State Dependence in a Multinominal-State Labor Force Participation of Married Women in Japan 1

Transcription

1 State Dependence in a Multinominal-State Labor Force Participation of Married Women in Japan 1 Kazuaki Okamura 2 Nizamul Islam 3 Abstract In this paper we analyze the multiniminal-state labor force participation of married women in Japan. We investigated statistically whether there is true state dependence, in which preferences, abilities or constraints for future decision are altered as a consequence of experiencing an event in a choice of regular and non-regular work. We estimate a dynamic multinominal logit model which allows the correlated random effect, and linear probability model to check the robustness of the multinominal logit model. The empirical results suggest that there is significant true state dependence for the choice of regular and non-regular work. Significant effect of true state dependence in regular work justified the stepping stone policy to regular work from non-participation or non-regular work or maintenance policy to support utilizing the temporal non-regular work. On the other hand, significant effect of true state dependence in non-regular work means that non-regular work constrains married women s ability or preference not conforming to regular work, and in this sense, non-regular work is exclusionary. This result suggests that the policy that supports the non-regular workers to step to regular work is needed. 1 Introduction It is well-known that there is M-shaped profile of labor force participation of married women in Japan, while it is inverse U-shaped pattern in other developed countries. One of reasons for differences in participation pattern lies in differences in women s transition pattern among regular work, non-regular work and non-participation. On the occasion of childbearing, the women who are employed as regular workers must go out 1 The authers thank Philippe Van Kerm (CEPS/INSTEAD, Luxenbourg) for helpful suggestions. We also thank the Institute for Research on Household Economics for access to the Japanese Panel Survey of Consumers. 2 Department of Economics and social Sciences, Kochi University, Japan. kokamura@cc.kochi-u.ac.jp 3 CEPS/INSTEAD, Luxenbourg nizamul.islam@ceps.lu 1

2 of labor market instead of transition to non-regular work if there are scarce opportunities for transition from regular work to non-regular work. Even if they are employed as non-regular workers, they tend to go out of labor market if non-regular work are not good job opportunities. As past researches have shown, the non-regular work plays several roles in labor market. As Blank(1989) suggests, the experiences of part-time work plays as a stepping stone for facilitating transition from out of labor market to full-time work. Alternatively, part-time work plays a maintenance role, which is a temporary alternative for full-time work during the childbearing to allow an individual to balance the working activity and family responsibilities. Finally, it has been claimed that part-time work are exclusionary. In this case, women are forced to be unstable employment and go out of labor market in the end. In a dynamic framework of participation, a choice of participation state among regular work, non-regular work and non-participation depends on the past experience of state. For example, an individual who participate in regular work at time t tend to participate in same state at time t + 1. Heckman (1981a) shows two explanations for this type of behavior. One explanation is that true state dependence in which preferences, prices or constraints for future decision are altered as a consequence of experiencing an event. And the second explanation is that spurious state dependence in which individual propensity to experience the event correlate over time. For the policy evaluation, it is important to distinguish true state dependence and spurious state dependence in a participation behavior. If there is positive true state dependence in regular work, it shows that the experience of regular work adjusts worker s ability or preference for conforming to regular work. This means that if some married women quit the regular work because of childbearing, it increases the economic disparity between women who continue to work as regular worker and who quit for childbearing, even if they have same observed and unobserved characteristics. As Ishawa (2001) remarks if in the market, we say that the genuine wage disparity that is problematic exists when there are workers who have the same ability and preference and yet do not obtain the same job opportunity (p.246), the existence of true state dependence in regular work produce the genuine disparity among married women whose abilities and preferences are same, if scarce opportunities are possible for childbearing for married working women. In this case, some policies that support the transition to regular work is effective for decreasing the economic disparity among married women. 2

