Gender differences in low pay labour mobility and the national minimum wage

! Oxford University Press 2008 All rights reserved Oxford Economic Papers 61 (2009), i122 i146 i122 doi:10.1093/oep/gpn045 Gender differences in low pay labour mobility and the national minimum wage By Euan Phimister* and Ioannis Theodossiouy *Centre for European Labour Market Research, University of Aberdeen; e-mail: e.phimister@abdn.ac.uk ycentre for European Labour Market Research, Business School, University of Aberdeen, Aberdeen AB24 3QY; e-mail: theod@abdn.ac.uk This paper examines gender differences in the duration of low pay employment spells prior to and after the introduction in the National Minimum Wage in 1999. The results suggest that the dynamics out of low pay differ by gender and that these differences change after 1999. These differences are driven by the differing impact of a number of covariates such as age and education on the baseline hazards. Overall, the effect of many covariates on expected duration is often less in absolute terms for women than men, although such differences frequently decline after 1999. At mean values, gender differences in expected duration of low pay effectively disappear and gender differences in the exit probability to high pay decline after 1999. However, for individuals with characteristics most associated with long periods of low pay, the high pay exit probability is substantially lower after 1999 for women than for men. JEL classification: J60. 1. Introduction The overall incidence of low pay in the UK remains significantly higher for women than men (Metcalf, 1999; Low Pay Commission, 2008). Empirical work shows that being a woman significantly increases the probability of remaining in the low pay category generally (Gregory and Elias, 1994; Sloane and Theodossiou, 1996; Cappellari, 2007). Gender differentials in the time spent in low pay are of interest as low pay is seen as a more serious problem if prime-age workers become trapped in low paid jobs than if the experience of low pay is a transitory phenomenon (Layard et al., 1991). While the link between low pay and poverty is relatively weak, part of the present UK government s strategy to combat poverty and inequality is to encourage individuals into work using policies to make work pay. However, as Stewart and Swaffield (1999) argue the success of this strategy depends on whether the starting jobs do not simply offer semi-permanent low pay, or a high probability of exit out

e. phimister and i. theodossiou i123 of employment. Therefore, gender differentials in time spent in low-paid jobs and the probability of progressing to a higher-paid job or other low-pay exit types could differentially affect the success of such policies for men and women in the UK. Finally, as upward wage mobility is lower for women, one might expect this to have a significant impact upon the gender earnings gap over time. Particular efficiency and equity concerns may also arise if the lower attachment to the labour force typically shown by women workers affects the time spent in low pay and their probability of gaining a higher paid job, e.g. if employers offer fewer training opportunities to women as a result. The introduction of the National Minimum Wage (NMW) in 1999 aimed to directly combat the worst excesses of low pay (Labour Party, 1997). Its introduction fundamentally changed the institutional arrangements for those on low pay. From a cross sectional perspective women as a group emerge as one of the primary direct beneficiaries of the NMW (Low Pay Commission, 2008), although a female low pay worker is more likely to be a recipient of the minimum wage for longer than an equivalent male worker (Jones et al., 2004; Bryan and Taylor, 2006). Evidence also suggests that the introduction of the NMW is associated with detrimental changes in low pay dynamics particularly for women (Stewart, 2002). Hence, it is natural to explore how the gender differentials in time spent in low-paid jobs and the probability of progressing to a higher-paid job might have changed since the introduction of the NMW. The aim of this paper is to examine gender differences in the experience of low pay. Although this study cannot claim any identification of casual mechanisms, it investigates whether such gender differences have changed after the introduction of the NMW. There is a substantial body of previous research on the dynamics of low pay, the impact of gender, and other observed and unobserved characteristics on the transitions into and out of low pay (Gregory and Elias, 1994; Sloane and Theodossiou, 1996; Stewart and Swaffield, 1999; Uhlendorff, 2006; Cappellari, 2007) and the interrelationship between low pay and unemployment dynamics (Stewart, 2007). However, there has been little research which explores the extent of gender differences both in terms of low pay duration and of where individuals move to at the end of a low pay spell, i.e. allowing for exits to high pay from other possible destinations (e.g. inactivity), and in duration of low pay. Both duration and the type of exit are potentially important sources of gender differences. As Royalty (1998) points out, explanations of wage gaps depend not just on differences in expected duration of a job but also on the destination state to which a worker exits. Thus, if job to job turnover is associated with pay increases while job to out of labour market or to unemployment resulting from family responsibilities or redundancies is not, one would expect differences in patterns and time to exit from low pay to be important in understanding gender differences in low pay. This study investigates the duration and exit type of low pay employment spells for men and women. Data on low pay employment spells from the British Household Panel Survey are constructed and used to estimate a competing risks

i124 gender, low pay mobility, and the minimum wage model (Lancaster, 1990) of low pay exits to three possible destinations, namely, to a high pay, to unemployment and to inactivity. Gender differences are identified by estimating separate specifications for men and women. The extent to which changes have occurred since the introduction in the NMW are illustrated by considering two separate inflow samples to low pay, covering the period 1992 98 and 1999 2005 respectively. Failing to account for the effects of unobserved personal characteristics which decrease (increase) the re-employment probabilities may bias the results in favour of negative (positive) duration dependence (Lancaster, 1990). This paper controls for unobserved characteristics by estimating models which allow for unobserved heterogeneity. The recent literature considering transitions both into and out of low pay has emphasized the need to control for selectivity associated with initial conditions (Uhlendorff, 2006; Cappellari, 2007; Cappellari and Jenkins, 2008). The approach used here considers the gender differences in the experience of individuals who have just begun a low pay spell and does not consider the process of entry into low pay. It also does not attempt to control for the initial conditions problem created by the differences in the joint distribution of observed and unobserved characteristics within the labour force and for those within the inflow sample (Meghir and Whitehouse, 1997). 1 Hence, the results obtained by this study relate to the specific experiences of those becoming low paid in the periods before and after the introduction of the National Minimum Wage (Chesher and Lancaster, 1983). 2. Low pay; a brief literature review Recent evidence in the UK suggests that wage growth for low pay workers is weak unless they experience substantial upward occupational mobility (Lucifora et al., 2005). This is consistent with the evidence which shows that relatively few people from the bottom of the earnings distribution escape into the top half (Gosling et al., 1997). Furthermore, many of those who exit low pay employment do not go to better-paid jobs but leave employment altogether (Sloane and Theodossiou, 1996; Uhlendorf, 2006). Thus, movement out of low pay is often likely to imply mobility out of employment rather than movement up the earnings distribution. Indeed, Stewart (2007) shows that being in low pay in the current period increases the probability of being in unemployment the next period by as much as a period of current unemployment. Thus, the combination of the fact that low-wage workers are more likely to move out of work and that those out of work are more likely to enter low-wage jobs produces a strong relationship between low pay employment and joblessness. The above emphasizes that any study of low pay dynamics must allow for both earnings mobility and the unemployment propensity for the individual worker.... 1 Unless strong assumptions are made concerning independence between unobserved and observed characteristics in the labour force.

