Main article

View the most recent version.

Archived Content

Information identified as archived is provided for reference, research or recordkeeping purposes. It is not subject to the Government of Canada Web Standards and has not been altered or updated since it was archived. Please "contact us" to request a format other than those available.

Archived

This page has been archived on the Web.

1. Introduction
2. A model of labour market transitions
3. Data
4. Estimation results
5. Concluding remarks

1   Introduction

The self-employment rate has often been used as a proxy for the level of entrepreneurship (Audretsch and Thurik 2001). However, not every self-employed individual has a well-established business. Some self-employed may not ever intend to incorporate or to start hiring employees. This fraction of the self-employed are only tentatively connected to the sector, and may be accumulating skills that are applicable more to wage-employment than to expanding a business. Other things equal, these "less entrepreneurial" self-employed would be on average more likely to become wage-and-salary workers than entrepreneurs in the future.

Using two panels of the Survey of Labour and Income Dynamics (SLID), 1993-1998 and 2002-2007, this paper examines whether the nature of self-employment has changed between the 1990s and the 2000s. It computes annual transition rates between non-employment, two types of self-employment, and paid employment, and asks what changes have taken place. The paper then estimates a mixed multinomial logit model with state dependence to ascertain why the transition probabilities have changed. An advantage of using this framework is that one of the interpretations of the degree of state dependence is sector-specific learning. Thus, it is possible to see whether the skills learned in less entrepreneurial types of self-employment have become more or less applicable to paid employment.

Understanding how the nature of self-employment in Canada has changed is important because the productivity performance of the self-employment sector has changed dramatically between the two decades. Baldwin and Rispoli (2010) showed that nominal labour productivity (nominal gross domestic product (GDP) per hours worked) for the unincorporated self-employed remained relatively constant over the 1993-1998 period, while nominal labour productivity increased in the rest of the business sector. Post 2000, the situation was reversed: labour productivity rose faster in the unincorporated sector than in the rest of the business sector. Comparing the labour markets in the 1990s and the 2000s can give some insight into why this turnabout occurred.

Chart 1 shows that the effects of the 1991 recession on the labour market lingered well into the mid-1990s. The unemployment rate was at 11.4% in 1993, and had declined to only 8.3% by 1998. In comparison, the average unemployment rate for the years 2002 to 2007 was 6.9%. The soft labour market conditions of the 1990s were accompanied by a sharp rise in self-employment, particularly in own-account self-employment, that is, self-employment where one does not hire any employees. The rate of self-employment with employees actually fell in the 1990s. Roy and Gauthier (1998) argued that much of the increase in the self-employment rate in the 1990s was due to cost-cutting by firms in the face of persistent weakness in overall demand, and that they expected the rate of self-employment to decrease as economic conditions improved. Thus, it may have been the case that workers were using self-employment as a stopgap in the 1990s. That is to say, self-employment was being used as an alternative to unemployment, and workers would easily switch to paid employment when the opportunity presented itself. Since workers did not intend to stay in self-employment, they may have put less emphasis on picking up skills that would make their businesses more profitable and productive and instead devoted more time to job search.

This is not the first paper to study self-employment dynamics in Canada. Kuhn and Schuetze (2001) compared the transitions between non-employment, paid employment, and unincorporated self-employment in the 1980s and the 1990s in order to examine what changes in the transition rates explained the increase in the self-employment rate in the 1990s. For men, they found that the increase in self-employment was due to decreased stability in paid employment. For women, they found that the increase in unincorporated self-employment was due to an increase in the retention rate in self-employment. The standard multinomial logit model that they used to explain the changes in the transitions themselves revealed that changes in demographics could not account for any of the changes in the transition rates. More recently, Esmaeilzadeh (2008) used the SLID to examine the labour dynamics (transitions between being out of the labour force, unemployment, paid employment, and self-employment) for males over the 1993-2004 period. That paper used an econometric framework similar to the one adopted here, but it focused on differences in the transition probabilities between Canadian-born and immigrants and did not look at differences between the 1990s and 2000s.

