Low wage employment in Poland

Transcription

1 Low wage employment in Poland Iga Magda This version: 8 April 2010 Abstract This paper analyses the low paid sector in Poland. Firstly we look at its composition and find that among personal characteristics, age, educational attainments and gender are the most important determinants of low wages, though characteristics of co-workers were also found significant. Importantly, once unobserved individual effects are taken into account, the estimated effects are significantly reduced. Next we analyse transitions out of low pay. We find that the probability of being low paid strongly depends on the earnings status in the previous year, whereas many of the individual characteristics associated with low wages are irrelevant for exits from low pay. The multivariate probit model indicates that the initial conditions of being low paid, as well as employment and panel attrition are endogenous processes impacting the estimates of mobility out of low pay. Keywords: earnings, low pay, Poland, wage mobility. JEL- Code: J31, J69. This paper is a part of my PhD thesis and is based on work carried out during a visit to the European Center for Analysis in the Social Sciences (ECASS) at the Institute for Social and Economic Research, University of Essex supported by the Access to Research Infrastructures action under the EU Improving Human Potential Programme. I would also like to thank the British Household Panel Survey principal investigator, the ISER - Colchester for allowing the access to the data. PhD student at the Warsaw School of Economics, researcher at the Institute for Structural Research, iga.magda@ibs.org.pl 1

2 1 Introduction Many of the OECD economies have witnessed a sharp increase in wage inequalities in the last decades, particularly in the 1980s and 1990s which, for many of them, brought an important rise in shares of employees earning relatively low wages (Katz and Murphy (1992), Acemoglu (2000), Autor, Katz and Kearney (2008), Blau and Kahn (1996), Card, Lemieux and Riddel (2004); Atkinson (2007), Machin (2008) summarizes most of the previous studies). Changes in wage inequalities have been particularly pronounced in most of the Central and Eastern European countries during and after the economic transformation. With a few exceptions, wage disparities remain on relatively high levels in the region (Rutkowski (2001), European Commission (2004), Keane and Prasad (2006)). The incidence and structure of low wage employment is important for the overall assessment of the country s socio-economic development. On the one hand, concerns are raised that larger incidence of low pay may translate into higher poverty rates among the working population. On the other hand, the creation of many low paid, low skilled jobs may be an important way of reducing unemployment for the most vulnerable groups on the labour market. Increasing proportions of workers with relatively low labour incomes raise concerns also about their welfare and revitalize interest in redistributive policy measures, such as minimum wages. The evaluation of the role and significance of low wage employment is limited if one looks at its static picture only. Mobility of individuals throughout the earnings distribution may offset the high or increasing wage inequalities. Incorporation of earnings mobility is of particular importance for the analysis of low pay. If the low paid move up to better paid jobs over time, the observed inequality is shared among many individuals, and its social and economic consequences are thus less severe. If, on the contrary, the pool of the low paid does not change much over time and those at the bottom of the earning distribution tend to persist there, cross sectional measures of wage disparities reflect life-time inequalities, calling for adequate policy response. Moreover, while in many cases low paid jobs may 2

3 be instrumental in helping some groups of workers to enter the labour market (eg. young people or parents after longer breaks for family related reasons), and as such do not raise long - term equity concerns, they may also entail a risk of a so-called low pay trap, whereby those who undertake low paid jobs find it more difficult to move up the earnings ladder. Therefore, it is also important to look at the characteristics of those who manage and fail to escape low wages over time, so as to point out the features, that contribute to low pay persistence and should be tackled with appropriate policy measures. The share of those who manage to move out from low pay into higher earnings over one year hardly ever exceeds 30 per cent of those low paid in the initial period, whereas approximately half of them remain at the same positions a year after, and these results seem to hold strong in various studies for many countries. (Stewart and Swaffield (1999), Cappellari (2000), European Commission (2004)). Furthermore, there is little variation in those transition rates over time. 1 from two factors (or a combination of both of them). This aggregate state dependence in low pay may arise First of all, as described above, it may result from individual heterogeneity, i.e. a different distribution of observable and unobservable characteristics and skills among the two groups of low paid: those who manage to move up and those who remain in this state. In this case it would be (the lack of) some personal or job attributes that drives persistence in low pay. The second hypothesis would suggest that there is a structural (true/ genuine) state dependence (Stewart and Swaffield (1999), Capellari (2007)): the sole fact of being low paid in a particular point of time increases the likelihood of being low paid in the next period, regardless of all other characteristics. The main potential explanations of real state dependence that have been put forward focus on the signalling hypothesis and supply side issues. The signalling hypothesis assumes that employers in case of asymmetric information between them and applicants view employee s low pay spells as a signal for his lower productivity (Layard et al. (1991)) and 1 For instance, European Commission has shown that the yearly transitions rates between low pay, high and no pay differed only slightly in compared to