e. phimister and i. theodossiou i125 This is particularly important in this context as gender differences in pay, labour market attachment, and job turnover are well documented (Altonji and Blank, 1999). For example, the loss of human capital and wage gains predicted by job matching due to discontinuous labour force participation suggest that the likelihood of an exit from low pay to inactivity or unemployment is likely to differ for males and females. Factors affecting wage growth, job turnover, and employment status change clearly affect the time spent in low pay and the ultimate exit destination. Unsurprisingly, previous research shows that not only gender but also human capital, job tenure and experience, firm size, trade union status, and part-time working, are all important in low pay mobility (Gregory and Elias, 1994; Sloane and Theodossiou, 1996; Gosling et al., 1997; Stewart and Swaffield, 1999; Uhlendorf, 2006; Cappellari, 2007). Given the observed gender differences in wages and labour market behaviour, it is also clearly important to consider whether the impact of such factors on low pay duration and exit type is different for men and women. Evidence also suggests that workers prospects of moving up the earnings ladder worsen as the duration of a low-paid employment spell lengthens. Hence, there is a scarring effect of low pay (Stewart and Swaffield, 1999). This implies that being in low pay employment itself traps people in low pay. There might also be gender differences in scarring or duration dependence effects if the more interrupted nature of female job histories compared to men is used by employers as a signal of female worker quality. The introduction of the UK National Minimum Wage in 1999 fundamentally changed the institutional arrangements for those on low pay. Its introduction has significantly compressed the lower end of the earnings distribution, but there appears little evidence suggesting that it had any effects further up the earnings distribution (Dickens and Manning, 2004; Metcalf, 2004; Lam et al., 2006). In terms of employment, there is no evidence the national minimum wage has reduced the overall adult employment rates or raised unemployment in the UK (Stewart, 2004a,b; Dickens and Draca, 2005; Metcalf, 2008). Furthermore, it has not affected the solvency of the low wage firms although their profitability decreased (Machin et al., 2005). However, there is some evidence suggesting that the minimum wage introduced negative employment effects for 16 to 17 years old (Frayne and Goodman, 2004; Neathey et al., 2004). Although the qualitative evidence on how firms reacted to the minimum wage suggest that very few firms reduced staffing levels (Grimshaw and Carroll, 2002; Cronin and Thewlis, 2004), some evidence does point to lower employment growth in sectors such as services (Galindo-Reuda and Pereira, 2004) and the care home industry (Machin and Wilson, 2004). Connolly and Gregory (2002) found no significant effect of the minimum wage on working hours for women. In contrast, Stewart and Swaffield (2004) did find that introduction of the minimum wage led to a reduction of working hours, mainly for men, although no such effects were observed for the 2000 or 2001 up-ratings.

i126 gender, low pay mobility, and the minimum wage Given the dramatic changes in the structure of the labour market for low paid workers associated with the introduction of the NMW, it is natural to explore whether low pay duration and exit type (and any gender differences) have changed post-1999. Evidence on overall low pay dynamics and in particular entry to low pay does indicate that the time of the introduction of the NMW is associated with a fall in the probability of a move into employment particularly for women (Stewart, 2002). At the same time cross-sectional evidence suggests that women in employment are one of the primary direct beneficiaries of the NMW (Low Pay Commission, 2008). The above discussion provides the general basis for the empirical work which follows. First, the need to model both low pay duration and type of exit is emphasized by the evidence on the gender differences in wages, job turnover, and labour market attachment, and the importance of the dynamics of earnings and employment status in low pay mobility. Second, previous research suggests a range of factors likely to affect the dynamics of low pay that may differ by gender. Third, there are also reasons why scarring effects or duration dependence may differ for men and women which should be taken into account. Finally, since the introduction of the minimum wage so dramatically changed the structure of the labour market for low pay workers, it is clearly important to explore evidence of any associated changes in the low pay duration and low pay exits. 3. Data The data are drawn from waves 1 to 15 of the British Household Panel Survey (BHPS). At each interview, respondents are asked for detailed information on their employment since the last interview. From this a complete sequence of labour market spells is constructed recorded to the nearest calendar month for all individuals with at least three consecutive interviews. A labour market spell is a job, a period of unemployment, a period of inactivity, or a period of self-employment. However, consistencies in the recorded spells may arise primarily from differences between what an individual recalls about their employment status at the previous interview and what is actually recorded at the previous interview. Following Upward (1999) these problems are reconciled by applying the principle that information recorded closest to any particular event is the most reliable. Information collected at each interview on personal and job characteristics, hours worked, and earnings is also used. Thus, the hourly wages are calculated by using information on usual gross monthly pay, regular work hours, and hours paid overtime per week. Then the hourly rate of pay is attributed to the current job for the current interview period. However, it is not possible to calculate wages for jobs that start and end between consecutive interviews. Hence, these spells are excluded from the analysis. As a result, the analysis is likely to underestimate the extent of low-paid jobs of very short duration. This also means that when employment is continuous, short high pay spells between interview dates are not captured, whereas any exit to a short spell of unemployment or to inactivity between interviews is captured as a low pay exit.