It is found that there is less movement between paid employment, own-account self-employment, and self-employment with paid help in the 2000s compared to the 1990s. This confirms that self-employment was more transitory in the 1990s. Moreover, estimates from the dynamic mixed multinomial logit suggest that increased state dependence, not cyclical factors (proxied by provincial unemployment rates), demographic change, or changes in the industrial and occupational structure, accounts for increased stability. State dependence can be due to a number of factors. High job search costs may prevent individuals from looking for jobs and moving into paid employment from other employment states, such as self-employment. Sunk costs associated with starting a business (e.g., the purchase of specialized equipment) may impede movement back into paid employment. The building of skills through learning-by-doing can lead to greater state dependence, if skills learned in self-employment cannot be easily transferred to paid employment. One possible explanation for the increased state dependence could be that many newly self-employed in the 1990s were concentrating on moving back into paid employment and consequently focused on accumulating skills for paid employment rather than on running a business even while they were self-employed. Additional research needs to be done in order to understand the reasons for the increased state dependence.

The more transitory nature of self-employment in the 1990s is likely related to the poorer productivity performance of the self-employment sector in the 1990s compared to the 2000s. Previous literature (e.g., Foster, Haltiwanger, and Krizan 2001, and Baldwin and Gu 2006) suggests that, on average, entrants are not as productive as incumbents and that it takes time for the productivity of entrants to improve. The fact that there were more newly self-employed in the 1990s and that their businesses did not survive as long as the businesses of the newly self-employed in the 2000s would have a negative impact of productivity in the self-employment sector.

The next section, Section 2, presents the conceptual and econometric framework used in this paper. This is followed by a description of the SLID and a presentation of the raw transition probabilities in Section 3. Section 4 presents the estimates of the model and an evaluation of why the transition probabilities changed. Concluding remarks are offered in Section 5.

2   A model of labour market transitions

2.1  Conceptual framework

Each year, individuals choose between non-employment, paid employment, a less entrepreneurial form of self-employment, and an entrepreneurial type of self-employment. ¹  In this paper, the presence of employees is used as an indicator of the latter type of self-employment. The hiring of employees is a milestone in the development of a business. It indicates that the business is expected to generate enough income to support not only the business owner but also at least one other person. The hiring of employees also indicates a certain level of commitment to a business. A person might report that he or she is self-employed, but that person's business may not be active. The presence of paid help generally indicates that there is some economic activity; the business owner is putting in at least enough effort to compensate his or her workers.

Other measures can also be used to distinguish the two forms of self-employment. Incorporation is costly and therefore also shows some level of commitment to a business. However, unlike the hiring of employees, incorporation is a one-time cost. It represents genuine interest or expectation of growth in the year of incorporation, but not necessarily in subsequent years. Furthermore, tax policy can influence the margin at which individuals choose to incorporate. Changes in the tax advantages of having a corporate business rather than an unincorporated business may affect the form of organization even when the business owner's expectations or level of effort have not changed. Another way to identify an entrepreneurial business is to look at the income of the business owner. If the amount of income is over a given threshold (the level of income the owner would expect to get if he or she were wage-employed, for example), this indicates a higher degree of commitment and effort on the part of the business owner. The main drawback of this measure is that it is an ex post facto indicator. Business incomes are volatile. Even when the level of effort and expectations are high, this does not necessary translate into higher incomes. As well, the period returns to the business owner are difficult to capture. They may come in the form of wages, unincorporated self-employment income, dividends, or expected future capital gains (expected at the time the business is sold). For the above reasons, the presence of paid help is used as a rough identifier of entrepreneurial businesses in this paper.

Individuals are assumed to choose between the four employment states according to which state gives them higher utility. This utility captures both the monetary returns and the various non-pecuniary benefits, such as the flexibility associated with self-employment, the social interaction associated with work, and the returns to home production, associated with employment. Let the utility or attractiveness of being in employment state j for individual i at time t be U^*_ijt, where j = non-employment (NE), paid employment (PE), own-account self-employment (OASE), and self-employment with paid help (ESE). Individual i chooses occupation j if U^*_ijt > U^*_ikt for all .

where: X_it is a k X 1 vector of observable individual characteristics; C_it is a cyclical indicator that varies over time and the province in which the individual resides; denotes unobservable time-invariant individual characteristics that may capture a preference for one sector over another or an innate ability; denotes the unobservable individual characteristics that vary over time; Y^PE_i,j,t-1 is an indicator variable that is 1 (one) if individual i was in paid employment at time t-1 and is 0 (zero) otherwise; Y ^OASE_i,j,t-1 is an indicator variable that is 1 (one) if individual i was own-account self-employed at time t-1 and is 0 (zero) otherwise; Y ^ESE_i,j,t-1 is an indicator variable that is 1 (one) if individual i was self-employed with paid help at time t-1 and is 0 (zero) otherwise.