3 On the other hand, if there is positive true state dependence in non-regular work, the experience of non-regular work adjusts worker s ability or preference for conforming to non-regular work. In this case, the experience of non-regular work is not necessarily good for worker s welfare, because it is more difficult to conform to the regular work even if she wants. Whether the persistent labor force participation of married women while childbearing succeeds depends on (1) the efficiency to promote the transition from non-participation to the regular or non-regular work and (2) the degree of true state dependence in regular and non-regular work. In this paper, we shed light on the second point. We examine the existence of true state dependence in regular and non-regular work in a multinominal-state dynamic model. The existence of positive true state dependence in regular work means that experience of regular work has some effects of changing worker s innate characteristics for conforming to regular work. Then it suggests that even if a married woman quit the job for childbearing, she is able to participate in regular work successively again by some stepping stone policy for labor force participation. On the other hand, if there is a positive true state dependence in non-regular work, there is some effect that changing worker s innate characteristics for conforming to non-regular work. Then it suggests that non-regular work is exclusionary, meaning that their abilities or preferences are constrained for being forced to engage in low-paid job continuously (ex. part-time trap). In this case, some stepping stone policy is needed to promote the transition from the non-regular work to regular work. The paper is organized in the following way. Section2 describes the data and variables, and in Section3 we show the observed pattern of transitions among three states, regular work, non-regular work and non-participation. Section4 provides the structure of model and empirical specification. In Section5 we present and discuss the results and conclude in Section6. 2 Data We use the data rounds of the Japanese Panel Surveys of Consumers (JPSC) surveyed annually. The JPSC includes panel surveys of 1500 women aged in 1993 (Panel A) and added 500 women of aged in 1997 (Panel B). To focus on the dynamic aspect of participation behavior, we used the sample included in Panel A which covers periods in The sample consists of 340 continuously married couples. 3

4 As a dependent variable, we used the employment status: regular, non-regular (part-time, contract work or another kind of non-regular work) and non-participation. Age, education, number of children, permanent and transitory non-labor income and non-regular work experience of post school (NWEP) are used as independent variables. Married women aged in 1993 are included, and years of education are imputed from the following category: junior high school = 9 years, special school or special training college (An entrance to school qualification is a junior high school graduate) = 10.5 years, high school = 12 years, special school or special training college (An entrance to school qualification is a senior high school graduate) = 13.5 years, junior college = 14 years, university = 16 years, graduate school = 19.5 years. Fertility variables are defined as number of children aged 0-2, 3-5, 6-17 (#Children0-2, #Children3-5, #Children6-17, respectively). Husband s annual earnings deflated by CPI (base year is 2005) is used as a proxy for non-labor income for married women. The 10 year s average of husband s annual earnings is used as permanent non-labor income (ymp) and deviations from the permanent income is used as transitory non-labor income (ymt) 4. In Japan, transition from school to work is the main port of entry for regular work. School plays an important role of matching between graduates and jobs. In other words, school bears a large amount of search cost. The experiences that fail to obtain regular work after graduating school can be a bad signal for ability of the graduates and it increase the search costs for them. The recognitions of such a circumstance affect the people s preference toward not to participate in work activity, or even if they participate, they participate in non-regular jobs, such as part-time or contract work. To examine such effects of transition from school to work on the subsequent participation, I used the dummy variable which indicate who had experienced part-time or temporary jobs or jobless after school. We call the variable as non-regular work experience of post school (NWEP) 5. 3 Observed Transition patterns Table1 shows the transition matrix among non-participation, non-regular and regular 4 We follow the definition of fertility variables (#Children0-2, #Children3-5, #Children6-17) and permanent and transitory income (ymp and ymt) used by Hyslop(1999). 5 The experience of non-regular work immediately after school may also delay the timing of marriage (See, Higuchi (2002)). By focusing on a sample of continuously married women, we can identify the effect of NWEP on subsequent participation behavior conditional on the effect on marriage. 4

5 work. It indicates that a substantial number of women remain in the same state. Greater amount of state dependence seems to lie in regular work: 92.6% of the women who are regular worker in year t are observed as regular worker in year t + 1. On the other hand, 87.3% of the women who are non-regular worker in year t are observed as non-regular worker in year t + 1. Almost 10% of the women who are non-regular worker in year t are non-participation in year t + 1, while almost 10% of the women who are non-participation in year t transit to non-regular work in year t + 1. There are few cases that transitions among regular and non-regular work. Table1 Transition Matrix of Married Women for the period state at time t + 1 Non-participation Non-regular Regular Non-participation state at time t Non-regular Regular Notes: Numbers are% of row. It is based on 3060 observations (340 women * 9 years) These transition patterns are partly influenced by observed characteristics such as age, education, and unobserved heterogeneity. In following sections, we check the degree of true state dependence for regular work and non-regular work statistically, controlling the observed characteristics and unobserved heterogeneity. 4 Empirical Model We estimate a dynamic multinominal logit model with random effects to analyze the transitions among non-participation, regular and non-regular work. The unobserved individual-specific effects are allowed to correlate with observed characteristics and to follow a discrete, non-parametric distribution. Dynamic multinominal logit model is used for welfare participation (Hansen and Lofstrom (2001)) and labor force participation (Prowse (2007)). Hansen and Lofstrom (2001) analyses the transitions among social assistance, unemployment and employment. Prowse (2007) analyses the transitions among full-time work, part-time work, and non-participation. For econometric modeling, we follow these papers 6. 6 Prowse (2007) uses the method of maximum simulated lelihood, but in this paper we didn t adopt simulation techniques. 5

