IS UNEMPLOYMENT REALLY SCARRING? EFFECTS OF UNEMPLOYMENT EXPERIENCES ON WAGES

Transcription

1 The Economic Journal, 111 (November), F585±F606.. Published by Blackwell Publishers, 108 Cowley Road, Oxford OX4 1JF, UK and 350 Main Street, Malden, MA 02148, USA. IS UNEMPLOYMENT REALLY SCARRING? EFFECTS OF UNEMPLOYMENT EXPERIENCES ON WAGES Wiji Arulampalam Joblessness leaves permanent scars on individuals. They not only lose income during periods of joblessness they are also further scarred by these experiences when they nd employment. A spell of unemployment is found to carry a wage penalty of about 6% on re-entry in Britain, and after three years, they are earning 14% less compared to what they would have received in the absence of unemployment. The scars are also carried into their second employment spell. The rst spell of joblessness is found to cause the most damage. Redundancy seems to be less stigmatising. There is plenty of evidence to suggest that an experience of unemployment increases the chances of a future unemployment experience; for Britain, see Narendranathan and Elias (1993) and Arulampalam et al. (2000). In addition, there is evidence to suggest that these individuals who experience repeat interruptions also go into jobs that are low paid and unstable (Stewart, 2000; Boheim and Taylor, 2000). These individuals not only experience loss of income during periods of interruptions, but are also further `scarred' by their experiences of unemployment. This scarring effect will not only contribute to wage inequality and poverty, but can also interfere with work incentives and therefore motivate a policy intervention. The current Labour government is concerned about this issue and has already implemented a series of policy reforms such as the National Minimum Wage, 10 pence starting rate of Income Tax, Working Families Tax Credit, and reforms to the National Insurance system under the headings `Welfare to Work' and `Making Work Pay'. This paper investigates this issue of scarring by estimating the magnitude of the effect of wage loss associated with interruptions using British data. In particular, this paper seeks an answer to the question of whether the wage loss associated with an interruption is temporary. Interruption brings about depreciation in acquired skills. In addition, `in-work' bene ts may also act as a disincentive for skill acquisition to the extent that these are not time limited. If scars associated with these interruptions are long lasting, then time-limited The BHPS data used in this paper were collected by the ESRC Research Centre on Micro-Social Change at the University of Essex, and made available through the ESRC Data Archive. I should like to thank Mark Taylor for providing help with the creation of some of the variables used in the study and to the Institute for Social and Economic Research Centre for providing the data on Travel-To-Work-Area unemployment rates. Any errors remain my responsibility. I am grateful to Alison Booth, Paul Gregg, Mary Gregory, Robin Naylor, Andrew Oswald, Mark Stewart, Mark Taylor, Jonathan Wadsworth, Ian Walker, Myrna Wooders, and participants at the Royal Economic Society Conference held in Nottingham in 1999, seminar participants at the Employment and Education Economics Group Meeting held in May 2000, Centre for Economic Performance, London School of Economics, and University of Warwick. This paper has also bene ted from comments from three anonymous referees and the editor, Steve Machin. This paper was produced as part of the project on Unemployment and Technical and Structural Change, which was funded by the Leverhulme Trust. The views in the paper are those of the author, and do not necessarily re ect those of the Leverhulme Trust. [ F585 ]

2 F586 THE ECONOMIC JOURNAL [ NOVEMBER welfare measures may be important to remove the dependency of these individuals on welfare. It is also important to provide/subsidise appropriate training for these individuals if unlimited welfare entitlements do adversely affect human capital incentives. If there was a penalty attached to an unemployment experience, this could be due to several reasons. 1 First, employers might use an individual's unemployment history to sort good workers from bad workers (Lockwood, 1991) and this might result in a relative loss of wage if the employer, at least initially, decides to pay the individual less than his/her marginal product. This will happen if employers perceive, rightly or wrongly, that spells out of employment result in some loss of skills (Pissarides, 1992). Second, an unemployed individual for various reasons (loss of unemployment bene ts, liquidity constraints, disutility from leisure, etc.) might be prepared to incur a wage loss. Nevertheless, it is also possible that the post-unemployment wage might be higher than the pre-unemployment wage if, for example, the individual voluntarily changes jobs with an intervening spell of unemployment so as to improve his/her wages. An unemployment spell might also be productive in enabling the individual to nd a job, which better matches his/her skills and preferences, and this might be associated with a wage gain over the preunemployment wage level. 2 Most of the earlier research on the effects of unemployment on subsequent wages used data from the Displaced Worker Supplements (DWS) to the American Current Population Surveys (CPS) where information is collected on pre- and post-displacement job characteristics and on the intervening period of non-employment. 3 Although the sample sizes are very large in these datasets, the DWS does have some disadvantages. First, information is collected retrospectively on one job loss in the past 5 years (3 years since 1994) prior to the interview date. The data might therefore be subject to recall bias. Second, one cannot address the effects of multiple job losses since this information is not collected in the DWS data. Third, since the data refers to a group of displaced workers, one cannot construct a control group of non-displaced workers to be able to make inferences on what would have happened to these individuals had not been displaced. To circumvent these problems, some researchers have used longitudinal or administrative datasets that contain 1 Atkinson et al. (1996), from an employer recruitment survey, nd that employers, when faced with a job application from the unemployed, do pay attention to reasons for unemployment, frequency of unemployment spells, duration of unemployment, what the unemployed had been doing during the spell, etc. Manning (2000) also provides evidence that employers do not select the unemployed for job interviews with the same frequency as those currently in work. Gregg and Wadsworth (2000) argue that labour supply shocks and cost shocks have been responsible for increases in the so-called `low-quality' jobs in the 1990s relative to the 1970s, and that these jobs are being taken up by individuals who are non-employed. 2 Interestingly, in a recent study using data for the United States as well as the Netherlands, Abbring et al. (1998) nd the post-displacement wages to be higher than pre-displacement wages. Displaced Worker Survey and the Current Population Survey for 1996 for the United States, and a combination of three separate datasets for 1992±96 for the Netherlands were used. 3 For excellent surveys, see Fallick (1996) and Kletzer (1998). For recent studies in a comparative context, see Kuhn (2000).

