Introduction

Similar to the United States, self-employment rates in Canada are substantially lower for women than for men. In 2013, approximately 36 percent of the self-employed were women, a figure that has changed little since the 1980s.Footnote 1Khun and Schuetze (2001) report that in 1998, only 7.77 percent of Canadian women were self-employed, relative to 11.05 percent of men. Moreover, exit rates are 27 percent higher for female self-employed over the 1981-1994 period in Canada.Footnote 2 Yet little is understood about what drives these gender differences in self-employment survival.

Some insight may be gained from the literature on gender differences in the credit market. For example, in 1994, the Canadian Federation of Independent Businesses report that women are 20 percent more likely to be refused financing than men, and those women who do obtain funds are charged higher interest rates.Footnote 3

This paper contributes to the literature on self-employment by exploring the determinants of self-employment survival for men and women using the Canadian Survey of Labour and Income Dynamics (SLID). Given the large gender gap in business financing in Canada, I will focus substantial attention to the issue of liquidity constraints and personal funds. In particular, I introduce a new measure to proxy for liquidity constraints; I consider the non-linearity of the relationship between personal funds and self-employment survival; and finally, I investigate whether this relationship differs if one exits self-employment for personal reasons versus business failure. Estimation is carried out with proportional hazards analysis, and I address the concern of unobserved heterogeneity using two separate techniques.

The paper is organized as follows: the first section, “Liquidity Constraints and Other Factors Associated with Self-Employment Survival,” briefly describes factors associated with self-employment survival in the literature; “Methodology” outlines the empirical methodology and proxies for liquidity constraints; “Data” describes the structure of the data; Regression results and sensitivity analysis are discussed in “Results”; “Discussion and Caveats” concludes.

Liquidity Constraints and Other Factors Associated with Self-Employment Survival

There is a large and growing empirical literature investigating whether liquidity constraints limit self-employment entry, and a somewhat smaller literature that considers business survival. Across both the entry and survival literature personal wealth is typically used as a proxy for liquidity constraints because, in the presence of binding constraints, individuals with personal funds will be among the few that have sufficient capital to start and succeed in business. Several of the entry studies (e.g., Evans and Jovanovic 1989; Holtz-Eakin et al. 1994a; Fairlie 1999; Dunn and Holtz-Eakin 2000; Kan and Tsai 2006), find results consistent with liquidity constraints inhibiting self-employment entry. However, Hurst and Lusardi (2004) report that the positive association between wealth and capital disappears when one controls for the endogeneity of wealth. Among the studies on self-employment survival, Taylor (1999) finds an insignificant relationship between a proxy for wealth and self-employment duration for men and women, while Evans and Leighton (1989); Bates (1990); Holtz-Eakin et al. (1994b); and Fairlie and Robb (2007, 2009), report evidence consistent with the theory that business survival is inhibited by the presence of liquidity constraints. Fairlie and Robb (2007, 2009) further estimate that differences in start-up capital can account for a substantial portion (over 40 percent) of the racial differences and gender differences in business closures in the United States.

In addition to capital, several other characteristics are associated with incidence and success in self-employment. For example, presence of children is a determining factor for self-employment selection among women (Hundley 2000), and Fairlie and Robb (2009) find that twice as many women report owning a business to manage family responsibilities, whereas a greater share of men own their business to have a primary source of income. Fairlie and Robb (2009) further note gender differences in the reasons for entering a business may be behind some of the male-female gap in business success.

Among the most influential characteristics associated with business survival versus failure are: age, education, capital, number of children, previous labor market experience (prior work experience, experience in a family business, business experience, unemployment, quitting one’s previous job), industry, marital status, and urban versus rural location (e.g. See Bates 1990; Holtz-Eakin et al. 1994b; Taylor 1999; Lin et al. 2000; Fairlie and Robb 2009; Haapanen and Tervo 2009; Millán et al. 2010). Moreover, Fairlie and Robb (2009) find that major contributors to the higher rates of business failure among women (relative to men) are startup capital, prior work experience in a family business, prior work experience in a similar business, education and marital status. This study employes a similar set of covariates in the proportional hazards and competing risks analysis.

Methodology

Proportional Hazards Analysis

Duration analysis is carried out on a sample self-employment spells. Because I have precise data on spell start and end dates, I use a continuous time duration analysis, the Cox (1972) proportional hazard rate model, to investigate spell duration.