e. phimister and i. theodossiou i127 Despite these limitations, the methodology used in this study arguably improves upon previous studies of low pay which assume that individuals spend the entire time between interviews either in or out of low pay. Observations with missing values on hours worked, earnings, or other variables used in the analysis are also dropped from the sample. Finally, since the study concerns low paid employees, individuals whose low pay spell ended in self-employment are also excluded. 3.1 Measurement of low pay and summary statistics Low pay thresholds are often defined as a relative measure with reference to the median wage, although the actual value chosen varies considerably (Sloane and Theodossiou, 1996; Gosling et al., 1997; Stewart and Swaffield, 1999; Cappellari and Jenkins, 2008). This study uses a relative measure selecting what might be considered the upper bound of this type of threshold namely, the bottom third of the earnings distribution of the BHPS sample in each year. To provide some indication as to the robustness of the results to this assumption, where appropriate, the use of an alternative relative low pay definition is made, namely, a threshold equal to twothirds the median hourly wage for each year. From the constructed dataset, two inflow samples of low pay spells were drawn covering two six-year periods prior to and post the introduction of the National Minimum Wage. Specifically, low pay spells starting between September 1992 and September 1998 are included in the pre-nmw sample, and spells starting between September 1999 and September 2005 are included in the post-nmw sample. All continuing spells are treated as censored if they continued beyond the end cut-off points, i.e. September 1998 and September 2005 respectively. The generated samples consists of 1716 low pay spells for men and 2353 spells for women for the period 1992 98, and 1964 spells for men and 2896 spells for women for the 1999 2005 period. A useful way to describe these spells is the survivor function, S(j). This is the probability that a low-paid job lasts beyond month j, estimated by counting the number of jobs which end on or before j compared to the total number of jobs, where j indicates the elapsed number of months that an individual has been in a low-paid job. Figure 1 shows the estimates of S(j) for males and females separately for the two samples. In the 1992 98 sample, the proportion of low-paid jobs remaining at time j is always smaller for men which is consistent with the evidence that low-paid jobs are more persistent for women. This is supported statistically with the equality of the survival functions rejected at 10% significance using the log rank test (p-value 0.056). However, in the post NMW period, this effect is less discernible, and for this period the equality of the survivor functions cannot be rejected (p-value 0.185). Table 1 reports sample means for key characteristics of the low pay spells. The low pay spells are also distinguished by the three exits types, namely, to a higher paid job, unemployment, and inactivity. Consistent with the survivor function evidence, the overall and conditional mean durations are somewhat higher for the females in both pre- and post-nmw samples. Notably, while the conditional

i128 gender, low pay mobility, and the minimum wage Survival rate 0.50 0.75 1.00 1991 98 0.00 0.25 Survival rate 0.00 0.25 0.50 0.75 1.00 0 20 40 60 80 100 Month Men Women 1999 2005 0 20 40 60 80 100 Month Men Women Fig 1. Kaplan Meier survival functions by gender means increase post-nmw, the decline in the percentage of censored spells brings about a decline in the overall mean for both male and females. Differences in a number of the characteristics of the male and female samples are also evident from Table 1. In both samples a higher proportion of low paid women work in the public sector, in small firms and in skilled occupations. Unsurprisingly, higher proportions are also employed part-time and in the service sectors. There are few differences in terms of education or training but a higher proportion of low paid females have a spouse who works. They also have more children and are more likely to be in the older age groups compared to their male counterparts.

e. phimister and i. theodossiou i129 Table 1 Summary statistics 1992 1998 1999 2005...... Men Women Men Women Number of spells 1716 2353 1964 2896 Of which Higher-paid job 549 (32%) 702 (30%) 670 (34%) 834 (29%) Exit to Unemployment 259 (15%) 210 (9%) 239 (12%) 235 (8%) Inactivity 72 (4%) 307 (13%) 211 (11%) 588 (20%) Censored 836 (49%) 1,134 (48%) 844 (43%) 1,239 (43%) Conditional duration (months) Exit to Higher-paid job 14.3 13.8 18.9 16.4 Unemployment 10.2 14.2 17.1 15.7 Inactivity 13.6 15.1 13.2 18.7 Overall mean 37.2 39.6 24.7 27.2 Characteristics Public sector 0.125 0.212 0.128 0.262 Covered by trade union 0.245 0.252 0.278 0.301 Firm Size < than 25 employees 0.361 0.458 0.377 0.450 Firm Size 25 99 employees 0.258 0.230 0.270 0.236 Part-time 0.074 0.396 0.116 0.408 Manufacturing 0.289 0.122 0.251 0.081 Services 0.546 0.763 0.654 0.836 Managerial and professional 0.158 0.142 0.148 0.142 Skilled 0.455 0.638 0.453 0.573 O-level/A-levels 0.523 0.541 0.482 0.506 Nursing etc 0.226 0.204 0.284 0.270 Degree 0.125 0.103 0.138 0.136 Vocational training 0.344 0.369 0.338 0.362 On-job training last 12 months 0.110 0.106 0.362 0.403 Married -spouse has job 0.363 0.538 0.342 0.484 Married-spouse no job 0.142 0.082 0.124 0.045 Number of children 0.444 0.696 0.449 0.659 Age 25 or less 0.420 0.327 0.444 0.377 25 < Age < 36 0.294 0.311 0.244 0.243 35 < Age < 46 0.146 0.209 0.159 0.230 45 < Age < 56 0.108 0.141 0.094 0.118 Private rented 0.162 0.151 0.151 0.137 Public rented 0.220 0.224 0.187 0.170 4. Modelling earnings mobility A discrete time competing risks hazard model with three exit types is estimated, namely, to high pay employment, unemployment, and inactivity. However, estimates of the effect of individual characteristics on the exit hazard may be biased when unobserved differences or individual heterogeneity is not taken into account. In this study, the unobserved heterogeneity is incorporated by assuming a proportional hazard model with Gaussian Mixing (Lancaster, 1990; Andrews et al, 2002). As described below, to capture unobserved heterogeneity across individuals, this model includes an unobserved random component, assumed to be normally distributed.

