Unemployment persistence

Transcription

1 Unemployment persistence Wiji Arulampalam 1, University of Warwick Alison L Booth, University of Essex and CEPR Mark P Taylor, University of Essex November 1997 Revised November 1998 Abstract We estimate dynamic panel data models of unemployment incidence for men using the British Household Panel Survey. Econometric issues concerning unobserved individual heterogeneity, genuine state dependence, and the initial conditions problem are addressed in detail. We find strong evidence of state dependence consistent with the scarring theory of unemployment - an individual s previous unemployment experience has implications for his future labour market experience. This suggests that policies reducing short run unemployment incidence will have longer-run effects by reducing the equilibrium or natural rate of unemployment. JEL Classification: C23, C25, J6, J64 Keywords: unemployment persistence, unemployment recurrence, state dependence, natural rate of unemployment. 1 Nee Narendranathan. Address for correspondence - Wiji Arulampalam, Department of Economics, University of Warwick, Coventry, CV4 7AL, UK. Tel: (44) ; Fax: (44) ; wiji.arulampalam@warwick.ac.uk

2 1 1. Introduction Despite the massive research effort that has gone into investigating the causes of the rise in European unemployment, the current high unemployment rates cannot be explained by either cyclical factors or by exogenous shifts on the supply side.while some of these have been found helpful in explaining particular episodes, neither singly nor as a group do they seem to be able to account for the continuous high unemployment levels. Rather there appear to be persistence mechanisms present that lead today s equilibrium unemployment rate to be positively related to yesterday s realisation of unemployment. As a consequence, temporary disturbances can have long-lasting effects. The presence of these mechanisms blurs the simple-minded distinction between demand and supply factors because demand shocks end up having longer-term supply consequences. Bean (1997: 90-1). In the first half of the 1990s, the unemployment rate in Britain was between 9 and 10%. It is important to find out if the same individuals experience unemployment year after year, or if instead unemployment incidence is spread more equally across the labour force with a different set of individuals observed in unemployment each year. The first implies highly persistent unemployment amongst a group of individuals, while the second suggests that everyone has an equal chance of being unemployed. If there is persistence in unemployment, it is important to find out the cause. Is persistence due to differences across workers in their propensity to work? Or is there instead a causal link between past unemployment and current unemployment? Understanding the cause of any observed persistence in unemployment occurrence will assist policy makers in determining the appropriate policy response. We explore the determinants of persistence in unemployment occurrence. We estimate dynamic panel data models of unemployment incidence, in order to disentangle the effects of unobserved individual heterogeneity and true state dependence after controlling for observable characteristics. 2 True state dependence - or scarring - is where there is a causal link between 2 Unobserved individual heterogeneity may affect the propensity of certain individuals to be unemployed. Individuals have characteristics unobservable to the survey statistician, but which may nonetheless be observed by the firm and affect the arrival rate of job-offers and individuals retention rates in current employment, thereby affecting an individual s propensity to be in unemployment.

3 2 past unemployment and current unemployment, so that an individual who does not experience unemployment now will behave differently in the future to an otherwise identical individual currently experiencing unemployment. Our investigation of unemployment incidence and persistence at the individual level yields some stylised facts, which we relate to current microeconomic theory underpinning macroeconomic unemployment persistence. We also investigate whether the relationship between previous and current unemployment is different for young, possibly more mobile workers, than for older workers. We argue that there are grounds for expecting this to be the case. Identification of the extent of true state dependence among men of working age is more than just an academic exercise. There is considerable diversity of government policies towards the unemployed in market economies throughout the world. 3 These policies rest on very different assumptions about the extent to which government intervention can alter the equilibrium, or so-called natural rate, of unemployment. If there is no state dependence in unemployment incidence at the micro level, then short run policies to reduce unemployment (such as job creation schemes and wage subsidies) will have no effect on the equilibrium aggregate unemployment rate. But if there is true state dependence, then policies reducing short run unemployment incidence will have longer run effects by reducing the natural rate of unemployment. The prevention of the initial unemployment experience becomes an important policy objective, perhaps indicating the need to focus on education and training initiatives. Thus it is crucial to know whether there is genuine state dependence, or if instead it is individual heterogeneity that causes individuals to be repeatedly observed in unemployment. The increasing flexibility of the labour market in Britain is also relevant. To the extent that high rates of labour turnover 3 See Snower (1997), and Snower and de la Dehesa (1997) and chapters therein, for an extensive analysis of unemployment policies.