4 thus are discouraged from making him a higher wage offer. They may also treat a low paid job as an indicator of higher turnover propensity (Stewart and Swaffield (1999)). On the labour supply side, low wage employment may decrease the rate of accumulation of human capital (or even lead to a depreciation of the unused one), resulting in a lower individual productivity and worse earnings prospects in the future. Furthermore, it is often emphasized that the spells of low pay may undermine employee s self-esteem and discourage him or her from applying for better paid jobs and/or reduce their reservation wages. Some of the existing studies of low pay attempted to asses the extent of state dependence in low pay. Sloane and Theodossiou (2000) found evidence that low wage employment may become a trap, as being a worker at the bottom of the earnings distribution at one point in time increases, ceteris paribus, the chances of being low paid again in the future. What is more, the longer one remains low paid, the harder it is to escape the trap. Other authors attempted to quantify the degree of state dependence in low pay: for example, Gosling et al. (1997) found that only approximately 50 per cent of the low pay persistence could be explained by individual characteristics. Stewart and Swaffield (1999) estimated the contribution of structural state dependence to be considerable: according to their estimates over half of the difference (58 per cent to 76 per cent) in aggregate probabilities is due to the fact of having been low paid at t-1, ceteris paribus. Cappellari (2000) found that the true state dependence ranged from 40 to 70 per cent of the aggregate state dependence, Cappellari (2001) reported the genuine state dependence to account for approximately 53 per cent of the aggregate figures. Furthermore, looking at the state dependence in a longer perspective he found that it emerges mainly at the beginning of the low pay spell, whereas the contribution of the subsequent low pay experiences is less pronounced. Also Uhlenndorff (2006) have stated that in case of the Swiss and German employees, remaining low paid is mainly due to state dependence, and this effect clearly dominates heterogeneity. The aim of our paper is to provide an investigation of low paid employment in Poland, both from a static as well as a dynamic perspective. The paper starts with a description 4

5 of the data we use in our study, focussing on challenges and potential pitfalls relating to the measurement and definitions of low pay. Next we look at the patterns of incidence of low pay in Poland over time and provide an in-depth investigation of the individual and job place characteristics associated with low wages. Section 4 looks at the mobility out of low pay and analyses individual characteristics that make employees less likely to persist at the bottom of the earnings distribution. The last section concludes. 2 Data and measurement issues The main challenge relating to the analysis of low pay in Poland refers to the availability, quality and representativeness of data on wages. In our analysis we exploit three sources of information: Labour Force Survey (LFS), Household Budget Survey (HBS) and Structure of Earnings Survey (SES), conducted by the Polish Central Statistical Office. The LFS and HBS are surveys based on households and individuals reporting their current and past situation. The LFS is designed mainly to track changes in the employment patterns of the population, whereas the HBS focusses on the level and the structure of household and individual incomes and expenditures. The HBS is conducted yearly on a sample of approximately 33 thousand households, whereas the LFS surveys approximately 45 thousand individuals quarterly. In both surveys the individuals can be observed only for two consecutive years. 2 The Household Budget Survey seems to be a more accurate source of information on wage data than the Labour Force Survey. In particular, the wage reporting rates for employees in the HBS are higher. 3 The respondents in LFS also tend to round up data (as a result, there are high peaks of responses at round numbers, such as 1000 PLN). Furthermore, in the LFS, the higher earnings are underreported: the reporting rate for employees with higher education is lower than for those with lower levels of education 2 Due to changes in methodology in the LFS in 1999 and a lack of two waves of the survey, it is impossible to construct the 1998/99 panel and the 1999/2000 is limited in size. 3 In particular in the mid 2000s, when the response rates in the LFS have fallen to below 70 % from the initial 95% in mid 1990s. Almost 100 % of employees in HBS report labour income. 5

6 by a few percentage points, which leads to a lower level of the average wage. As a result, the LFS is slightly more biased towards lower earnings, its median to average ratio is 0.86, with 0.82 in the HBS. At the same time, the Labour Force Survey has many advantages over the Household Budget Survey: much more accurate labour market information, higher number of observations and, importantly, lower panel attrition rates among employees. The third source of data, the Structure of Earnings Survey, is an establishment-based survey of employees, which focusses on their demographic and occupational characteristics and their monthly wages (linked to information on their workplaces). This survey is conducted on a random sample of companies (excluding those which employ less than 10 people) 4 every two years (in particular: 1996, 1998, 1999, 2001, 2002, 2004, 2006). It is representative for ca. 6-7 million of employees. 5 Up until 2004, it covered full-time workers only. The earnings data is reported in gross terms and is in fact registered on the firm-level, which determines its main advantage (i.e. lack of problems with non-response, rounded responses etc.). However, its main drawback comes from the fact that it does not cover unregistered wages (i.e. both those of people with no formal job contracts and those, whose wages are under-declared in official registers) and wages of the employees in the smallest firms. This might be of particular relevance for the analysis of low pay, though it is impossible to assess the impact of this phenomenon. Given the advantages and drawbacks of each of these data sources, we use them complementarily in my analysis in order to fully assess the picture of low pay in Poland. In particular, we use the HBS, LFS and SES datasets. The wages are defined as monthly earnings net of deductions (social contributions and tax) 6 in the LFS and the HBS, and monthly gross earnings in the SES. As in the some of the HBS waves there is no information on hours worked (and the information available in the LFS may pose difficulties to derive reliable hourly earnings), we use monthly wages of full time 4 Less than 6 in the early waves. 5 It constitutes approximately % of all employees. 6 In the LFS, these are the earnings in the main job. 6