The vector of observable individual characteristics includes controls for the following: education; age; marital status; other family income; the presence of young children; immigrant status; and industry and occupation. Education and age (a proxy for experience) are directly related to an individual's productivity and hence to the monetary returns to each employment state. As shown in Rees and Shah (1986), the effect of education and age on the choice between paid employment and self-employment works through non-pecuniary channels as well. They argue that the more educated are better informed and thus better able to assess business opportunities. This ability to better assess business opportunities lowers the riskiness of self-employment and increases the probability of self-employment if individuals are risk-averse. They also argue that there may be age-related shifts in a person's preference over employment states. For example, older individuals may be more risk-averse and less likely to choose self-employment. Marital status and the education and income of other family members have been used by others, such as Borjas and Bronars (1989) and Hamilton (2000), to capture the degree of family support one has. The idea here is that more family support may make self-employment less risky or less demanding. Likewise, more family support could come in the form of having more children, but the presence of young children may shift an individual's preference toward non-employment or to a form of self-employment that allows for a more flexible work schedule. ²  With respect to being an immigrant, it has been argued that enclave effects and discrimination in paid employment predispose immigrants toward self-employment. ³  Moreover, the existence of a Business Immigration Program in Canada, which seeks to attract investors, entrepreneurs, and the self-employed, may make immigrants more likely to choose some type of self-employment over paid employment. Finally, industry and occupation dummies are included in the choice model for the purpose of capturing the differences in self-employment opportunities that may exist across sectors and occupations.

The cyclical indicator captures the impact of economic conditions on the returns to the employment state. For example, higher business bankruptcy rates in poorer economic times would lower the expected returns of starting a business.

The indicator variables that show the individual's employment state in the previous period capture the amount of state dependence in the labour market. There are a number of possible interpretations for this state dependence. High job-search costs may lead to state dependence in non-employment (see, for example, Hyslop 1999). The high costs of market research and other information costs may prevent the non-employed and the paid-employed from setting up businesses. Sunk costs associated with capital equipment and buildings and the duration of a lease may dissuade employers or own-account self-employed from becoming paid-employed. Alternatively, state dependence can be the result of the building of human capital through learning-by-doing (see for example, Eckstein and Wolpin 1989). For example, if an individual spent the previous year in paid employment, he or she undoubtedly would have picked up skills that would increase his or her wage in his current job and/or raise his or her employment prospects. This increased skill or knowledge may not be picked up in the usual set of observed worker attributes. Moreover, these skills may not be state-specific. Skills learned from managing part of a business in a large corporation may be applicable to running one's own business. Skills accumulated in running a business with no employees would be transferable to running a business with employees, but could also find application in paid employment.

2.2  Econometric framework

Utilities are generally not observed, and the fact that an individual is observed in paid employment reveals only that his or her utility from paid employment is greater than that from the other employment states. Therefore, the parameters in (1) are not identified, and some normalization is needed. In this paper, non-employment is chosen as the base category; ⁴  the effect of the observed covariates on non-employment and the individual time-invariant effect for non-employment are subtracted from the utility from paid employment, own-account self-employment, and self-employment with paid help. For j = PE, OASE, andESE, the transformed utilities are:

Whereas, for NE, the transformed utility is:

This formulation changes the interpretation of the parameters. For example, would be the marginal effect of X_it on the utility of paid employment less the marginal effect of X_it on the utility of non-employment (i.e., ).

To proceed from here, it is necessary to make a distributional assumption about the random shock, and to decide how to estimate the time-invariant individual effect. In this paper, it is assumed that follows a type I extreme value distribution for all i, j, and t, that is uncorrelated across individuals, employment states, and time, and that is uncorrelated to the explanatory variables and to . Furthermore, it is assumed that:

Together these assumptions result in an example of a mixed multinomial logit (MML) model, or random parameters logit model. Unlike the regular multinomial logit model used by Kuhn and Schuetze (2001), the MML model allows for correlation in the error components across alternatives. Allowance for this correlation is important because it relaxes the assumption of independence of irrelevant alternatives made by the multinomial logit. With independence of irrelevant alternatives, the probability associated with choosing between paid employment and non-employment is not affected by the introduction or removal of self-employment as an option. The MML model is a thus a more flexible framework.

