The Impact of Self-Employment Experience on Wages and the Risk of Unemployment

Transcription

1 The Impact of Self-Employment Experience on Wages and the Risk of Unemployment Michaela Niefert Centre for European Economic Research, Mannheim (competing for Young Economist Award) Abstract: The paper analyses the impact of self-employment experience on earnings in the wage and salary sector and on the probability of unemployment using a fixed effects model and a random effects probit model, respectively. The results reveal that women working in wage employment receive positive returns to self-employment experience. The returns for men may be positive or negative depending on the industry of self-employment and the current occupation. The probability of finding a job after leaving self-employment is generally higher for men than for women. In any case temporary self-employment seems to be a good alternative to unemployment for both genders. Keywords: Self-employment experience, wage and salary work, earnings, unemployment risk JEL-Classification: J23, J31, J64, C23

2 1 Introduction There is extensive empirical work on the determinants of earnings in the wage and salary sector (hereafter wage sector ) and the risk of unemployment. A lot of research has been dedicated to the labour market consequences of educational attainment, labour force attachment, job tenure and unemployment experience. Moreover, it exists a comprehensive literature on the transition from wage employment to self-employment. There are, however, still only relatively few studies on the transition from self-employment into other labour market states and on the returns to self-employment experience in wage work. This is in so far surprising as the experience gained from leading an own company presumably differs considerably from the experience acquired as an employee and could therefore have a different impact on earnings and employment opportunities in the wage sector. The present study wants to contribute to fill this gap. The question addressed here is in how far human capital specific to self-employment can be transferred to the wage sector and whether former self-employment works as a signal to employers. The paper analyses how self-employment experience is rewarded by employers in comparison with wage work or unemployment experience. It further dwells on the chances of those who give up self-employment to find wage employment. The topic is currently of special relevance since policy in Germany increasingly promotes self-employment as an alternative to unemployment by granting start-up assistance to unemployed people becoming selfemployed. For the evaluation of this policy it is important to know about the consequences of choosing self-employment over unemployment for those who return to the wage sector later on. The next section describes the possible influences that self-employment experience may have on wages and the risk of unemployment and relies on labour market theory. The relevant empirical literature is surveyed in section 3. Section 4 discusses some methodological issues. The econometric models used for the empirical analysis are explained in section 5. A description of the underlying data set is given in section 6. Section 7 presents the results, section 8 concludes. 2 Theoretical Considerations Labour market theory and research on the gender earnings gap give important hints to how experience in self-employment may affect earnings and employment prospects in the wage sector. There are several kinds of potential influences. Most of them can be attributed to hu- 2

3 man capital effects. Firstly, during the time being self-employed a person foregoes growth in human capital specific to the wage sector which she could have gained if she had worked as an employee. Secondly, possible wage sector-specific human capital which a person possesses when entering self-employment may depreciate during the absence from the wage sector. Thirdly, a self-employed person foregoes the accumulation of employer-specific capital which she could have acquired during working for the same employer. All these detrimental effects are supposed to occur also in case of unemployment or labour force intermittence (Mincer and Ofek 1982) and to reduce wages and the chances of finding employment. But different from those who are unemployed or out of the labour force, self-employed persons acquire work experience. The knowledge gained in self-employment can possibly be transferred to the wage sector, enhance job opportunities and be rewarded by an employer. The transferability presumably depends on the occupation, the sector of origin and the sector of destination. It may also be affected by company sizes and by the position which the former self-employed person if she finds a job takes in the new company. It could be hypothesised that returns to self-employment experience are the higher, the more similar the jobs in self-employment and paid employment are with respect to these terms. For example, one might think that what distinguishes self-employment experience are management capabilities which are particularly helpful in a leading position. These skills are certainly less useful in a low-qualified job at the lower scale of the company hierarchy. However, owners of very small firms in the service or retail sector can be supposed to have engaged not only in management tasks but also in the production process. So there is less difference between the work of the owner and of the employees in these firms, and a self-employed person can easily transfer its knowledge to the wage sector. Independent of the actual effects of self-employment on human capital, the fact of having been self-employed may work as a signal to employers. These considerations are based on statistical discrimination theory which has been used to explain the gender wage differential. Since employers have only limited information about the potential employee, they tend to ascribe to her the pertinent characteristics of the group she belongs to. In the case of selfemployed workers employers must probably rely even more on this strategy because they do not have any information from references of previous employers for the periods of selfemployment (Trzcinski 1999). If employers consider the characteristics of self-employed workers to be less advantageous than those of workers who have exclusively been employed in the wage sector, they are probably not apt to employ them or they pay them less. Employers might stigmatise the event of self-employment, particularly if the potential employee has 3