7 workers only. We also restrict the sample to employees aged Defining low pay To analyse the low wage employment one needs to define the particular group of employees concerned by it and determine the cut-off threshold, below which earnings can be assumed to be low. Many definitions have been suggested in the literature: low pay can be measured in relative or absolute terms 7, and it is not surprising that these definitions yield different outcomes, both of the incidence of low pay and the interpretation of its changes. The absolute measures may relate the pay to a previously determined level, for instance the social minimum and as such may be more useful in analysing the relations between low pay and poverty. However, the relative measures are more common, also because they are more appropriate for international comparisons. They usually relate the pay to the median pay in the economy or its percentage. Sometimes one uses also a definition based on a fixed percentage of employees from the wage distribution (for example, the first two quintiles). The choice of the method and the low pay threshold depends mainly on the subject of the analysis. If one wants to investigate poverty among the working population, the absolute measures are used more often, although it makes international comparisons more difficult. In case of the relative measures, the most common thresholds that are used in the empirical research refer to 60 per cent and 2/3 of the median wage (hourly, or less frequently, monthly) for full time workers and the latter will be used in our analysis. 8 It is important to emphasize that our choice of the cut-off points for low pay is not based on any strict level of labour income, that would draw the line between low and higher paid workers, as such a delimitation does not exist. We use different measures in order to get a coherent and robust assessment of the characteristics and patterns relating to low wages, they allow us also to compare our results to other findings in the literature, which 7 Subjective measures could be another alternative. They are used rather rarely, mainly in poverty studies, where respondents are asked about the adequate or minimum incomes. 8 The results of the analysis based on the former threshold (60 per cent of median) are available from the author upon request. 7

8 use similar levels of reference. 3 Low pay in Poland 3.1 The extent of low pay The existing research suggests that the extent of low pay varies considerably across countries. In some economies its incidence does not exceed 10 per cent (Sweden, Italy or Portugal) whereas in other (Poland being one of them) it concerns more than a fifth of the employees. The latest evidence presented by the European Commission (2004) has shown that low wage employment (defined as wages below 2/3 of median gross hourly pay) concerned approximately 15 per cent of the European workforce (in 13 of the EU-15 countries included in the analysis). Also the evolution of the incidence of low wages over the recent years provides a rather mixed picture (OECD (2006), European Commission (2004)). Some of the countries experienced a decline in its incidence between 1994 and 2004 (e.g. Spain, Portugal or Ireland, where it has fallen from a relatively high level). It has increased in others, both in those with low starting levels (Denmark, Sweden) as in those where it already exceeded several per cent (Korea, Hungary, Poland or the United Kingdom). The share of employees with low earnings has fallen in the Unites States (although only slightly), a country with the highest incidence of low pay in mid 1990s. Figure 1 plots the incidence of low pay in Poland against time dimension. We can see that it kept increasing since the transition, and the changes have been particularly abrupt between 1992 and 1994, and later, between 1998 and The determinants of low pay This section aims at determining the individual (personal, household - related and firmlevel) characteristics that impact the probability of being low paid. There are several econometrical problems that relate to the proper analysis of this phenomenon, which in- 8

9 Figure 1: The extent of low pay in Poland, , in %. Source: Own calculations based on SES data and Rutkowski (2001). clude also the choice of data used. Therefore, in our analysis we will make use of different types of models, basing both on cross-sectional and longitudinal data. These models are presented below and their results compared and summarised at the end of this section. We describe earnings in year t with the following equation: y it = β 0 + β x it + ɛ 1i for i = 1,..., n (1) where y i are the monthly earnings of full-time employees and x i is the vector of personal and job-place characteristics that impact the individual wages. The ɛ 1i are normally distributed with N(0,1) and independent of x. If we define the low-pay threshold as λ, an individual will be low-paid (y it = 1) if his earnings fall below it: { yit = 1 if y it < λ t y it = 0 if y it λ t 9

10 The probability of being low paid is thus given by: P [y it = 1] = P [y it < λ t ] = P [β 0 + β x it λ t < ɛ 1i ] = (2) = 1 Φ(β 0 + β x it ) = Φ( β 0 β x it ) Where Φ is a standard normal cumulative distribution function. The model can be estimated as: P [y it = 1] = Φ(βx it ) (3) with coefficients β j = βj and the intercept β 0. Basic model In the first five sets of estimates we analyse the characteristics associated with low wage employment basing on the LFS and SES datasets for In Model (1) the x (cf. equation 3) is a vector comprising the following explanatory variables: female, age and age squared, experience and experience squared, tenure, dummies for educational levels (tertiary, post-secondary, general secondary, basic vocational and primary; vocational secondary being the reference category); dummies for part time work, fixed term contracts, public sector, firm size, occupations held, sector of firm s main activity as well as regional dummies (for the 16 voivodeships). The main reason for the specification of this model was its comparability (in terms of available information) with a model based on the SES data (Model 3). Model (2), also based on the LFS 2006 data, includes additional explanatory variables (which are not available in SES and hence were not included in Model 1), i.e. individual s disability status; dummies for the size of place of residence in which he or she lives (countryside, smaller towns, big cities), dummies for household s main type of income. 10