i130 gender, low pay mobility, and the minimum wage s¼1 h rj r¼1 Assuming that each person exits low pay in an interval ½ j 1,jÞ to one of three states r = 1,..,3, the three latent variables represent the potential time in low pay with an exit of type r, i.e. T r, r = 1,.., 3. Low pay duration is then the random variable T where T ¼ minðt 1,T 2,T 3 Þ. 2 For any period j, the hazard to state r, h rj is the probability of an exit to state r during period j given that the low pay duration lasted to j-1. In each period j the overall low pay exit hazard is the sum of the hazards to all possible states, i.e. h j ¼ P3 h rj. Similarly, the survivor function is r¼1 defined as S j ¼ Qj 1 P3. The influence of observed covariates and unobserved heterogeneity are captured by modelling each exit type using a proportional hazard model with Gaussian mixing, i.e. h rj ðx 0 i,v iþ¼ h m rj0 1 d f þ df h f 0 rj0 v i exp 1 d f x i m r þ d f x 0 i f r ð1þ where h m rj0, hf rj0 are the male and female baseline hazards for the exit type r, v i is a random variable capturing unobserved heterogeneity such that u ¼ logðvþ is normally distributed, x i is the vector of covariates. The dummy variable d f equals one for females and hence both the baseline hazards and the impact of each characteristic are allowed to vary across the male and female samples. The vector of covariates x i contains the individual and job characteristics summarized in Table 1 plus regional and time dummy variables. Hence both sources of potential gender differences in low pay dynamics discussed in Section 2 are allowed for. If low pay duration dependence differs for men and women then this should be captured by differences in the underlying baseline hazards. If gender differences in factors affecting wages, job turnover, and labour market attachment bring about differences in low pay dynamics, this should be captured via differences in the impact of the covariates on the baseline hazard. As the overall survival function depends on all the transition intensities or exit hazards, the estimated coefficients, k r, provides no information on overall effects of a covariate. To compute these, the probability of exit via type r, r and the expected low pay duration given an exit of type r, E r are required. These can be shown to be: r ¼ X 1 j¼1 h rjs j 1 E r ¼ 1 r X 1 j¼1 jh rjs j 1 ð2þ ð3þ (Lancaster, 1990; Andrews et al., 2002) and the overall expected low pay duration is P 3 r¼1 re r. Using the above formulae, the effect of any covariate on exit probabilities, conditional expected waiting times by exit type and unconditional expected waiting time can be approximated numerically.... 2 Censored exits are treated in a symmetric manner in this framework.

e. phimister and i. theodossiou i131 In line with Lindeboom and Van den Berg (1994) it is assumed that the unobserved heterogeneity is independent across exit types and therefore the model parameters can be estimated separately by exit type. As discussed in the introduction, the methodology used does not control for all potential selection issues if the unobserved heterogeneity components are correlated with selection into the low pay state. The sampling scheme produces a random sample for the specific populations of those falling into low pay for the two periods. This model is then estimated using standard software once the data is reorganized into sequential binary response form (Han and Hausman, 1990). A piecewise constant baseline hazard is estimated, with the baseline hazards allowed to vary across 16 different intervals, namely, one for each of the first 12 months of a low pay spell, then constant within six month intervals until month 24 of the spell, then for 12-month intervals until month 60, and constant thereafter. 5. Econometric results Since the estimated k r convey no information on the overall impact of each covariate, these estimates are not reported in the text. 3 However, to provide an indication of the validity of the estimations and to explore the sources of any statistical differences, Table 2 reports a number of summary statistics from the estimations. 4 Overall, the regression evaluation measures provide some validation for the modelling approach used. For all regressions, the dummies and the covariates are each jointly significant. In the majority of cases, the test of no duration dependence is rejected, (female exits to unemployment or inactivity in the earlier sample is an exception to this). The reported estimate of u 2 provides an indication as to whether unobserved heterogeneity is important. The results show little evidence of unobserved heterogeneity for any type of exit and one possible explanation might be that the assumed form of unobserved heterogeneity is too restrictive. To explore this, a version of the model is estimated using the mass-point approach with two points of support for each exit type allowing for correlations between the unobserved heterogeneity for each exit type (Heckman and Singer, 1984). The results of this approach show no evidence of unobserved heterogeneity. While somewhat puzzling, other studies which estimate competing risks models show that when a flexible baseline hazard function is estimated, unobserved heterogeneity effects are often unimportant (Dolton and van der Klaauw, 1999; Andrews et al., 2002). 5... 3 They are available from the authors on request. 4 The models are estimated with a common set of regional and time dummy variables (as this restriction could not be rejected statistically). To partially control for the upgrading of the minimum wage, the time dummy variables are not defined on calendar years but run from 1 October to 30 September. 5 Previous studies find that unobserved heterogeneity is often important when a parametric baseline hazard is used, e.g. Weibull, but these effects disappear when baseline hazards are more flexibly estimated. This perhaps suggests that unobserved heterogeneity effects are captured via differences in the