4 been unemployed before entering self-employment. But they might also attribute qualities like entrepreneurial spirit and willingness to perform to the self-employed. To summarise, the relationship between self-employment experience, earnings in the wage sector and risk of unemployment is determined by human capital and signalling effects. The direction of these effects is not clear ex ante. They can be assumed to depend very much on the characteristics of the job in self-employment as well as of the job (aimed at) in paid employment. 3 Survey of Empirical Literature The empirical evidence regarding the wage returns to self-employment experience is rather mixed. Evans and Leighton (1989) do not observe any significant differences between the returns to wage employment and self-employment in the wage sector analysing the National Longitudinal Survey (NLS) of Young Men. They conclude that workers leaving selfemployment return to wage work at roughly the same wages they would have received if they had not tried self-employment. They note, however, that it is unclear whether this is due to the value of business experience or because those with the best wage opportunities tend to switch. Using the NLS Youth Cohort and the NLS of Young Women, Williams (2000) finds smaller returns to self-employment as compared to wage employment only for women. The result does not hold for women working in sales occupations. In a further analysis of the Youth Cohort, Williams (2004) discovers that youth self-employment experience in contrast to wage employment experience is not rewarded in the wage employment market at age 27. Bruce and Schuetze (2004) observe smaller, negative returns 1 to self-employment for both genders in their analysis of the Panel Study of Income Dynamics (PSID). They explain the difference between their results and those of Williams (2000) by the fact that they focus on short-term self-employment, namely on those who are wage workers at the start and the end of a fiveyear-period. This would imply that short spells of self-employment have a more detrimental effect on wage earnings than longer ones. The returns to self-employment, however, are larger than the returns to unemployment. Thus, the negative labour market consequences associated with unemployment seem to be more severe than for self-employment. Trzcinski s (1999) study on basis of the PSID and the German Socioeconomic Panel (GSOEP) is partly contradictory to the findings of Bruce and Schuetze. Controlling for time spent in unemployment and out of the labour force she can detect a smaller, negative effect of 1 The sign of the self-employment experience coefficient depends on which other employment states are included in the regression. Bruce and Schuetze control also for the years spent in unemployment. 4

5 self-employment experience in comparison with experience in wage work only for American women, not for American men. In Germany, in contrast, this result is only valid for men. For women, the coefficient of self-employment experience is negative but insignificant. In addition to experience, Trzcinski uses the number of spells spent in self-employment as an explanatory variable in order to capture the signalling effect. The corresponding coefficient turns out to be positive for American women and German men. It seems that for these two groups self-employment deteriorates the human capital relevant for wage work, but that at the same time frequent self-employment spells work as positive signal to employers. Trzcinski concludes that these individuals fare better when moving frequently into and out of selfemployment than when spending extended periods of time in self-employment. The returns to self-employment seem to be a mixture of human capital and signalling effects. Her study also indicates that the consequences of self-employment are less detrimental than those of unemployment and time out of labour force. Williams (2003) also uses data from the GSOEP but does not differentiate by sex. He finds that the return to self-employment experience is lower than the return to wage employment experience. Not controlling for other labour market states, however, the return is still positive. Bruce and Schuetze s (2004) study is the only one known to the author which deals also with the effect of self-employment experience on the chances to find wage employment. Focussing on a sample of individuals who are wage workers at the start and not self-employed at the end of a five-year-period, Bruce and Schuetze find small positive effects of self-employment experience on the probability of unemployment which are insignificant for most periods. In any case, self-employment experience increases the risk of unemployment by far less than unemployment experience. To sum up, there is no clear picture whether the wage returns to self-employment experience are lower than those to wage employment experience. Some of the results support this hypothesis. The effects seem to differ by sex, country and length of self-employment spell. The finding that self-employment experience produces higher returns than time spent in unemployment or out of the labour force, however, is consistent across all studies. There is only very weak and scarce evidence that self-employment experience could increase the probability of unemployment for those who want to return to the wage sector. 4 Methodological Issues There are three potential sources of bias when estimating an earnings function. Firstly, wages typically depend on factors like motivation or talent which are unobservable. If these factors 5