Other assumptions could have been made to make the model operable. The time-varying unobservable individual characteristics (the random shock term) could be assumed to follow a multivariate normal distribution, which would lead to a multinomial probit. The advantage of assuming a multivariate normal distribution is that the errors terms could be freely correlated across individuals and employment states. In contrast, the MML model allows for only correlations in the error terms, through the correlation between the individual specific components. While the multivariate normal assumption is appealing theoretically, in practice, it is very time-consuming to estimate the multinomial probit model without any restrictions on the covariance matrix. A drawback that is of greater concern to the application in this paper is that identification is fragile—parameter estimates are unstable, and standard errors are large—when no alternative specific covariates are available. Keane (1992) showed this in the context of a single period multinomial probit, and Rendtel and Kaltenborn (2004) extended Keane's findings to the case of multiperiod multinomial probits. On the other hand, Prowse (2006) demonstrated that the MML model can yield well-behaved and reliable parameter estimates without alternative specific explanatory variables. ⁵ 

Consequently, while the multinomial probit is a more flexible model, identification is more difficult, especially without alternative specific explanatory variables. ⁶ 

Additionally, the individual time-invariant effect could be treated as a fixed effect, not as a random effect, as in Honoré and Kyriazidou (2000), or the random effect could be allowed to vary over time in a particular way, as in Prowse (2007). ⁷  These approaches are not taken here mainly because of the short-time-series aspect (six years) of the data.

Conditional on , the logit assumption means that the probability of individual i to be in state j (for j = PE, OASE, and ESE) in period t is:

Whereas the probability of individual i to be in non-employment in period t is:

Conditional on , , and , the probability of observing an individual's sequence of choices in the sample is:

The unconditional probability of observing individual i's sequence of choices is obtained by integrating (4) with respect to the joint probability density function of , , and  :

where F is the cumulative density function of a multivariate normal distribution. The log-likelihood function for an independent sample of N individuals each observed during T periods is:

Since Monte Carlo integration is used to evaluate the integral in (5), the model is said to be estimated by means of simulated maximum likelihood.

The remaining complication in estimating the model is the problem of initial conditions. The presence of lagged state variables means that the model is dynamic. The employment outcomes observed in the data depend on unobserved employment outcomes in the past. The paper addresses this problem by following Wooldridge's (2005) suggestion of defining the likelihood of observing the individual's outcomes in t = 3, 4, …, T, conditional on the explanatory variables and the initial condition in t = 1.

3   Data

3.1  Description and sample restrictions

The comparison of labour market dynamics in the 1990s and the 2000s is carried out by means of the Survey of Labour and Income Dynamics (SLID). The SLID is a series of overlapping six-year panels. This paper uses the first and last of these six-year panels. The first spans 1993 to 1998, while the last covers 2002 to 2007. The target population for the sample survey includes almost all individuals in Canada; residents of the territories (Northwest Territories, Yukon, and Nunavut), residents of institutions, and persons living on Indian reserves or in military barracks are not included. The panels are representative of the population at the start of each panel. Labour market and demographic information is collected in January for the previous year, and income information comes primarily from administrative sources. In this paper, individuals are non-employed if they did not have a job at any point in the year, and are paid-employed, own-account self-employed, or employers according to the "class of worker" characteristic of their main job. The main job is determined by the number of hours worked and tenure.

Sample restrictions are made in order to abstract from a few labour market issues. First, attention is restricted to males. ⁸  Second, as in Kuhn and Schuetze (2001), the analysis is limited to individuals aged 25 to 54 at the beginning of each panel, in order to abstract from the movement of older individuals into self-employment as a way of easing into full retirement. It has been shown that the self-employed are more likely to continue working at older ages, and that the paid-employed are increasingly likely to move into self-employment as they age (see Fuchs 1982). Third, although the model includes the non-employed, some individuals do not have any labour force attachment. These individuals are eliminated from the analysis by dropping all individuals who never had a job in any of the six years during which they were in the survey. Fourth, the amount of labour engaged in agricultural production has declined over time. Since a large fraction of these workers are self-employed farmers, the decline in agricultural labour would translate into increased movement from self-employment to paid employment or to non-employment. In order to eliminate the effect of this trend on the transitions from the analysis, individuals who have worked in agriculture at any time over the six-year period are dropped from the sample. Finally, one of the explanations for the dramatic increase in self-employment in the 1990s was weak labour market conditions. Moore and Mueller (2002) showed that laid-off workers had a higher probability of moving into self-employment over this time period. One sector that continually exhibited declining employment as self-employment rose was public administration. In order to remove the effect of government downsizing on the transitions, individuals who have worked in public administration at any time over the six-year period are not included in the analysis. ⁹ 