4 3 increase the risk of experiencing unemployment, then - if there is genuine state dependence in unemployment - unstable early years in the labour market could have far-reaching repercussions for an individual s career and for the social security bill. What causes true state dependence or scarring? Past unemployment experiences may change preferences, prices and or constraints that help determine current unemployment (Heckman and Borjas, 1980). Firms may judge workers productivity by their past history of unemployment. Thus workers with a history of job mobility and unemployment incidence may be offered less secure jobs because they lose valuable work experience or human capital while unemployed, or because their unemployment experience is used as a signal by employers that they are low productivity (Phelps, 1972; Lockwood, 1991; Pissarides, 1992; Blanchard and Diamond, 1994). Alternatively, individuals in unemployment may lower their reservation wage with the passage of time, and accept poorer quality jobs that are more likely to be destroyed, and for this reason may be more likely to experience unemployment in the future. However, it is difficult to distinguish between unobservable heterogeneity and state dependence. 4 Even when careful control is taken of unobservable heterogeneity using panel data estimation techniques, the initial conditions problem must still be dealt with in order to disentangle further the effects of state dependence and unobserved heterogeneity. The initial conditions problem arises when the start of the observation period does not coincide with the start of the stochastic process generating individuals unemployment experiences. For example, our analysis is based on the British Household Panel Survey, which began in But a large proportion of the men in our sample entered the labour market prior to 1991, and these men have already been at risk of unemployment before the survey period began. Thus an 4 In addition, any observed persistence in unemployment incidence may be an artifact of the data. For example, in panel data that are collected annually, a single unemployment spell may span two consecutive years. This would yield a spurious estimate of unemployment persistence if, for example, a very large percentage of men in our sample had experienced an unemployment spell which had lasted for more than one year. We investigate a number of ways of dealing with this problem in our estimation.

5 4 individual who is observed in the state of unemployment in 1991 may be there because of an earlier history of unemployment (state dependence) or because of some observed and/or unobserved characteristics affecting the job-offer or job retention rates facing that individual. In order to unravel these effects, the initial condition needs to be explicitly modelled rather than assumed exogenously given. This is what we do in the present paper. These problems can only be properly addressed with panel data. 5 The remainder of this paper is set out as follows. Section 2 describes the data set, while Section 3 presents the dynamic econometric model. Our estimates of the determinants of unemployment incidence are presented and discussed in Section 4. Some calculations from the estimated models of unemployment probabilities and the contributions due to true state dependence are presented in the next section. The final section concludes. We find strong evidence of state dependence for British men. This state dependence effect is especially pronounced for men aged 25 and over in Our findings are consistent with the scarring theory of unemployment. 2. The data The data are from the first five waves of the British Household Panel Survey (BHPS), a nationally representative survey of households randomly selected south of the Caledonian Canal. 6 The first wave of the BHPS was conducted from September to December 1991, and 5 Most work in this area has investigated state dependence in durations, ie whether there is duration dependence in the exit rate from unemployment. Exceptions are the following. Corcoran and Hill (1980) use US data for male heads of households age from and Heckman and Borjas (1980) use US data for young males from the National Longitudinal Survey Both studies find no evidence of state dependence in unemployment occurrence. In contrast, Narendranathan and Elias (1993) find strong evidence of state dependence in unemployment occurrence for a cohort of young British men aged 23 in 1981, as do McCulloch and Dex (1996) for men aged between 16 and 55 in Flaig et al (1993) and Muhleisen and Zimmermann (1994), using the first 6 waves of the German Socio-economic Panel, find strong evidence of state dependence for men (aged in 1984 for the former study, and in 1984 in the latter study). Another interesting study is Meyer and Rosenbaum (1996) which looks at repeat use of unemplolyment insurance. Our definition of an unemployed individual does not distinguish those who receive unemployment benefits from those who do not. 6 Thus the north of Scotland is excluded.

6 5 annually thereafter (for details, see Taylor, 1996). The sample has been chosen according to Wave 1 characteristics. In particular, it includes all men who were (i) aged 16 or over in 1991, (ii) born after 1935 (and thus under 60 at wave 5), and (iii) active in the labour market at Wave 1. Individuals remain in the sample at subsequent waves until they (i) exit the labour force, (ii) have missing relevant information, or (iii) are not interviewed at a particular wave. We therefore use an unbalanced panel, allowing individuals to exit the sample, but not allowing individuals to be added. The reason for this is that our econometric analysis requires consecutive observations (to allow for lagged unemployment) and a common date of entry to the panel (to facilitate estimation of initial conditions). We also estimated all our models on a balanced panel (defined as the sample only including men in the labour force in all five waves), and our conclusions remain unaltered. We use the usual survey definition of unemployment - currently out of work and looking for employment at the date of each interview. Table 1 shows the distribution of unemployment spells across waves, for all men with valid data for variables used in the analysis, and the transition probabilities from the raw data. At Wave 1 (1991), 8.0% of the initial sample of 2524 men were unemployed and looking for work. At Wave 2, the size of the sample falls to 2408, of which 7.1% are unemployed. The sample size continues to fall at Waves 3, 4 and 5, to 1995, 1818 and 1657, while the percentage unemployed also falls, to 5.8%, 5.3% and 3.6% respectively. 7 Some of this fall is due to attrition, but the downward trend since 1992 also reflects the business cycle, with aggregate unemployment falling over the period. 8 (See Appendix Table A1 for details of reasons for sample attrition.) Table 1 Unemployment incidence and persistence 7 When the sample is stratified into men aged under 25 in 1991 at the date of interview, and men aged 25 and over, the total unemployment rates are 5.7 % for the former and 6.5% for the latter. 8 Paull (1996) and Elias (1996) have investigated the extent and effects of recall error among unemployment spells in the BHPS employment histories. However, recall error is not an issue here as we are concerned only with labour market status at each date of interview.

