Life satisfaction and self-employment: a matching approach
- First Online:
- Cite this article as:
- Binder, M. & Coad, A. Small Bus Econ (2013) 40: 1009. doi:10.1007/s11187-011-9413-9
- 1.5k Views
Despite lower incomes, the self-employed consistently report higher satisfaction with their jobs. But are self-employed individuals also happier, more satisfied with their lives as a whole? High job satisfaction might cause them to neglect other important domains of life, such that the fulfilling job crowds out other pleasures, leaving the individual on the whole not happier than others. Moreover, self-employment is often chosen to escape unemployment, not for the associated autonomy that seems to account for the high job satisfaction. We apply matching estimators that allow us to better take into account the above-mentioned considerations and construct an appropriate control group (in terms of balanced covariates). Using the BHPS dataset that comprises a large nationally representative sample of the British populace, we find that individuals who move from regular employment into self-employment experience an increase in life satisfaction (up to 2 years later), while individuals moving from unemployment to self-employment are not more satisfied than their counterparts moving from unemployment to regular employment. We argue that these groups correspond to “opportunity” and “necessity” entrepreneurship, respectively. These findings are robust with regard to different measures of subjective well-being as well as choice of matching variables, and also robustness exercises involving “simulated confounders”.
KeywordsSelf-employmentHappinessMatching estimatorsUnemploymentBHPSNecessity entrepreneurship
Self-employment is something highly valued by individuals for the self-determination and autonomy it entails (Benz and Frey 2008a). Being one’s own boss has been shown to increase individuals’ satisfaction with their job, despite drawbacks such as initially often decreased incomes through self-employment (Hamilton 2000). But are self-employed individuals happier in a broad sense not only related to their job? Does their attraction to self-employment possibly crowd out other pleasures of life, leading to overall unhappy “workaholics”? And how happy are self-employed that are forced to go into self-employment to escape unemployment? These are questions that have only incompletely been addressed in the literature so far (e.g., Andersson 2008).
The aim of our article is thus to assess the satisfaction of the self-employed with their life in general. Our contribution to the literature is fivefold. First, we focus our analysis on the satisfaction with life of the self-employed, something as of yet not sufficiently understood (Dolan et al. 2008, p. 101). Most of the previous literature on the other hand has focused on the relationship between self-employment and the narrower concept of job satisfaction. By making life satisfaction the dependent variable, which implicitly considers the trade-off between total income and job satisfaction, we are more interested in the more global well-being of the self-employed (it is a broader indicator of “total utility”).
Second, we use a large, nationally-representative dataset with a relatively large panel dimension, where annual responses are recorded for the period 1996–2006. This in itself is a useful contribution because early work on the topic has often focused on small samples (Brockhaus 1980; Cromie and Hayes 1991), and even in more recent work the data analyzed is often merely cross-sectional data (Hyytinen and Ruuskanen 2007; Block and Koellinger 2009) or data with a limited panel dimension (Bradley and Roberts 2004; Andersson 2008). In particular, smaller, specialized samples can be considered to be much more prone to problems of common method variance, something that seems less problematic in cases of large-scale multipurpose surveys (Chang et al. 2010).
Summary statistics. For variable definitions, read Sect. 4. T-tests can reject the null hypothesis of equal means in all cases but log income (for the self-employed) at levels of significance below 0.01 (Levene’s tests for unequal variances suggested that variances are not equal in all cases, also at highly significant levels)
(1) Employed mean
(3) Unemployed mean
Fourth, we distinguish between opportunity and necessity entrepreneurship in our analysis of self-employment and life satisfaction. This distinction stands to be one of the most important causes for heterogeneity in the group of self-employed, since the former are going into self-employment voluntarily to pursue entrepreneurial opportunities, while the latter are forced into self-employment to escape unemployment. As important as this distinction a priori seems, few studies account for it when analysing the impact of self-employment on individuals’ satisfaction (see, e.g., Block and Koellinger 2009).
A fifth feature of our article is that we focus on the years of transition into self-employment. Most previous studies into job satisfaction and the self-employment decision have tended to pool together new entrants into self-employment and senior self-employed individuals, implicitly grouping together individuals who have spent greatly different periods of time in self-employment, an approach which has recently been criticized (Bradley and Roberts 2004). In this article, we focus on the periods of transition into self-employment, thereby focusing on nascent entrepreneurship (as opposed to individuals who have been self-employed for many years). 1
The article is structured as follows. Section 2 gives the literature background on different employment types (inter alia self-employment) and their effects on life satisfaction. Section 3 introduces our matching estimators in more technical detail. We then present our dataset in Sect. 4, and present some preliminary regressions in Sect. 5, before moving on to our matching estimates in Sect. 6. The robustness of our findings is explored in a variety of ways. Section 7 concludes.
2 Literature review
Work is an important facet of human life and it has strong effects on individuals’ satisfaction with life or happiness (which we will use synonymously here). This relationship is especially strong and clear for unemployment, which makes individuals unhappier than can be explained by only the effect of loss of income. Effects are consistently negative across a wide range of studies (e.g., Clark and Oswald 1994; Di Tella et al. 2001; Helliwell 2003). Moreover, males are more strongly affected by unemployment and there seems to be only incomplete adaptation to continued unemployment for them (Clark 2003; Lucas et al. 2004). These effects are robust in panel studies that control for selection effects, i.e. the relationship is not due to unhappy individuals that self-select into unemployment (Winkelmann and Winkelmann 1998; Lucas et al. 2004; Oswald and Powdthavee 2008).