3 2001] IS UNEMPLOYMENT REALLY SCARRING? F587 information on both displaced as well as non-displaced workers. 4 Although one is able to construct a comparison group, the sample sizes of displaced workers are typically quite small in most of the available longitudinal data sets. Some problems of recall bias might also be present in these data sets. Administrative data sets on the other hand do not allow one, in general, to identify the nature of joblessness nor the reason for the interruption (Gregory and Jukes, 2001). 5 It is generally found that wages of workers who had separated from long tenure jobs had already started to fall well before the occurrence of actual separation, where separation is usually de ned as due to plant closures or mass layoffs (Jacobson et al., 1993; Stevens, 1997). In addition, when separated individuals are re-employed, the recovery of wages relative to pre-separation wages is found to take a very long time. 6 Additional job losses are also found to carry some scarring (Stevens, 1997). The number of separations experienced, the length of separations as well as the time in employment up to the point of the wage measurement are thus important as both will affect the measurement of wage loss associated with the separation. Hence, unless one has longitudinal information on pre- and post-separation wages, along with information on timing of these separations, comparison of just one pre-separation wage with one post-separation wage may underestimate the loss associated with the separation. To address some of the above issues, the approach taken in this study is to use a generalisation of the `difference-in-differences' technique. 7 A comparison group is used to estimate the wage changes that would have occurred in the absence of interruptions, using a rich set of controls and allowance for unobserved individual speci c effects. This study uses a sample of men drawn from the British Household Panel Survey data (BHPS) over the period 1991±97. Unlike most of the data sets that have been used in the literature, the BHPS has retrospective information on labour market history. This has a very big advantage since the analysis need not be con ned to interruptions that took place during the sample observation 4 See, for example, Ruhm (1991) and Stevens (1997) for studies that have used longitudinal data for United States. Jacobson et al. (1993) used an administrative data set for the State of Pensylvania in the United States. 5 DWS data also does not distinguish spells of unemployment from spells of economic inactivity (Abbring et al., 1998). 6 Jacobson et al. (1993) found that, for high tenure prime age men who were part of mass layoffs in 1982, quarterly earnings had started to decline substantially three years prior to their separation and recovery was very slow with losses still around 25% of their pre-displacement earnings even after ve years from separation. But note, Jacobson et al. (1993) assigned a value of zero for periods of joblessness. In contrast, using administrative data on registered unemployment and wages of British men for the years 1984±94, Gregory and Jukes (2001) found unemployment incidence to have only a 10% temporary effect on hourly earnings that disappeared over two years on re-engagement. 7 This approach is similar to those used in the programme evaluation literature (Heckman et al., 1999). To obtain unbiased effects, one has to assume that the shocks faced by all individuals are common and that the observables as well as unobservable characteristics which may affect the incidence of interruptions are all controlled for in the model.

4 F588 THE ECONOMIC JOURNAL [ NOVEMBER period 1991±97. 8 One is able to trace back and date the last separation and also other separations using the retrospective information. Hence, the effects of multiple spells of unemployment and also the duration of the spell on subsequent wages can be addressed. In addition, more importantly, information on the cause of job loss or change is available. This enables one to look at the effect of voluntary versus involuntary separation. The BHPS does have some shortcomings. Although one is able to increase the number of observed interruptions by the use of retrospective information, the number of separations recorded is still not very large. Only about, 10% (722 out of 7,291) of the person-year observations in the sample used for the wage analysis had a recorded previous labour market status that was nonemployment. 646 of the 722 observations (almost 90% related to a nonemployment spells that had occurred during 1990±97 and therefore only required a recall period of less than 12 months because the survey was carried out annually. Bias due to recall errors is still a possibility because some of the short duration interruptions may have been forgotten altogether. However, the longitudinal nature of BHPS along with very rich information on individual, employer and job characteristics as well as retrospective job history information allows one to control for unobservable individual and job speci c effects and thereby take account of recall errors that are time-invariant. 9 In summary, this study, while looking at the effects of an interruption on reentry wages, also addresses these questions: Does the type of interruption play a role? Are the effects of an interruption temporary and, if so, how long do they last? Do multiple spells of interruptions have additional wage losses associated with them? Is it only the incidences that matter, or does the actual duration of the interruption also have an impact on re-entry wages? Data and the sample used for the analysis are described in Section 1. The econometric model speci cation and estimation results are presented in Section 2. Some model extensions are considered in Section 3. Section 4 concludes. 1. The Data and Sample The data used for the analyses are from the rst seven waves of the British Household Panel Survey (BHPS), which is a nationally representative survey of 8 Although Borland et al. (1999) used the BHPS data set for their analysis, they restricted their unemployment spell to those that occurred during the observation period 1991±96 and looked at the effect of an interruption on wage growth between pre- and post-interruption weekly earnings. Nickell et al. (1999) imposed the same restriction. On the other hand, Gregory and Jukes (2001) imposed the restriction that the effects of registered unemployment only have a time-dependent effect up to about 2 years so as to minimise on the loss of data because of lack of retrospective information. 9 Paull (1997) notes that there is a tendency for individuals to rede ne time in unemployment as time out of the labour force in the BHPS data set. Some of the estimation carried out in this study combines the state of unemployment and inactivity into non-employment.

