We use longitudinal data from the OSA. The OSA panel study is targeted at a representative sample of 4000 to 5000 respondents in each wave, first drawn in 1985 and then in 1986 with further biannual waves until 2000. For our analyses, we limit our sample to respondents (men and women) between 21 and 54 years old who have valid observations on their labour force status. The advantage of this dataset is that it provides detailed information about workers' labour market situation at the time of interview distinguishing between the following labour market states: (a) employed, (b) self-employed, (c) unemployed, (d) non-participating, (e) in military service and (f) in education. Labour force information between the interview dates is also traceable through a series of retrospective questions about the start and end dates of labour force changes between the current and previous waves. This data structure allows us to investigate workers' time in unemployment after the expiration of their employment contract. Another advantage of this dataset is that workers have been asked to report the reason behind their labour force changes, which allows us to differentiate between workers who are involuntarily unemployed due to plant closings, massive lay-offs or reorganizations from those who were laid off due to other reasons which may relate to their own personal failures. In this study, unemployment is explicitly defined as ‘currently out of labour and searching actively for a job’, while fixed-term contracts are defined as ‘contracts with a known expiration date’.
Our analyses focus on the unemployment spells of workers who were employed in the previous wave (i.e. at time t−1) but are unemployed at the date of interview of the present wave (i.e. at time t). In these analyses, we exclude those who lost their jobs due to uneasily defined reasons and seasonal employment (191 respondents). An implication of the design of our study is that respondents who have experienced a first unemployment spell but have reported no temporary or permanent employment previously are not considered in our analyses (63 respondents). Spells interrupted due to a withdrawal from the sample are recorded as truncated. These restrictions leave us with a total of 2912 unemployment spells, 14.5 per cent of which are right censored (remain unemployed), 67.8 per cent end with a transition to employment (dependent worker), 5.7 per cent of the spells ends in self-employment, 7.3 per cent enter non-participation, 1.8 per cent enter military service and 2.8 per cent make the transition into education.
Figures 1 and 2 depict workers' course of labour force participation conditional on whether they were employed in a regular or a fixed-term contract in the previous wave (t−1). As expected, Figure 1 depicts a stable career trajectory for those with a previous regular contract. Specifically, a large share of workers (around 80 per cent) with a regular contract in the previous wave (in 1985) remains in employment in the following wave. A small share of this group either disappears out of the labour force (1 per cent), becomes unemployed (4 per cent) or self-employed (3 per cent) in the following wave. This trend remains slightly constant over time.
Figure 1. The Labour Force Distribution of Workers with Regular Contracts in the Previous Wave.
Source: Authors' calculations, based on the OSA panel 1980–2000.
Download figure to PowerPoint
Figure 2. The Labour Force Distribution of Workers with Fixed-term contracts in the Previous Wave.
Source: Authors' calculations, based on the OSA panel 1980–2000.
Download figure to PowerPoint
Another picture emerges for workers with fixed-term contracts in Figure 2. Specifically, from those in fixed-term contracts in 1985, less than the half (44 per cent) remains in the labour force in the following wave. The remainder of this group disappears out of the labour market (42 per cent), becomes unemployed (11 per cent) or chooses to start their own businesses (3 per cent) in the next wave. Interestingly, there appears a turning point in the labour force distribution of these workers from 1994 and onwards. Specifically, from those workers with fixed-term contracts in 1994, around 55 per cent remains in the labour force in the next wave (as opposed to 44 per cent) and a lower proportion of workers (32 per cent) disappears out of the labour force as inactive (as opposed to 42 per cent). This change in the distribution may reflect policy effects with regard to the prescription of the equal labour rights of workers (regardless the type of the employment contract) that was implemented through the Civil Code in 1996 (Heerma van Voss 2000).
Who are the workers utilizing fixed-term contracts? Descriptive statistics in Table 1 show that these are more often single women in their early 30s with relatively fewer (home living) children, with slightly higher attained education compared to workers with a regular contract before unemployment. Despite their slightly higher education level, workers with fixed-term contracts have a shorter tenure compared to those in regular contracts.