6 are correlated with the explanatory variables used in the regression, the estimated coefficients will be biased due to unobserved heterogeneity. For example, unobserved characteristics which are relevant for the level of earnings in the wage sector might also influence the choice of self-employment over paid employment. In this case, individuals select non-randomly into the different employment states. The estimated effect of self-employment experience on wages then is biased if it is not controlled for this non-random selection. Secondly, the variables related to the employment history may be endogenous. Wages are likely to affect labour force participation and unemployment. They might also influence the choice of selfemployment as an alternative to wage employment. As a result, the estimation will be fraught with an endogeneity bias and it is not clear whether the coefficients reflect the effect of the regressors on wages or the reverse. Both types of bias may also occur when estimating the probability of unemployment. Unobserved factors affecting wages are likely to influence the unemployment risk, too. In addition, the probability of unemployment may exert an influence on the decision to try self-employement. Thirdly, there is the sample selection problem. The wage equation can only be estimated for individuals in paid-employment, not for those who are in one of the other labour market states. As a consequence, the effect of self-employment experience, for example, can only be estimated for those who leave self-employment and enter the wage sector later on. These self-employed persons might be a non-random group of the self-employed. Individuals who remain in self-employment and therefore do not enter the earnings estimation might have had higher or lower wages upon returning to the wage sector. The self-employment-related coefficients then only reflect the labour market consequences of self-employment for those who leave self-employment and not of self-employment in general. The empirical work so far has only addressed the heterogeneity and endogeneity issues. In an attempt to correct for non-random selection into self-employment, Williams (2000) and Bruce and Schuetze (2004) include the wages received by self-employed workers before leaving the wage sector into the regression, thereby accounting for a possible lower productivity of these workers. Their results suggest that in some cases self-employment experience has a more positive effect on wages when controlling for such productivity differentials. This would imply that individuals negatively select into temporary self-employment. Evans and Leighton (1989), Trzcinski (1999) and Williams (2003, 2004) use a standard two-step procedure to control for selectivity. Only Trzcinski and Williams (2003) are able to detect any correlation between selection into self-employment and earnings. Williams selectivity-adjusted estimate of the return to self-employment is of about the same magnitude as the simple OLS estimate. To correct for endogeneity bias, Williams (2003) applies an instrumental variables approach 6

7 using a GMM estimator. But he only instruments the education variable, not self-employment experience. All of the studies although using panel data apply only cross-sectional methods to investigate the effect of self-employment experience on wages. Thus they miss the opportunity to control completely for constant unobserved individual effects which would reduce all the biases considerably (see next section). Moreover, none of them controls for sample selection bias. Evans and Leighton (1989), Williams (2000) and Bruce and Schuetze (2004) admit that the methodology they use does not take into account the possible selectivity emanating from observing only those self-employed workers who left for paid employment. In an attempt to address these methodological problems more fully and to reduce all three kinds of biases, the present paper will apply a panel data method to account for heterogeneity and use a procedure correcting for sample selectivity. 5 Econometric Model 5.1 Fixed-Effects Estimation of the Wage Equation The analysis of wages is based on the typical Mincerian earnings function which has been adapted to the panel data context by Kim and Polachek (1994): w = β + α + ε, i = 1,2,..., N t = 1,2,..., T. 1 it x it i it i is an index for the individual, t is an index for the year. w is the logarithm of monthly wages deflated by the consumer price index (1990=100), x is a vector of time-varying regressors, α is a vector of unobservable individual, time-constant effects, and ε is the error term reflecting time-varying unobservable factors. The heterogeneity and endogeneity bias are closely interrelated. Both the omission of variables and a simultaneity between the explanatory variables and the dependent variable lead to a correlation of the error term with the regressors, so that the exogeneity assumption E( α ) = 0 (in the case of heterogeneity) respectively E( α ε x ) = 0 (in the case of i x it i + it it endogeneity) is violated. Heterogeneity usually implies endogeneity (Kim and Polachek 1994, Polachek and Kim 1994). Applying panel data estimation techniques decreases the importance of both biases. Panel data allow to transform equation 1 by mean deviation so that the unobserved constant effect α i nets out: w it w = x x ) β + ( ε ε ), 2 i ( it i it i 7

8 and the exogeneity condition relaxes to E( ε ε x x ) = 0. Estimating equation 2 by OLS i corresponds to the fixed-effects approach and eliminates the heterogeneity bias as far as it arises from time-constant unobservables. Likewise, the endogeneity problem diminishes because invariable individual factors affecting wages are controlled for so that they cannot exert any influence on the regressors via wages. Since employment history and wages can be supposed to be largely determined by individual factors like motivation or talent, the biases are presumably reduced considerably by applying this panel data approach. Sample selection, the third type of bias, is caused by the fact that i it i w it can only be observed if individual i is working in the wage sector. Selection into the wage sector can be described by the probit model * sit = X it γ + υ t it υ it X it ~ Normal(0,1) 3 s = 1 if s 0, and 0 otherwise. it * it > s it takes value 1 if i works in the wage sector in period t. X should include at least one significant explanatory variable which is not part of x in order to avoid collinearity among the regressors. The sample selection problem arises if υ is correlated with ε in equation 1 so that E( ε, υ ) 0. Using the fixed effects approach reduces the sample selection bias to it x it it the extent that the selection process is determined by time-constant factors. However, the decision and possibility to work as a dependent employee are probably also influenced by timevarying factors like household income, family background, regional economic situation or existence of a promising self-employment alternative. Therefore, an extension of the two-step selection correction procedure by Heckman (1979) to the panel data context is used. This is done by estimating equation 3 for each t and calculating the inverse Mills ratio λˆ it for all i and t (Wooldrige 2002). Then equation 2 is estimated by OLS using the Mills ratios as additional regressors: w it w = x x ) β + ρ d1 ˆ λ ρ ˆ λ + ( ε ε ) 4 i ( it i 1 t it T it it i where d1 t through of H : ρ 0 for t = 1,..., T. 0 t = dt t are time dummies. Sample selection bias can be tested by a joint test 5.2 Random-Effects Probit Estimation of the Probability of Unemployment The probability of unemployment is estimated by a random effects probit model. It can be written 8