5 2001] IS UNEMPLOYMENT REALLY SCARRING? F589 around 5,500 households randomly selected South of the Caledonian Canal. The rst wave of the BHPS was conducted in the autumn of 1991, and annually thereafter (Taylor, 1996). Information on labour market status is recorded at each interview, and for the period beginning on 1 September a year prior to the interview. In addition to the normal set of questions, at the second wave of interviews (in 1992), respondents were asked to detail their complete labour market histories since leaving full-time education for the rst time. At the third wave, complete job history data were also collected. The analyses presented in this paper uses information from regular questions as well as the two additional retrospective lifetime data. At each interview, the respondent is asked about the current labour market status. If the individual is currently employed, details about the current job are collected. If the individual started this particular job after the 1 September of the previous year, only then is the information about the previous labour market status collected. As an example, consider an individual who was present in waves 1 and 3 but not in wave 2. Supposing this individual started a job sometime between wave 1 interview date and 1 September 1992, then there is no information about what s/he did in the year prior to the job s/he was doing at the time of the third wave interview. Hence, the analyses on wages presented in this paper had to be restricted to those individuals who were interviewed in consecutive waves starting with the initial wave in Selecting men who were aged between 16 and 58 in 1991 and who were directly interviewed by an interviewer in 1991 gave a sample of 3,516 men. 11 These are the Original Sample Members. To avoid problems associated with the measurement of self-employment hourly wages, current employment status is de ned to exclude self-employment. Restricting the sample to those who were continuously present in the survey and who said that they were in paid employment in at least two of the waves gave a sample of 2092 men and 7,291 person-year observations. Descriptive statistics on a set of characteristics for the original as well as the selected sample are given in Table 1. There are more men, aged 25±44, with partners in employment in 1991, with dependent children and well quali ed in the selected sample compared to the original sample. There are also fewer men who are council house tenants and who have some sort of disability in the selected sample. The 1991 Travel-to-Work-Area (TTWA) unemployment rate, which can be thought of as a proxy for local labour demand, seems to be similar across both samples. Given the differences in the characteristics of the original and the selected sample members, to allow for the possibility that the selected sample may not be a random sample, a correction for selectivity is carried out in the estimation of wages equations (see Section 2.1 for details). 10 Analysis of wages of women is more complicated because of interruptions related to family formation and is therefore omitted from this work. 11 To avoid problems associated with early retirement decisions, the sample was restricted to those men who were less than 59 in Individuals who gave either a telephone interview or a proxy interview were not included in the analysis as many of the variables of interest had missing values.

6 F590 THE ECONOMIC JOURNAL Table 1 Mean (Standard deviation) [ NOVEMBER Sample used in the Original sample wage analysis Age less than ±34 years ±44 years White Married Spouse employed Has dependent children School type Private Grammar no fee Highest quali cation Degree and above Other higher (teaching, nursing) `A' levels or equivalent `O' levels or equivalent Has completed an apprenticeship Housing tenure Owns outright Has a mortgage Council tenant Health disability affects the type of work done Registered disabled Travel-To-Work-Area unemployment rate (1991) (0.02) (0.02) Number of observations (men) 3,511 2,092 See Section 1 for de nitions of these samples. 2. The Model Speci cation and Estimation There are two important related econometric issues that need to be dealt with in this type of analysis. The rst is to do with unobserved heterogeneity. In the standard human capital model, returns to tenure and experience are interpreted as returns to speci c and general human capital, respectively. In these models, high tenure workers are better paid because their productivity rises with time on the job and more experience implies more acquired general skills. However, search models have been put forward as an alternative. Here, the explanation given for the returns to tenure and experience is in terms of selectivity and matching arguments. The selectivity argument is that more able workers will face better promotion prospects and thus will have high tenure. The matching argument is that a match between a rm and an individual will last longer if it is a `good' match and also more experienced workers would have had more time to nd a good match. Tenure and experience variables are thus functions of past quits and lay-off decisions. As a result, tenure and experience variables will be correlated with unobservable job speci c or match speci c variables. Hence, ordinary least squares (OLS) estimation of a standard wages function using cross-sectional data will produce biased estimates of