6 The model can be described as follows. Married woman (indexed by i, i = 1,..., N ) belong to any of the three mutually exclusive states k at time t ( t = 2,3,..., T ) : being non-participation ( k = 1), being employed as non-regular worker ( k = 2 ), and being employed as regular worker ( k = 3). The value, for married woman i, of belonging to state k, at time t ( V t ) be specified as follows. V t = X β + Z β + L β + ε it 1 k i 2 k it 3 k t (1) where, ε = µ + v. (2) t t X it is a vector of time varying observable characteristics, including time dummies, age, number of children (aged 0-2, aged 3-5, aged 6-17), transitory non-labor income. Z i is a vector of time invariant observable characteristics, including, educational attainment, permanent non-labor income, and dummy variable for NWEP. L it is a vector of previous participation state. The error terms ε t is composed of two terms, of which µ it represents an unobserved individual specific and time invariant component, and v t represents a serially uncorrelated error term. In equation, parameter vectors β = 1,..., 3 l k l are to be estimated, and for identification purposes, we normalize l β 1 l = 1,...,3 and µ i1 to zero. The probability of observing individual i in state k at time t ( t > 1), conditional on X it, Z i, L it and µ it, can be written as: P ( k t t µ ) exp( X it β k + Z iβ k + Lit β k + µ ) = exp( X it β s + Z iβ s + Lit β s + µ is ) = 3 s 1 (3) 6

7 It is possible that unobserved individual specific and time-invariant component µ is correlated with number of children, transitory non-labor income, and NWEP. Hence, following Mundlak (1978) and Chamberlain (1984), we estimate the correlated random effects model (CRE) which is specified as follows. µ = δ mean(# Children0 2) + δ mean( Ymt) δ mean( NWEP) 5 + δ mean(# Children3 5) 2 + η + δ mean(# Children6 17) 3 (4) We assume that unobserved individual specific and time-invariant component µ is correlated with mean values (for t ) of number of children, transitory non-labor income, and NWEP. A person initially observed may be endogenous, that is initial observations are correlated with unobserved heterogeneity, in a dynamic model. We adopt a procedure suggested by Heckman (1981) to deal with the initial condition. Let the value, for individual i in state k at initial period ( t = 1), be specified as: V = X θ + Z θ + ε i1 k i k 1 (5) where, ε = τ + v. (6) 1 1 The parameters θ = 1, 2 l k l are to be estimated, and as earlier, we normalize l θ 1 l = 1,2 and τ i1 to zero. The probability of observing individual i in state k at initial period ( t = 1), conditional on X i1, Z i and τ, can be written as: P( k 1 1 τ ) 1 2 exp( X i1θ k + Ziθ k + τ ) 1 2 = exp( X i1θ s + Ziθ s + τ is) = 3 s 1 (7) The unobserved individual specific effects, τ are assumed to be correlated with η, and they can be identified through serial correlation in ε t. It is straightforward to estimate the model with maximum lelihood techniques. The sample lelihood for the 7

8 multinominal logit with random intercepts has the following form: L = N T d ( P k τ )) ( P ( k η) ) 3 d t i= t= k= t t f ( ϑ) 1 ( dϑ where d = 1 if individual i chooses alternative k at time t and zero otherwise. (8) ϑ i is a vector with η and τ elements. As ϑ i is not observed, we have to integrate out this term for above lelihood to obtain the unconditional lelihood function. To do this, we specify a distribution for ϑ i following the Heckman and Singer (1984). We approximated the distribution for ϑ i by a discrete distribution with a finite number ( J ) of support points. Associated with each support point is a probability, π j, J j = 1 where = 1 π and π 0 j j. We assume that there J types of individuals and each individual is endowed with a set of unobserved characteristics, j ϑ i (consist of τ j and L = η for k = 1,2, 3 ), for j 1,..., J j N J π =. The lelihood is then: T 3 d1 dt ( ( P1 ( k τ ) ( Pt ( kt η) ) i= 1 j = 1 j t= 2 k= 1 1 We report estimates based on this model where J = 3. (9) 5 Empirical Results Table2 shows the regression results of standard multinominal logit model. The negative effects of number of children are larger in non-regular work than regular work. If the childbearing has fixed time cost, negative effects should be larger for regular work. On the other hand, negative effects of husband s earnings are large in regular work than non-regular work and the effects are larger for permanent non-labor income than transitory non-labor income. It seems that married women s work behavior depends mainly on the husband s earnings. On the other hand, NWEP does not affect the multinominal-state choice significantly. We check the robustness of these results by estimating several models controlling the unobserved heterogeneity and dynamic structure. Table2 8