The hazard rate, λ(t;X), is the conditional probability of spell end at time t, where t depicts the amount of time that has passed since the start of a job spell. In the proportional hazards model, observable characteristics, X, are assumed to affect a baseline probability of failure at an equal proportion for any time t. The model is thus:

$$ \lambda(t;X) = \lambda_{0}(t) e^{X \beta} $$
(1)

where λ 0(t) is the baseline hazard at time t, that is, the probability of exiting self-employment in time t, conditional on survival up to time t. Estimation involves a partial likelihood technique and, as such, the functional form of λ 0(t) need not be specified. Thus, no restrictions are imposed on duration dependence. Rather than reporting coefficients, β, I report hazard ratios: exp (β). A positive coefficient, β, results in a proportional hazard contribution greater than 1; therefore a hazard greater than 1 indicates an increased probability of business exit. For example, a hazard ratio of 1.10 for a particular characteristic is interpreted as a ten percent increase in the probability of exit, at any time t, given a one unit increment of that characteristic. The variables in X include characteristics which are associated with business survival. Thus, X includes a proxy for liquidity constraint, indicator variables for age categories, marital status, previous labor force status, immigrant status, and home ownership, log(income), log(income) 2, tenure at previous job, tenure 2 number of self-employed in family, number of children under 15, the provincial unemployment rate, and regionFootnote 4 and annual fixed effects. Values for these characteristics are recorded the year prior to the spell start.Footnote 5 This baseline specification is similar to that of Taylor (1999), with the exception that he includes education, industry, and involuntary job end, which I incorporate in the sensitivity analysis.

Sensitivity analysis includes both alternative specifications and estimation techniques. A discrete analogue to Cox proportional hazards, the Prentice and Gloeckler (1978) model is implemented to ensure that the assumption of continuous time does not affect the results. Further, I consider the sensitivity of these results to potential unobserved heterogeneity by including gamma distributed frailty in the Cox framework, and a discrete mixture distribution, recommended by Heckman and Singer (1984), in the Prentice and Gloeckler (1978) framework.Footnote 6 Alternative starting values for two mass points are considered in the discrete mixture model. Finally, I investigate duration by exit type in a competing risks framework, where competing risks are analyzed by treating all self-employment exits, other than the one being analyzed, as censored. (See Jenkins (2005) for a detailed discussion of competing risks models).

Liquidity Constraint Proxies

A common method used to ascertain the presence of liquidity constraints is to investigate the effect of wealth on self-employment outcomes. A positive coefficient on wealth is consistent with binding liquidity constraints. However, measures of wealth are not always available. In such cases, investment income is often used. Fairlie (1999), Georgellis and Wall (2005), Taylor (1999), Bruce (1999), and Cowling and Taylor (2001) all use investment income as a proxy for wealth (liquidity). Taylor (1999) and Cowling and Taylor (2001) use a binary cutoff rather than a linear format. Evans and Leighton (1989) subtract labor from total earnings, to obtain an estimate of capital income, while Holtz-Eakin et al. (1994b) impute asset holdings from investment income by assuming constant rates of interest. Both Fairlie (1999) and Holtz-Eakin et al. (1994b) note that investment income, as a proxy, may be noisy. The latter authors employ inheritance as a preferred proxy, while Fairlie (1999) uses cash payments (which include inheritances) in this regard. Hurst and Lusardi (2004), however, demonstrate that inheritances are not random events since the impact of an inheritance received after starting self-employment is as much a predictor as receiving an inheritance beforehand.

Theoretically a cut-off, rather than a continuous variable, is a reasonable measure. Once a person has sufficient capital to fund a business, additional capital should have little impact on ability to enter and succeed. As such, the first liquidity constraint proxy that I employ is a flag equaling one if an individual’s own investment income is at or above $200 annually. The cut-off of $200 is chosen because it implies (at five percent interest) an amount of capital that is large enough to cover minimal start-up costs, $4,000.Footnote 7 This cut off is also consistent with Taylor’s (1999) wealth proxy, a flag for £100 in investment income.

Investment income includes (net of carrying charges): interest received on bonds, deposits and savings certificates, trust funds, estates, or other investment income, dividends, net rental income, as well as interest on mortgages, and loans. It also can include some income from partnerships and incorporated businesses, which is why it is essential that this variable be measured prior to the start of a business. Investment income does not include pension income, nor any support payments or scholarships. This measure is quite close with that used by Taylor (1999), Georgellis and Wall (2005) and Fairlie (1999).