7 2001] IS UNEMPLOYMENT REALLY SCARRING? F591 the returns to tenure and experience. 12 It is therefore important to account for unobservable individual heterogeneity and unobservable job-match heterogeneity in the estimation. Although the longitudinal nature of this data set enables one to control for unobservables, one is unable to distinguish the cause of positive returns to tenure. Information on rm and job characteristics is used to control for this. 13 Thus, all correlations are assumed to be adequately captured by the use of extensive individual, rm, and employer characteristics, as well as allowance for unobservable time-invariant individual speci c characteristics. Consider the following log-linear wage equation: w it ln(w it ) ˆ x it 9â (d it z it )9ã á i u it i ˆ 1,..., n and t ˆ 1,..., T i (1) where w it is the natural logarithm of wage at time t for individual i; x it is a vector of observable variables on both individual as well as rm characteristics which can be time varying or invariant; d it is a dummy variable which takes the value of one if the individual had come into this employment via a spell of interruption; z it is a vector of observable variables on individual characteristics; á i is a time-invariant individual speci c error term capturing the effects of unobservable characteristics; u it is the equation error term. The possible correlation between the unobservables and the observables needs to be accounted for in the estimation of the parameters of interest, â and ã. Hence the above model is estimated using within-group (WG) estimation, which is equivalent to a simple least squares estimation of the model in which the variables are de ned as deviations from their individual means. 14 This is the generalisation of the `difference-in-difference' estimation that will enable us to recover the effect of an interruption by removing the common macro effects as well as the unobservable individual speci c effects. This procedure requires at least two wage observations and therefore requires individuals to be found in employment in at least two of the waves with nonmissing wage information. This is the second econometric issue, the possibility of sample selection bias, that is addressed in the next section. 12 Some examples of studies that have tried to address the issue of endogeneity of tenure are Abraham and Farber (1987), Altonji and Shakotko (1987), Topel (1991), Altonji and Williams (1997), and Dustmann and Meghir (1999). 13 If there are omitted individual-job speci c unobservables in the equation, correlation between these and the tenure variables can either result in an upward biased estimate (Altonji and Shakotko, 1987), or in a downward biased estimate of the returns to tenure (Topel, 1991). A good match implies a lower turnover because the individual will be less likely to quit and hence a positive correlation between tenure and job-match variable. On the other hand, individuals who have high turnover are the individuals who move to obtain a better job match and hence turnover will be negatively correlated with tenure. This is a very dif cult issue to resolve which requires strong assumptions regarding the nature of correlations, and access to a very rich data set with information on individuals' full labour market history along with associated wages (Dustmann and Meghir, 1999). 14 Taking rst differences to eliminate the unobservables is another technique that can be used. Since the wage observations do not necessarily refer to two consecutive years, this study uses WG estimation.

8 F592 THE ECONOMIC JOURNAL [ NOVEMBER 2.1. Selection Equation As discussed earlier, to deal with unobservables one needs at least two wage observations per individual. Selecting a sample like this can create a nonrandom sample of individuals. For example, all those individuals who have had a spell of interruption but have not reported two wages information will be dropped from the analysis. If proper account is not taken, the estimated wage loss can be understated. The standard technique employed in these circumstances involves two steps. First, a model to explain the probability of an individual being in the selected sample used in the estimation of the wages equation is estimated using a reduced form probit. Second, a correction term (Heckman's) is constructed using the generalised residuals (inverse Mill's ratio) from the probit and used as an additional regressor in the wage equation to correct for the selection. The dependent variable in the selection equation is a dummy that takes the value of 1 if an individual was continuously present from the beginning of the survey (1991) for at least two waves and found to be in employment in at least two of the waves. The dependent variable is set to 0 if the above is not satis ed. Men who were only interviewed in 1991 and never after, and men who were not in employment in at least two waves, were all coded as having a value of 0. Thus, the men who are in the wage equation sample are those with the value of 1 for this dummy variable. The selection into the estimation sample will generally depend on characteristics that affect employment probabilities, mobility decisions, attitudes to participation in surveys, etc. Among the set of variables entering the selection equation, one also requires some variables that in uence the probability of being in the sample but not the observed wages conditional on being in the sample. This is the identifying restriction one requires to identify the parameters of the wages equation using the selected sample. The choice of this set of identifying variables is, as usual, somewhat ad hoc. Travel-to-Work-Area unemployment rate, employment status of the partner if the individual was married or co-habiting (`spouse employed'), a dummy for whether the individual has any children (also interacted with the marital status), type of school attended, housing tenure, father's occupational dummies, etc. were used as identifying variables. 15 These identifying variables are assumed to in uence the probability of being in the selected sample but are assumed not to in uence the wages conditional on being in the sample. All variables in the selection equation refers to pre-sample information and characteristics recorded in the rst interview in In addition to the above set of variables, dummies for various age groups, ethnicity, disability, highest educational quali cations and standard region of residence in 1991 were all included in the selection equation. Some of the 15 TTWA unemployment rate, which is used to proxy local labour demand is likely to be a poor proxy. For example, Webster (2000) nds there to be large errors in the 1981-based de nition of TTWAs because of commuting imbalances. Since workless households are spatially very concentrated in high unemployment areas, the variables `spouse employed', regional dummies, housing tenure, etc. in the model should enable one to pick up additional effects. I am grateful to a referee for this point.