9 * y it = X it β + uit u it X it ~ Normal(0,1) 5 where y = 1 if y 0, and 0 otherwise, it * it > y it is an indicator of unemployment. u it can be decomposed into α + ε, where the a i and ε it are random drawings which are assumed to be independent and normally distributed, both with mean zero and respective variances 2 σ α and σ 2 ε i it. The Butler and Moffitt (1982) formulation of the model restricts the correlation of the u it to be equal across different time periods in order to avoid problematic computation of joint probabilities from a T-variate normal distribution. The likelihood function then reduces to a single integral which can be evaluated by Gaussian quadrature. The correlation of the error term over time is given by Corr [ u it, u is ] = ρ = σα /( σ ε + σα ) for t s. 2 Accounting for the serial correlation of observations on the same person reduces the variance as compared to the pooled probit model. It does not eliminate a possible heterogeneity or endogeneity bias, though, since it assumes the α and X to be uncorrelated. The random-effects probit model is still preferred to the fixedeffects logit model because the latter would lead to a drop out of all individuals which do not change between employment and unemployment during the observation period. This would imply a heavy loss of information. 6 Data Description The data base used for the empirical analysis in section 7 are the first 20 waves of the German Socioeconomic Panel (GSOEP). 3 The GSOEP is a longitudinal survey of private households and persons which started in 1984 and included 12,245 persons in the first wave. It provides annual information on various individual and household characteristics, e.g. in the areas of population and demography, education, labour market and occupational dynamics, and earnings and income. Its monthly employment and income calendars provide an informative database on the duration in various labour market states. Several extensions of the GSOEP have taken place since In the 20 th wave (2003) the GSOEP comprises nearly 24,000 persons. All available waves ( ) and samples (A-G) are used for the empirical analysis. Peri- 2 For more details on the estimation see Greene (1997). 3 See for a detailed description of the GSOEP. 4 Starting with a West German (A) and a foreign sample (B) (the latter consisting of the so-called Gastarbeiter ) in the first wave, an East German (C) sample was added after the German reunification in 1990; an immigrant sample (D) containing persons who immigrated into Germany after 1984 started in 1994/95, a refreshment (E) and an innovation (F) sample were selected in 1998 and 2000, respectively, and a high income sample (G) was drawn in

10 odical information, calendar information as well as life history information contained in the GSOEP has been exploited to compile the work history of individuals. The panel estimations comprise the waves 1990 to Variable definitions are given in the appendix. Table 1 gives some descriptive statistics by employment history and sex using the most recent wave of the GSOEP. It compares those who are currently employed in the wage sector and have never been self-employed ( always paid-employed ) with those who are currently selfemployed and have never been paid-employed ( always self-employed ) and those who are currently paid-employed but have once been self-employed ( formerly self-employed, now paid-employed ), further differentiating between men and women. The relative quantities of these groups reveal that, while there are only half as much females continuously selfemployed as males, there are at least as much females choosing temporarily self-employment as males. The formerly self-employed are older than the always paid-employed but younger than the always self-employed. The always self-employed are married more often and have more children on average than the always paid-employed. The same holds for formerly selfemployed females, but to a lesser extent. Wage employment apparently makes it more difficult for women to have a family than self-employment. This finding is consistent with studies by Boden (1996), Connelly (1992) and MacPherson (1988) according to which women often choose self-employment because it provides them with more flexibility in combining market work and household production. Women who have worked only in the wage sector have spent a somewhat higher proportion of their time in the labour force in unemployment than men. For women who were formerly self-employed, however, this is not the case. Formerly self-employed employees have spent more time in unemployment than always paid-employed workers. Their working careers seem to be more unstable because they experience more spells of all kinds of employment states. This result holds for both genders. Although formerly self-employed women do not seem to be disproportionately hit by unemployment when looking at the sample of wage workers, they have a relatively high probability of unemployment based on an analysis of the full sample of women in the labour force excluding the self-employed. The unemployment rate among females with self-employment experience amounts to 17.9% whereas it is only 12.3% for those without self-employment experience. There is no such difference for males; their corresponding unemployment rates are 14.1% and 13.2%. The always self-employed work more hours per week than both groups of paid-employed workers. Formerly self-employed males work in smaller companies on average than always paid-employed males, maybe because smaller companies are more similar in structure to the 10