7 6 Wave 1 (1991) Wave 2 (1992) Wave (year) Wave 3 (1993) Wave 4 (1994) Wave 5 (1995) Incidence Percent unemployed Total Conditional Probabilities Prob(y t =1 y t-1 =1) Prob(y t =1 y t-1 =0) Sample size Note: y t =1 if the individual is unemployed in time t and zero otherwise. The second half of Table 1 gives the transition probabilities, where y t =1 if the individual is unemployed in time t, and zero otherwise. The conditional probabilities in the final column are calculated using pooled data from all five waves (that is, over the four pooled transitions). The first row of conditional probabilities shows the probability of being unemployed at time t, conditional on being unemployed at time t-1, while the last row presents the probability of being unemployed at time t, conditional on being employed (either as an employee or selfemployed) at time t-1. Some 54% of men who were unemployed at the interview date in year t-1 are also unemployed at the interview date one year later. For men who were employed at the interview date in year t-1, the probability of being unemployed at the interview date one year hence is just 2.7%. 9 The probability of being unemployed in Wave 2 (1992) conditional on unemployment at Wave 1 (1991) was 53%. In subsequent waves, this conditional probability has increased, reaching a maximum of 57% in Wave 4, after which it fell by 8 percentage points by Wave 5. Table 1 suggests that, based on the raw data, there is considerable persistence in the state of unemployment measured at the interview date. 9 These probabilities are little changed by splitting the sample by age. See the top half of Table 4 for further details.

8 7 Of course, while the focus of our analysis is on state dependence in unemployment occurrence, individuals observed as unemployed at consecutive interview dates may be in the same single, unbroken spell of unemployment. Table 2 shows the frequency distribution of elapsed (or incomplete) unemployment spells recorded at each of the dates of interview. From the last column, it can be seen that 21% of unemployment spells observed at the dates of interview have an elapsed duration of under two months, while about 43% have an elapsed duration of under 6 months. Some 36% of unemployment spells have an elapsed duration exceeding 12 months (the average time between interviews), and thus some individuals observed as unemployed in consecutive waves are in one continuous spell of unemployment. However, the median spells are considerably shorter than 12 months, as Table 2 reveals, especially for men under the age of 25. The median elapsed unemployment spell for the under 25s is over two months shorter than for those aged at least 25, perhaps reflecting the higher labour market mobility and flexibility of younger workers. Nonetheless, the fact that just over one third of unemployment spells have an elapsed duration exceeding 12 months is of some concern. Since we are interested in true state dependence in unemployment occurrence, we explore a number of different ways of netting out these longer spells, as will be further described in the next section.

9 8 Table 2 Frequency distribution for elapsed duration of unemployment spells Elapsed duration Wave 1 Wave 2 Wave 3 Wave 4 Wave 5 Total N % N % N % N % N % N % Under 2 months months months months months Total Median elapsed duration in months: whole sample men aged < for men aged Mean elapsed duration in months: whole sample men aged < for men aged The Econometric Model The observed dependent variable is binary, taking the value of one if the individual is unemployed at the time of the interview, and zero otherwise. This variable is observed at most at five separate interview dates. We specify the model for individual i at the interview date at time t as y * it = x it β + γy it-1 + v it, i=1,2,...,n and t=2,...,t i (1) where y * denotes the unobservable individual propensity to be unemployed, x is a vector of observable characteristics affecting y *, β is the vector of coefficients associated with x, and v is the unobservable error term. Since the sample is an unbalanced panel, the total number of observations per individual is T i - 1. An individual is observed to be unemployed when his

10 9 propensity to be unemployed crosses a threshold (zero in this case), that is, if y * it > 0 and = 0 else. Modelling State Dependence In equation (1) y * it is a function of the observed unemployment status of the individual in the previous period y it-1 ; that is, it is the actual experience of an unemployment spell, rather than the propensity to be unemployed, that affects the current incidence of unemployment. The inclusion of the lagged dependent variable on the right hand side of (1) allows us to test for the presence of genuine state dependence, or the so-called scarring effect. However, a positive sign to the coefficient of lagged unemployment can arise from spurious correlation resulting from (i) a single unemployment spell overlapping two (or more) consecutive time periods, or (ii) inadequate controls for individual characteristics correlated with individuals propensities to experience unemployment (Heckman, 1981a, 1981b). To control for the first of these problems, we estimate the model over a number of sub-samples and with differing lag structures, as will be explained in the next sub-section. To deal with the second, we control for observable and unobservable individual characteristics. Modelling Unobserved Heterogeneity Assuming the unobservable individual-specific heterogeneity is time-invariant, we can decompose the error term v it in (1) as: v it = ε i + u it (2) where ε i denotes the individual-specific unobservable effect and u it is a random error. We treat the ε i as random, and use the random effects probit models estimated under the common assumption that u it ~ IN(0, σ u 2 ) and the u it are independent of the x it for all i and t IN refers to Independent Normal distribution.