On the other hand, the relationship between self-employment and happiness is less clear. 2 We have “rather robust finding[s] across the nations on which data are available” that self-employment is related to higher job satisfaction (Blanchflower 2004), this being the case, e.g., in the United States (Blanchflower and Oswald 1998; Kawaguchi 2008) and for other OECD countries (Blanchflower 2000; Blanchflower et al. 2001). In contrast to this finding, however, one must also take into account the robust finding that the returns to self-employment are lower, on average, than those obtained from employment (Hamilton 2000). In addition, it has been shown that individuals in small businesses have fewer fringe benefits compared to their counterparts in large firms (Storey 1994, p. 6). The self-employed generally have lower pay than the employed, but this does not mean that the self-employed are not interested in financial rewards—in fact, it has been observed that financial success is the single most important variable associated with start-up satisfaction among a group of self-employed individuals (Block and Koellinger 2009). In addition to lower pay, there is also evidence that the self-employed have longer working weeks than paid employees (Hyytinen and Ruuskanen 2007). Interestingly enough, it has even been observed that, among the self-employed, the number of hours worked for the start-up business is positively correlated with start-up satisfaction (Block and Koellinger 2009). Taken together, these results suggest that the self-employed derive utility from their job (known as “procedural utility”, Benz and Frey 2008a, b) that cannot simply be expressed in terms of the “output” (remuneration, hours worked) associated with their jobs. 3
Self-employed individuals obtain satisfaction from leading an independent lifestyle and “being their own bosses”. Hundley (2001) finds that the self-employed are more satisfied with their jobs mainly because of greater autonomy, but also because of more flexibility, skill utilization and, to some extent, higher (perceived) job security. Relatedly, empirical work has shown that employees have a lower job satisfaction in large firms compared to small firms (Idson 1990; Benz and Frey 2008a), and this can be explained to a large extent by “procedural” aspects of work such as the nature of the work tasks and the ability to use one’s own initiative (Benz and Frey 2008a). 4
Other researchers have found that self-employment can be associated with a dissatisfaction with previous circumstances. For example, Kawaguchi (2008) observes that job quitting tends to follow low job satisfaction. Noorderhaven et al. (2004) observe that the levels of “dissatisfaction with life” observed in a society are positively associated with self-employment rates.
Having a higher job satisfaction, however, does not necessarily translate into self-employed individuals being overall more satisfied with their lives as a whole. Life satisfaction in itself is a much more global evaluation of individual’s actual state of being, being influenced not only by job satisfaction but a complex and interacting web of factors (Binder and Coad 2010; Binder and Coad 2011). Since individuals might be able to compensate high achievement in some domains of life with low achievements otherwise, a high job satisfaction might be counterbalanced by lower satisfaction in the family domain, or social life more generally, or, as mentioned above, in the income domain (etc.). If the satisfying work the self-employed enjoy crowds out pleasures from other domains of life, the overall life satisfaction of the self-employed could actually be not as high as one might expect based on their job satisfaction assessment alone. And indeed, there is scant evidence so far on the relationship between happiness and self-employment (Andersson 2008, p. 231). Blanchflower and Oswald (1998) report for cross-sectional data from the United States that young self-employed are happier and in a similar vein Craig et al. (2007) provide some evidence for this relationship from Australian small businesses. Looking at European countries, Blanchflower (2004) fails to find overly strong effects of self-employment on life satisfaction (only for subgroups, self-employment is significantly related to life satisfaction, and strongly depending on the dataset used). In a recent survey of the happiness literature, Dolan et al. (2008, p. 101) thus conclude that evidence on a relationship between self-employment and happiness is insufficiently clear.
The empirically weak association between happiness and self-employment, however, could also be explained by a different, methodological phenomenon: it could be due to the fact that the self-employed are quite a heterogeneous group (Santarelli and Vivarelli 2007). 5 It has been argued that, while some individuals would gladly self-select into self-employment, others who are forced into self-employment might not appreciate the self-employed lifestyle (Fuchs-Schundeln 2009). Vivarelli (1991) writes that entry into self-employment cannot be seen merely in terms of pull factors such as expected profits (as predicted by the traditional industrial economics perspective), but there are also important push factors such as previous unemployment. In other words, in the light of the terminology in Reynolds et al. (2005), we should distinguish between necessity entrepreneurship (such as the flight from unemployment) and opportunity entrepreneurship (such as the exploitation of new business opportunities). Block and Koellinger (2009) distinguish between necessity entrepreneurship and opportunity entrepreneurship, and observe that necessity entrepreneurs have a lower average satisfaction with their startup than opportunity entrepreneurs, and also that a long period of unemployment is negatively related to startup satisfaction.
If one does not make this distinction of types of self-employment, one thus might lump together widely different individuals in the regression exercise and thus not be able to find any robust relationship between life satisfaction and self-employment. In order to separate the two possible explanations for the scant empirical evidence for a (positive) relationship between the two variables, we thus differentiate between opportunity and necessity self-employment and use a matching methodology that is better suited to deal with this individual heterogeneity than regression techniques usually employed in this context. We now turn to an exposition of this matching technique.