As cited above, investment income may be a noisy measure of wealth. Moreover, neither wealth nor investment income is a perfect proxy for liquidity constraints.Footnote 8 For one, these proxies may be endogenous. Individuals with higher ability may succeed in both wealth accumulation and in self-employment survival. There are several additional concerns associated with the interpretation of the coefficient on the liquidity proxy. Cressy (1999) notes that a positive coefficient on wealth may be capturing risk aversion because higher wealth implies higher risk tolerance, which would increases the likelihood of riskier labor market choices. Furthermore, Treichel and Scott (1987) remark that women may have a preference to self-finance. This preference would generate a positive correlation between liquidity and self-employment, even in the absence of credit market constraints. Interpretation of gender differences in coefficients is also complicated. If the coefficient for men is zero, and the coefficient for women is positive, then results are consistent with the hypothesis that the liquidity constraint is binding for women, but not men. The interpretation is not straightforward if both sexes’ coefficients are positive, even if one coefficient is significantly larger than the other.Footnote 9 The size of the coefficient indicates a magnitude of response, but does not necessarily identify the severity of the liquidity constraint itself.

In order to address all but the last of these issues, I employ a novel proxy for liquidity constraints: I construct an indicator variable which equals 1 if the person withdraws funds from their Registered Retirement Savings Plans (RRSPs). Unlike wealth, which represents an accumulated stock that is potentially correlated with ability, RRSP withdrawals are flows out of savings, and given the tax penalties involved, withdrawals are not likely to occur in the absence of a financial constraint. As discussed in Rybczynski (2013), approximately 65 percent of the population holds RRSPs. Moreover, one’s own business is not an eligible RRSP investment. Thus, one must typically make a withdrawal to obtain RRSP funds for business. However, a major drawbacks of the RRSP proxy is that a relatively small fraction of the population makes an RRSP withdrawal in any given year. This low rate of withdrawal, coupled with the fact only a small fraction of individuals are self-employed, means that estimates will be based on the actions of a small number of observations. Thus, one must be cautious in using the results from this proxy to make inferences on the general population.

Data

I use Panel 1 (1993–1998) of the Canadian Survey of labor and Income Dynamics (SLID). I focus on the first panel because of the historically high sources of private funding during this period. Private funding and self-employment rates declined substantially after the dot-com bust in early 2000, and continued to fall for several years thereafter. Thus, the 1993–1998 period represents a rather stable period of easy access to credit, whereby if there is discriminatory lending, it is likely concentrated amongst marginalized groups. Moreover, RRSP contribution limits remained stable over this time period, and the absence of a recession during this time frame means that withdrawals are less likely to be made for income smoothing.

A major benefit to using the SLID is that it is a nationally representative, longitudinal survey, with detailed information on personal and job related characteristics for individuals (aged 15 and older), and their family. Information is collected for up to six jobs per year per person. The baseline sample is restricted to those who are over 20 at the start of the panel. Moreover, persons who are out of scope, or leave the panel due to institutionalization, or international migration are excluded. Because the numbers of self-employed are small, I drop observations for which data is missing, except when doing so would significantly diminish the sample. In such cases I include an indicator for the missing variables. Aside from stronger predicted associations and slightly lower statistical significance, results are not sensitive to this decision.

Duration analysis is carried out on self-employment job spells. Thus, each observation represents a self-employment job spell, not a person.Footnote 10 Individuals with more than one self-employment job spell will have more than one observation in this sample. As such, standard errors are adjusted by clustering on the person identifier. A job is categorized as self-employed if the reported class of worker is self-employed, incorporated or unincorporated, with employees or without.Footnote 11 Because I observe exact start and end dates in the SLID, I can retain multiple spells within the same year. Self-employment job duration is calculated as the difference between spell end and start dates. However, a fraction of spells do not end within the panel or have missing end dates (these are treated as right censored), and a fraction of spells start before 1993 or have missing start dates (these are treated as left censored).Footnote 12 Left censored spells are dropped, as is standard practice in the duration literature. Moreover, in order to observe the characteristics of the self-employed in period t-1, I only consider self-employment spells that start on or after 1994. Finally, home ownership information was not collected for 1993, so I extrapolate that those who owned a home in 1994 did so in 1993 as well. Results are not sensitive to these decisions.

The final sample of self-employment job spells contains 987 male and 734 female spells. I conduct robustness checks on alternative sample criterion, and results are substantively similar.