i132 gender, low pay mobility, and the minimum wage Table 2 Discrete competing risks hazard model: estimation summary 1992 1998 d.f. Higher-paid jobs Unemployment Inactivity Variance 0.001 (>0.99) 0.37 (0.27) 0.37 (0.28) Log likelihood 5811.2 2538.3 2122.5 Wald tests men Duration dependence 15 61.2 (<0.001) 57.3 (<0.001) 23.9 (0.07) All covariates 23 227.6 (<0.001) 140.7 (<0.001) 86.3 (<0.001) Wald tests women Duration dependence 15 83.6 (<0.001) 19.4 (0.19) 21.4 (0.12) All covariates 23 335.4 (<0.001) 149.4 (<0.001) 182.0 (<0.001) Equality male-female coefficients All coefficients 39 40.1 (0.41) 95.3 (<0.001) 63.8 (<0.001) Baseline hazard dummies 16 18.7 (0.28) 21.4 (0.17) 16.9 (0.39) All covariates coefficients 23 23.1 (0.45) 57.2 (<0.001) 34.1 (0.06) Education coefficients 3 2.8 (0.42) 7.8 (0.05) 6.3 (0.10) Age Coefficients 4 4.5 (0.34) 4.7 (0.32) 8.1 (0.09) Housing coefficients 2 1.7 (0.43) 2.3 (0.32) 1.9 (0.38) 1999 2005 d.f. Higher-paid jobs Unemployment Inactivity Variance 0.3 (0.31) 0.001 (>0.99) 0.001 (>0.99) Log likelihood 7145.0 2790.0 4192.0 Wald tests men Duration dependence 15 45.3 (<0.001) 36.6 (0.001) 48.1 (<0.001) All covariates 23 165.5 (<0.001) 85.6 (<0.001) 201.8 (<0.001) Women Duration dependence 15 84.0 (<0.001) 27.6 (0.02) 71.5 (<0.001) All covariates 23 304.9 (<0.001) 94.1 (<0.001) 253.9 (<0.001) Equality male-female coefficients All coefficients 39 62.7 (0.01) 41.2 (0.37) 83.2 (<0.001) Baseline hazard dummies 16 11.3 (0.79) 10.4 (0.85) 16.9 (0.39) All covariates coefficients 23 45.5 (0.0) 21.3 (0.56) 55.5 (<0.001) Education coefficients 3 0.6 (0.90) 0.6 (0.89) 10.3 (0.02) Age coefficients 4 12.6 (0.01) 6.3 (0.18) 8.6 (0.07) Housing coefficients 2 1.6 (0.45) 2.7 (0.26) 0.7 (0.72) All estimations included separate baseline dummies and job, individual and household characteristics reported in Table 1 for men and women plus a common set regional and time dummies. Full results available on request. In the statistical modelling, two potential sources of male-female differences in exiting low pay are allowed for. First, there are differences which may occur in the underlying baseline hazards for men and women. Second, the impact of observed characteristics on the baseline hazards may differ. Table 2 reports the results of a series of the Wald tests exploring whether these types of differences are statistically significant. The tests do not suggest gender differences in duration dependence,... flexible baseline hazards. Alternatively, if unobserved heterogeneity is observed in fully parametric studies this may simply reflect the restrictive nature of the parametric assumptions rather than identifying a true individual level effect.

e. phimister and i. theodossiou i133 with the hypothesis of no gender difference not rejected in any sample or exit type. However, there is evidence of significant gender differences in the impact of the covariates on the baseline hazards, with the hypothesis of no difference rejected (at least at 10%) in four out of the six cases. Further testing for gender differences in the impact of particular subsets of variables suggests that education and age effects appear to be the main driving force for the overall statistical male-female differences in exiting low pay. 5.1 Marginal effects Table 3 reports approximate marginal effects for each characteristic, where the estimated impact is allowed to vary by gender. An approximate bootstrapped standard error is also reported (below each coefficient). 6 These provide an evaluation of the overall impact of male-female differences in the estimated coefficients for males and females in the two samples. They also show the different possible combinations of marginal effects on expected duration in low pay and the probability of a high pay exit. For example, consider the impact of being in part-time employment reported in the first part of the table. Before the introduction of minimum wage, the marginal effect of being in part-time employment decreases the probability of an exit to a higher paid job by 0.25 for women (0.32 for men). This can be compared with the period after the introduction of the minimum wage where the relevant marginal effects are 0.27 for women (and 0.29 for men). Being a female employed in a parttime job decreased the overall expected duration of the low pay spell by 7.6 months before the introduction of the minimum wage (compared to 22 months for men), and by 2.7 months after its introduction (compared to 12 months for men). Comparing the coefficients with the standard errors suggest that the marginal effects associated with the part time variable are generally well determined. As is described in more detail below, it is difficult to establish clear patterns about the direction of gender differences from the reported marginal effects. However, as in the part-time case, the marginal impacts of the covariates on expected duration are often less in absolute terms for women than men. This difference tends to decline after 1999 when the impact of all covariates on expected duration tends to fall. Characterizing an overall pattern in the marginal impact of the covariates on the exit probabilities is not so clear cut, as little systematic pattern is observed in either the gender differences or the changes between the two periods. First, consider the job characteristics marginal effects. Bell and Pitt (1998) show that the decline in trade union membership can explain part of the widening in the distribution of earnings in the UK over the 1980s. The results here suggest that union coverage may reduce mobility out of low pay perhaps due to employment contract arrangements in unionized firms. Trade union coverage increase the... 6 Due to the time required for these computations bootstrapped standard errors are only available for the model without unobserved heterogeneity. However, these are consistent given the overall results that unobserved heterogeneity is not statistically significant.