3 Matching methodology
“How happy would I be if I had not chosen to be self-employed?” To answer this kind of question, one must consider a counterfactual. The main problem is that if an individual chooses to be self-employed, then there is no data on exactly what would have happened had they not chosen to be self-employed.
In the case of a randomized laboratory experiment, such as a clinical trial, an accurate counterfactual can be established by referring to a control group that was not exposed to the treatment of interest. Randomized laboratory trials are often regarded as the “gold standard” (Oakes and Kaufman 2006; Imbens 2010) because subjects are randomly allocated into treatment and control groups, and as such there is no systematic difference between treatment and control groups. In randomized experiments, individuals cannot self-select into groups, and as a result treatment and control groups can be expected to be similar in terms of both observable and non-observable variables. However, establishing a counterfactual is much harder when the researcher is not dealing with randomized experimental data but instead observational data. The problem is that individuals are prone to self-select into their preferred employment category, which implies that comparing individuals from different employment categories is prone to a selection bias (this is like “comparing apples with oranges”). Conventional regression analysis is not suitable to dealing with this kind of selection bias. The approach we take is to carefully match individuals from the treatment group with individuals from the control group, to obtain more accurate estimates of the counterfactual. By comparing the treatment group with the control group, we can thus identify the causal effect of the self-employment decision on happiness. We here understand “causal effect” as defined by Rubin (1974), viz. “the causal effect of one treatment, E, over another, C, for a particular unit and an interval of time from t1 to t2 is the difference between what would have happened at time t2 if the unit had been exposed to E initiated at t1 and what would have happened at t2 if the unit had been exposed to C initiated at t1” (p. 689). For a more detailed introduction to matching, see the surveys in Imbens (2004) and Caliendo and Kopeinig (2008).
The first assumption, CIA, is a strong assumption, and it cannot be verified directly. Matching techniques aim to recreate an appropriate control group with respect to observed variables, but there is no way of knowing if the treatment and control groups differ with respect to unobserved variables. 6 There are ways in which the robustness of the matching estimator τATT can be investigated, although unfortunately most previous work that applies matching estimators has not verified the robustness of the estimates in a satisfactory way (Caliendo and Kopeinig 2008). We examine the robustness of our estimates in Sect. 6.1, inter alia using the procedure described in Ichino et al. (2008) and Nannicini (2007), which explores the sensitivity of the matching estimates to simulated confounding variables.
We use two different matching procedures in this paper and begin our matching analysis by using the nearest-neighbour matching estimator outlined in Abadie et al. (2004), which finds the nearest neighbour from the control group for each of the dimensions of X. If we have many matching covariates X, however, it becomes prohibitively difficult to find good matches for individuals in all dimensions simultaneously. On the one hand, it has been argued that omitting important variables can seriously increase bias in the resulting estimates (Heckman et al. 1997; Dehejia and Wahba 1999). On the other hand, however, including too many variables should also be avoided, because it becomes more difficult to find suitable matches, and the variance of the estimates increases. Caliendo and Kopeinig (2008, p. 39) write that “there are both reasons for and against including all of the reasonable covariates available”, and suggest that the choice of matching covariates be undertaken with reference to theory and previous empirical findings.
One alternative to nearest neighbour matching, that does not suffer from dimensionality problems when a large number of matching covariates are considered, is propensity score matching, which matches individuals by collapsing the vector of individual characteristics into a scalar propensity score. This synthetic propensity score can then be used as the single matching criterion (Almus and Czarnitzki 2003). Matching according to a propensity score implies that there is a (data-driven) tradeoff between the different dimensions—one observation might be matched to another observation that scores higher in one dimension but this is compensated for by a lower score in another dimension. These sorts of compensation are not done in nearest neighbour matching.
For our analysis, we use a dataset that is not primarily concerned with self-employment but which offers a rich variety of employment status information for a representative sample of the British populace. The British Household Panel Survey (BHPS) is a longitudinal survey of private households in Great Britain that contains information on various areas of the respondents’ lives, ranging from income to household consumption, education, health, but also social and political values. The survey is undertaken by the ESRC UK Longitudinal Studies Centre with the Institute for Social and Economic Research at the University of Essex, UK (BHPS, 2009). Its aim is to track social and economic change in a representative sample of the British population (for more information on the dataset, see Taylor 2009). The sample comprises about 15,000 individual interviews. Starting in 1991, up to now, there have been 17 waves of data collected with the aim of tracking the individuals of the first wave over time (there is a percentage of rotation as some individuals drop out of the sample over time and others are included, but attrition is quite low; see Taylor 2009).
We are using unbalanced panel data from 1996 to 2006 (waves f to p) and have a total of 78,664 observations after cleaning the panel; during the time period, two waves had to be deleted since not all of our variables have been asked in them, leaving us with a total of 9 waves. 7 We will now discuss the indicators chosen for our analysis as well as characteristics according to which we later match our individuals. While our main analysis will focus on the matching methodology described in Sect. 3, a benchmark will be a set of preliminary regressions, where we analyze the impact of different job situations on life satisfaction, job satisfaction and mental well-being.