11 Models (3), (4) and (5) are performed using the SES data. Model (3) aims at being comparable with Model (1), therefore the set of explanatory variables includes the same variables (although some of them differ slightly, e.g. the categories of firm size, owing to differences in the original datasets). In Model (4) additional regressors have been introduced, including a more detailed information on section of activity (on the NACE 2 level) and on type of collective bargaining within the company, in which the employees work. Model (5) is further extended and includes variables describing co-workers (such as the share of highly educated, females, younger and older workers within each employee s firm). The results of the estimations are presented in Table 1. As one could expect basing on the experience of other countries, also in Poland gender is one of the most important determinants of low earnings. Women are significantly more likely to be low paid, even if we control for occupations and sectors they are employed in. As for the employee s age, it has a significant, though non-linear impact on individual s earnings. Each additional year of age decreases the probability of low pay by 3 to 4% on average, but this rate becomes smaller for older workers. Length of service matters for the chances of higher earnings only to a small extent, similar to that of tenure within a particular firm. Not surprisingly, educational attainments have a significant impact on the probability of being at the lower end of the earnings distribution. The differences are not large among employees with basic vocational and primary (or lower) levels of education. However, these estimates may be biased due to the fact that those with lowest levels of education are most likely to withdraw from the labour market (cf. Bukowski 2005). It is interesting to see the remarkable difference between the two levels of tertiary education i.e. the bachelor and master s degrees. Employees with the latter are on average twice less likely to be low paid than their colleagues with bachelor s diplomas, other things being equal. It is also worth noting that the exclusion of the occupational dummies from the specification (Model 2) significantly increases the educational coefficients. It suggests that on one hand, adding the information on the type of job one performs may be a better proxy of skills one possesses, 11

12 Table 1: Estimated marginal effects of the probability of being low paid, basic specifications Model Model 1 Model 2 Model 3 Model 4 Model 5 Female 0.166** 0.193** 0.084** 0.070** 0.062** Age ** ** ** ** ** Age ** 0.000** 0.001** 0.001** 0.001** Age ** ** ** ** ** Education (ref: general secondary) tertiary (master) ** ** ** ** ** tertiary (bachelor) ** ** ** post-secondary ** ** ** ** secondary vocational * basic vocational 0.093** 0.158** 0.052** 0.048** 0.044** at most lower secondary 0.154** 0.273** 0.053** 0.052** 0.047** Experience ** ** ** ** ** Experience ** 0.000** 0.000** 0.000** 0.000** Tenure (within firm) ** ** ** Part time work 0.037** 0.027** 0.027** Fixed term contract 0.124** 0.105** 0.083** 0.077** 0.075** Public sector ** ** ** ** Firm size (ref: employees) 1-10 employees 0.059** 0.064** 0.062** 0.058** 0.055** employees ** ** ** ** ** more than 100 employees ** ** ** ** ** Disability 0.214** Household s main source of income (ref: unemployment benefits) Employment ** Work in agriculture ** Self employment ** Pension Having a second job Place of residence (ref: small towns) countryside medium sized towns large cities ** Collective bargaining (ref: none) supra-level work agreements ** ** firm-level agreements Co-workers characteristics Share of female 0.088** Share of young workers ** Share of older workers Share of workers with tertiary education ** Additional controls Regional dummies (16 ) yes yes yes yes yes Occupations yes no yes yes yes Sectors (NACE 1) yes yes yes no no Sectors (NACE 2) no no no yes yes No of observations Log pseudolikelihood R Standard errors clustered at the individual (LFS) and firm (SES) level significance: **1%; *5%; +10% Marginal effects calculated at means of the independent variables Firm sizes are different in SES data, with the following thresholds: (ref.), 11-19, , 250+ Source: Own calculations based on LFS and SES data

13 than the sole level of education. On the other hand, demand side factors (captured by occupational dummies) may also significantly determine the level of individual wages. It is interesting to see that job characteristics play an important role for the probability of being at the lower end of the earnings distribution. Firstly, part time workers are more likely to be low paid than their colleagues who work on a full time basis. This is partly related to the tax progression in Poland, as differences in net earnings between part and full time workers are smaller than in gross values (cf. Magda and Szyd lowski (2008)). Secondly, there exists also a remarkable wage penalty (8-12%) related to fixed term contracts. Temporary labour contracts are very often concluded with workers whose skills demand screening. They are also more likely to reflect fluctuations in demand and economic instability, which has an impact on wages offered by employers and by the propensity to accept them among employees. Employment in the public sector has a significant negative effect on the probability of being at the bottom of the earnings distribution. The estimated coefficients are lower in case of the LFS data, and one might expect that this discrepancy results from the differences in the datasets, rather than particular model specifications. 9 this issue in the next models we analyse. We will look at Other workplace characteristics we identified as playing a role for individual probabilities of low earnings include firm s size, its wage setting mechanism and type of workforce. Our analysis shows that employees in large companies have a much lower probability of being low paid than those in SMEs. 10 The same situation applies to employees covered by supra-level (mostly sectoral) agreements. These are present above all in large state-owned companies, suggesting that the existing agreements create wage floors and thus limit the incidence of low earnings. 9 Higher probability of low pay in the private sector in the SES data (compared to the LFS) might support the anecdotal evidence of underreporting wages in the private sector official data and under the table payments (mainly in smaller firms). 10 Again, the differences between the coefficients for small and large firms are much smaller in case of the LFS data, which might probably reflect the problem presented in the previous footnote. 13