9 2001] IS UNEMPLOYMENT REALLY SCARRING? F593 above time-invariant variables such as ethnicity and highest educational quali- cations, will also have an affect on wages, but the wage equation already allows for individual speci c unobservables. The selection analysis is based on 3,511 men. 16 The results from this probit equation estimation are presented in Appendix Table A.1. The set of identifying variables is jointly signi cant (likelihood ratio test gives a 2 (11) value of 77.8) in the selection equation. This selection analysis is a cross-section one. To allow for possible time-varying biases, the selection term (Heckman's correction for sample selectivity) was interacted with time dummies in the models estimated. 17 Any biases that are time invariant due to sample selection and any possible recall errors will be accounted for in the WG estimation Wage Equations The dependent variable in the wage equation is the log of the real hourly wage that was calculated by dividing the usual weekly gross pay by the usual weekly hours, de ated using the Retail Price Index, and measured in 1991 prices. 18 Unfortunately, the usual pay includes usual overtime pay, as information excluding overtime payment is not available. 19 The main issue addressed in the paper is the effect of an interruption on reentry wages. To carry this out, previous labour market states considered are employment, unemployment and out-of-the labour force (OLF) (or economic inactivity). Unlike the current labour market status of employment that excludes self-employment, the previous labour market status of employment includes self-employment. 20 Individuals in the sample categorise themselves. The retrospective information in the data set does not allow one to distinguish between those who were searching and those who were not searching during these spells. The OLF category includes individuals who had been long-term sick, at home looking after family, in education or training, etc. The tenure variable used is the current employer tenure as recorded at the interview. This variable was grouped into dummy variables instead of using the standard quadratic speci cation. This was carried out to minimise the effects of possible measurement error due to problems of recall and also to pick up any higher-order nonlinearities in the effects of tenure on current wages. The 16 This gure is 5 less than the original gure of 3,516 due to the deletion of cases with missing values for variables used in the selection equation. 17 The reduced form probit equation for sample selection was estimated as one cross-section so as to account for the fact that there was an overall selection of at least two wage observations per individual over the sample period. This is not just an employment selection alone and hence a multinomial probit or a logit model which models the observed labour market status of the individual is not used for this purpose. 18 If `hours worked' is measured with errors, using hourly earnings can suffer from measurement errors. But the use of weekly wages will not allow one to control for the fact that some individuals also might be `scarred' in terms of going back to jobs with fewer hours. 19 Results presented are not sensitive to the assumption that over-time is paid at the rate of 1.5 times the usual hourly rate. 20 This restriction was not rejected by the data.

10 F594 THE ECONOMIC JOURNAL [ NOVEMBER distribution of tenure by the previous labour market status is given in Table 2. Since this study looks at the effect of an interruption on the immediately following spell of employment, the employer tenure is the same as the time since interruption for people who had an interruption. Looking at it from the perspective of current employment, 90% of current employment refers to an employment-to-employment transition. Only about 8% of current employment refers to unemployment-to-employment transition and about 2% to an OLF-toemployment transition. Although the time period covered by the actual survey was seven years, it is very interesting to note that only about 20% of those who had come into this employment spell after a spell of unemployment have recorded current employer tenure to be in excess of 3 years. Long tenure seems to be a very rare occurrence among this group of individuals. Individuals coming into employment after a spell of economic inactivity also have low observed tenure. There are at least two possible reasons for low tenure among these individuals. One possibility is that these individuals enter into unstable low-paid jobs and are back in unemployment soon afterwards. The other possibility is that some individuals nd better jobs and change employers. To shed some light on this issue, the paper extends the basic model by looking, in Section 3, at the effects of multiple spells of interruptions. The variables which are commonly thought to have an effect on wages such as age, marital status, type of job, sector of the rm, rm size, union coverage, whether the individual was in receipt of any work-related training in the last 12 months, etc. were included as controls. Regional and industry dummies as well as a dummy for those who had entered this employment directly after leaving full-time education were also included. Descriptive statistics for all the variables used in the analyses are presented in Table 3. As expected, there are more men aged less than 25, single, and with lower educational quali cations in the sample of men who had experienced an interruption prior to this employment Employer tenure Employment (Col %) Table 2 Previous spell Unemployment (Col %) Economic inactivity y (Col %) Less than one year 818 (12.5) 230 (40.8) 59 (37.3) 1±2 years 415 (6.3) 142 (25.2) 43 (27.2) 2±3 years 309 (4.7) 81 (14.4) 19 (12.0) 3±4 years 249 (3.8) 49 (8.7) 11 (7.0) 4±5 years 244 (3.7) 30 (5.3) 8 (5.1) 5±10 years 1,193 (18.2) 24 (4.3) 12 (7.6) 10 or more years 3,341 (50.9) 8 (1.4) 6 (3.8) Total 6,569 (90.1 % of Total) 564 (7.7 % of Total) 158 (2.2 % of Total) 7,291 person-observations in total 1991±97. y `Economic inactivity' category includes men who were in full-time education prior to this spell of employment.