To examine an individual’s life satisfaction, we use the BHPS’s life satisfaction question. It covers the response to the question “How dissatisfied or satisfied are you with your life overall?” It is effectively tracking an individual’s life satisfaction ordinally on a seven point Likert scale, ranging from “not satisfied at all” (1) to “completely satisfied” (7). Comparatively more studies on the BHPS center on the GHQ-12 measure of mental well-being, but recent work took up using the life satisfaction question too (Binder and Coad 2011; Clark and Georgellis 2010; Powdthavee 2009). Nevertheless, to explore the robustness of our findings, we also use the broader GHQ-12 “mental well-being” variable, which is more encompassing as it also relates to mental health. It is an index from the “General Health Questionnaire” of the BHPS, composed of the answers to 12 questions that assess happiness, mental distress (such as existence of depression or anguish), and well-being. This subjective assessment is measured on a Likert scale from 0 to 36, which we have recoded so that high values denote high mental well-being. The GHQ-12 measure of mental well-being is a remarkably valid instrument that is widely used in the medical literature. Validity and reliability have been established for many different contexts, languages and so on (see, e.g., Goldberg et al. 1997; Gardner and Oswald 2007, and the references therein). In our case, Cronbach’s α for the 12 questions is 0.8997 and well above the usual threshold values for α.
Some remarks on the validity and reliability of the life satisfaction measure seem in order; the validity of this construct has been established by an impressive psychological literature (Diener et al. 1999). There is a strong correlation between well-being measures and emotional expressions like smiling (Fernandez-Dols and Ruiz-Belda 1995) and brain activity (Shizgal 1999). Moreover, individuals tend to discontinue unsatisfactory behaviors (Kahneman et al. 1993; Shiv and Huber 2000), thus also relating (low) satisfaction scores to choice behavior (a more drastic example is the negative relation to suicide, see Helliwell 2006). Lastly, studies found that individuals are to a certain extent able to (ordinally) compare and assess other individuals’ levels of satisfaction or happiness (Sandvik et al. 1993; Diener and Lucas 1999). While there are certainly difficulties related to the practical elicitation of subjective well-being measures, a broad consensus emerged within the literature that these quite reliably measure the intended individual well-being. The test-retest reliability of subjective well-being constructs lies between 0.5 and 0.7 (over 2 weeks, both for cognitive and affective measures, see Krueger and Schkade 2008), somewhat lower than some other economic variables’ reliability.
A related issue for our study might be common method variance (CMV). 8 CMV refers to a bias in answers to survey questions when individuals systematically distort their replies, e.g. according to social desirability or their lay theory of the purpose of the questionnaire (Chang et al. 2010). These distortions might pose problems if researchers use only data from the survey in question, hence the “common method”, and it might bias results if the distorted replies are used on both sides of the regression equation. No survey is immune to the suspicion that there might be bias in replies and best practice would require researchers to take steps to minimize CMV by appropriate survey design and (post survey) model specifications. We do not think that CMV might be a big issue in our case, however. The BHPS is a large-scale anonymous survey that covers so many areas of the respondent lives, that it is difficult for respondents to outguess the survey goals and give socially desirable answers (pre survey design). While we of course cannot completely exclude the possibility to CMV, we nevertheless think this might be much less of an issue than in the usual small-scale entrepreneurship literature surveys that are often asked with very specific purposes in mind. Also note that only life satisfaction or income measures might be subject to this sort of bias in our case, since it is not really conceivable how questions regarding socio-demographic characteristics or job and marital status might be answered in distorted ways. While we have discussed the validity and reliability of the life satisfaction measure above, we think using a household income measure might also be a step in minimizing CMV as opposed to, for instance, using the respondent’s answer to the question of how much income his job or self-employment has generated. Such a specific question then might prompt especially the self-employed to overreport their income, as is usually observed in the named specialized small-scale entrepreneurship surveys (see also below). In sum, while it would be, of course, optimal to assess the given answers of individual respondents by an independent source of information, this is generally infeasible in the case of such large-scale surveys.
In this article, our main focus lies on analysing the effects of self-employment on life-satisfaction, but we also include a variable for job satisfaction in our preliminary regressions. Our variable for job satisfaction is based on the question “How dissatisfied or satisfied are you with your job (if in employment)?” and also ranges from “not satisfied at all” (1) to “completely satisfied” (7). Pairwise correlation between life satisfaction and job satisfaction in our sample is ρ = 0.4763 overall. It is higher for the self-employed (ρ = 0.5432) than for the employed (ρ = 0.4788).
Our main explanatory variable is the job status of individuals. A variety of job conditions are detailed in the BHPS, the three most important of which are being unemployed, employed and self-employed. Pooled over all sample years (n = 78,664), 41,085 (52.23%) individuals have been in employment, 5,547 (7.05%) were self-employed and 2,386 (3.03%) were unemployed. The rest were either retired (16,711; 21.24%), in some form of schooling or studies (3,490; 6.82%), had maternity leave (362; 0.46%), were long-term sick (3,259; 4.14%) or in family care (5,362; 6.82%). A total of 462 (0.59%) fell into none of these categories (the “d_other” variable. Individuals in this category were enjoying government training programs or were waiting to take up a job, inter alia). Except for these different employment types, we control for some important individual characteristics in our analysis. The most prominent of our control variables are detailed in Table 1, where we have disaggregated them for the three most important employment categories.