11 company they led when they were self-employed and, consequently, the match between qualification and job requirements is higher. For females this relation is less apparent. Continuously self-employed people earn considerably more than paid-employed individuals. Formerly self-employed men earn little more than men working in the wage sector, whereas women have a lower income than their always paid-employed counterparts. Both the proportion of employees with a university degree and of those without any professional qualification is higher among the formerly self-employed than among the always paid-employed workers. Formerly self-employed women are less qualified than formerly self-employed men. Interestingly, formerly self-employed males are working in the occupation they are trained for more often than those who were always paid-employed whereas for females holds the reverse. Table 1: Descriptive Statistics by Employment History, GSOEP 2003, weighted mean always paid-employed always self-employed formerly self-employed, now paid-employed male female male female male female total age married kids partner_full mon_unemp mon_wage mon_self spells_unemp spells_wage spells_self hours comsiz20_ comsiz200_ comsiz income (DM) university no degree occ_trained voctrain_req college_req occ_exec occ_scien occ_tech occ_office occ_serv occ_craft occ_mach occ_unskil

12 Formerly self-employed females have less often a job for which a college degree is required as compared to males and to always self-employed females. In contrast, the highest proportion of males having such a job is found among the formerly self-employed. All in all, paidemployed women who have once been self-employed earn less, are less qualified and work less often in the occupation they are trained for than men with a comparable employment history. Among the always paid-employed and the always self-employed this difference by gender is less pronounced. This is also confirmed by the distribution by occupation. The share of executives among the always paid-employed is much higher for males than for females. But among the formerly self-employed this difference is still by far larger. However, as compared with their female colleagues who have never been self-employed, it cannot be said that formerly self-employed women have on average less prestigious jobs. They are more often office workers and craftsmen, but they also work more often as a scientist and work less often as a machinist or unskilled worker. Moreover, they have more often a job for which a college degree is required. Consequently, their job quality does not compare as unfavourable with always paid-employed women as with men who were also formerly self-employed. The distribution of self-employment by industry sector is given in table 2. It differentiates between the currently self-employed and the currently paid-employed who were selfemployed in their last spell (where the industry sector refers to this last spell of selfemployment). Comparing currently self-employed males and females, the most striking differences are that women run less often a business in agriculture, construction and business services but engage more often in personal services, education, health and social work. Thus women tend to be self-employed in less-rewarding industries than men. Comparing currently self-employed males with formerly self employed males it shows that men relatively seldom give up a business in agriculture, business services, and health, education and social work whereas they leave relatively often the manufacturing, trade and personal service sector. In contrast, women leave disproportionally seldom manufacturing and services. They often leave business services and most of all the retail sector. 43% of all formerly self-employed females come from this sector. In absolute terms, self-employed males leaving for the wage sector mostly worked in manufacturing, business services, construction and retail trade (in descending order). Females mostly come from the retail and business services sector. 12

13 Table 2: (Former) Self-Employment by Industry Sector, GSOEP 2003, weighted share (%) self-employed paid-employed, last spell self-employed male female male female agriculture, hunting, forestry; fishing manufacturing construction wholesale and intermediate trade retail trade transport, storage and communication financial intermediation business services personal services education, health and social work Empirical Results 7.1 Impact of Self-Employment Experience on Wages Table 3 shows the results of the fixed-effects estimation of the earnings function without and with selection correction (equation 2 and 4, resp.) for males and females. Besides other regressors typically used in earnings functions, the estimation contains the months and spells spent in different labour market states including self-employment. Following Trzcinski (1999), the coefficients referring to the months are interpreted as human capital effect whereas the coefficients referring to the spells are interpreted as signalling effect. The human capital effect can be supposed to depend on the time spent in a labour market state. The signalling effect probably depends rather on the fact whether an individual has entered a labour market state and, if so, how often. Further, industry and occupation dummies are included and some variables characterising possible former self-employment. These are income in last selfemployment (as an indicator of success), the time since last self-employment, size of last selfemployment, and various interaction terms between the industry sector of last selfemployment and the current occupation. In addition to these variables, the probit equations which are estimated in the first step of the correction procedure include some wealth-related variables, family-related variables and information on household income (results not reported). Firstly, the results of the estimation without selection correction are considered (column 2 and 4). Comparing males and females, it turns out that being married increases men s earnings and decreases women s earnings, the latter being also negatively influenced by the number of children in the household. Being a single reduces men s wages. An explanation might be that 13