11 2001] IS UNEMPLOYMENT REALLY SCARRING? F595 Mean (Std. Deviation) Table 3 Descriptive Statistics by Previous Labour Market Status 1991±97 Previous labour market status Variables Employment Nonemployment Personal Characteristics Age 25± Married Spouse is in employment Has dependent children School type Private Grammar no fee Highest quali cation ± Degree ± Other higher (teaching, nursing, etc.) ±`A' levels or equivalent ±`O' levels or equivalent Has completed an apprenticeship Housing Tenure Owns outright Has a mortgage Council tenant Health disability affects work Registered disabled Job and Workplace Characteristics Working in a public company Working for a charity Workplace size ± 50±99 employees ±199 employees >200 employees Had work-related training in the last 12 months Workplace union presence Current job is part-time Job ± Professional/managerial ± Skilled non-manual ± Skilled manual ± Semi-skilled manual Current employer tenure ± months 149 (124) 24 (34) Full-time experience ± months 202 (149) 147 (156) Industrial af liation Energy and water supplies (11 to 17) Extraction of minerals, etc. Manufacture of metals (22 to 26) Metal goods, engineering & vehicles (31 to 37) Other manufacturing (41 to 49) Construction (50) Distribution, hotels & catering, repairs (61 to 67) Transport & communication (71 to 79) Banking, nance, etc. (81 to 85) Other Services (91 to 99) Hourly Wage (1991 prices) 8.74 (5.6) 5.85 (4.7) Weekly hours (7.1) (9.5) Sample size-person year observations 6, The sample refers to the sample used in the estimation of the wage equation.

12 F596 THE ECONOMIC JOURNAL [ NOVEMBER compared to those who had not. These men also seem more likely to be working in private companies and doing manual jobs. Not surprisingly, they are also less likely to have received any work-related training in the last 12 months. The estimates of the log real hourly wage equations for various models estimated are summarised in Table 4. There are 7,291 person-year observations used in the analysis of the wages. The previous labour market status variables are entered as dummies in Column [1]. In this speci cation, the type of interruption the individual had prior to current employment is assumed to Table 4 Log Real Hourly Wage Equations for Men 1991±97 Within-Group Coef cient Estimates (Robust Std. Errors) [1] [2] [3] Tenure in current employment base is,1 year 1±2 years (0.012) (0.013) (0.013) 2±3 years (0.012) (0.015) (0.015) 3±4 years (0.014) (0.017) (0.017) 4±5 years (0.014) (0.017) (0.017) 5±10 years (0.013) (0.014) (0.014). 10 years (0.015) (0.015) (0.015) Previous labour market status Unemployment (0.020) Inactivity (0.041) Came from u/e ± time since interruption,1 year (0.020) 1±2 years (0.025) 2±3 years (0.032) 3±4 years (0.039). 4 years (0.040) Came from other state ± time since interruption,1 year (0.034) 1±2 years (0.042) 2±3 years (0.058). 3 years (0.052) Came from non-employment ± time since interruption,1 year (0.018) 1±2 years (0.023) 2±3 years (0.029) 3±4 years (0.036). 4 years (0.035) Adjusted R Number of observations ± person years ˆ 7,291. The sample consists of all men who were continuously interviewed (until they exit) and were found to be in employment at least twice over the period 1991±97. Full set of control variables are: age dummies, time dummies, a dummy for men whose current job is the rst job since leaving full-time education, labour market experience dummies, marital status, health disability, temporary/ xed-term contract, part-time job, employment sector, rm size, received training in current job, job type, regional dummies and industry dummies. Correction for selectivity interacted with time dummies is also included. See text for further details. ± signi cant at 5% or less; ± signi cant at 10%