3.2  Transition rates

The transition rates for the estimation samples are presented in Table 1. As explained in the previous section, the dynamic nature of the model means that the model is estimated for the 1995-1998 and 2004-2007 periods. For example, since the data for 1993 are used for the initial conditions, the first employment outcomes to be modeled are those for 1995 (this outcome depends on the lagged state in 1994). Table 1 shows the transition rates for all the person-year observations in those two time periods. There are 17,570 person-year observations for the 1990s and 13,754 person-year observations for the 2000s.

The 1990s and 2000s columns of Table 1 show the distribution of individuals over the employment states of non-employment, paid employment, own-account self-employment, and self-employment with paid help, conditional on the individual's previous state. For example, the first row in the first column of Table 1 shows that, of the males who were non-employed in the previous year, 61% remained in non-employment in the following year. Subsequent rows in the same column show what happened to the remainder of the non-employed in the previous year: 33% found work in paid employment; 5% became own-account self-employed; and 1% started a business with employees. There is persistence in all employment states. For example, in the 1990s, 95% of the paid-employed, 77% of the own-account self-employed, and 85% of employers did not change employment states from year to year. Despite the fact that there may be setup costs associated with starting a business, there is less persistence in the self-employment states than in paid employment. This is likely due to higher variability in business income compared to wage income. The argument that hiring employees is a sign of greater commitment to running a business is in line with the finding that there is more persistence in self-employment with paid help than in own-account self-employment.

The fifth column of Table 1 shows how the transitions rates changed between the 1990s and the 2000s. The persistence in all employment states, except non-employment, rose between the two time periods. Persistence increased by only about 1 percentage point in paid employment, but it increased by 13 percentage points in own-account self-employment and by 8 percentage points in self-employment with employees.

Interestingly, not only do the transition rates between paid employment and both types of self-employment fall, but so do the transition rates between the two types of self-employment. One might have expected more transitions from own-account self-employment to self-employment with paid help in the 2000s because fewer individuals were using self-employment as a stopgap. Thus a higher percentage of own-account self-employed might be expected to progress to self-employment with paid help in the 2000s. The decline in the transition rates between the two types of self-employment may instead indicate that own-account self-employment is becoming more distinct from self-employment with paid help. For example, a computer programmer/consultant who works on projects by himself at home might have no intention of ever hiring paid help.

There is also an increased flow out of non-employment, but such results are not statistically significant. This is probably due to the fact that the number of males that remain in non-employment for the entire year is generally small. The likely explanation for the decrease in persistence in non-employment is straightforward: the improved labour market conditions in the 2000s made it possible for individuals to find work more easily. The reason for the increases in the persistence in the other employment states is less certain and is explored by means of regression results.

A major difference between the 1990s and the 2000s is that the 1990s were a period of transition; own-account self-employment rose sharply. In contrast, own-account self-employment was relatively stable in the 2002-2007 period. The higher transition rates of the 1990s could be due simply to the fact that this study compares a transition period to a period that more closely approximates a steady state. Increasing own-account self-employment driven by higher entry rates into self-employment need not be accompanied by higher exit rates. However, previous evidence shows that younger firms are more likely to fail than older firms. ¹⁰  More generally, it might be argued that the increased transition rates in the 1990s may be associated with lower tenure in each employment state. In terms of the estimation model, tenure in the employment state is an omitted variable that may bias the results, especially the estimates of the lagged state variables. The lagged state variables are meant to pick up the effect of an individual's entire past employment history, including the length of time in each employment state. Unfortunately, the six-year panels available in the SLID mean that more lagged states cannot be included and that the length of time in each employment state cannot be determined. ¹¹ 

3.3  Summary statistics

Possible explanations for the observed changes in the pattern of labour market transitions include changes in demographics and changes in industrial and occupational structure. Tables 2 and 3 present these summary statistics by employment outcome. Compared to males in the 1990s, males in the 2000s are older, more educated, and more likely to be immigrants. ¹²  If age, education, and immigration status increase the attractiveness of being own-account self-employed vis-à-vis the other employment states, for example, then this observed trend would mean that the probability of moving to own-account self-employment, conditional on any previous employment outcome, would increase.