9 Standard Multinominal Logit model Regular Non-regular Permanent non-labor income (Ymp) (0.170) (0.052) Transitory non-labor income (Ymt) (0.042) (0.063) No. of Children Aged 0-2 years (#Children0-2) (0.237) (0.759) No. of Children Aged 3-5 years (#Children3-5) (0.992) (0.306) No. of Children Aged 6-17 years (#Children6-17) (0.121) (0.204) Non-regular work experience of post school (0.921) (0.622) Log-lelihood Note: Estimated standard errors are in parenthesis. All Specifications include unrestricted time effects, a quadratic in age, years of education. Table3 shows the regression results of random effect and correlated random effects multinominal logit model, which controls the unobserved heterogeneity. Column (1) and (2) show the results of random effects multinominal logit model. Number of children decreases the probability of choice for regular work and non-regular work as compared to the choice for non-participation, and the younger children have larger effect than older children. Comparing the coefficients of regular work and non-regular work, negative effects of number of children are larger for non-regular work. This result can be explained by the difference in the availability of family policy between regular work and non-regular work. If regular workers have more options for work-life balance (such as aid for childcare) than non-regular workers, number of children decreases the 9

10 probability of choice for non-regular work than regular work. The effect of NWEP is significantly positive for choice of regular work, while its effect is significantly negative for choice of non-regular work. In the definition of our data set, regular work does not exclude the self-employment or help for familiy bussiness. Then it is possible that if she experienced the NWEP, it is difficult to be employed. Then she employs herself or helps for family business as a regular worker. The effects of non-labor labor income are insignificant or significantly negative, and its effects, especially of permanent income, are larger for non-regular work than regular work. These results suggest that non-regular workers are more dependent on husband s earnings. The estimated support points and accompanying probabilities for the model indicate unobserved heterogeneity in individual-specific preferences. For example, first column shows that first support point is with probability 0.228, second support point is with probability 0.358, and third support point is with probability These results suggest that 23% of the sample has a strong innate preference for work; 41% of the sample has a weak preference for work; and, 36% of the sample has a noncommittal preference. Column(3) and (4) in Table3 shows the regression results of correlated random effects (CRE) multinominal logit model. If the unobserved heterogeneity is endogenous, in which married women s preference is influenced by number of children, husband s income and NWEP, parameters in simple random effect model have some biases. In a simple random effect specification, the effect of NWEP is significantly positive for choice of regular work, while its effect is significantly negative for choice of non-regular work. But in a CRE specification, the effects of NWEP are significantly positive for both regular work and non-regular work. Why the experience of NWEP increases the probabilities of choice for regular and non-regular work than non-participation? It is possible that the women belonging to non-regular work tend to marriage with men belonging to non-regular work, so they must work after marriage because of low level household income. NWEP is negatively correlated with unobserved heterogeneity. This indicates that NWEP depresses the innate ability or preference for work. In a CRE specification, negative effects of number of children are larger in regular work except for the child aged 0-2. Childcare has fixed time cost, number of children is more costly for working as regular worker than non-regular worker. For the children aged 0-2, there are more opportunities for official aid, then there is not much difference between regular and non-regular work. 10