One important control variable is an appropriate measure of income, for which we have decided to use net equivalised annual household income (in British Pound Sterling), before housing costs and deflated to the price level of 2008, as provided and detailed by Levy and Jenkins (2008). As equivalence scales, we have opted for applying the widely accepted McClements scale (McClements 1977). In accordance with the consensus in the happiness literature, we use the logarithm of the income measure as a matching variable in our analysis, assuming that a given change in the proportion of income leads to the same proportional change in well-being (Easterlin 2001, p. 468). A remark is in order on self-reports of income in the context of self-employment and entrepreneurship: it is quite well-known that self-reports of income are quite unreliable in the context of entrepreneurs and the self-employed (e.g., Block and Koellinger 2009; Blanchflower and Oswald 1998), leading to biased estimates when controlling for income in standard regressions. A second problem lies in theoretical concerns whether one should control for income at all: “if the hypothesis is that the self-employed have higher job and life satisfaction but at the same time, they receive lower incomes, it is not certain that we want to control for income. Including income as a control variable in the fixed-effects models is even more problematic. Since becoming self-employed for the average individual means a decrease in income, it can be hard to disentangle the effect of becoming self-employed from the effect of receiving a lower income on the outcomes” (Andersson 2008, p. 218).
Using a matching approach, we are immune to both kinds of problems since we are not regressing income on life satisfaction but using the variable to match individuals who report similar incomes. Our approach thus allows us to use the information contained in the income variable without having to fear that the negative effect of lower income on happiness is entangled with the positive effect of being self-employed.
To measure individuals’ health, we focus on an individual’s subjective assessment of health (during the last 12 months). This is ordinally scaled on a five-point Likert scale, ranging from “excellent” (five) to “very poor” (one). 10 Subjective assessments of health seem to predict objective health quite well in some cases (e.g., regarding morbidity). Whether objective health is sufficiently well captured by subjective health assessments is still debated (Johnston et al. 2009). In order to account for more objective aspects of individual health, we also included a dummy variable to account for disability of an individual.
Besides income and health, our control variables also comprise the usual set of gender, age, and age2 (we use the squared difference between age and mean-age instead of age2 in order to avoid problems of multicollinearity), as well as some dummies regarding marital status (e.g., being married, being separated, divorced or widowed). We have also added a regional control variable, dummies for different ethnicities and years (which we do not report, however). Also included is an educational control variable, viz. an individual’s highest level of education, as measured by the CASMIN scale. This is measured ordinally, ranging from one (“none”) to nine (“higher tertiary”), giving intermediate values to the middle education levels. Of our sample, 53.29% were female. The mean age was 46.14 years (SD 17.94) with maximum age at 98 years and minimum age at 15 (younger individuals were not interviewed in the BHPS).
In Table 2, we report pairwise correlations between our variables. A look at these correlations offers first insights. Most of the correlations are highly significant, with the exception of life satisfaction and gender as well as log(income) and the self-employment dummy. Inspection of the correlations shows us that multicollinearity seems to be not an issue here (and in the same vein, we computed the variance inflation factors (VIF) for our regression model (1) below, which were all well below 4, with a mean of 1.45). 11 Beside ascertaining the absence of multicollinearity, the simple correlation exercise can but give a first impression of our dataset, since it does not control for any confounding factors.
Concerning the life satisfaction distribution, the self-employed generally report higher life satisfaction, although we do not strictly observe that the life satisfaction distribution for the self-employed stochastically dominates the corresponding distribution for the employed because of a handful of very dissatisfied self-employed individuals. 12
Transition matrix: mean change in life satisfaction and change in job status (“E”, “SE” and “UE” denoting employment, self-employment and unemployment, respectively) from t − 1 to t
Mean change in life satisfaction
5.1 Preliminary regressions
Ordered probit and fixed effects regressions for life satisfaction and job satisfaction; standard errors clustered at the individual level
Life satisfaction (FE)
Job satisfaction (FE)
Mental well-being (FE)
Δ Life satisfaction
Let us conclude by observing that by removing time-invariant characteristics (be that via FE or FD), we may be “throwing out the baby with the bathwater”, and effectively we may be removing some effects of interest (in particular, slow-changing character traits) by “over-controlling”. Also, fixed-effect regression suffers from other drawbacks of regression models discussed above (in particular, lack of a common support for treatment and control groups). In order to come to more reliable estimates of the effect of being self-employed on life satisfaction, we turn now to our matching estimates.
6 Matching estimates
We contribute to the literature by making use of recent developments in matching econometrics to create an accurate control group, and thus identify the causal effect of self-employment on happiness. Our dataset has comprehensive information on individual characteristics, which allows us to recreate an appropriate control group in our analysis of how satisfied individuals are with self-employment by finding a reasonably similar, if not a “perfect twin” (Almus and Czarnitzki 2003), for each self-employed individual.