11 10 In order to marginalise the likelihood, we also assume that ε i ~ IN(0, σ 2 ε ) and is independent of the u it. In the simplest of the models, it is customarily assumed that the ε i is also independent of x it for all i and t. If this assumption is violated, maximum likelihood estimates will be inconsistent. That is, the estimated β coefficients will pick up some of the effects of the unobservable ε. As an example, suppose that ε represents individual commitment or responsibility, which makes the individual both more likely to be able to find a job and more likely to be owning a house. Thus any model which does not allow for the correlation between housing tenure and ε i will suffer from omitted variable bias. In this simple example, the estimates of the impact of owner occupation on the unemployment probability will be negatively biased. To avoid this problem, we relax the assumption that ε i is independent of the observable time-varying characteristics in x it. Following Chamberlain (1984), we model the dependence between ε and x by assuming that the regression function of ε i is linear in the means 11 of all the time-varying covariates and therefore we can write this as ε i = a 0 + a 1 x i + α i (3) where we also assume that α i ~ IN(0,σ 2 α) and is independent of the x it and the u it for all i and t, a 0 is the intercept, and x i refers to the vector of means of the time-varying covariates for individual i over time. Note, the coefficients in a 1 corresponding to the time-invariant variables in equation (1) are set equal to zero. Thus equation (1) becomes y * it = x it β + γy it-1 + a 1 x i + α i + u it, i=1,2,...,n and t=2,...,t i (4) where we have absorbed the intercept a 0 into the β. This is equivalent to the random effects probit model with additional regressors x i. 11 The results from the model where the regression function was specified in terms of leads and lags of all the time varying covariates were very similar to the one that is reported here where it is specified in terms of the means.

12 11 In the above specification the correlation between two successive error terms for the same individual is a constant, given by: r = corr(v it, v it-1 ) = σ2 α σ 2 σ 2, t=2,...,t i (5) + α u The Initial Conditions Problem We now consider the initial observation y i1 and the initial conditions problem that occurs if y i1 is correlated with the unobservable α i. This problem arises because the start of the observation period (1991) does not coincide with the start of the stochastic process generating individuals unemployment experiences. A large proportion of men in our estimating subsample entered the labour market prior to 1991, and these men have already been at risk of unemployment before the survey period began. Thus an individual observed in the state of unemployment in 1991 may be there because of an earlier history of unemployment (state dependence) or because of some observed and/or unobserved characteristics affecting his unemployment propensity. To account for this problem we follow Heckman (1981c) and first specify a reduced form equation for the initial observation as follows: y * i1 = λ z i + η i (6) where z i is a vector of strictly exogenous instruments, var(η i )= σ η 2 and corr(α i, η i )=ρ. The vector z includes variables relevant in period 1, some pre-sample information affecting the probability of unemployment in period 1 and, the vector of means x i. As noted above, the vector of means is included to pick up possible correlation between the time-varying regressors and any unobservable heterogeneity. The next step involves a linear specification, in terms of orthogonal error components, to account for the possibility of non-zero ρ:

13 12 η i = θ α i + u i1 (7) By construction, α i and u i1 in (7) are orthogonal to one another, θ=ρσ η /σ α and var(u i1 )= σ 2 η 2 ( 1 ρ ). We assume that the initial observation y i1 is uncorrelated with u it and also that u i1 is uncorrelated with the x it for all i and t. From (4), (6) and (7) we obtain: y it * = x it β + γy it-1 + a 1 x i + α i + u it i=1,...,n and t=2,..,t i (8a) y i1 * = λ z i + θ α i + u i1 i=1,...,n and t=1 (8b) Equations (8a) and (8b) specify a complete model for the unemployment process. As shown in Heckman (1981a, 1981b), this model is easily estimated by noting that the distribution of y it * conditional on α i is independent normal. One can then marginalise the likelihood with respect to the α to obtain the appropriate likelihood function for the maximisation. However, the estimation of the full model requires special software which needs to be written. Orme (1997) suggests a two-step method of estimation, in the spirit of Heckman s standard sample selection correction method which is an approximation in the case of small values of ρ. 12 To account for the correlation between the initial condition and the unobserved heterogeneity α, a correction term is added to the conditional model which is then easily estimated using standard software packages that enable estimation of random effects probit models. The motivation is as follows. We begin with equations (8a) and (6), and capture the correlation corr(α i, η i )=ρ by now specifying a different form to that of (7). This new specification is: 12 Although the asymptotic validity of this approach is based on fairly restrictive distributional assumptions and on the appropriateness of a locally valid approximation to the true model in which ρ is small, this procedure was found to produce very similar results to that of the full model estimates in the investigations presented in Arulampalam (1998). Other procedures based on less restrictive distributional assumptions have been developed by Arellano and Carrasco (1996) for random effects, and Honore and Kyriazidou (1996) for fixed effects. These approaches are, however, more demanding to estimate.

