3-33** 8 years after entry31**-47**-12-50** Cumulated unemployment 6 months after entry-43** 815 8 years after entry-50**271722 Cumulated employment 6 months after entry20 69 8 years after entry52**-37**-4-22 Note: For the uncumulated outcomes the unemployment rate at outcome measurement is the rate measured in the respective month after program start. For the cumulated outcomes the unemployment rate at outcome measurement is the average unemployment rate over the respective period. Newey-West autocorrelation-robust standard errors: ** significant at the 1% level, * significant at the 5% level.

19. © Michael Lechner, Conny Wunsch, p. 18 But: participants change over time

20. © Michael Lechner, Conny Wunsch, p. 19 But: participants change over time Characteristics of program participants Unemployment rate at program start Woman-52** Foreigner-24* No professional degree-67** University/college degree7 Duration of last unemployment spell-51** Fraction of months employed in the last 6 years82** Fraction of months unemployed in the last 6 years-46** Note: Correlation of the monthly mean of the respective variable (six-month moving average) with the corresponding unemployment rate. ** significant at the 1% level, * significant at the 5% level....in a way which is correlated with labour market conditions If effects are heterogenous w.r.t. these variables, then correlation may be due to them! Adjust participants and nonparticipants to the same distribution of characteristics over time

21. © Michael Lechner, Conny Wunsch, p. 20 Reference population which is held constant over time pool of all participants 1986-1995 for which there is common support for all 6-month windows

22. © Michael Lechner, Conny Wunsch, p. 21 Effects with characteristics of participants held constant Unemployment 6 months after program start Employment 8 years after program start Employment 6 months after program start Unemployment 8 years after program start

23. © Michael Lechner, Conny Wunsch, p. 22 Correlations with UE rate rather increase! Outcome Unemployment rate at Previous specification: Unemployment rate at program start outcome measurement program start outcome measurement Unemployment 6 months after entry-49**-45** -43**-33* 3 years after entry-48**19 -36*21 8 years after entry1915 26 Employment6 months after entry36**24 25*5 3 years after entry45**-56** 45**-45** 8 years after entry31*-30** 31**-47** Note: For the uncumulated outcomes the unemployment rate at outcome measurement is the rate measured in the respective month after program start. Newey-West autocorrelation-robust standard errors: ** significant at the 1% level, * significant at the 5% level.

24. © Michael Lechner, Conny Wunsch, p. 23 But: composition of programs also changes over time

25. © Michael Lechner, Conny Wunsch, p. 24 But: composition of programs also changes over time Note: Correlation of the monthly mean of the respective variable (six-month moving average) with the corresponding unemployment rate. ** significant at the 1% level, * significant at the 5% level....in a way which is correlated with labour market conditions Characteristics of programs Unemployment rate at program start Participants“Stable” participants Fraction of participants in short training33**24* Fraction of participants in long training33**25** Fraction of participants in retraining6 Fraction of participants in job search assistance-42**-27* Planned program duration81 Planned duration of short training-17-20* Planned duration of long training-25**-21* Planned duration of retraining-20*-50** Planned duration of job search asssistance40** keep program shares and planned duration constant over time (drop participants in job search assistance because of lack of support)

26. © Michael Lechner, Conny Wunsch, p. 25 Effects with constant population and program mix Unemployment 6 months after program start Employment 8 years after program start Employment 6 months after program start Unemployment 8 years after program start

27. © Michael Lechner, Conny Wunsch, p. 26 The correlations are still there! Outcome Unemployment rate at Previous specification: Unemployment rates at program start program start outcome measurement participants constant participants change Unemployment 6 months after entry-61**-60** -49**-43** 3 years after entry-33*33** -48**-36* 8 years after entry614 19 Employment6 months after entry51**39** 36**25* 3 years after entry27-58** 45** 8 years after entry42**-20 31*31** Note: For the uncumulated outcomes the unemployment rate at outcome measurement is the rate measured in the respective month after program start. Newey-West autocorrelation-robust standard errors: ** significant at the 1% level, * significant at the 5% level.

28. © Michael Lechner, Conny Wunsch, p. 27 Merely seasonal patterns captured? - No evidence! Outcome Average monthly unemployment rate program start 1986-19901991-1995 Unemployment 6 months after entry34-21 3 years after entry-552 8 years after entry-5520 Employment6 months after entry-23-21 3 years after entry-9-43 8 years after entry649 1986-19901991-1995 but only 12 observations!

29. © Michael Lechner, Conny Wunsch, p. 28 Regional variation? – Not much! Low UE regionHigh UE region Outcome Monthly unemployment rate program start Low UE rate regionsHigh UE region Unemployment 6 months after entry-56**-32* 3 years after entry-46**-38* 8 years after entry-21-20 Employment6 months after entry36**13 3 years after entry39*38* 8 years after entry48**49** more noisy but still some correlation left

30. © Michael Lechner, Conny Wunsch, p. 29 Stable over time? – Perhaps (more noisy) Outcome Monthly unemployment rate at program startoutcome measurement 1986-19901991-19951986-19901991-1995 Unemployment 6 months after entry-36*-79**-44**-67** 3 years after entry-63**-1356**-20 8 years after entry97734 Employment6 months after entry33*63**39*47* 3 years after entry59**-3-47**3 8 years after entry2259**13-24 Note: The last month in the first period is September 1990. The first month in the second period is October 1990. For the uncumulated outcomes the unemployment rate at outcome measurement is the rate measured in the respective month after program start. Newey-West autocorrelation- robust standard errors: ** significant at the 1% level, * significant at the 5% level.

31. © Michael Lechner, Conny Wunsch, p. 30 Sensitivity checks (final specification) merely seasonal pattern captured? -> no variation between low/high UE regions -> no contradictions stable over time? -> some changes depending on when outcome is measured but overall conclusions unchanged pooling of observations over 4/9 months: reduced/increased precision, correlation somewhat smaller/larger, conclusions unchanged no program in 6/12/24 months before program: no change since in common support no program before nonparticipants no program for 6/24 months after start date: conclusions unchanged future participation of nonparticipants uncorrelated with labour market cond. operational characteristics of matching estimator: LMW (2005) -> robust

32. © Michael Lechner, Conny Wunsch, p. 31 Conclusions negative effects in the short run over the whole period positive employment effects in the long run most of the time almost no long-run effects on unemployment  confirms findings of previous studies

33. © Michael Lechner, Conny Wunsch, p. 32 Conclusions short and long-run employment effects are positively correlated with unemployment at program start short-run unemployment effects are negatively correlated holding the composition of participants and program constant over time sharpens this finding seasonal correlation does not contradict this finding regional correlation does not contradict this finding patterns over time do not contradict this finding other sensitivity checks do not question this finding

34. © Michael Lechner, Conny Wunsch, p. 33 Conclusions Possible explanations for our findings: short run: lock-in effect less severe if UE is high or worsening long run: long-term consequences of lock-in effect? - worse employment record - human capital depreciation during longer UE