Since we are also interested in the dynamics of well-being, we have chosen to examine whether there is a lagged effect of self-employment on life satisfaction, i.e. the possible impact of self-employment on life satisfaction at period t + 2. A robust finding emerging from the happiness literature is that individuals adapt to changes in their life circumstances. Hedonic adaptation, the hedonic dulling of repeated or constant affective stimuli (Frederick and Loewenstein 1999), is highly domain-specific and varies with the concrete stimulus (for example, hedonic adaptation to marriage is faster and more complete than hedonic adaptation to repeated unemployment, see, e.g., Clark et al. 2008b). The yearly structure of our panel dataset suggests the inclusion of a second year to check for hedonic adaptation but additional lags might be added in future work. Our dynamic approach, according to which we match individuals at time t and observe them again at times t + 1 and t + 2, additionally takes into account the fact that failure to control for lagged outcomes can lead to bias in matching estimators (see, e.g., González and Pazó 2008). 14
We are carrying out our analysis for two different types of matching, viz. nearest neighbour matching as well as propensity score matching. Both methods differ with respect to how individuals are matched. Nearest neighbour matching finds a match in many dimensions simultaneously while propensity score matching collapses all covariates into one composite variable (the “propensity score”). This difference has the consequence that adding too many covariates according to which one matches individuals in nearest neighbour matching results in a dimensionality problem, i.e. one is not likely to find good matches in each of the dimensions simultaneously. Therefore, for the nearest neighbour matching, we matched individuals according to a smaller number of criteria, namely, previous life satisfaction, log(income), gender, age, education, subjective health assessment as well as dummies for ethnicity and being married. Adding more criteria would have made it harder to get good matches in our context.
For the propensity score matching, we did not have pressing concerns of dimensionality (since the matching covariates are collapsed into a synthetic propensity score, and matching is performed with reference to the propensity score only). Therefore with propensity score matching, we matched individuals according to the above mentioned factors but added also the following list of covariates: year dummies, regional dummies for the different former Metropolitan counties and Inner and Outer London, dummies for being separated, divorced or widowed, a dummy for being disabled and a quadratic age term. 15
Nearest neighbour matching estimates of the sample average treatment effect (SATE)
E to SE vs. E to E
UE to SE vs. UE to E
Propensity score matching estimates of the average treatment effect on the treated (ATT), obtained using the kernel option (using the “pscore” command in Stata 11, developed by Sascha Becker and Andrea Ichino)
E to SE vs. E to E
UE to SE vs. UE to E
Interestingly enough, both matching estimators show a larger effect at the second lag compared to the first, which suggests that the positive impact on well-being is not only not transitory, but it even increases with time. Individuals who move from employment to self-employment will appreciate this transition even more two years afterwards, once they have become more accustomed to the self-employed lifestyle. The causal effect self-employment has on life satisfaction is thus (at least in the first years) not only exempt from hedonic adaptation, it seems to show the opposite—an increasing antiadaptive effect.
It is important to point out, however, that there is a marked difference between this positive effect of self-employment on happiness for individuals who switch from employment (i.e., opportunity entrepreneurs) and those who become self-employed to escape unemployment. In all these cases, there is no significant difference between the well-being of individuals who switch from unemployment to self-employment, compared to those who switch from unemployment to regular employment. In each case, the estimated effect is actually negatively signed, but far from significant (perhaps due to the lower number of observations). In the case of those leaving unemployment (i.e., when it comes to necessity entrepreneurship), self-employment has no advantage over a regular job in terms of individuals’ life satisfaction. Note also that we could not present propensity score results for the second time lag for necessity self-employed due the very small number of individuals who went from unemployment to self-employment and stayed there for more than 1 year. This in itself is interesting because it suggests that those entering self-employment from unemployment are unlikely to persist in self-employment, presumably because of business failure.
6.1 Robustness of matching estimates
While such a matching approach offers a robust way of identifying appropriate control and treatment groups, it can be quite sensitive to identification bias. In particular, problems might arise if the conditional independence assumption (CIA) is not valid. This aspect is often ignored in the literature on matching (Caliendo and Kopeinig 2008). In order to account for this sensitivity, we conduct various robustness tests.
Robustness of treatment effect estimates. Sensitivity analysis investigating the effect of calibrated confounders using the simulated approach presented in Nannicini (2007)
E to SE: 1 lag
E to SE: 2 lags
UE to SE: 1 lag
Bias reduction for matching estimates, for the “opportunity” case (E to SE)
Bias reduction for matching estimates, for the “necessity” case (UE to SE)
P > |t|
A third robustness test we conducted was to repeat our analysis with mental well-being as the dependent variable instead of life satisfaction. Mental well-being is a broader concept of well-being that includes more affective and mental health related aspects of human life (the correlation between life satisfaction and mental well-being is ρ = 0.5552). We thus repeated the matching analysis with mental well-being and obtained similar results, for both models, and for both lags, and for both matching techniques (nearest neighbour matching, and propensity score matching). 17 In other words, using mental well-being instead of life satisfaction we observed that individuals switching from regular employment enjoyed higher levels of mental well-being (at t + 1 and t + 2) than those who remained in regular employment. On the other hand, individuals who moved from unemployment into self-employment were not significantly different from those who moved from unemployment into regular employment; the coefficient was also always negative but not significant. The one notable difference between life satisfaction and mental well-being is that in lag 2, we find smaller coefficients for the opportunity self-employed than in lag 1. This could point to different dynamics of hedonic adaptation, where individuals would adapt to the increased well-being from self-employment when measured more affectively while the opposite is true with regard to a more cognitive evaluation of their individual well-being.