14 13 α i = δ η i + w i (9) Again, by construction, η i and w i in (9) will be orthogonal to one another, δ = ρσ α /σ η and var(w i )= σ α 2 ( 1 ρ 2 ). 13 We next substitute (9) into (8a) to obtain: y it * = x it β + γy it-1 + a 1 x i + δ η i + w i + u it i=1,...,n and t=2,..,t i (10) Before we can proceed with the estimation of equation (10), note the following two points (Orme (1997)). First, there are now two individual specific random error components, η i and w i. Second, an assumption of bivariate normality of (η i,α i ) implies that E(w i y i1 )=0 but E( η y ) = e where e = (2 y 1) ϕ( λ' z ) / Φ ({2 y 1} λ' z ) by construction. 14 Since i i1 i i i1 i i1 u it is also assumed to be orthogonal to the regressors, we can treat the w i as the usual error i component in a random effects probit model, provided we can take care of the unobservable η i. Since e i is a probit generalised error in the probit equation given by (6), we can replace η i by this conditional expectation. This will make equation (10) a random-effects probit model equation with an additional regressor e i under suitable normality assumptions. Thus the first step of Orme s two-step procedure involves estimation of (7) to generate this regressor, and the second-step involves estimation of (10) by the usual random effects probit estimation technique where e i is replaced by the probit residual. Unfortunately, the assumption of bivariate normality of (η i, α i ) implies that w i (the new error component that enters the random effects probit model specification) has var(w i y i1 ) = σ (1- ρ ζ ) α i where ζ i = ϕ( λ' z ) Φ( λ' z ) Φ(- λ' z ) i i i which is now heteroskedastic. But Orme (1997) shows that, from 13 These moments are NOT conditional on y i1. 14 Note the variables in x it are assumed to be exogenous and therefore by assumption, uncorrelated with the error terms in equation (8).

15 14 Monte-Carlo results, it is not necessary to worry about this heteroskedasticity producing inconsistent parameter estimates in the case of small values of ρ. In summary, the two-step estimation involves the following steps. First, using a simple probit model, we estimate the reduced form model for the initial observation y i1, as given in equation (6). Second, we generate the probit generalised error as given above, and use this as an additional regressor in equation (8a). A test of the null hypothesis that ρ=0 is given by the standard t-test of the coefficient of this additional regressor. It is interesting to note that, in the case of a small ρ, an estimate may be obtained as follows. Consider equation (10). We know that δ = ρσ α /σ η,. Given the binary nature of the unemployment variable, the first period probit equation estimation requires the normalisation that σ η = 1, and thus ρ= δ /σ α. For small ρ, var(w i ) var(α i ). We also know that ρ = σ 2 ω/(σ 2 ω + σ 2 υ). The software package we used, Limdep 7.0 (Greene, 1995), uses the 2 normalisation σ u = 1. Thus σ α r/(1- r) and an estimate of ρ may be obtained as ρ δ [(1- r)/r] 1/2 where δ is the coefficient attached to the probit generalised error variable and r is the proportion of the variance attributed to the unobserved individual heterogeneity in the total variance of the error term. 4. Empirical results In our model of unemployment incidence, we include the usual set of control variables that reflect individual search intensity, and job-offer arrival or job-retention rates. Since the model is a reduced form, we are unable to distinguish between the independent roles of search intensity or of arrival and retention rates. Demographic and family variables (age, family size, ethnicity etc.) are likely to affect both search intensity and marginal productivity (which will in turn affect the arrival of job offer and retention rates). The level of education is likely to affect

16 15 both search intensity and marginal productivity. Only the labour market tightness variable will proxy job-offer arrival and retention rates alone. Because just over one third of unemployment spells have an elapsed duration exceeding 12 months, we estimate the two-step econometric model outlined in the previous section on a number of different sub-samples and with different lag structures. The three variants of the random effects probit model are: (1) one year lags on the full estimating sample; (2) one year lags on the sub-sample excluding observations which last longer than the period between a pair of consecutive interviews; and (3) two-year lags on the sample including only Waves 1, 3 and 5. Table 3 displays the results from two-step estimation of the full model for each of these variants. The first column of results in Table 3 reports the estimates of the unemployment probability in Wave 1 (the initial condition), used in estimation of the full model. All models were estimated by including time means of the time varying covariates in the equation to pick up possible correlation between the xs and the unobservable heterogeneity. Variant 3 and the initial conditions used for variant 3 were estimated using means calculated over Waves 1, 3 and 5 only. The results for the 1991 initial conditions for variant 3 are reported in the penultimate column. Definitions of the explanatory variables are given in Appendix Table A2, while sample means and standard deviations are in Appendix Table A3. The proportion of the variance explained by unobserved individual heterogeneity, r, varies across model variants, from 9% (variant 2) to 25% (variant 1). Note that the value of the log-likelihood of the model with r=0 is smaller than for the reported model. Estimated values of ρ in the three variants are 0.555, and respectively, where ρ is calculated as described at the end of Section ΙΙΙ. The test of the null hypothesis that ρ=0 is rejected by the data for variant 1 at the 1% level of significance. We now consider the results of the full model.