Instead, a more localized type of change is needed. For example, assume again that more education increases the attractiveness of being in own-account self-employment more than it does the attractiveness of any other state. If only the workers who were own-account self-employed in the previous period were to become more educated, then only the probability of being in own-account self-employment in period t+1, conditional on having been in own-account self-employment in period t, would rise. The probability would rise, and the other probabilities in the same row, , , and , would fall. These employment-state-specific changes in the demographics, however, are not evident. In most cases, the demographic changes are occurring in the same direction.

Tables 2 and 3 also show the distribution of individuals across industries and occupations by employment state. ¹³  One reason why persistence in paid employment, own-account self-employment, and self-employment with employees may have increased is that paid employees may be increasingly concentrated in some industries, while the own-account and employer self-employed may be increasingly concentrated in others. To the extent that industry- or occupation-specific human capital exists (see Neal 1995 and Kambourov and Manovskii 2009, for example), the differences in industrial/occupational concentration across employment states would mean that the paid-employed and the two-types of self-employed are accumulating human capital that is not easily transferable between employment states. If the differences in the industrial/occupational structure increased, then it would be increasingly difficult for workers to move between employment states. In the extreme, if the paid-employed were found only in certain industries and occupations, own-account self-employed only in others, and employers only in yet others, moving between employment states would be as difficult as moving between industries and occupations.

As well, Tables 2 and 3 show that, for males, the distributions across industries for the paid-employed, the own-account self-employed, and employers are quite different. In the 1990s, roughly one-quarter of paid-employed could be found in manufacturing, compared to 4% for the own-account self-employed and 9% for employers. In contrast, roughly one-quarter of the own-account self-employed could be found in construction, compared to 9% for paid employment and 22% for employers. Employers are also more highly concentrated in trade (21%) than the own-account self-employed (9%) and paid-employed (16%). The differences in the distributions across occupations tell a similar story. The paid-employed are more highly concentrated in processing occupations; own-account self-employed are more concentrated in the trades; and employers are more likely to be managers. One way of measuring how similar the distributions are is to calculate the absolute values of the differences in the distributions between two employment states and then take the sum of those absolute values. The use of this metric indicates that the industry and occupation distributions have generally become more similar. The exception is the difference in the occupation distributions for the wage-employed and employers. Therefore, for males, changes in the industry and occupation distributions likely did not cause the increase in persistence.

The final variable in Tables 2 and 3 is other family income. This is the total income of the family from all sources minus the total income of the individual. The consumer price year (base year 2000) is used to deflate the amounts. It is expected that higher other family income is negatively correlated with work. Other family income is included as an explanatory variable to aid in the identification of the model. Prowse (2007) showed that a statistically significant continuous variable is one of the conditions that need to be fulfilled before it can be said that the identification of the model does not rely solely on the parametric assumption for the error term.

4   Estimation results

4.1  Model estimates

The model estimates for males are presented in Tables 4 and 5. Explanatory variables typically found in earnings regressions and labour force participation equations are included, namely controls for age, education, immigrant status, marital status, the presence of school-age children, other family income, and industry and occupation. Provincial unemployment rates are used to account for changes in the business cycle, and year dummies are used to pick up other unidentified aggregate effects. A set of dummies indicating the employment state of the individual at the beginning of the panel are included in order to address the initial conditions problem. ¹⁴ 

Focusing first on the 1990s, Table 4 shows that the coefficients on each of the lagged employment state dummies are positive and statistically significant. This means that any form of previous work for compensation raises the attractiveness of being paid-employed, own-account self-employed, or an employer in the current period relative to the returns associated with being non-employed. The effect of having been paid-employed, own-account self-employed, and an employer in the last period has the greatest impact on the attractiveness of paid employment, own-account self-employment, and self-employment with paid help, respectively. For example, having been paid-employed last period raises the attractiveness of being paid-employed in the current period (net of the attractiveness of being non-employed) by 3.875. Having been paid-employed in the last period has less of an effect on the attractiveness of being own-account self-employed (2.324) and an employer (3.048) in the current period. When one interprets the lagged employment state variables as picking up human capital accumulation, the results show that, while skills are transferable, there is some degree of specificity.