14 We took advantage of the SES linked employer-employee data and identified additional workplace characteristics, which may have an impact on individual earnings. First of all, we find that a higher share of employees with tertiary education in a particular workplace significantly decreases the probability of low pay. This result is in line with other theoretical and empirical findings of spill over effects in firms, pointing to a wage premium related to the average level of education in the workplace (e.g. Battu et al. (2004)). 11 Secondly, younger establishments tend to decrease the probability of low pay, ceteris paribus. Again, this result may reflect unobserved heterogeneity in firm s productivity, as younger firms may either have been heavily restructured and / or privatised, or are green field investments, all of which may have a positive impact on work efficiency and wages. Also the share of females in the workplace turned out to play a role for the individual probability of low pay. This problem has been recognized in the literature. For example, Simon and Russell (2008) provide evidence on the existence of female segregation into low paying firms, which has an outstanding detrimental effect on female wages. Furthermore, most of the estimated coefficients on sector and occupation turned out significant for the probability of low earnings. Interestingly, some of the low-skilled occupations in fact decrease the probability of low pay, which is the case of e.g. metal and machinery workers or plant operators. Also regional differences appear to be an important determinant of low pay, reflecting the existing disparities in productivity and wages (cf. Magda and Zawistowski (2007)). We have also tested the robustness of the results against our third source of data on wages and estimated Model (1) and Model (2) on the HBS data. 12 Overall, the great majority of estimated coefficients in our basic model appear robust, regardless of the data and specification used. Moreover, most of these results correspond 11 It may also signal workers unobserved ability, under the assumption that workers with higher unobserved skills tend to have co-workers with higher average skill levels, see Card and de la Rica (2006). 12 The results are available upon request from the author. The HBS dataset allowed us to include fewer explanatory variables. Nevertheless it is clear that the patterns remain the same for the estimated coefficients (both in terms of size and direction of impact). 14

15 to findings for other countries. Models capturing unobserved individual heterogeneity The models presented above allowed for the identification of the significance and size of the impact of different observable variables on the probability of being low paid. However, if there are any personal characteristics that we are not observing in the data (or they are in general difficult to observe), that are relevant for the individual earnings, the estimated results may be biased. Models based on panel data allow to control for the time invariant unobserved heterogeneity of observations (employees in our case). At the same time, these models, having both time and cross sectional dimension, raise new econometrical requirements that have to be considered. Among other, the individual unobserved effects may be fixed or random across time (an issue which we discuss in detail below). One must also take into account the potential problems resulting from panel attrition. As within the framework of binary choice models, we assume that a worker has a low wage (y it = 1) if an unobservable latent variable y* is positive (y > 0), and the y* is defined as: y it = βx it + ν i + u it (4) ν i is the unobserved individual effect (i.e. time-invariant characteristics not included in the vector of regressors x i ), and u it is the random error. i = 1, 2,..., n ; t = 1, 2 There are two basic strategies of estimating panel data non linear models, i.e. treating the individual effects either as fixed or random ones. In the fixed effects model, the ν i are treated as an additional set of parameters. However, in our case, where the number of observations is large but the T (time dimension of the panel dataset) is small (two periods of time), this will result in an incidental parameters problem and inconsistent estimates Another fixed effects approach is based on a conditional likelihood (Anderson (1970), Chamberlain (1980)). However, this approach cannot be used in our data either, since it would in fact rely on only a small part of our sample, not large enough to draw reliable conclusions (the model is identified on the switchers i.e. those who change from low into higher pay). 15

16 In the random effects model the individual effects are treated as a random variable with a specified distribution (normal in case of a probit model) and as such eliminated by integrating them over this distribution. The key weakness of this approach lies in the assumption that the individual effects are independent of the regressors. We estimated a random effects probit model, its results are presented in Table 2. First of all, the estimate of rho is positive and statistically significant, which is an evidence of the presence of unobserved individual heterogeneity. Moreover, the individual effects turn out substantial, as they account for more than a half of the observed conditional variance. Regarding the estimated coefficients of the individual characteristics, one can see that their impact is substantially lower compared to the models based on cross sectional data. The differences are particularly striking when it comes to educational levels and gender. The effect is also reduced for those under fixed term contracts. Overall, the estimated impact of all the characteristics is much lower once we control for the unobserved heterogeneity between workers with the random effects specification, though one should keep in mind we made a strong assumption of no correlation between the unobserved and observed individual characteristics. To sum up, gender and education are important determinants of the probability of low earnings. A third, relatively strong effect found in our analysis is related to the fixed term job contracts, which significantly increase the probability of low pay among their holders. Each additional year of age, public sector employment and living in large cities were found to decrease the individual probability of low wages, though their impact is relatively small. The disability status, on the contrary, increases the chances of being at the bottom of the earnings distribution, which reflects the impact of health conditions (and / or their perception by employers) on individual wages. Finally, firm - level job characteristics also proved to play an important role for the individual chances of low earnings. Importantly, models based on cross sectional information only tend to overestimate the impact of personal and job level features on individual wages and low pay probabilities. 16