Results

Spell durations by gender are presented in Table 1. On average, female spells are shorter than male spells. This result is echoed in the Kaplan-MeierFootnote 13 survival curves, depicted in Fig. 1. Note that survival time is scaled to months, rather than days, for ease of viewing. Women have slightly higher survival probabilities over the first few months of a job spell, but for all later months men have substantially higher survival probabilities. A log rank test for the equality of the survivor functions rejects gender differences in duration. Table 2 shows summary characteristics of self-employment spells by gender. Provincial unemployment rate, squared terms, missing value, industry, regional, and annual indicators are omitted for brevity. The values of characteristics are taken from the year prior to the start of the spell. Of note is that women in self-employment jobs spells are less likely than men to have high investment income, to own a home or to withdraw RRSPs prior to the start of the job.

Fig. 1
figure 1

Kaplan-Meyer survival curves by gender

Full size image
Table 1 Self-employment spell characteristics
Full size table
Table 2 Sample characteristics (year prior to spell start)
Full size table

Cox proportional hazard estimates for the baseline specification are reported in Table 3.Footnote 14 Rather than coefficients, the hazard ratio contributions are presented, due to the ease of interpretation. A value less than one indicates a negative impact on the hazard ratio, and thus a lower propensity to exit self-employment. The hazard rate contribution of 0.69 for investment income 200 + implies that those with over $200 in investment income have a 31 percent lower probability of exit than those with less than $200 in investment income. Because I report hazard ratios, and because their relationship with the coefficients is non-linear, P-values are presented rather than standard errors.

Table 3 Self-employment exits, by gender, Cox proportional hazard rate model
Full size table

Estimating the proportional hazard model for each sex individually, I note that different characteristics correlate with male versus female survival probabilities. In particular, additional children and self-employed in family are associated with an increase in the probability of exit for women, but a decrease for men. Moreover, the hazard ratios on the liquidity proxy are distinct. Unlike Taylor (1999), I find results consistent with the credit constraint hypothesis for women. A self-employed woman who has $200 or more in investment income, the year before she starts a self-employed spell, is associated with approximately 31 percent lower probability of exit. The same is not true for men. A self-employed man, with high investment income, faces a half percent greater probability of self-employment exit. Home ownership indicator variables are insignificant for both sexes.

I consider whether estimates on the liquidity constraint proxy are robust to alternative specifications, subsamples, and proxies. Table 4 presents this sensitivity analysis. The first row re-states the baseline specification estimates. In the second row, I present the hazard ratios for spells of unmarried observations only, since the financial resources of a married person may greater than their individual resources. However, I find that the hazard ratios for the unmarried subsample are not significant.

Table 4 Cox proportional hazards results on the liquidity constraint proxy, by gender, for alternative specifications, samples
Full size table

Industry Canada (2003) reports that new businesses are less likely to obtain commercial financing than existing firms. My results suggest the same. When I omit spells of those who were previously self-employed (those who would most likely have developed banking relationships) high investment income is even more strongly associated with survival for female spells. Female self-employment spells with high investment income have nearly a 60 percent lower probability of failure than those without high investment income. Moreover, this estimate is significantly different across the sexes.Footnote 15

Alternative proxies, in the form of alternative investment income covariates (rows 4–6), are consistent with the liquidity constraint hypothesis, with the exception of investment income over $400.Footnote 16 Similarly, estimates remain consistent with liquidity constraints when I include additional controls for education, industry, management and self-employment experience, involuntary end to last job, full-time, worked at home, spousal self-employment, spousal log income, and unemployment insurance receipt (row 7). In row 8, I consider whether estimates are robust to dropping spells with missing values, older and younger observations, as well as agricultural and professional workers (restricted sample), results are unchanged.

Although my duration data is at the daily level, daily is still inherently discrete. As such, I consider a discrete proportional hazard model popularized by Prentice and Gloeckler (1978). Results, in row 9, are similar to the continuous time model. I also estimate models with unobserved heterogeneity for both the continuous and discrete cases. For the continuous case, I use a gamma distributed frailty model, and for the discrete case, the Heckman and Singer (1984) approach (rows 10 and 11).Footnote 17 In both cases, the hazard ratio remains less than one and statistically significant for female spells. In other words, high investment income is associated with a lower probability of failure among women. Finally, in row 12, I report the estimated hazard ratio on the RRSP withdrawal indicator. The estimate for female spells is less than 1, but it not significant; however, such a result should not be surprising given the small sample of self-employed spells coupled with the small fractions that withdraw RRSPs. For example, in the base sample, only 36 female spells involve a withdrawal of RRSPs. Additional robustness tests include Cox proportional hazards regressions with alternative methods for ties (the Efron and exact methods), and the baseline specifications on a sample where I remove the restriction of spell starts after 1994. These results are omitted because the coefficients from these models were substantively similar with significant hazard contributions around 0.6–0.7 (on the liquidity proxy) for female spells. The ratios are typically over 1 and insignificant for male spells.