17 16 State dependence estimates: We interact the lagged unemployment variable with age, in order to investigate whether the relationship between previous and current unemployment is different for young, possibly more mobile workers, than for those who are likely to have more responsibilities. 15 The positive and significant coefficient on the lagged dependent variables suggests that there is state dependence in unemployment for both mature (aged 25 years and over) and young (aged less than 25 years) male workers for all variants of the model. The magnitude of the estimated coefficients to the lagged unemployment variables is similar across variants 1 and 3 of the model, and provides clear evidence of state dependence both from the entire sample (variant 1) and from the sub-sample obtained by dropping Waves 2 and 4 (variant 3). Estimates from the latter are thus picking up the effect of state dependence across 2 years (since the dependent variable is lagged 2 years). 16 For variant 2, while the lagged unemployment variables are significantly positive, there is little difference between the magnitude of the coefficients for young and more mature men. 17 This result of state dependence is consistent with the scarring hypothesis, with unemployment experience having a significant effect on future labour market behaviour. Perhaps employers use an individual s unemployment history as a screening device (Phelps, 1972; Lockwood, 1991; Blanchard and Diamond, 1994), or unemployment experience results in depreciation of human capital (Pissarides, 1992). Persistence during a demand contraction might be quite long, because the reduction in the perceived average quality of the unemployed 15 Various interactions of age in 1991 with lagged unemployment status were tried, including aged 20, 22, 25, 27 and 30. The effect of the lagged dependent variable was robust across these various interactions. Age 25 at 1991 was chosen in the preferred specification based on the likelihood value. 16 Variant 3 was estimated as a way of netting out the impact of very long spells, since here we allow the interval between waves to be 24 months rather than the usual 12 (by dropping waves 2 and 4). 17 Since the models are estimated as probabilities using cumulative distribution functions, there is no restriction on the values that can be taken by the coefficient on lagged unemployment. This is in contrast to time-series studies, where the absolute value of the coefficient needs to be less than unity for stationarity.

18 17 during a demand contraction leads firms to open fewer vacancies, thus exacerbating the problem (Pissarides, 1992; Bean, 1997). Another alternative is that unemployed workers are more likely to accept low quality jobs characterised by high rates of job destruction. For variants 1 and 3 of the model, the state dependence effect for more mature workers is much larger than that for young men. 18 Why should the state dependence effect vary with age? According to the signalling hypothesis, unemployment incidence (and duration) may be interpreted by firms as proxies for unobservable worker productivity. Suppose some firms obtain information about prospective employees through testing and/or interviews. Workers passing (and seen as high productivity) are hired - so high productivity workers move from job to job without any experience of unemployment, or exit unemployment faster. Thus the frequency of unemployment occurrence (or the length of time unemployed) conveys information to non-testing/non-interviewing firms about a given worker s productivity, and these firms may free-ride by conditioning hiring on workers unemployment histories. Lockwood (1991) couches his model in terms of unemployment duration, rather than unemployment occurrence. But there is no obvious reason why the arguments should not carry through with unemployment incidence, as outlined above. The signalling hypothesis might work differently for young and more mature men for the following reason. To the extent that job-shopping is an acceptable form of behaviour for the young, who move from job to job until they find a suitable match, then firms may be less likely to use young men s past unemployment history as a signal to the same extent as for older workers. Why should a reduction in worker quality lead to higher unemployment, rather than lower wages? Blanchard and Diamond (1994) assume firms use unemployment history to rank workers in order of quality. This effectively confers greater bargaining power on the more desirable workers. 18 This result remains when controlling for new entrants to the labour market, suggesting that unemployment persistence among young men is not just reflecting the high levels of job mobility typically found among school leavers.

19 18 Wages are higher with ranking because the currently employed and the more recently unemployed have a better chance of being re-employed than the more longer-term unemployed. The skills depreciation rationale for state dependence in unemployment occurrence might also be extended to allow for age effects. Suppose the vintage of human capital is proxied by age. Thus younger workers will be characterised by a more recent vintage of human capital than older workers, but will have lower levels of on-the-job training acquired through labour market experience. However, in a period of rapid obsolescence of skills, the vintage effect may be more important than the experience effect. For this reason, the state dependence effect may be stronger for older than younger workers, and we find this to be the case. Of course, it is also possible that age differences in the magnitude of state dependence effects are a result of the new flexible labour market of the 1990s, with mature workers being unemployed (perhaps for the first time) and then struggling to regain a foothold in employment by taking temporary or short-term jobs. Indeed, our finding of age effects in state dependence contrasts with earlier comparative cross-sectional findings for Britain, where recurrent unemployment was more prevalent among the young, and long-term unemployment more prevalent among the old (see for example Pissarides and Wadsworth, 1992, who use crosssectional Labour Force Survey data to examine recurrent unemployment among men and women in 1979 and 1986). Labour market tightness: To pick up possible demand-side effects, we matched in data on the unemployment-to-vacancy (U-V) ratio, defined as the unemployment stock over the vacancy stock in an individual s