17 Table 2: Estimated marginal effects of the probability of being low paid, panel data model with random effects Model Model 6 Female 0.023** Age ** Age ** Age ** Education (ref: general secondary) tertiary ** post-secondary ** secondary vocational basic vocational 0.014** at most lower secondary 0.045** Tenure 0.00 Fixed term contract 0.031** Public sector ** Firm size (ref: employees)* 1-10 employees 0.009** employees * more than 100 employees ** Disability 0.062** Household s main source of income (ref: unemployment benefits) Employment ** Work in agriculture ** Self employment ** Pension * Having a second job Place of residence (ref: small towns) countryside medium sized towns large cities ** Additional controls Regional dummies (16 ) Yes Sectors (NACE 1) Yes No of observations Log pseudolikelihood Rho significance: **1%; *5%; +10% Source: Own calculations based on LFS data 17

18 4 Mobility out of low pay This section looks in more detail at earnings mobility of those at the bottom of the wage distribution in Poland. In order to investigate the potential state dependence in low pay we must in first place disentangle it from other processes that may possibly play a role for the persistence at the bottom of the earnings distribution, observed in the aggregate data. This requires a proper modelling of the dynamics of individual earnings. However, there are several potential caveats - in form of endogenous selection mechanisms - that need to be taken into account when analyzing transitions out of low pay. These include the problem of initial conditions and different sorts of sample retention (arising from panel attrition, employment retention or non response). Each of these processes leads to selecting a subset of the surveyed population, which may potentially be a non-random one. Endogeneity may arise, if the unobserved individual characteristics related to the selection mechanism (e.g. personal features that impact the probability of dropping out of the sample) are correlated with the unobserved individual characteristics in the main process, that is the transitions out of low pay in our case. 4.1 Model specifications Our empirical strategy is based on three steps. We start with the simplest approach, a conditional probit model. We will treat is as a reference, allowing us to verify the degree of the potential bias introduced by the endogenous selection mechanisms. In the next step, basing on an approach commonly used in the literature (suggested firstly by Stewart and Swaffield (1999)), we estimate a bivariate probit with endogenous selection, allowing us to control for the initial conditions. Both of these models account for transitions within the wage distribution only. In the last step, we extend the analysis, adding additional equations that allow us to control for panel attrition, employment retention and wage non-reporting. We believe this approach enables us to obtain, on one hand, a broad picture of low pay transitions in Poland, and on the other, results comparable to other existing analyses. 18

19 Conditional probit model Following the analysis of low pay in t (cf. equation 3) we assume earnings in year t+1 depend on whether the individual was low paid in year t. For individuals low paid in t (y it = 1) the earnings equation in t+1 is determined by: y it+1 = α 0 + α z it + ɛ 2i for i = 1,..., n (5) The vector α may differ for individuals who were not low paid in t, however, the distribution of the error terms is assumed the same for the entire sample. 14 The error terms from equations 1 and 5 have a bivariate standard normal distribution with correlation ρ. The probability of individual i to be low paid in year t and t+1 is therefore given by: P [y it+1 = 1 y it = 1] = Φ 2 (βx it, αz it, ρ) (6) with coefficients α j = αj and the intercept α 0. Φ 2 is a cumulative distribution function of the bivariate normal distribution. The conditional probability of being low paid in year t+1 given being low paid in t is given by: P [y it+1 = 1 y it = 1] = Φ 2(βx it, αz it, ρ) Φ(βx it ) We start with a basic assumption that ρ = 0, in which case the conditional probability of remaining in low pay can be modelled as a simple univariate probit, over the sample of those low paid in t. 15 However, this way we do not take into account the problem of initial conditions (we assume them to be exogenous), which most probably leads to a selection bias in the results, due to the correlation between the unobservables in the two models (for years t and t+1 ). The estimates of this model (Model 1) are presented in Table The characteristics in the z it vector are measured in t n order to avoid endogeneity between changes in individual characteristics and earnings. 15 A similar model can be constructed for those high paid in t. (7) 19