The final analysis in this paper is a consideration of self-employment duration by exit type. I use the competing risks categories employed by Taylor (1999), but further disaggregated. I divide business exits into three categories: failure, non-failure and personal. Failure includes self-employment job ends due to: going out of business, business slowdown, dismissal, movement of company and workplace conflict. The latter three purport to capture partnership issues. Non-failure is applied to those who find a new job or who will focus on another job (which may or may not be self-employment). Personal reasons include illness/disability, moving, school, and retirement. There are too few observations which indicate family related reasons for exit to include these as a separate category from “Personal.”

The primary motivation for disaggregating analysis by exit type is that an individual whose business ends for personal reasons or due to a better job offer, should not be considered the same light as those who exit involuntarily. In the case of better job offers, optimal capitalization may even increase the probability of an exit from self-employment. Selling one’s business and/or accepting an upper management position for a larger company are more probable if current business is successful. Exiting for personal reasons might also be positively correlated with liquidity (if wealth enables loose labor attachment). However, sufficient funds are likely to reduce exits when exit implies a non-voluntary business failure. Indeed, results, presented in Table 5, support this hypothesis among male spells. For involuntary or failure exits, investment income $200 + reduces the probability of exit for both male and female spells, significantly so for the former. For non-failure exits, men with high investment income are significantly more likely to exit a self-employment job spell, women are not. Indeed, women’s response to the liquidity proxy is similar across all exit types, suggesting that if the self-employment job provides at least sufficient earnings, women may not seek to use entrepreneurship to leverage salary. This interpretation is consistent with Cliff (1998) who suggests that women entrepreneurs typically prefer smaller businesses than men and do not seek growth beyond a fairly low threshold.

Table 5 Liquidity Constraint Proxy Hazard Contribution, by Gender and Exit Type
Full size table

Discussion and Caveats

In sum, I find differences in the determinants of self-employment duration among Canadian men and women that are consistent with the work-family balance hypothesis and, to some extent, the liquidity constraint hypothesis. The probability of exiting a self-employment spell increases in the number of children for female spells, but decreases in the number of children for male spells. Being out of the labor force prior to spell start, is associated with a higher probability of failure for female spells, while the result for male spells is insignificant; however the gender difference in hazard ratios is not statistically significant. Evidence consistent with liquidity constraints is found using a series of proxies for personal capital. High levels of investment income are correlated with a lower probability of exit for female self-employment spells. While the hazard for the liquidity proxy is insignificant for men across most specifications. The gender difference is statistically significant for the sample that is restricted to those who were not previously self-employed. However, when I investigate duration by exit type, results support the credit constraint hypothesis for men as much as women, for the “failure” exit type. In the case of failure, high investment income is associated with a significant reduction in the probability of exit for male spells. Whereas investment income $200 + correlates to an increase in exit rates to ‘non-failure’ outcomes for male spells, but not female spells, indicating that men may utilize self-employment to obtain better wage jobs, while women do not. Note that gender differences are only significant for the non-failure exit type.

This paper reports estimates consistent with the theory that female self-employment spell motivated by family responsibilities (children) may have lower rates of survival. Moreover, results are consistent with the hypothesis that removing liquidity constraints would increase female self-employment duration, regardless of exit type. For men, liquidity constraints also appear to inhibit business failure, but increase the conditional probability of non-failure exits. These results are apparent during the peak of small business lending, and as such represent conservative estimates.

One caveat is that caution should be applied in the interpretation of the investment income results. First, because the investment income variables are likely to suffer from endogeneity. Second, a positive correlation between wealth and self-employment may result from decreasing relative risk aversion instead of credit constraints. However, results with the alternative proxy, the RRSP withdrawal flag, suggest that risk aversion is not the entire story. A hazard ratio below one on the RRSP withdrawal flag suggests that many women withdraw funds prior to starting a business. The RRSP withdrawal flag cannot distinguish whether this estimate is picking up liquidity constrains or a preference to self-fund. Yet, whether the issue is credit constraints or a preference to self finance, one should note that both have at least one similar policy implication: improving access of credit to women can increase credit utilization even if the preference for self-financing is strong. Outreach and micro-credit programs could be effective, both in reducing constraints and in helping women become more comfortable with external funding.