i134 gender, low pay mobility, and the minimum wage Table 3 Marginal effects: job characteristics 1992 1998... Exit probability... Higherpaid Unem- Inactivity Expected job ployment duration 1999 2005... Exit probability... Higherpaid Unem- Inactivity Expected job ployment duration Public sector Men 0.003 0.0004 0.002 6.1 0.08 0.01 0.07 5.9 (0.10) (0.04) (0.11) (7.0) (0.06) (0.04) (0.04) (3.1) Women 0.02 0.04 0.02 0.3 0.06 0.07 0.01 0.6 (0.04) (0.04) (0.04) (3.4) (0.03) (0.02) (0.03) (1.9) Trade union coverage Men 0.004 0.03 0.04 79.8 0.04 0.03 0.07 35.6 (0.13) (0.06) (0.16) (30.6) (0.05) (0.04) (0.06) (8.8) Women 0.05 0.00 0.05 60.3 0.08 0.04 0.04 36.9 (0.06) (0.04) (0.05) (14.2) (0.04) (0.03) (0.04) (5.1) Firm size less than 25 employees Men 0.01 0.01 0.01 18.3 0.003 0.03 0.03 8.7 (0.06) (0.04) (0.08) (7.2) (0.03) (0.03) (0.03) (3.2) Women 0.10 0.03 0.07 6.4 0.02 0.01 0.01 7.8 (0.04) (0.03) (0.03) (3.2) (0.03) (0.02) (0.03) (2.1) Firm size 25 99 employees Men 0.04 0.01 0.03 11.7 0.02 0.03 0.05 11.1 (0.07) (0.04) (0.09) (5.9) (0.04) (0.03) (0.04) (4.1) Women 0.07 0.01 0.08 3.5 0.03 0.01 0.02 4.3 (0.05) (0.03) (0.04) (3.8) (0.03) (0.02) (0.03) (2.2) Part-time Men 0.32 0.24 0.08 22.4 0.29 0.01 0.28 12.1 (0.11) (0.09) (0.14) (7.7) (0.04) (0.04) (0.05) (2.9) Women 0.25 0.01 0.23 7.6 0.27 0.001 0.27 2.7 (0.04) (0.03) (0.04) (2.9) (0.03) (0.02) (0.03) (1.5) Manufacturing Men 0.05 0.10 0.05 22.2 0.06 0.07 0.01 5.0 (0.10) (0.06) (0.11) (8.3) (0.06) (0.06) (0.05) (4.0) Women 0.05 0.09 0.04 21.0 0.29 0.06 0.23 9.0 (0.08) (0.07) (0.06) (5.1) (0.06) (0.05) (0.08) (2.7) Services Men 0.07 0.02 0.05 29.9 0.0002 0.002 0.002 7.1 (0.08) (0.05) (0.11) (12.2) (0.06) (0.05) (0.04) (5.2) Women 0.02 0.003 0.01 30.5 0.15 0.003 0.15 11.2 (0.07) (0.06) (0.05) (9.3) (0.05) (0.03) (0.04) (3.8) Managerial and professional Men 0.04 0.06 0.02 2.5 0.03 0.05 0.02 0.7 (0.07) (0.04) (0.08) (3.5) (0.05) (0.04) (0.04) (3.6) Women 0.22 0.08 0.14 4.0 0.04 0.03 0.01 3.2 (0.04) (0.04) (0.03) (4.5) (0.05) (0.03) (0.04) (2.3) Skilled Men 0.09 0.07 0.02 3.5 0.02 0.02 0.03 1.1 (0.05) (0.04) (0.07) (3.7) (0.04) (0.03) (0.02) (2.4) Women 0.12 0.02 0.10 1.7 0.04 0.03 0.01 1.4 (0.04) (0.03) (0.04) (3.4) (0.03) (0.02) (0.03) (2.0) (continued)

Table 3 Marginal effects: individual characteristics e. phimister and i. theodossiou i135 1992 1998... 1999 2005... Exit probability... Exit probability... Higherpaid Unem- Inactivity Expected Higher- Unem- Inactivity Expected job ployment duration paid job ployment duration O-level/A-levels Men 0.13 0.16 0.03 11.4 0.003 0.03 0.03 3.0 (0.13) (0.05) (0.14) (6.0) (0.06) (0.05) (0.05) (4.3) Women 0.09 0.04 0.05 5.3 0.05 0.04 0.00 1.0 (0.06) (0.04) (0.04) (4.4) (0.06) (0.04) (0.05) (2.9) Nursing etc. Men 0.04 0.08 0.04 20.0 0.02 0.04 0.06 8.8 (0.13) (0.04) (0.15) (7.1) (0.07) (0.05) (0.06) (4.1) Women 0.10 0.03 0.08 7.8 0.10 0.02 0.08 0.9 (0.06) (0.05) (0.05) (4.3) (0.06) (0.04) (0.05) (3.4) Degree Men 0.18 0.18 0.0001 11.5 0.05 0.08 0.03 12.3 (0.14) (0.04) (0.15) (5.0) (0.07) (0.04) (0.07) (3.4) Women 0.13 0.03 0.10 13.2 0.19 0.05 0.14 4.8 (0.06) (0.05) (0.05) (4.0) (0.06) (0.03) (0.05) (2.7) Vocational training Men 0.01 0.04 0.03 5.4 0.02 0.01 0.01 0.5 (0.06) (0.03) (0.07) (4.8) (0.04) (0.03) (0.03) (2.5) Women 0.02 0.02 0.03 2.4 0.02 0.03 0.05 0.7 (0.03) (0.03) (0.02) (2.6) (0.03) (0.02) (0.02) (1.5) On-job training last 12 months Men 0.07 0.02 0.05 26.7 0.03 0.03 0.001 6.1 (0.13) (0.06) (0.16) (14.3) (0.03) (0.03) (0.03) (2.2) Women 0.10 0.13 0.03 4.1 0.04 0.01 0.05 5.3 (0.06) (0.07) (0.05) (8.0) (0.03) (0.02) (0.03) (1.4) Age 25 or less Men 0.01 0.06 0.07 10.2 0.06 0.03 0.09 7.6 (0.14) (0.11) (0.11) (7.9) (0.11) (0.11) (0.05) (7.1) Women 0.06 0.06 0.00 15.5 0.05 0.04 0.02 6.8 (0.14) (0.08) (0.10) (10.8) (0.11) (0.07) (0.10) (5.2) 25 < Age < 36 Men 0.06 0.06 0.12 5.0 0.17 0.003 0.17 7.7 (0.17) (0.13) (0.17) (9.4) (0.11) (0.11) (0.05) (5.4) Women 0.11 0.05 0.05 20.0 0.08 0.03 0.05 3.3 (0.13) (0.08) (0.10) (11.9) (0.10) (0.07) (0.10) (5.3) 35 < Age < 46 Men 0.07 0.02 0.08 6.1 0.10 0.05 0.15 8.5 (0.18) (0.13) (0.16) (9.7) (0.12) (0.12) (0.04) (5.8) Women 0.12 0.03 0.09 35.6 0.18 0.05 0.13 3.7 (0.13) (0.09) (0.08) (17.6) (0.10) (0.07) (0.09) (5.7) 45 < Age < 56 Men 0.01 0.06 0.06 3.8 0.06 0.06 0.13 8.3 (0.15) (0.13) (0.12) (10.1) (0.12) (0.12) (0.04) (4.8) Women 0.15 0.09 0.07 35.5 0.08 0.03 0.11 5.6 (0.11) (0.06) (0.08) (15.8) (0.11) (0.09) (0.08) (7.0) (continued)