14 married women with children specialise in household production whereas married men specialise in market work. Empirical evidence by Coverman (1983) and Hersch and Stratton (1997) indicates that housework and hourly earnings are negatively correlated. It corresponds to Becker s (1985) effort/energy hypothesis according to which devoting energy to household production necessarily reduces the amount of effort available for market work. Living with a partner who is full-time employed decreases the earnings of both genders. As to be expected, earnings increase with age, hours worked, tenure and company size. Working in the occupation trained for and having a job for which vocational training or college is required also has a positive effect on wages. The time spent in unemployment has a detrimental influence, whereas wage sector experience increases earnings (at a decreasing rate). The coefficients of the training and wage sectors experience variables are somewhat larger for women than for men indicating that the returns to education, training and work experience are larger for women. The coefficient of months spent in self-employment is negative for males and positive for females but insignificant in both cases. According to the theoretical considerations in section 2 one might conclude that self-employment experience does not increase human capital specific to the wage sector but that this capital does not significantly depreciate during selfemployment, either. It is also possible that both effects cancel out each other. Consequently, time spent in self-employment has not such a detrimental effect on wages as time spent in unemployment. As the spells coefficients indicate, the signalling effect of self-employment is positive for females, particularly if the last spell has been self-employment. For males the signalling effect seems to depend on the number of spells of self-employment. Basically, having been self-employed in the last spell has a positive effect. However, the overall signalling effect becomes negative if the individual has experienced more than two selfemployment spells. This would imply that employers associate frequent self-employment spells with negative characteristics for men but with positive characteristics for women. This result conflicts with Trzcinski s (1999) finding that frequent self-employment episodes have a more beneficial effect on German men s earnings than extended periods of self-employment. For both genders the positive effect of having been self-employed in the last spell becomes the smaller the longer the spell is ago. The income earned in self-employment and a company size in last self-employment larger than nine indicators of success and scope of former selfemployment increase only the wages of men. 5 5 The company size variable drops out of the estimation for females because there are only very few observations for women with more than nine employees. 14

15 The industry and occupation dummies contribute significant explanatory power to the estimations. In addition, interaction terms between the industry of last self-employment and current occupation give important insights into the relationship between self-employment experience and wage level. 6 Apparently the effect of self-employment depends heavily on the sector the former self-employed operated in and the occupation she has subsequently in the wage sector. Particularly for males the coefficients of the interaction terms are by far larger than the main effect of self-employment and may reverse the overall effect. The returns are especially high for them if they were self-employed in the business service sector or to a lesser extent in the construction sector. It seems that returns are the larger the more qualified the occupation in the wage sector. For example, the positive effect of a self-employment spell in business services is largest for executives, followed by scientists and technicians. For males coming from manufacturing the returns are mostly negative; however, they are positive if males are occupied as scientists. The result corroborates the hypothesis put up in section 2 that management capabilities acquired in self-employment can best be transferred to the wage sector if a leading position or a job implying a certain responsibility is achieved. For females the interaction effects are quite different. The relationship between returns and occupation is less clear. Former self-employment can have a positive effect on wages even in a lower-qualified occupation (e.g. for craftsmen or unskilled workers who were self-employed in manufacturing). For females coming from the retail sector, however, returns depend positively on the qualification level of the occupation: Returns are positive for technicians and office workers and negative for unskilled labourers. The presumption made in section 2 that self-employment experience in the retail or service sector might be particularly useful for related (i.e. service) occupations in the wage sector is not confirmed. However, it is supported by the apparent positive impact of self-employment in the service sector on wages in service occupations. Turning to the estimations with selection correction (column 3 and 5), the joint significance of the inverse Mills ratios indicates the presence of sample selection bias. Although the results do not substantially alter as compared to the estimations without selection correction, there are some changes in the magnitude of the self-employment related coefficients. The negative coefficient of spells of self-employment for males is no longer significant now. Thus there is no more evidence that frequent self-employment spells would work as a negative signal to employers and decrease wages. But the positive effect of self-employment in the last spell 6 Some interactions drop out of the estimation because the corresponding transition from self-employment into the wage sector cannot be observed. In addition, interaction terms are only included if they are significant in at least one of the estimations. 15

16 Table 3: Fixed-Effects Estimation of Earnings Function males females selection correction no yes no yes married *** *** *** *** [0.0052] [0.0052] [0.0069] [0.0069] single * * [0.0055] [0.0055] [0.0072] [0.0072] kids * *** *** [0.0020] [0.0020] [0.0035] [0.0035] partner_full *** *** ** ** [0.0030] [0.0030] [0.0052] [0.0052] age *** *** *** *** [0.0036] [0.0036] [0.0032] [0.0033] age *** *** *** *** [1.92e-05] [1.95e-05] [2.69e-05] [2.73e-05] hours *** *** *** *** [0.0003] [0.0003] [0.0003] [0.0003] tenure *** *** *** *** [0.0003] [0.0003] [0.0004] [0.0004] comsiz20_ *** *** *** *** [0.0044] [0.0050] [0.0058] [0.0065] comsiz200_ *** *** *** *** [0.0052] [0.0058] [0.0067] [0.0074] comsiz *** *** *** *** [0.0057] [0.0062] [0.0073] [0.0079] occ_trained * *** *** [0.0035] [0.0035] [0.0057] [0.0057] college_req *** *** *** *** [0.0070] [0.0070] [0.0116] [0.0116] voctrain_req *** *** *** *** [0.0036] [0.0037] [0.0057] [0.0057] mon_unemp *** *** *** *** [0.0005] [0.0005] [0.0006] [0.0006] spells_unemp [0.0040] [0.0040] [0.0061] [0.0061] mon_wage *** *** *** *** [0.0003] [0.0003] [0.0002] [0.0002] mon_wage2-3.13e-06*** -3.13e-06*** -4.71e-06*** -4.76e-06*** [1.52e-07] [1.53e-07] [2.59e-07] [2.62e-07] spells_wage *** *** *** *** [0.0029] [0.0030] [0.0040] [0.0041] mon_selfem * [0.0007] [0.0007] [0.0012] [0.0012] spells_selfem ** ** ** [0.0126] [0.0127] [0.0141] [0.0142] lastsp_self * *** *** [0.0330] [0.0331] [0.0464] [0.0464] ylast_self *** *** *** *** [0.0352] [0.0352] [0.0531] [0.0531] income_self *** *** [0.0021] [0.0021] [0.0033] [0.0033] comsiz_self * * - - [0.1015] [0.1014] manu_scien * [0.1454] [0.1452] manu_tech ** ** [0.0906] [0.0905] [0.2863] [0.2860] manu_craft ** *** *** *** [0.0681] [0.0681] [0.1286] [0.1285] manu_unskil * * [0.0881] [0.0881] [0.1374] [0.1373] 16