A comparison of the 1990s with the 2000s (Tables 4 and 5) reveals that the coefficients on the lagged employment state variables have generally fallen. However, the coefficients that give the cross-sector effects have declined more in percentage terms than have the coefficients representing the own-sector effects. The impact of having been in paid employment last period on the attractiveness of paid employment in the current period fell from 3.875 in the 1990s to 3.700 in the 2000s, a decline of 4.5%. In contrast, the effect of having been in paid employment last period on the attractiveness of own-account self-employment in the current period fell by 50% (from 2.324 in the 1990s to 1.152 in the 2000s), and the effect of having been in paid employment last period on the attractiveness of being an employer in the current period fell by 42% (from 3.048 to 1.669). The declines in the other own-sector effects are also small in comparison with the other cross-sector effects. Being in own-account self-employment last period increased the attractiveness of own-account self-employment in the current period by 5.278 in the 1990s and 4.954 in the 2000s, a 6.1% decline. This is a much smaller decline than the 36% (from 2.689 in the 1990s to 1.714 in the 2000s) decline in the effect of lagged own-account self-employment on the attractiveness of paid employment and the 47% (from 4.246 in the 1990s to 2.251 in the 2000s) decline in the effect of lagged own-account self-employment on the attractiveness of being an employer. Similarly, the 12% decline (from 8.032 in the 1990s to 7.052 in the 2000s) in the lagged effect of having been an employer in the past on the attractiveness of being an employer in the current period is smaller than the effect on the attractiveness of being in paid employment and in own-account self-employment, at 51% and 49%, respectively. Again, in the context of human capital accumulation, this suggests that skills have become less transferable across employment states.

The superscript number 2 (²) in Table 5 denotes the coefficients in the 2000s that are statistically different from their counterparts in the 1990s. ¹⁵  Many of the changes in the lagged state variables are statistically significant. Aside from the coefficients on the lagged state variables, few other coefficients have changed over time.

Further analysis of the coefficients from the model is omitted, as the sign and size of the coefficients themselves do not give all the information necessary in order to predict the sign and size of the effect of the explanatory variables on the probability of being in one employment state or another. ¹⁶ 

The average marginal effects from the model are presented in Tables 6 and 7. The cell in the first row shows the average marginal effect of having moved from non-employment to paid employment in the previous period on the probability of being non-employed in the current period.

In the 1990s (Table 6), having moved from non-employment to any other employment state in the previous period lowers the probability of being in non-employment in the current period substantially, by over 40 percentage points. The effect of having moved to paid employment in the previous period raises the probability of being in paid employment in the current period by 44 percentage points and lowers the probability of being in own-account self-employment and in self-employment with paid help in the current period as well. These negative effects, however, are not statistically significant. The results are similar for moving to the two forms of self-employment in the previous period. ¹⁷  State dependence raises the probability of remaining in the same sector, but it has little or no effect on the probability of moving to other sectors, save non-employment. These average marginal effects reflect the coefficient estimates presented in the previous table. For example, since having been in paid employment last period raises the returns to being in paid employment in the current period more than it raises the returns to being in any other employment state, it is not surprising that the marginal effect of moving from non-employment to paid employment in the last period raises the probability of being in paid employment in the current period the most.

Turning to the 2000s (Table 7), one can see that the effect of the lagged employment state variables has become larger. The effect of having moved from non-employment to self-employment with paid help in the last period on the probability of being an employer in the current period rises from 55 percentage points to 71 percentage points. Likewise, the effect of moving from non-employment to own-account self-employment in the previous period on the probability of being own-account self-employed in the current period rises 10 percentage points. Whereas the cross-sector effects were statistically insignificant in the 1990s, they are negative and significant in the 2000s. Having moved from non-employment to self-employment with paid help in the previous period lowers the probability of being paid-employed in the current period by 28 percentage points, and having moved from non-employment to own-account self-employment in the previous period lowers the probability of being paid-employed in the current period by 17 percentage points. Moreover, having been paid-employed last period lowers the probability of being own-account self-employed in the current period by 7 percentage points in the 2000s.

Both the coefficient estimates and the estimates of the average marginal effects suggest that the change in the effect of an individual's employment state in the previous period plays a role in the change in the transition probabilities. This is confirmed in the experiments discussed in the following section.

4.2  Predicted changes in transition rates

Once the model in the 1990s and the 2000s has been estimated, it is possible to conduct experiments in order to determine which factors are important in terms of explaining the changes in the labour market dynamics. The non-linear nature of the model makes a perfect decomposition unfeasible; however, by changing one element of the 1990s model at a time to that found in 2000s model, one can get a sense of which factors are key.