20 Bivariate probit model with endogenous selection In the second step we restrict the assumption on ρ, allowing it to be nonzero and thus control for the initial conditions. For that purpose, following Stewart and Swaffield (1999) we estimate a bivariate probit model with endogenous selection with partial observability, using information in year t+1 only for individuals who were low paid in t (Meng and Schmidt (1985), Maddala (1987)). The likelihood function for the individual i low paid in t and t+1 (equation 7) is given by: L i = Φ 2 (βx it, αz it, ρ) yty t+1 Φ 2 (βx it, αz it, ρ) yt(1 y t+1) Φ( βx it ) (1 yt) (8) The estimates of this model (named Model 2) are presented in Table 4. Accounting for panel attrition, employment retention and non response The above specification allows us to control for initial conditions, but does not take into account other potential biases due to sample retention, like the ones arising from panel attrition, employment retention or wage non reporting. We cannot assume that the either of them is random in our case and that no selectivity bias occurs in the estimates of low pay persistence based on Model 2. We undertake therefore an attempt to include these potential selection mechanisms in our analysis. There have been relatively few studies of labour market transitions which account for more than one selection process (Cappellari and Jenkins (2004, 2007)). These demand an estimation which takes into account more than one endogenous binary variable in binary response models, which is more complex compared to models, in which at least one of these is continuous. We decided to adopt a three- variate probit approach and add to Model (2) an additional equation, which allows us to take into account a second selection mechanism describing retention in our sample. 16 different definitions of sample retention: Below we estimate two models with 16 Accounting for more than one selection at the same time (i.e. having more than three outcome equations in the Model) would be computationally too demanding in our analysis. 20

21 t+1.: We define rit+1 as a variable describing retention in our sample between years t and r it+1 = γ 0 + γ w it + ɛ 3i for i = 1,..., n (9) The ɛ 3i are normally distributed with N(0,1) and independent of w. We define a certain threshold χ, below which employees drop out of our sample: { rit+1 = 1 if r it+1 < χ t+1 r it+1 = 0 if r it+1 χ t+1 Low pay status in t+1 can be observed only for those with a retention status equal to one (r it+1 = 1). Our final specification of the three- variate probit model with two endogenous selection processes is described by the following set of equations 17 : yit = β 0 + β x it + ɛ 1i y it+1 = α 0 + α z it + ɛ 2i (10) rit+1 = γ 0 + γ w it + ɛ 3i The error terms in the three equations (10) are assumed to be jointly distributed as a three- variate normal with unrestricted correlations, which allows for a parameterization of the unobserved heterogeneity (Cappellari 2007): ɛ 31 ɛ 31 ɛ = N 0 ; ρ ρ 31 ρ 32 1 The parameter ρ 21 describes the correlation between unobservables in the equation for 17 We assume these processes take place simultaneously and there are no nesting sequences between them. 21

22 retention (eg. 9) and in the initial low pay status (eq. 1), which shows whether the initially low paid are more likely to drop out of the sample; ρ 31 measures the correlation between unobservables in the retention and low pay transition equations, whereas ρ 32 measures the correlation between unobservables in the equations for low pay status in t and t+1 (thus it determines, whether the low paid are more likely to persist in low pay compared to those above the low pay threshold). The likelihood contribution for individual i can be written as: L i = Φ 3 (βx it, αz it, γw it, ρ 21, ρ 31, ρ 31 ) y t+1r t+1 Φ 3 (βx it, αz it, γw it, ρ 21, ρ 31, ρ 31 ) (1 y t+1)r t+1 Φ 2 ( γw it, βx it, ρ 21 ) (1 r t+1) (11) where Φ 3 is the three-variate normal distribution function. The evaluation of the above likelihood function requires computation of multivariate normal distributions. We perform it by using the GHK (Geweke-Hajivassiliou-Keane) simulator, which gives us a maximum simulated likelihood estimator (Lee (1997), Cappellari and Jenkins (2003)). 18 We experiment with two types of sample retention processes described by equation 9. In the first step (Model 3) we account for those who are not retained in the earnings distribution because they stop reporting wages in t+1. In this case r t+1 = 1 if an individual is a full time employee reporting his earnings in both time periods, and r t+1 = 0 if he remains within the group of full time employees over the year, but decides not to reveal his or her labour income in t+1. In the second approach we estimate a model (named Model 4) which accounts for those who are not selected into the balanced panel sample because they become unemployed, 18 To speed up the (heavy) computational processes we use Stata programs -mvdraws- and -mvnp-, developed by Capellari and Jenkins (2006). The first one produces Halton draws which can be next used by the -mvnp- to calculate multivariate normal probabilities using the GHK simulator (the detailed information and description of the programme can be found in the above publication and the accompanying stata ado file). We base our analysis on 100 Halton draws. 22

23 inactive or are not surveyed in t+1 (they drop out of the panel). 19 Therefore we will control for the fact that individuals who do not retain employment, and those subject to panel attrition (r t+1 = 0), may be a non-random sample of those low paid in t. 4.2 Results The presentation of the empirical results proceeds in three steps. Firstly, we present and discuss results of the conditional probit model assuming no endogeneity. Next we analyse results of the bivariate probit model with endogenous selection into the initial state of low pay. Finally, we contrast those results with the estimates of the three-variate probit, with different selection mechanisms considered. Conditional probit model We estimate the conditional probit model using a sample restricted only to those, who have been employed and reported wages in each of the two time periods. The variables included in the z t vector (equation 6) describe worker s personal characteristics (gender, age, level of education, his tenure in the firm, a dummy variable for being disabled), job characteristics (public or private sector, occupation on 1-digit SOC level), type of contract (fixed term or open-ended), firm s size, section of economy (agriculture, industry, business services and public services 20 ), and dummies for the region (voivodeship, in which it is located), as well as a dummy for variable describing whether the individual has a second job and whether he is seeking new employment. Table 3 presents the results of the conditional probit (Model 1) in terms of estimated coefficients (columns [1], [3], [5], [7]), and the marginal effects (columns [2], [4], [6] and [8]). Model 1 is estimated for the entire sample of (we report the estimates in 19 We experimented also with (1) merging the cases of employment retention and wage non response (and not accounting for panel attrition) and (2) accounting for all three selection mechanisms together. Not surprisingly, the results did not add interesting observations to those we present in section 5.5. (also because the individual attributes relating to the probabilities of dropping out of work and wage non reporting are different, e.g. the better educated are less likely to become unemployed and more likely to hide the information on their earnings). 20 Industry includes NACE sectors C - F, business services - NACE sectors G - K, public services - NACE sectors L - O. 23