Table 1. Summary of Sample Characteristics for Workers with Regular versus Fixed-Term Contracts
| ||Regular contract||Fixed-term contract|
|Mean %||SD||Mean %||SD|
|Demographics|| || || || |
|# Home living children|| || || || |
|Human capital and labour force history|| || || || |
|Tenure (in months)||29.01||17.74||21.3||16.81|
|# Working hours||33.62||10.04||32.77||10.38|
|Lower intermediate education||36.09||48.03||32.26||46.76|
|Higher intermediate secondary education||36.79||48.23||39.57||48.92|
|Total observations||4,558|| ||1,468|| |
To model how a previously held job with a fixed-term contract influences the duration of subsequent unemployment spells (i.e. re-entry to employment), we rely on survival or event history methods (Blossfeld et al. 2007). In the first set of our analyses, we produce parameter estimates in the form of the piecewise-constant exponential models. The advantage of these models over semi-parametric or parametric models is that it allows the time span, during which the workers re-enter the labour force, to be split into several intervals where for each interval a baseline hazard is estimated. This flexible approach does not impose a functional form of the baseline hazard, but leaves the data to speak for themselves (Cleves et al. 2008). The piecewise-constant exponential model yields an overall hazard (hj) of:
Where h0 refers to the baseline hazard rate that is assumed to be constant within each time interval (λij) (where j = 6, 12, … , J months) for worker i (i = 1, … , N workers in the sample) with (t) representing the elapsed unemployment duration. χi,t−1 refers to a vector of explanatory variables previous to the current unemployment spell that may affect a worker's current unemployment duration. Finally, βj refers to a transposed vector that accounts for coefficients associated with the observables characteristics.
Estimation of our piecewise-constant exponential models faces a methodological challenge with regard to the issue of the sample selection. For an individual's unemployment duration to be observed a worker should: (a) be unemployed within the observation period; (b) report the type of contract before unemployment. To correct for the non-randomness related to the sample selectivity, we use a two-stage equation, where in the first stage, we run a probit model on the probability of being part of the sample. The additional variable at this stage, which is necessary for the identification of the equation, is the dummy variable ‘ever unemployed during the observation period’, which strongly determines workers' likelihood to end up in a job with fixed-term contract but that may not directly influence the current spell of unemployment. To test for the instrument's validity, we employ a test for exogeneity as proposed by Green and Heywood (2011) elaborating on the work of Stock and Yogo (2005). The test statistic (F-test = 16.38) as outlined by Stock and Yogo (2005), yields a value above the critical value that is necessary to detect a weak instrument of (F-test = 10.57) and implies that we have a valid instrument for our analyses.
Next, to examine how a previous held job with a fixed-term contract influences the likelihood of subsequent unemployment spells, we apply random-effect probit models that include lagged (independent) variables on the right-hand side as used by Heckman and Willis (1976) and by Chamberlain (1985). Consider the following linear reduced form equation for the latent dependent variable unemployment occurrence in time periods t (where t = 1, 2, … , T) for worker i (i = 1, … , N workers in the sample):
where the value of yit refers to the unemployment occurrence of individual i at time t, conditional on workers' observable characteristics (xi,t−1) in the previous wave t-1. The symbol Φ refers to the cumulative density function of a standard normal distribution. The value β refers to a transposed vector that accounts for coefficients associated with the observables characteristics whereas αi refers to the unobserved time-invariant and individual-specific effect, while eit refers to the error term of the model. So far, our models treat the individual-specific error (αi) as random and assume that the error term of the model (eit) is normally distributed, with zero mean, a fixed variance (eit ∼ IN (0, σ2ε) and independently distributed for all individuals across time periods. A danger occurs when this assumption is violated. To account for this problem, we relax the assumption that αi is independent of time-varying characteristics by using a model as proposed by Chamberlain (1985). The model, assumes that the regression function of αi is linear in the means of all time-varying covariates. This implies that using the mean of time-varying variables in the model as additional regressors, allows the random effects to depend on the current, future and past X's. In doing so, the correlation between two successive error terms for the same individual is constant over time, implying that the effect of one year's unemployment on the next year's unemployment does not change over time and is constant across individuals.