20 19 travel-to-work area. 19. This variable is an inverse measure of labour market tightness. Weak local labour demand (proxied by a high U-V ratio) is likely to be associated with higher job destruction rates and lower job creation rates, and therefore a greater individual probability of being unemployed. Our estimates show that the U-V rate has a significant and positive impact for men aged 25 and over: the unemployment probability for these men moves procyclically. While the estimated coefficient is also positive for young men and very similar to that of men aged 25 and over, the effect is not statistically significant. Time dummies were also included to pick up economy-wide trends over the period , but these have an insignificant impact on unemployment incidence. Observed individual heterogeneity: The remainder of the explanatory variables have been separated into two broad categories: personal attributes and highest educational qualifications. Turning initially to the personal attributes, men 45 years and over are more likely to experience unemployment than the excluded category of under 25s, in variants 1 and 2 of the model. For variant 3, this variable is insignificant. For all three variants, home owner-occupation has a significantly negative effect (albeit only at the 10% level in variants 1 and 2) on the probability of being unemployed (where the base is individuals renting privately). The negative association is as expected, since owner-occupation is typically associated with a long-term financial commitment that may induce individuals to search intensively for work and to work hard when in employment thereby increasing the job retention rate. Of course, owner occupation may also measure accumulated wealth, but any wealth effect is not apparent here. 20 Our results contrast with 19 These are obtained from the National On-line Manpower Information Service (NOMIS). Since the date of interview within each wave varied across individuals, the NOMIS data have been matched in at the appropriate month. For example, if an individual was interviewed in November, that individual had matched in the November unemployment and vacancy data for his local area. 20 An analogous result was found by Booth, Jenkins and Garcia-Serrano (1997) in a study of labour force participation of men and women in 1990s Britain.

21 20 those of Oswald (1996), who uses more aggregated data to estimate a significantly positive relationship between unemployment and owner-occupation. The unemployment probability for local authority tenants is not significantly different to that for the base of private tenants. 21 Bad health may reduce search intensity, or signal lower productivity to employers and thereby lower job-offer arrival and retention rates. Employers may be less willing to hire individuals if the scope of possible employment is reduced. We find that a health condition that limits the type and/or amount of work possible increases the probability of unemployment. 22 Ethnicity and country of birth have no significant effects on the unemployment probability. The insignificance of ethnicity is particularly interesting, since in popular discussion it has often been held that these variables are important correlates of unemployment. 23 It is interesting that none of the family variables were found to have a significant effect on the unemployment probability. More highly educated men are likely to face more job openings than less educated men, either through being more attractive to potential employers or because they search more effectively, and are thus less likely to be unemployed. The coefficients on the highest educational qualification variables confirm that more highly educated men are less likely to experience unemployment (although the impact of A-Levels and other qualifications below O-Levels is insignificant), relative to the base of no qualifications. This may be because voluntary unemployment is not necessary for the highly qualified, as higher-status and professional jobs are typically advertised in journals and newspapers, and job search activities 21 Other studies have suggested that local authority housing reduces labour mobility across local authority boundaries, thereby preventing workers moving to find new jobs (Nickell, 1980; Hughes and McCormick, 1981). 22 It could be argued that health and housing tenure are endogenous, in that unemployment is likely to have adverse effects on health and the unemployed are less likely to be offered mortgages etc. However the estimates displayed in Table 3 are robust to omitting these variables. 23 These ethnicity variables remain insignificant even when dropping the education and housing tenure measures.

22 21 can be efficiently undertaken while still employed. Also, the more highly educated and skilled are less likely to be laid off, since these workers are more expensive to hire and fire and have higher productivity. Now consider the reduced form equation for the initial conditions, reported in the first column of estimates in Table 3, for variants 1 and 2. The variables chosen for the initial conditions equation include the set of pre-determined attributes at Wave 1, plus some presample information about parental background, school type, characteristics of the first job, and whether or not the first labour market experience was unemployment. Interactions of education with age were also included. The estimated effects of many of the variables are similar to those of the random effects probit. In particular, men aged 45 or over are more likely to be unemployed, and a health condition limiting the type and/or amount of work possible increases the probability of unemployment. There is a significant positive association between local authority rental tenancy and unemployment (albeit only at 10% level), and an insignificant association between owner-occupation and unemployment. Marriage by 1991 does not significantly affect the 1991 unemployment probability, but the presence of children under the age of five increases the unemployment probability. Holding qualifications of any sort reduces the probability of unemployment. We also allowed for interactions of education with age dummies, and find that ceteris paribus effects of various qualifications on the unemployment probability are larger for younger men than older men. For example, for a man who is aged 25 and over and who also has a degree, the estimated coefficient is given by ( ) relative to someone of the same age without any qualification. But if the man had been under 25, this education effect is estimated as There are also some interesting cohort effects: men whose first job was manual and who entered the labour market