i136 gender, low pay mobility, and the minimum wage Table 3 Marginal effects: household and other characteristics 1992 1998... 1999 2005... Exit probability... Exit probability... Higherpaid Unem- Inactivity Expected Higher- Unem- Inactivity Expected job ployment duration paid job ployment duration Married -spouse has job Men 0.09 0.10 0.01 11.9 0.14 0.07 0.07 10.7 (0.07) (0.04) (0.08) (5.30) (0.04) (0.03) (0.03) (3.81) Women 0.14 0.12 0.02 17.4 0.02 0.06 0.03 6.5 (0.04) (0.03) (0.03) (3.40) (0.02) (0.02) (0.03) (1.75) Married-spouse no job Men 0.05 0.02 0.03 2.4 0.02 0.04 0.01 5.0 (0.08) (0.05) (0.09) (4.54) (0.06) (0.04) (0.05) (5.22) Women 0.01 0.03 0.04 0.5 0.04 0.08 0.04 2.3 (0.05) (0.03) (0.05) (3.78) (0.06) (0.03) (0.06) (3.02) Number of children Men 0.00 0.01 0.01 2.1 0.01 0.03 0.01 0.5 (0.04) (0.02) (0.04) (2.42) (0.02) (0.02) (0.02) (1.54) Women 0.00 0.04 0.04 0.7 0.04 0.00 0.04 1.1 (0.02) (0.02) (0.02) (1.74) (0.02) (0.01) (0.02) (1.07) Private rented Men 0.04 0.001 0.04 10.8 0.03 0.03 0.001 1.6 (0.07) (0.04) (0.09) (4.96) (0.04) (0.04) (0.03) (2.88) Women 0.06 0.04 0.02 9.3 0.01 0.04 0.03 2.4 (0.04) (0.04) (0.03) (3.27) (0.03) (0.02) (0.03) (1.91) Public rented Men 0.03 0.01 0.02 2.7 0.14 0.14 0.004 6.5 (0.05) (0.03) (0.07) (3.68) (0.05) (0.04) (0.03) (2.21) Women 0.01 0.04 0.04 5.4 0.09 0.08 0.01 0.2 (0.04) (0.03) (0.03) (2.67) (0.04) (0.03) (0.03) (1.80) Notes: Derived from Table 2 estimation results. Approximate marginal effects calculated by simulating probabilities and expected durations for variable equal to zero and one, all other variables held at mean values. Bootstrapped standard errors in brackets. expected duration of a low pay spell for both men and women in the 1992 98 period (although by less for women). However, the extent of this effect and any gender difference declines substantially after the introduction of the NMW. This pattern of a gender difference in the 1992 98 period which significantly declines after the introduction of the NMW, is also evident in the case of small firms. In this case there is a gender difference in both the expected duration effect and the high pay exit probability in the early period, although these effectively disappear in the later period. In contrast for employees in manufacturing and services, there is an overall reduction in the expected low pay durations after the introduction of the minimum wage (but no gender difference). However, there is an increase in the gender difference in the high pay exit probability, with the marginal probability of women escaping to high pay substantially more negative after 1999, while the male marginal probabilities show less change.

e. phimister and i. theodossiou i137 The second part of Table 3 focuses on the impact of individual worker characteristics. In the 1992 98 period, the results are generally consistent with human capital theory that educational attainment enhances wage growth. Hence, compared to low paid workers both male and females who have obtained O-levels or A-levels face a reduction in overall expected duration of the low pay spell by 11.5 and just over five months respectively, with higher probabilities of exit to high pay (although these probabilities are relatively poorly determined). The reduction in expected low pay duration is 20 months for males and eight months for females who have a higher qualification such as nursing, and also 11 months for males and just over 13 months for females who have obtained a university degree or equivalent. Higher qualifications also increase the probability of exit to a high pay job (although not all probabilities are well determined). After the introduction of the minimum wage, again the gender differences apparent in expected duration fall. Also the effect of education in increasing the probability of moving to a high pay job declines for men and women, although this effect is more pronounced for men with O-level or A-level qualifications. Although the evidence is mixed (Denvir and Loukas, 2006), this effect may reflect some compression in pay scales associated with the introduction of the minimum wage, and hence with a decrease in the impact of experience on an individual s pay over time. The results for vocational and work-related training qualifications are not on the whole well determined, either before or after the introduction of the minimum wage. This may reflect the greater value of accumulation of general rather than specific human capital in helping individuals escape low pay employment. There are strong age effects in line with Gregory and Elias (1994) for the period before the minimum wage. In this period, expected duration increases somewhat with age for men, while for women expected duration is least for the oldest age group. After the introduction of the minimum wage, the age effects for women fall substantially (and become relatively poorly determined). However, for men the change is less pronounced with the estimated expected duration approximately the same for all ages (a reduction of around eight months). Finally, Table 3 provides rather mixed evidence on the role of household characteristics in low pay duration and exit probabilities. In the pre-1999 period, having a spouse who is employed increases low pay duration and the probability of an exit to higher pay. In the post-1999 period, the effect of having a spouse who is employed on expected duration and the probability of high pay exit decline substantially for women. In contrast, no such differences are observed for the effects of having a spouse who does not work, or, surprisingly, for the effect of the number of children on expected duration and the probability of high pay exit. Research has shown that housing variables are important in labour market mobility (Boheim and Taylor, 2002). The results of this study suggest that living in privately rented accommodation has a negative impact on expected low pay duration for both men and women for the earlier period, although these effects