13 2001] IS UNEMPLOYMENT REALLY SCARRING? F597 have a constant effect that persists over time, ceteris paribus. Relative to making an employment-to-employment transition, entering employment via a spell of unemployment is estimated to carry a wage penalty of about 7% ceteris paribus (Column [1]). Entering employment via a spell of economic inactivity is estimated to carry an even larger penalty. For these individuals, the penalty is estimated to be about 11% ceteris paribus. Excluding the controls (results not reported in the table) did not change these estimates signi cantly. 21 Columns [2] and [3] allow the effect of an interruption to vary with the interruption as well as with time since interruption. Unlike the model in Column [1], these models allow one to check whether the `scarring effect' of a spell of interruption diminishes over time. After a spell of unemployment, an individual is estimated to earn about 6% less during his rst year with the employer, ceteris paribus. 22 This loss is estimated to increase over time up to about 4 years to a value of 14% and starts to diminish very slowly after that. After 4 years with the current employer, the penalty is estimated to be about 11% (Column [2]). If these individuals enter into unstable and short-duration jobs, the above result implies that these scarred individuals never have an opportunity to recover from the loss in this employment spell. This has important implications for labour market policies since the results imply that these individuals are not able to acquire enough job speci c skills. It is therefore important to match individuals to good jobs and to make sure that they do stay on in these jobs. An individual coming into employment after a spell of economic inactivity is estimated to be worse off compared to someone coming from a spell of unemployment. As seen in Table 2, the number of observations relating to this category is very small. It is also possible that the distinction between unemployment and inactivity is very fuzzy (Paull, 1997). Column [3] therefore looks at the effect of amalgamating the unemployment and out-of-the labour force statuses into one category of `non-employment'. The estimated `average scarring' effects in the rst year is about 6%; thereafter it increases to about 10% and remains at this level for a long time. This is consistent with the fact that alternative prospects for these individuals must be even worse than what they are experiencing in terms of scarring since they are still in the same employment. As discussed earlier, job changes are not purely random. Indivi- 21 These models include a correction for possible selectivity bias (see Section 2.1 for details). The selection correction term which is time invariant, was interacted with time dummies to pick up possible time-varying biases as the model already allows for possible time-invariant biases via the á i term (see (1)). This speci cation only allows one to estimate the differential impact of the selection term over time with respect to the base year, which is In all of the models estimated, only the 1997 selection correction effect was signi cantly different from the base year effect. In addition, exclusion of these terms did not signi cantly affect the estimated coef cients of the other variables. For example, the effects of previous unemployment and economic inactivity were estimated to be and with associated standard errors of and respectively when the correction terms were excluded. These are no different to those reported in Column [1] of Table Gregory and Jukes (2001) estimated this penalty to be about 10% and Nickell et al. (1999) estimated it to be about 12% but the de nition of unemployment used in these two studies refers to registered unemployment. The model speci cations are slightly different in terms of the assumption made regarding the maximum length of time the unemployment is expected to have a differential effect. The sample periods used are also different.

14 F598 THE ECONOMIC JOURNAL [ NOVEMBER duals who had a choice of a better alternative outside of this employment (say a better job) would have chosen to leave. In this case, the above estimated wage loss would be over-stated. However, it is also possible that the above estimated wage loss is under-stated if some individuals had left this job and had reentered non-employment. It is therefore very important to look at the issue of scarring caused by repeat incidences of non-employment spells. Another interesting related question is whether those individuals who had changed employers are also facing a wage penalty in the next employment. This and some other related issues are addressed in Section Model Extensions Unlike most of the other data sets that have been used in the past to look at the effects on wages of individuals who have been displaced, BHPS has some extra information about the interruption itself. In particular, the type of interruption, the length of interruption and also the number of previous spells of unemployment experienced by the individuals, are available. For most of the sample members, these are derived from information collected retrospectively. Given the possibility of recall errors, this information is used in an `aggregate' manner to extend the previous analyses. The data also has many missing values. Observations with missing values were not omitted from the analysis. Instead, appropriate dummy variables for missing cases were used as additional regressors. The model presented in Column [3] of Table 4 was extended for these analyses and the results are presented in Table 6. Some relevant frequencies are presented in Table Effect of an Interruption Caused by a Redundancy The most important reported reason for leaving the last employment was redundancy with the second being voluntary changes (Table 5). Of the 1,430 men who were made redundant from their last job, 81% had managed to nd employment without an intervening spell of `non-employment'. The estimated results that include a dummy for men who lost their last job because of redundancy are presented in Column [1] of Table 6. This dummy was also interacted with the dummy for individuals who had made a nonemployment to employment transition. Inclusion of dummies for reasons such as termination of employment because (i) a temporary job came to an end, (ii) of dismissal, did not produce any signi cant results because of small cell sizes. The omitted category therefore includes individuals who had voluntarily left their job, been sacked or dismissed, and those who had temporary jobs that had come to an end. Experiencing a spell of non-employment after being made redundant from a job is estimated to be less scarring compared to an interruption caused for a different reason. The estimated wage loss is only about 3.4% (9:9±6:5), ceteris paribus. There are at least two possible reasons for this. First, there is a period of notice that has to be given by the employer prior to making someone

15 2001] IS UNEMPLOYMENT REALLY SCARRING? F599 Table 5 Frequencies Previous Labour Market Status Employment (% of 6 569) Non-employment (% of 722) 1. Reason for leaving previous job Made redundant 1,155 (17.6) 275 (38.1) Sacked/dismissed 151 (2.3) 20 (2.8) Temporary job ended 234 (3.6) 61 (8.5) Voluntarily quit 3,894 (59.3) 99 (13.7) Missing 379 (5.8) 111 (15.4) 2. Length of previous interruption Less than 6 months 126 (17.5) 6 to 12 months 84 (11.6) More than 12 months 104 (14.4) Missing 405 (56.1) 3. Has had a spell of unemployment One or more previous unemployment spells 2,630 (40.0) 582 (80.6) Missing 5 (0.00) 65 (9.0) Number of person-year observations 6, For the unemployed, this is in addition to the spell of unemployment under consideration as an interruption. redundant (Borland et al., 1999). Second, some of these individuals may also be eligible for redundancy payments and unemployment bene ts. Individuals who quit their jobs voluntarily are disquali ed from receiving bene ts for up to 26 weeks. Given these reasons, it is possible that these individuals start looking for a job early (because of the period of notice) and are able to wait until they nd something suitable (because of redundancy payments). Unfortunately, with the exception of Wave 5, information on whether the individuals who were made redundant were in receipt of any redundancy payments is not available in this data set. Previous existing work on scarring has not been able to estimate this effect due to lack of data. Another possible reason for the estimated results is that individuals who were made redundant were those possessing general skills and therefore the redundancy itself does not result in scarring due to loss of rm speci c skills Effect of the Length of the Interruption It is often believed that acquired human capital will depreciate during a spell of non-employment, and the longer the interruption, the more the decline. As a result, many labour market policies have been aimed at preventing long spells of unemployment. In particular, there has been a programme of `RESTART' in existence in this country since the Initially, all individuals who had been registered as unemployed for twelve months or more ± and, since 1987, all those who had been registered unemployed for six or more