11 In a simple random effects specification, negative effects of non-labor income are larger for non-regular work than regular work. But in a CRE specification, negative effects of non-labor income are larger for regular work. These results suggest that women s work behavior depends on the household earnings, then as the husband earns more, they devote less time for work activity. The estimated support point and accompanying probabilities for CRE model are not so different from uncorrelated random effect model. We checked goodness of fit for CRE model by lelihood ratio test. The test statistic for correlated to uncorrelated random effect model is 12.08, then the null hypothesis that there is no difference between correlated and uncorrelated random effect model can be rejected. Table3 Uncorrelated and Correlated Random Effects Multinominal Logit model Uncorrelated random effect Correlated Random Effect Regular Non-regular Regular Non-regular (1) (2) (3) (4) Permanent non-labor income (Ymp) (0.003) (0.002) (0.005) (0.003) Transitory non-labor income (Ymt) (0.006) (0.003) (0.006) (0.003) No. of Children Aged 0-2 years (#Children0-2) (0.011) (0.008) (0.015) (0.012) No. of Children Aged 3-5 years (#Children3-5) (0.009) (0.006) (0.013) (0.009) No. of Children Aged 6-17 years (#Children6-17) (0.003) (0.005) (0.014) (0.009) 11

12 Non-regular work experience of post school (NWEP) (0.022) (0.013) (0.011) (0.050) Correlated with Unobserved Heterogeneity Transitory non-labor income (Ymt) (0.007) (0.004) No. of Children Aged 0-2 years (#Children0-2) (0.047) (0.055) No. of Children Aged 3-5 years (#Children3-5) (0.045) (0.043) No. of Children Aged 6-17 years (#Children6-17) (0.016) (0.014) Non-regular work experience of post school (NWEP) (0.011) (0.050) Log-lelihood First support point (θ1) (0.058) (0.020) (0.039) (0.031) Second support point (θ2)

13 (0.054) (0.022) (0.024) (0.044) Third support point (θ3) (0.050) (0.019) (0.027) (0.031) Probability(π1) Probability(π2) Probability(π3) Note: Estimated standard errors are in parenthesis. All Specifications include unrestricted time effects, a quadratic in age, years of education. The CRE model expresses individual-specific component as a linear function of ymt, # Children 0-2, # Children 3-5, # Children 6-17, and NWEP. Table4 shows the regression results of dynamic uncorrelated and correlated random effects multinominal logit specification. Column(1), (2) show the results of dynamic random effects model. Our main concern is whether there are true state dependence in regular work and non-regular work. Empirically, we check the existence of true state dependence based on the coefficient of lagged dependent variable with controlling for unobserved heterogeneity. If the coefficient of lagged dependent variable is significantly positive, we inference that there is true state dependence in participation behavior of married women. As table shows, coefficient of lagged dependent variable is significantly positive in both regular work and non-regular work, and its size is larger for regular work. This means that if married women experience the regular (non-regular) work in last period, the probability to participate in regular (non-regular) work activity in this period increases as compared to the women who experienced non-participation in last period. If the experience of regular (non-regular) work changes married women s ability and preference conforming to the regular (non-regular) work, state dependency exists in married women s participation behavior controlling the unobserved heterogeneity. The degree of true state dependence depends on the skill level required for work and preference for work activity. If regular work requires more skills or more strong preference than non-regular work, the probability to participate should be less for regular work than non-regular work. 13

14 These results have important policy implications. If some active labor market policy for married women such as training program can be a stepping stone to the regular work, it increases their aptitude for regular workers and their participation continues successively. But, if the reason for the true state dependence in regular work lies in scarce availabilities of non-regular work for childbearing, it is necessary to support temporary opportunities for non-regular work. The existence of true state dependence in non-regular work suggests that the experience of non-regular work alters the married women s ability or preference not conforming to the regular work. In that case, it is difficult to transit from non-regular work to regular work, so some kind of policy is needed to support the opportunities for non-regular workers to experience the regular work. The accompanying probabilities of estimated support points are first support point with probability 0.396, second support point with probability 0.250, and third support point with probability 0.355, which is more biased for high ability or strong preference than static model. Column(3) and (4) in Table4 show the regression results of correlated dynamic random effects multinominal logit specification. The negative effects of number of childen aged 3-5 and aged 6-17 are large for regular work, except the case for aged 0-2. The effects of non-labor income are more negatively correlated with participation for regular work than non-regular work. The effects of NWEP are significantly positive for choice of regular work and non-regular work, and its coefficient is larger for regular work than non-regular work. Similar to the result of dynamic uncorrelated random effects specification, coefficient of lagged dependent variable is significantly positive in both regular work and non-regular work, and its size is larger for regular work. These results can be interpreted as same way as the correlated random effects multinominal logit specification. The estimated support point and accompanying probabilities for CRE model are not much different from uncorrelated dynamic random effects model. We again checked the goodness of fit for CRE model by lelihood ratio test. The test statistic for correlated to uncorrelated random effects model is 15.18, then the null hypothesis that there is no structural difference can be rejected as static model. Table4 Dynamic Uncorrelated and Correlated Random Effects Multinominal Logit model Uncorrelated random Correlated Random 14