The difference between the intertemporal structure of life satisfaction and mental well-being could be fruitfully explored in further work, using longer time horizons. In the case of necessity self-employment, we already noticed the few cases our dataset offers for analysis so that we were not able to extend the time lag here. For the opportunity case, however, we also added a third time lag as additional robustness test and found a yet increasing life satisfaction score in both matching estimates (the coefficient for the third lag nearest-neighbour matching estimate was 0.407, SE of 0.163, significant at the 5% level; and the propensity score coefficient was 0.251, SE of 0.155, just falling short of the 10% level of significance). For the mental well-being variable, here also well-being scores are higher in year t + 3 than in both years before, so that the decrease in mental well-being in the second time period (a slump as it seems) should probably not be over-interpreted. Nevertheless, longer time horizons for the analysis would be desirable.
Finally, we repeated model 1 (the transition from employment to self-employment) and added SIC-92 industry dummies to the propensity score (no industry information was available for the previous position of the unemployed, however). Results here are virtually identical to the model estimated without the dummies, and so we do not present them separately here. Taken together, we are reasonably convinced of the soundness and robustness of our matching estimation strategy.
In previous studies, the self-employed were found to be more satisfied with their jobs than employed control groups. This finding has proven robust even though the self-employed often earn less and work more hours than individuals in regular employment. Explanations suggest that the autonomy enjoyed by “being one’s own boss” more than compensates entrepreneurs for the hardships otherwise associated with self-employment. The present article has investigated to what extent higher job satisfaction of the self-employed also translates into a more global assessment of well-being, namely, their satisfaction with life in general. Few studies so far were able to present empirical evidence on higher life satisfaction of the self-employed, either due to methodological difficulties since the self-employed are a very heterogeneous group, or due to a lack of a causal connection between the variables. Following the latter line of reasoning, self-employed individuals might not enjoy higher life satisfaction than the employed because their high job satisfaction could result in the self-employed focussing so strongly on their work that they crowd out other activities that contribute to high life satisfaction, such as social relations or health.
To account for these methodological difficulties and to explore the above-mentioned theoretical intuition, we have applied a matching methodology in order to better identify treatment and control groups, and thus enable us to estimate the causal effect that a transition into self-employment has on life satisfaction. Since individuals go into self-employment for quite divergent reasons, we have broadly distinguished two motivations in our regressions: individuals going into self-employment to escape unemployment (necessity self-employment) differ from individuals who go into self-employment to exploit a business opportunity (opportunity self-employment). In our analysis we found that individuals moving from regular employment into self-employment (the case of “opportunity entrepreneurship”) experience a positive and significant increase in life satisfaction, that actually increases from the first year of self-employment to the second. However, we also observed that individuals moving from unemployment to self-employment were not better off than those moving from unemployment to regular employment (the case of “necessity entrepreneurship”). Those moving from unemployment to self-employment actually had lower life satisfaction scores than the control group, but these differences were not statistically significant.
With these results, we can now shed some light on the initial puzzle found in the literature: self-employment increases life satisfaction for those who pursue entrepreneurial opportunities and who leave their employment to pursue self-employment. Considering the strength of the effect (and increase over time) of life satisfaction for these individuals, we deem it quite unlikely that there is some crowding out taking place, where the job domain’s pleasures lead self-employed individuals to neglect other domains of life and the pleasures found there. Of course, crowding-out could still play a role even finding an effect as we have described. While it is difficult to completely rule out the crowding-out hypothesis, a look at different satisfactions with life domains for a subsample of our sample was possible (these domains included inter alia satisfaction with social life, spouse, leisure time, house, etc.). The self-employed are in almost none of the domains less satisfied than the employed, the exception being the amount of leisure time (t-tests find no difference for amount and use of leisure time and satisfaction with spouse; otherwise the differences are significant). The absolute difference here is quite minimal, especially compared to the difference when job satisfaction is considered. Of course, this is only a simple exercise (a FE regression for self-employed versus rest of the population, where we included the domain satisfactions, also led to rather similar coefficients for the life domains, but we omit this regression here for space reasons). In sum, our analysis gives credibility to our hypothesis that the usual absence of effect of self-employment on life satisfaction is due to the limitations of standard regressions that have to rely strongly on extrapolation in the case of this very heterogeneous group of self-employed individuals. In order to further explore the welfare effects of self-employment, it is thus highly called for to differentiate the various reasons for which individuals pursue their own businesses.
Self-employment is indeed a remarkably heterogeneous category (Santarelli and Vivarelli 2007), and policy-makers would do well to recognize this important fact. Our results indicate that individuals moving from regular employment into self-employment experience an improvement in their life satisfaction scores, while individuals moving from out of unemployment express no such preference for self-employment, and might even prefer regular employment to self-employment (although our results on this latter issue are not statistically significant, no doubt due to a small number of transitions out of unemployment into employment or self-employment). This suggests that while a transition into self-employment is preferable on average, unemployed individuals might be consistently less enthusiastic about self-employment. This has implications for the design of government schemes to encourage unemployed individuals to move into self-employment (for a discussion of such schemes in the German context, see Caliendo and Kritikos 2010).