17 cons_tech *** *** - - [0.1361] [0.1359] cons_serv *** *** [0.0925] [0.0926] [0.2336] [0.2334] cons_craft ** ** - - [0.0743] [0.0745] cons_mach * * - - [0.1054] [0.1053] retail_tech *** *** [0.0868] [0.0874] [0.1432] [0.1431] retail_office *** *** *** *** [0.2372] [0.2378] [0.1190] [0.1189] retail_serv [0.0865] [0.0865] retail_unskil ** ** [0.1039] [0.1043] buserv_exec *** *** [0.2451] [0.2451] [0.2092] [0.2091] buserv_scien ** ** - - [0.2192] [0.2191] buserv_tech * [0.1918] [0.1919] [0.1312] [0.1311] serv_serv *** *** * * [0.0740] [0.0739] [0.1327] [0.1326] serv_craft *** *** [0.1462] [0.1460] [0.2180] [0.2177] industry dummies *** *** *** *** occupation dummies *** *** *** *** tmills90-tmills03 - *** - *** constant *** *** *** *** [0.0935] [0.0967] [0.0842] [0.0883] number of observations number of individuals R 2 within R 2 between *** (**,*) indicates a significance level of 1% (5%, 10%); standard errors in brackets. loses significance, too. Noteworthy signalling effects can yet only be observed for specific interactions between industry of self-employment and occupation. For females, the coefficient of months of self-employment increases slightly and becomes weakly significant, indicating that self-employment increases human capital specific to wage employment. Altogether the general effect of self-employment experience as revealed by the estimations with selection correction seems to be more favourable than the one obtained from the estimations without selection correction. The coefficient changes can be explained by the non-consideration of two groups in the latter estimations: Firstly, unemployed workers and individuals who are not part of the labour force are not included. They might self-select into these employment states due to low earnings prospects in the labour market. If these inactive individuals as to be supposed have less self-employment experience than average, the returns to self- 17

18 employment will be underestimated. Secondly, self-employed workers are not part of the wage regressions. The estimation of the impact of self-employment on wages relies only on those self-employed persons who leave self-employment for paid employment. If the other self-employed workers, who could be assumed to be more successful in self-employment because they are still in this employment state, are also more successful in wage employment the returns to self-employment will be underestimated, too. The regression results confirm at least for men that success in self-employment (as measured by the income in selfemployment) has a positive impact on earnings in the wage sector. It should be noted, however, that there is evidently no general negative selection out of self-employment into wage employment for all interactions between industry of former self-employment and occupation. The contrast between the favourable general effect of self-employment on wages of females according to the multivariate analysis and the relatively low earnings of formerly selfemployed women as revealed by the descriptive statistics can be explained as follows. The multivariate analysis accounts for being married and having children something which is particularly prevalent among (formerly) self-employed women, usually implies a specialisation into housework for them, and tends to reduce their earnings. It further controls for the fact that women who change from self-employment to paid employment are less qualified on average, less often exert the occupation they are trained for and less often have a qualified job than men. Empirical evidence suggests that women select negatively into self-employment whereas selectivity is positive for men (Clain 2000). This possible selection bias is largely corrected by the fixed-effects approach. In addition, the selection correction procedure accounts for selectivity out of self-employment which seems to affect the general returns to selfemployment negatively. After considering all these issues it is possible to isolated the pure effect of self-employment on wages. 7.2 Impact of Self-Employment Experience on the Probability of Unemployment The results of the random effects probit estimations of the probability of unemployment (equation 5) are given in table 4. In addition to the explanatory variables used in the fixed effects estimation, some time-constant variables are included in the regression. They refer to nationality, education, professional training, age at first job and kind of first job. No industry or occupation dummies concerning the current job are included, but industry dummies referring to the last spell if it was self-employment are used. Two samples are selected for each gender. The first sample only includes individuals who are currently either wage-employed or unemployed and who have wage employment experience. 18