15 effect Effect Regular Non-regular Regular Non-regular (1) (2) (3) (4) Permanent non-labor income (Ymp) (0.007) (0.004) (0.007) (0.003) Transitory non-labor income (Ymt) (0.008) (0.004) (0.009) (0.005) No. of Children Aged 0-2 years (#Children0-2) (0.017) (0.011) (0.027) (0.018) No. of Children Aged 3-5 years (#Children3-5) (0.015) (0.008) (0.022) (0.015) No. of Children Aged 6-17 years (#Children6-17) (0.013) (0.007) (0.022) (0.013) Non-regular work experience of post school (NWEP) (0.041) (0.026) (0.314) (0.058) Lagged Participation (0.021) (0.009) (0.017) (0.009) Correlated with Unobserved Heterogeneity Transitory non-labor income (Ymt) (0.010) (0.006) No. of Children Aged 0-2 years (#Children0-2) (0.153) (0.057) 15

16 No. of Children Aged 3-5 years (#Children3-5) (0.103) (0.035) No. of Children Aged 6-17 years (#Children6-17) (0.026) (0.016) Non-regular work experience of post school (NWEP) (0.336) (0.060) Log-lelihood First support point (θ1) (0.142) (0.073) (0.223) (0.865) Second support point (θ2) (0.148) (0.075) (0.215) (0.253) Third support point (θ3) (0.143) (0.072) (0.209) (0.257) Probability(π1) Probability(π2) Probability(π3) Note: Estimated standard errors are in parenthesis. All Specifications include unrestricted time effects, a quadratic in age, years of education. The CRE model expresses individual-specific component as a linear function 16

17 of ymt, # Children 0-2, # Children 3-5, # Children 6-17, and NWEP. Table 5 shows the results of dynamic uncorrelated and correlated random effects model in initial period. The effects of number of the children aged 3-5 and 6-17 are different between initial period and whole period for the choice of both regular and non-regular work. The effects of NWEP are negative for the choice of non-regular work in initial period, although its effect is positive in whole period. Table5 Dynamic Uncorrelated and Correlated Random Effect Multinominal Logit model (Initial Period) Uncorrelated random effect Correlated Random Effect Regular Non-regular Regular Non-regular (1) (2) (3) (4) Permanent non-labor income (Ymp) (0.007) (0.008) (0.007) (0.008) Transitory non-labor income (Ymt) (0.012) (0.012) (0.013) (0.018) No. of Children Aged 0-2 years (#Children0-2) (0.017) (0.023) (0.018) (0.024) No. of Children Aged 3-5 years (#Children3-5) (0.016) (0.019) (0.017) (0.027) No. of Children Aged 6-17 years (#Children6-17) (0.017) (0.017) (0.017) (0.015) Non-regular work experience

18 of post school (NWEP) (0.035) (0.041) (0.035) (0.039) First support point (θ1) (0.001) (0.001) (0.001) (0.001) Second support point (θ2) (0.001) (0.001) (0.001) (0.001) Third support point (θ3) (0.001) (0.001) (0.001) (0.001) Note: Estimated standard errors are in parenthesis. All Specifications include unrestricted time effects, a quadratic in age, years of education. We also estimate the linear probability model to check the robustness of dynamic multinominal logit model. We estimated the Instrumental Variable model, in which first difference of lagged dependent variable is used as instrument for lagged dependent variable 7. We made the two sub-sample that (a) including only regular worker and non-participated women and (b) including only non-regular worker and non-participated women, and estimated linear probability model, separately. Table6 shows the results of linear probability model. We estimated the linear probability model by instrumental variable method using first difference of lagged dependent variable as instrument for the lagged dependent variable 8. The qualitative results are almost same other than the case of non-regular work experience of post school, which is insignificant in linear probability model. The effects of lagged participation are significantly positive both in regular and non-regular work, it confirms the main results of our dynamic multinominal logit model are robust. Table6 Linear Probability 7 See, Hyslop(1999). We don t estimate the Arellano and Bond estimate, because we don t assume the AR(1) in transitory errors in our dynamic multinominal logit model. 8 We also estimated the model in which the out-of period realizations of the covariates are included as instrument, it did not changed the results. See Hyslop(1999). 18