Further research might fruitfully centre on extending our findings from the British Household Panel Survey (BHPS) dataset to other countries as well as extending the analysis to cover longer horizons in order to explore the longer-term causal effects of self-employment on life satisfaction. This might be a worthwhile undertaking, considering that it has been observed that the self-employed not only have lower pay, but also that their pay increases at a lower rate over time (Hamilton 2000). It might also be worthwhile to examine why the effect of going into self-employment differs with respect to hedonic adaptation when looking at the mental well-being variable: are life satisfaction measures as cognitive assessments of one’s life perhaps more prone to the same biases of overconfidence that “entrepreneurs” exhibit, overstating thus their cognitive assessments of well-being? Or is it perhaps that it is much more difficult to attain lasting increases of affective well-being as this is more closely related to human biological functioning? These questions point to interesting future research possibilities towards better understanding the entrepreneurial mind.
Our focus on nascent entrepreneurship bears similarities to some previous work (Bradley and Roberts 2004; Schjoedt and Shaver 2007; Fuchs-Schundeln 2009); see also Andersson (2008) who focuses on changes between two cross-sections (1991 and 2000).
Van Praag and Versloot (2007, pp. 375–376) provide a brief overview over some contributions entrepreneurship has on the utility levels of entrepreneurs and their employees.
Cooper and Artz (1995) found that entrepreneurs with initially high expectations for their business venture performance turned out to be more satisfied than other entrepreneurs, suggesting that these more satisfied individuals have some more optimistic personality traits that influence their subsequent job satisfaction. This finding does probably only pertain to those entrepreneurs that create their business out of opportunity, not to escape unemployment.
The positive effect of being self-employed on job satisfaction diminishes markedly when taking into account the heterogeneity of the control group of the employed in terms of the size of the firm they are working in (Benz and Frey 2008a, p. 374).
Another source of heterogeneity might stem from distinguishing entrepreneurship from self-employment, of which we abstract here for reasons of data limitation.
Note however, that this problem does not apply to randomized experiments, because treatment and control groups have no systematic differences with respect to both observed and unobserved variables.
Waves before 1996 did not elicit life satisfaction measures and had to be discarded thus.
We are grateful that a referee made us aware of this issue that might affect survey data.
A Levene’s test for unequal variances was conducted before the t-tests, suggesting in all cases significantly differing variances.
We have reversed the numerical order of the Likert scale to consistently use higher values for better outcomes. The original coding in the BHPS codes a value of one to be excellent health and five to be very poor health.
To be precise, computing VIF is only possible in the context of OLS models, so that we reestimated model (1) as a simple OLS regression.
We observe that 0.63% of the self-employed report a life satisfaction score of 1, compared to 0.50% of employed individuals.
The rationale for model (5) lies in some econometric reservations one could have in our using an ordinal scaled life-satisfaction variable in a fixed effects OLS regression, thus implicitly treating life satisfaction as a cardinal variable. This is in part motivated by the absence of a commonly agreed-on method to account for fixed effects in an ordered probit framework. However, econometric research on happiness shows that there are no substantial differences between both approaches in terms of the results they generate (Ferrer-i-Carbonell and Frijters 2004) and a cardinal treatment of life satisfaction is common in the psychological literature on well-being. One reason for the robustness of the life satisfaction measure to being treated as cardinal could lie in the fact that individuals seem to convert ordinal response labels into similar numerical values such that these cardinal values equally divide up the response space (Van Praag 1991; Clark et al. 2008a). Nevertheless, model (5) with a 37-point scale alleviates the possible objection to using life satisfaction in an FE framework, since treating a 37-point scaled mental well-being construct as a cardinal variable seems much more uncontroversial.
Cooper and Artz (1995) focus in their analysis on third year business because prospects can fluctuate and in the beginning, uncertain prospects might lower job satisfaction. This initial uncertainty offers thus an additional reason for taking into account the intertemporal structure of life satisfaction following one’s decision to go into self-employment.
We also wanted to match individuals according to industry in which they are employed (or had their last employment), but since data reporting definitions changed over the sample period and were only available for the subgroup of employed individuals, we were prevented from doing this except for the robustness exercise reported below.
We only present results from the “pscore” command (Becker and Ichino 2002) since these two estimators are seen to give virtually identical results.
The authors will provide the detailed results of this exercise on request.
No individuals were mistreated during our matching procedures. We are grateful to Tom Broekel, Rob Byrne, Jan Fagerberg, Steffen Künn, Ben Martin, Maria Savona, Josh Siepel, Jagannadha Pawan Tamvada, Dagmara Weckowska, Ulrich Witt and seminar participants at SPRU (University of Sussex) and Kingston University, and also to Bram Timmermans for some interesting suggestions, comments, etm. We also want to thank two anonymous referees for many invaluable comments and suggestions. The authors are grateful for having been granted access to the BHPS data set, which was made available through the ESRC Data Archive. The data were originally collected by the ESRC Research Centre on Micro-Social Change at the University of Essex (now incorporated within the Institute for Social and Economic Research). Neither the original collectors of the data nor the Archive bear any responsibility for the analyses or interpretations presented here. Any remaining errors are ours alone. We are grateful for financial support from the ESRC, TSB, BIS and NESTA on grants ES/H008705/1 and ES/J008427/1 as part of the IRC distributed projects initiative. Alex Coad also received funding from the AHRC as part of the FUSE project.