23 22 pre-1960, or after 1980, have a significantly higher probability of being unemployed in 1991, relative to the base of non-manual men who entered the labour market prior to Predicted probabilities In this section we calculate various unemployment probabilities from variant 1 of our estimated models and compare these to the raw aggregate probabilities, in order to see how much of the estimated unemployment probabilities are attributed to pure state dependence. The calculations, carried out following Chamberlain (1984), are presented in Table The first panel in Table 1 records the raw aggregate probabilities, and the second and the third panels record the predicted probabilities as calculated from the base model (variant 1). Since the distribution of characteristics amongst the two age sub-samples can be very different, all calculations presented below are based on averages over the different age sub-samples. For panel 2 calculations, we take each individual and calculate the predicted probabilities conditional, first on unemployment in the previous period, and secondly on employment in the previous period, and then average over the separate age sub-samples. These calculations therefore keep characteristics constant. Our calculations indicate (for a randomly chosen individual from a particular age group, conditional on the previous labour market status) the probability of observing this individual in unemployment in the current time period. Since each calculation is carried out for each individual by changing the previous labour market status, the difference between rows (6) and (7) probabilities gives the contribution due to true state dependence, and is given in row (8). The main points to note about the predicted probabilities reported in Table 4 are the following. 25 First we find that, for young men, only about 20% of observed persistence in the 24 See Arulampalam and Booth (1997) for similar calculations with more details. 25 Note that the raw data in Panel 1 show that, for young men, there is a fall in the probabilty of being unemployed in wave 5 conditional on being unemployed in Wave 4. Yet the conditional probabilities in previous waves are not all that dissimilar. We believe that the reason for this is likely to be the fact that some of these workers may have left the sample.

24 23 unemployment probability is accounted for by state dependence (calculated by expressing row (8) as a percentage of row (4) and averaged over the 4 years). However, for more mature men, roughly 38% of observed persistence in the unemployment probability is accounted for by state dependence. This finding has important implications for policy. It suggests that policies aiming to reduce unemployment occurrence that are targeted at more mature unemployed men will have a greater impact on the natural rate of unemployment than policies targeted at younger workers. Nonetheless, the magnitude of the state dependence result is large for both age groups, suggesting policy action for all age groups to reduce the natural rate. 6. Conclusions This paper provides some answers to the important question about whether or not there is state dependence in unemployment occurrence by estimating binary panel data models of unemployment incidence that also control for unobserved individual heterogeneity. Our results suggest that strong state dependence effects exist with respect to previous unemployment incidence, especially for more mature men (defined as those aged 25 and over in 1991). For young men, we find that less than one quarter of observed persistence in the unemployment probability is accounted for by state dependence, as compared with roughly 40% for more mature men. This finding is consistent with the scarring theory of unemployment - an individual s previous unemployment experience has implications for his future labour market behaviour. This may be because of depreciation of human capital, or because employers use an individual s previous labour market history as a signal of productivity, or because unemployed workers are more likely to accept low quality jobs characterised by high rates of job destruction. We find that local labour market conditions appear to have little effect for the unemployment probability of young men, which suggests that unemployment among the young

25 24 is independent of the business cycle. The unemployment probability of men aged 25 and over is increasing with the local unemployment to vacancy rate. Age, health and qualifications emerge as statistically significant determinants of unemployment incidence. The finding that previous unemployment experience increases the probability of current unemployment has important implications for policy, since it suggests that there is scope for intervention to alter the equilibrium, or natural rate, of unemployment. Our evidence suggests that policies reducing short-term unemployment incidence for adult British men will have longer run effects by reducing the natural rate. Moreover, to the extent that high rates of labour turnover increase the risk of experiencing unemployment, then employment instability could have far-reaching implications for the natural rate of unemployment and for aggregate unemployment persistence. Our findings also suggest that policies aimed at preventing unemployment, such as promoting education and training, would have a long-term impact on aggregate unemployment. Our results, taken in conjunction with the findings of the few other micro-econometric studies in this area, are suggestive as to why there are observed differences in the natural rate of unemployment between the US and Europe. Corcoran and Hill (1980) and Heckman and Borjas (1980) find no evidence of state dependence in unemployment occurrence for US men, but their evidence is for the late 1960s to early 1970s. In contrast, Narendranathan and Elias (1993) find strong evidence of state dependence in unemployment occurrence for a cohort of British men aged 23 in 1981, as do McCulloch and Dex (1996) for adult men in Britain in the 1990s. There is also strong evidence of state dependence in Germany for adult men in the 1980s (Flaig et al, 1993; Muhleisen and Zimmermann, 1994). However, it is clear that much more research is required in order to compare inter-country differences in state dependence in unemployment occurrence. It is necessary to have evidence on the extent of state dependence