In the first experiment, the predicted changes in the transitions rates, when only the coefficients on the lagged employment states are allowed to change from their values in the 1990s to their values in the 2000s, are calculated. The results of this experiment can be viewed in the first column of Table 8. Since there is no coefficient for having been non-employed last period, the probabilities conditional on non-employment in the previous period are unaffected. The effects of the changes in the coefficients on the lagged employment states on the other transition rates are significant. A large fraction of the changes in the remainder of the transition matrix are accounted for. For example, the predicted increases in persistence in paid employment, own-account self-employment, and self-employment with paid help are 1.3 percentage points, 8.5 percentage points, and 6.8 percentage points, respectively. These figures can be compared to the actual changes of 0.9 percentage points, 12.9 percentage points, and 7.7 percentage points, respectively.

The second experiment allows the remaining coefficients to change, holding everything else constant at 1990 values (including the coefficients on the lagged employment states). The second column of Table 8 shows that changes in the other coefficients have a minimal impact on the probabilities, conditional on being paid-employed, own-account self-employed, and self-employed with paid help. The effect of changes in the other coefficient does, however, account for all of the changes in the transition rates conditional on being non-employed. This makes sense, as better labour market conditions in the 2000s raised the returns to all types of work relative to non-employment.

The final experiment, shown in the third column of Table 8, calculates the predicted probabilities when the estimated coefficients in the 1990s are applied to individuals in the 2002-2007 panel. ¹⁸  These are then compared to the predicted probabilities from the 1990s. As predicted in Section 3.3, the effects of changes in demographics and changes in industrial and occupational structure are small.

5   Concluding remarks

This paper compares the self-employment dynamics for males in Canada in the 1990s and the 2000s. It is found that there was less movement between paid employment, own-account self-employment, and self-employment with paid help in the 2000s compared to the 1990s. The more transitory nature of self-employment in the tougher labour market conditions of the 1990s suggests that many individuals were using self-employment as a stopgap, and may not have been fully prepared to start a business. This lower commitment to, or reduced preparedness for, self-employment in the 1990s is likely related to the poorer productivity performance of the unincorporated self-employment sector in the 1990s compared to the 2000s. Previous literature suggests that, on average, entrants are not as productive as incumbents and that it takes time for the productivity of entrants to improve. The fact that there were more newly self-employed in the 1990s and that their businesses did not survive as long as those of their counterparts in the 2000s would have a negative impact on productivity in the self-employment sector. Although this paper draws a distinction between the self-employed with employees and the self-employed without employees, not between the unincorporated self-employed and the incorporated self-employed, the vast majority of the newly self-employed are unincorporated. ¹⁹  The individuals who had just transited into self-employment in the 1990s likely had lower attachment to their own businesses and a lower commitment to running a business.

Estimates from a dynamic mixed multinomial logit model suggest that increased state dependence accounts for the increased persistence in paid employment, own-account self-employment, and self-employment with paid workers. Demographic change and changes in the industrial and occupational structure had little effect, while changes in the returns to worker characteristics and improving economic conditions (picked up by the decline in unemployment rates) accounted for the increased flow out of non-employment.

One possible explanation for this increased state dependence could be that greater specialization has made skills learned in one employment state less transferable to another. Another possible explanation is greater costs of moving, such as job search costs and information costs. This latter explanation, however, is improbable given that improvements in information and communications technology should have lowered such costs. Another possibility is that the extensive use of self-employment as an alternative to non-employment in the 1990s made recruiters increasingly wary of reported experience in self-employment. Yet another possibility is that the increased state dependence is picking up the effect of missing variables, such as the length of time one expects to be not-employed and the length of time one has already spent in a particular employment state. Some would argue that the greater entry into self-employment in the 1990s led to a greater number of young businesses, and that this resulted in higher self-employment exit rates since young businesses are more likely to fail than older ones.

Future research could involve a more in-depth investigation of the reasons behind the increase in state dependence. One approach could be to interact the lagged employment state variables with the other explanatory variables. Perhaps older individuals are less likely to make switches in employment state, and the aging of the labour may explain the greater state dependence. Another approach could be to use a longer panel, such as the Longitudinal Worker File, to examine the effect of time spent in each employment state. Another area of possible research is a comparison between Canada and the United States. This might shed light on the differences in the productivity performance of the self-employment sectors of the two countries.