19 The corresponding estimation results are given in column 2 and 4. The intention behind the definition of this sample is to analyse the effect of self-employment on the probability of unemployment exclusively for those who were only temporarily self-employed and have also experience in wage employment. The second sample additionally includes the self-employed and the unemployed without wage employment experience, i.e. it comprises the full sample of individuals being part of the labour force (see column 3 and 5 for estimation results). Thus, it contains also those individuals who have spent their whole working life in self-employment. To begin with, those coefficient estimates are discussed which do not differ very much between the two samples. Similar to the earnings estimations, family related variables have contrasting effects on the unemployment probability of genders. Being married and having children increases the unemployment risk of females. In contrast, married men have a comparatively low probability of unemployment. This corroborates the hypothesis that when having a family, women usually specialise in household production and men in market production. Being in charge of housework and child rearing reduces timetable flexibility and therewith the number of acceptable jobs. Moreover, employers are often reluctant to employ women with small children. The negative age coefficient probably reflects a tenure effect. Having no occupational degree increases the unemployment risk of males, and having a university degree decreases the unemployment risk of females. Independent of gender, having completed an apprenticeship is connected with a relatively high probability of unemployment. The occupational status in the first job has an impact, too. Compared with the reference category first job civil servant, all other professions have a higher risk of unemployment. The risk is highest for individuals who entered the labour market as a blue collar worker. They are followed by those who started with self-employment and those who started with a white-collar job. Time spent in unemployment increases the probability of being unemployed, whereas time spent in wage employment reduces it. There are some marked differences between the estimation results of the two samples. Being a foreigner increases the unemployment risk of females only when accounting for those who have exclusively work experience in self-employment. Likewise, a high school degree significantly reduces it for males only when the always self-employed are included. For the full sample, self-employment experience as measured in months and spells reduces the probability of unemployment by far more than for the reduced sample. Self-employment experience certainly decreases the risk of business failure and therewith of unemployment for the currently self-employed included in the full sample. In addition, the difference in coefficients might 19

20 Table 4: Random Effects Probit Estimation of the Probability of Unemployment males females always self-employed included no yes no yes married *** *** *** *** [0.0539] [0.0508] [0.0566] [0.0538] single [0.0587] [0.0554] [0.0611] [0.0581] kids *** *** [0.0213] [0.0199] [0.0238] [0.0222] partner_full *** *** *** *** [0.0360] [0.0342] [0.0431] [0.0404] age *** *** *** *** [0.0120] [0.0109] [0.0133] [0.0122] age *** *** *** *** [0.0001] [0.0001] [0.0002] [0.0001] foreign *** [0.0585] [0.0542] [0.0706] [0.0646] noschool [0.0790] [0.0739] [0.1021] [0.0941] highschool ** [0.0848] [0.0775] [0.0888] [0.0804] nodegree *** *** [0.0586] [0.0549] [0.0671] [0.0608] apprentice *** *** ** ** [0.0471] [0.0444] [0.0561] [0.0508] university ** *** [0.0724] [0.0665] [0.0802] [0.0721] agefjob ** ** [0.0086] [0.0079] [0.0075] [0.0069] fjselfe ** *** *** *** [0.2337] [0.2037] [0.3892] [0.2987] fjblue *** *** *** *** [0.1587] [0.1423] [0.2682] [0.2178] fjwhite * * *** *** [0.1586] [0.1430] [0.2629] [0.2133] mon_unemp *** *** *** *** [0.0016] [0.0013] [0.0014] [0.0016] spells_unemp *** *** *** *** [0.0264] [0.0239] [0.0298] [0.0288] mon_wage *** *** *** *** [0.0005] [0.0005] [0.0007] [0.0006] mon_wage2 7.82e-06*** 9.76e-06*** 1.03e-05*** 1.58e-05*** [1.23e-06] [1.17e-06] [1.88e-06] [1.77e-06] spells_wage *** *** *** *** [0.0295] [0.0269] [0.0329] [0.0302] mon_selfem * *** *** [0.0013] [0.0010] [0.0026] [0.0020] spells_selfem *** *** *** [0.0785] [0.0772] [0.0741] [0.0665] lastsp_self * [0.2915] [0.2670] [0.3374] [0.3646] income_self [0.0165] [0.0126] [0.0210] [0.0162] ylast_self *** ** [0.0785] [0.0683] [0.1061] [0.1333] comsiz_self [1.1169] [1.1998] nace_agri ** [0.9680] [0.8874] [3.1633e+07] [1.4185e+06] nace_manu [0.4479] [0.4325] [1.1358] [0.7624] 20