24 columns [1] and [2]) and the three sub-periods (columns [3] - [8]), under the assumption that the coefficients α and β remain constant within the time periods (and ρ is equal to zero). The univariate conditional probit estimates suggest that being a woman or a disabled person significantly increase the probability of being low paid in t+1, given being low paid a year before, while being employed in the public sector (in particular in the period of economic downturn), as well as higher levels of education, reduce the probability of remaining at the bottom of the earnings distribution. Age is another significant determinant of keeping the low pay status, although interestingly, the estimates suggest the direction of that impact has changed over time. Many of the occupations in which one works are significant for the chances of escaping low pay. Some, like elementary occupations, significantly increase the probability of remaining in low paid employment (compared to being employed as a sales or service worker), other things being equal. The results suggest also that workers in large firms are less likely to remain low paid. Most of the estimated coefficients hold strong regardless of the time spell taken into account. Bivariate probit model with endogenous selection The above estimates are based on a strong assumption of ρ = 0 and no sample selection bias. This assumption is relaxed in Model (2), based on a bivariate probit distribution, with the likelihood function presented in equation (8). Vector z t includes the same variables as in Model (1). The x t vector includes characteristics from vector z t and additional variables, which can be used as instruments for the endogenous selection into the initial state of low pay. In other words, these are attributes, which have a significant impact on the probability of being low paid, but are not significant for the probability of being low paid conditioning on the state in the previous period (thus do not alter the chances of escaping or falling into low pay). These variables include the individual s marital status (being married or single), size of place of residence (rural areas, small towns, towns and cities), the household s main source of income (employment, self- 24

25 Table 3: Model 1: Low pay transitions (conditioning on low pay status in t) coeff (s.e.) mfx coeff (s.e.) mfx coeff (s.e.) mfx coeff (s.e.) mfx [1] [2] [3] [4] [5] [6] [7] [8] female 0.386** 0.105** 0.365** 0.133** 0.375** 0.095** 0.449** 0.103** (0.027) (0.033) (0.047) (0.048) education (ref: secondary) tertiary ** ** ** ** ** ** ** ** (0.069) (0.101) (0.117) (0.101) basic vocational 0.083** 0.026** 0.121** 0.046** ** 0.038* (0.03) (0.037) (0.051) (0.054) primary 0.230** 0.067** 0.366** 0.134** 0.142* 0.040* 0.160* 0.042* (0.039) (0.046) (0.071) (0.079) age * 0.002* (0.001) (0.002) (0.002) (0.002) tenure (0.002) (0.002) (0.003) (0.004) disability 0.325** 0.091** 0.250** 0.093** 0.364** 0.093** 0.263* 0.066* (0.069) (0.096) (0.118) (0.114) fixed term job * * * (0.036) (0.065) (0.058) (0.051) firm size (ref ) less than * 0.025* 0.169** 0.064** (0.036) (0.045) (0.06) (0.059) (0.036) (0.043) (0.066) (0.066) ** ** * * (0.03) (0.034) (0.054) (0.057) public sector * * * (0.033) (0.036) (0.064) (0.073) occupation (ref. Service and sales workers) Managers ** ** * * (0.118) (0.147) (0.21) (0.258) Professionals ** ** ** ** ** ** ** ** (0.07) (0.076) (0.148) (0.135) Technical, associate professionals ** ** ** ** ** ** (0.051) (0.057) (0.089) (0.104) Clerical workers ** ** ** ** ** ** * * (0.049) (0.054) (0.088) (0.091) Skilled agricultural workers (0.113) (0.125) (0.209) (0.274) Craftsmen * * * * (0.046) (0.053) (0.082) (0.09) Plant, machine operators ** ** ** ** ** ** (0.052) (0.063) (0.092) (0.092) Elementary 0.117** 0.036** 0.137** 0.052** (0.042) (0.049) (0.074) (0.079) Seeking a new job (0.074) (0.061) Having a 2nd job (0.038) (0.044) (0.069) (0.075) Constant 0.606** 0.337** 0.638** 0.629** (0.101) (0.095) (0.147) (0.136) Additional controls NACE sections yes yes yes yes yes yes yes yes year dummies yes yes yes yes yes yes yes yes Regional dummies yes yes yes yes yes yes yes yes Observations Observations weighted with sampling weights. Standard errors clustered at individual level Marginal effects evaluated for males with general secondary education, at age 35 and 3 years of tenure in the workplace, holding permanent contract, working in a small firm (11-49) in private sector as a service and sale worker 25