16 F600 THE ECONOMIC JOURNAL Table 6 Log Real Hourly Wage Equations for Men 1991±7 Within-Group Coef cient Estimates (Robust Std. Errors) Model Extensions [ NOVEMBER Was made redundant in last employment [1] [2] [3] (0.024) Previous labour market status Non-employment (0.018) (0.027) (0.030) Was made redundant and had a spell of non-employment (0.033) Was made redundant and had a spell of non-employment and age (0.041) Previous labour market status was non-employment and the spell duration Was > 6 and,12 months (0.053) Was > 12 months (0.044) One or more spell of (0.026) unemployment y One or more spell of unemployment and previous spell is non-employment (0.035) Adjusted R See notes to Table 4 on page F596. These are extensions to model in Column [1] of Table 4 with the previous labour market status entered as non-employment. In addition to all the previous variables, these models also have a dummy for cases with missing information on the reason for leaving previous employment in Column [1], a dummy for cases with missing previous labour market status duration in Column [2], and a dummy for cases with missing number of previous spells of unemployment in Column [3]. y For the individuals who were unemployed in the previous state, one or more spell means that they have had at least one other spell of unemployment prior to the last one. months ± have been invited for counselling. These individuals are then introduced to available training and support in order to match them into jobs that are better suited to their skills. Bene t entitlements are also re-evaluated at the same time (Borland et al., 1999). This important feature of the programme carries a threat element in the sense of bene ts withdrawal if the individual does not attend the RESTART interview or if s/he is deemed not to be making genuine attempts to nd a job. Dolton and O'Neill (1996) found that the RESTART interview scared individuals into signing off and the timing of exits out of unemployment coincided with the rst RESTART interview. If individuals are under pressure to take up any employment, it might be the case that these individuals go into low-pay jobs. 23 Gregory and Jukes (2001) and Nickell et al. (1999) found for Great Britain that long unemployment spells were associated with greater wage losses. 23 Boheim and Taylor (2000) found that individuals who spend more time unemployed and searching for work are rewarded with a better worker± rm match in their subsequent job.

17 2001] IS UNEMPLOYMENT REALLY SCARRING? F601 To test the above in our sample, the base model (Column [3] Table 4) was re-estimated with the inclusion of two dummies for the length of the spell of interruption. The results from this extension are presented in Column [2] of Table 6. Unfortunately, there was no information on the length of the interruption for just over half the sample members who had experienced an interruption (Table 5). As before, a dummy for missing values was included in the regressions. The omitted duration category was a non-employment spell of length less than 6 months. No signi cant effect of the actual spell duration was found in addition to the incidence effect. Differentiating unemployment spells from those of economic inactivity also did not produce any signi cant effects via the spell duration, ceteris paribus. This is in contrast to what was found by Gregory and Jukes (2001) and Nickell et al. (1999). One possible explanation for this is that the unemployment variable used in these two studies refers to registered unemployment. It is possible that the negative effect found in these studies was due to the RESTART programme that is targeted at individuals who are signing on. The unemployment variable in this study includes both registered as well as unregistered unemployment and hence the estimated effects refer to the average over both types of spells. There is also another possible explanation for lack of signi cant effects from the actual length of the spell. There was a big shift of unemployed people on to sickness bene ts approximately over the decade 1987±97 and the sample period used here in this study covers part of this period. In contrast, Gregory and Jukes (2001) who look at the period 1984±94 could have therefore picked up more long-duration unemployed Effect of Previous Spells of Unemployment Nearly 81% of those who had experienced a spell of interruption have also had another spell of unemployment prior to the last one (Table 5). Among the men who had moved to this spell of employment via another spell of employment without an interruption, 40% had had a spell of unemployment sometime in the past. One implication of the earlier base model was that individuals with better alternative outside would have changed employers. If this is the case, then an interesting question to ask is whether a previous spell of unemployment still has an effect after a subsequent change in employer, or is it the case that only the immediately preceding spell of unemployment carries a scar. 25 The results of this investigation are presented in Column [3] of Table 6. Three different binary indicator variables were included in this regression. These were for whether (i) previous labour market status was non-employ- 24 I am grateful to a referee for this point. 25 The model estimated looks at the effect of a previous spell of unemployment and not of `nonemployment'. The reason for this is that it was felt that it would not be correct to include spells of inactivity where some spells included education and training.