i138 gender, low pay mobility, and the minimum wage dissipate after 1999. In contrast, the impact of being in publicly rented accommodation becomes better determined after 1999 with a reduction in the exit probability to high pay for men ( 0.14) and women ( 0.09) and, at least for men, a decrease in the expected low pay duration. 5.2 Predicted low pay duration and exit probabilities The marginal effects presented in Table 3 do not provide any information on the level of male female differences in low pay exits and the expected duration of the low pay spell. Further, since the marginal effects are calculated at the mean, they do not reflect real world individual differences. To address these issues, the model estimates are used to calculate overall approximate exit hazard and expected low pay durations for male and female individuals at both mean characteristics and for some specific characteristics of interest. These results are reported in Table 4 (bootstrapped standard errors are reported in brackets). The first panel of Table 4 shows the predicted exit probabilities and expected durations using the overall male and female sample mean values. This provides some indication of the extent of male-female differences in low pay dynamics on average after controlling for differences in observed characteristics. At these values, the exit probabilities from a low pay spell to high pay and unemployment are higher for men both prior to and after the introduction of the minimum wage. However, overall expected low pay duration is smaller for women (25 months) compared to that of men (29 months) for the earlier period, although this difference disappears in the period after the introduction of the minimum wage. 7 A further insight into the predictions of the estimated models at the mean values, is given by the overall predicted hazard rate for men and women at mean values illustrated in Fig. 2. The shapes of the predicted baseline hazards in Fig. 2 reflect the shape of the estimated underlying baseline hazards for the two samples. In the 1992 98 period, women face lower overall low pay exit hazards in the beginning of the spell (roughly the first six months) compared to men but as the low pay spell lengthens, the baseline hazards become similar. In contrast in the period after 1999, the overall low pay exit hazards for men and women follow a roughly similar pattern, with both facing higher baseline hazards initially. To explore how the experience of low pay of specific individuals might differ from the average, the impact of different combinations of characteristics in the third panel of Table 4 are also reported. The first set of results is for a low paid individual (Type 1) with no qualifications, in part-time employment, unmarried with no children, aged less than 25 years old, living in the South West region in privately rented accommodation, employed in a small firm in the non-unionized service sector. As expected, for Type 1 men and women the probability of a high... 7 Table 4 also reports the predicted probabilities and expected durations when the models are estimated using a low pay threshold of 2/3 of the median wage. These results indicate that differences found are relatively robust to the threshold chosen.

e. phimister and i. theodossiou i139 Table 4 Predicted exit probabilities and expected durations 1992 1998 1999 2005...... Exit probability Exit probability...... Higherpaid job Unemployment Inactivity Expected duration Higherpaid job Unemployment Inactivity Expected duration Mean values Men 0.69 0.25 0.06 29.1 0.62 0.21 0.17 19.0 (0.14) (0.07) (0.19) (7.6) (0.05) (0.04) (0.06) (4.5) Women 0.63 0.16 0.21 25.5 0.51 0.14 0.34 18.4 (0.04) (0.04) (0.04) (4.1) (0.04) (0.03) (0.04) (3.5) Mean values 2/3 of median wage Men 0.72 0.24 0.05 22.6 0.65 0.19 0.16 17.3 (0.14) (0.08) (0.18) (4.9) (0.12) (0.09) (0.12) (7.8) Women 0.63 0.17 0.20 21.5 0.55 0.14 0.31 15.4 (0.04) (0.04) (0.04) (3.3) (0.04) (0.03) (0.04) (4.0) Type 1 Individual* Men 0.13 0.76 0.11 6.0 0.34 0.28 0.38 13.5 (0.05) (0.16) (0.19) (1.7) (0.08) (0.09) (0.11) (3.8) Women 0.14 0.38 0.48 7.9 0.25 0.20 0.56 13.6 (0.03) (0.10) (0.10) (2.1) (0.06) (0.07) (0.07) (3.2) Type 2 = Type 1 except 25 35 years old, with 2 children, living in public rented accommodation Men 0.17 0.79 0.04 10.1 0.42 0.36 0.22 11.3 (0.06) (0.14) (0.17) (3.1) (0.09) (0.10) (0.09) (3.4) Women 0.20 0.30 0.50 11.7 0.17 0.24 0.60 17.5 (0.05) (0.09) (0.10) (3.0) (0.04) (0.07) (0.08) (3.9) Type 3 = Type 2 except married, spouse without job, workplace covered by union Men 0.42 0.52 0.06 134.3 0.64 0.31 0.06 66.9 (0.14) (0.17) (0.26) (36.0) (0.11) (0.10) (0.11) (23.1) Women 0.35 0.22 0.43 66.3 0.40 0.11 0.50 71.7 (0.09) (0.08) (0.10) (16.3) (0.08) (0.05) (0.09) (15.4) Type 4 = Type 2 except has vocational qualifications + work related training in last 12 months Men 0.17 0.83 0.01 22.0 0.44 0.34 0.23 8.3 (0.08) (0.13) (0.14) (8.3) (0.10) (0.10) (0.09) (2.9) Women 0.14 0.43 0.43 15.2 0.20 0.32 0.48 11.3 (0.05) (0.13) (0.14) (5.8) (0.05) (0.09) (0.09) (3.0) Type 5 = Type 3 except aged between 46 and 55, spouse has job, no children, and has had work related training in last 12 months Men 0.27 0.70 0.04 42.2 0.45 0.38 0.17 14.0 (0.11) (0.19) (0.24) (14.2) (0.09) (0.10) (0.08) (4.1) Women 0.27 0.41 0.32 33.6 0.26 0.30 0.44 28.7 (0.07) (0.11) (0.11) (10.2) (0.07) (0.08) (0.09) (6.6) Notes: *Type 1 Individual no qualifications, aged less than 25, part-time, living in south west, unmarried, employee of firm <25 employees, in services, no union coverage, no children, living private rented accommodation. Bootstrapped standard errors in brackets. pay exit and expected low pay duration is substantially lower relative to the values obtained using the mean characteristics. Further, for the post 1999 period, there is a noticeable increase of the exit probabilities from a low paid to a higher job and to inactivity but a substantial decrease of the probability of a low pay spell ending to