19 Model Regular Non-regular (1) (2) Permanent non-labor income (Ymp) (0.008) (0.008) Transitory non-labor income (Ymt) (0.004) (0.006) No. of Children Aged 0-2 years (#Children0-2) (0.012) (0.022) No. of Children Aged 3-5 years (#Children3-5) (0.012) (0.020) No. of Children Aged 6-17 years (#Children6-17) (0.012) (0.016) Non-regular work experience of post school (NWEP) (0.038) (0.050) Lagged Participation (0.051) (0.040) R-squared Number of Observation Note: Estimated standard errors are in parenthesis. All Specifications include unrestricted time effects, a quadratic in age, years of education. We estimated the instrumental variable model, in which first difference of lagged variable is used as instrument for 19

20 lagged dependent variable. 6 conclusion This paper analyzes the multinominal-state labor force participation of married women in Japan. We investigated whether there is true state dependence, in which preferences, abilities or constraints for future decision are altered as a consequence of experiencing an event in the choice of regular work and non-regular work. If true state dependence exists in the choice of regular work, giving married women the opportunities to experience the regular work opens the way for the successive participation as regular workers. On the other hand, if true state dependence exists in the choice of non-regular work, the experience of non-regular work constrains the possibilities of career advancement for married women. We used the Japanese panel data set JPSC, for the years 1993 to 2002 and estimated the dynamic multinominal logit model which allows the correlated random effects, and linear probability model to check the robustness of the multinominal logit model. The empirical results suggest that there are significant true state dependence for the choice of regular work and non-regular work. Significant effect of true state dependence in regular work justified the stepping stone policy to regular work from non-participation or non-regular work or maintenance policy to support utilizing the temporal non-regular work. On the other hand, significant effect of true state dependence in non-regular work means that non-regular work constrains married women s ability or preference not conforming to regular work, and in this sense, non-regular work is exclusionary. This result also suggests that the policies that support the non-regular workers to step to regular work. Prowse(2007) shows the presence of true state dependencies for full-time and part-time work using British Household Panel Survey. She estimates both dynamic multinominal logit model and linear probability model, and shows that the degree of true state dependency depends on the specifications of unobserved heterogeneity. In our model, we assume that transitory errors are serially uncorrelated, it is a strong assumption. Okamura and Islam (2007) estimated the binary choice model using same data set, and they found there are strong serially correlated transitory errors, while there is no true state dependence. Estimating the effects of true state dependence, controlling serially correlated transitory errors components, in a multinominal-state model may confirms our results which is our future concern. 20

21 References Blank, R (1998) Labor Market Dynamics and Part-time Work Research in Labor Economics, Vol.17, pp Chamberlain, G Panel data, in Z. Griliches and M. Intrilligator (eds.). Handbook of Econometrics. Amsterdam: North-Holland. Heckman, James J. (1981a) Statistical Models for Discrete Panel Data, Chapter 3 in C. Manski D. McFadden, ed., Structural Analysis of Discrete Data, MIT Press, Cambridge, MA, pp Heckman, James J. (1981b) The Incidental Parameters Problem and the Problem of Initial Conditions in Estimating a Discrete Time-Discrete Data Stochastic Process, Chaper4 in C. Manski and D. McFadden, ed., Structural Analysis of Discrete Data, MIT Press, Cambridge, MA, pp Heckman, James J. and B. L. Singer (1984) A Method for Minimizing the Distributional Assumption in Econometric Models for Duration Data, Econometrica, 52, pp Higuchi, Yoshio (2001) Women s Employment in Japan and the Timing of Marriage and Children, The Japanese Economic Review, Vol.52, No.2, June, pp Hyslop, Dean R. (1999) State Dependence, Serial Correlations and Heterogeneity in Intertemporal Labor Force Participation of Married Women, Econometrica, Vol.67, No.6, pp Ishawa, Tsuneo (2001) Income and Wealth, Oxford University Press. Mundlak, Y On the pooling of time series and cross section data. Econometrica 46: Okamura, K. and Nizamul Islam (2007) Intertemporal Labor Force Participation of Married Women in Japan unpublished manuscript. Prowse, Victoria (2007) Modeling Employment Dynamics with State Dependence and Unobserved Heterogeneity Department of Economics Discussion Paper Series, No.337, August, Oxford University. 21