Heterogeneous risk attitudes
Estimates from the German Socio–Economic Panel
july 2014
and job separation patterns
Universiteit van Amsterdam
Faculty of Economics and Business
Abstract
The decision to separate jobs involves uncertain outcomes. Consequently risk–attitudes may pose a significant factor when undertaking such decisions. Yet, there have been no enquiries into the explicit relationship between job separation and risk. We conduct a first investigation of the relationship between job separation and self–reported willingness to take risks in general using data from the German Socio–Economic Panel (soep). Our estimation results suggest that, at means, there is evidence of a small, positive, and statistically significant correlation at the five to one percent level between the number of job separations and the self–reported willingness to take risks in general.
Heterogeneous risk attitudes and job separation patterns Estimates from the German Socio–Economic Panel
Introduction
Risk is a ubiquitous feature of economic decision making under uncertainty.
Examples of decisions with risky outcomes include choices regarding insurance (e.g. Pratt, 1964; Ross, 1981), immigration (Jaeger et al., 2010), and which employment opportunities to pursue. It is in the last–mentioned case that existing economic research has focused on the relationship between risk and unemployment (Feinberg, 1977), and risk and
occupational sorting (Bonin, Dohmen, Falk, Huffman, & Sunde, 2007). However, to our knowledge, there is a lack of literature linking risk and job separation patterns. Instead, both empirical and theoretical economic literature concerned with job separation patterns have treated individuals as homogeneous agents with risk-neutral preferences whose decisions to separate from jobs depend upon factors which are specific to the job search process.1 These factors consist mainly of experience, wage (growth) and current job
duration.
Yet, we may enquire as to whether a paradigm which ignores risk preferences and individual characteristics can adequately describe the job separation process. First, let us assume that an agent does not possess perfect information about her employment future and she is faced with a decision where she can either stay or separate from her job. In this case we speak of a decision making problem with uncertain outcomes in which risk–attitudes may play a role. Secondly, within organizational psychology there has been evidence since the prototypical work Organizations by March and Simon (1958) that individual characteristics, including among others biological factors, education, household composition and job satisfaction correlate with job separation.
The view that job separation decisions are carried out within a context of
uncertainty and the existence of theory, albeit outside of economics, which links together individual characteristics and job separation behavior serves as enough justification for us to motivate the exploration of a potential link between risk and job separation patterns. In particular, in this thesis we directly investigate whether individuals who report that they are more willing to take risks in general switch jobs more often throughout a period that consists of up to nine years at the beginning of their career.
In an attempt to answer the above hypothesis, we examine the number of job separations for each individual during the 2004-2012 period conditional upon their 2004 self-reported willingness to take risks in general. The data has been obtained from the German Socio–Economic Panel (soep). To start we describe our sample in terms of risk–attitudes and job separation characteristics. Next we regress the total number of job
1See e.g. Bartel and Borjas (1981); Blumen (1955); Jovanovic (1979); Jovanovic and Mincer (1982);
separations on self–reported willingness to take risks in general using both Poisson and negative binomial models while controlling for behavioral characteristics and employment effects.
The structure of this thesis is as follows: In the next section the available literature on job change patterns, and risk-attitudes is reviewed. After this sample selection is discussed and the key variables are described. Then the data is inspected along the dimensions of risk–attitudes and job separation patterns. Thereafter, estimation results are presented and discussed. The last section concludes.
Literature Review
The first 8-10 years of labor force participation mark a transitory period towards stable employment relationships. On average, workers hold 6-7 jobs.2 After the initial ten year period the number of job separations subsides (Hall, 1982; Jovanovic & Mincer, 1982). Job turnover is highest in the first half of this initial period (Light & McGarry, 1998; Topel & Ward, 1992). In particular, Light and McGarry (1998) have shown that roughly one quarter undergoes no job changes at all during the second four years of the eight year observation period and that almost half of all participants switch jobs at most once.
The most documented factors that result in decreased job turnover are age or experience, current job tenure (duration), wage and intra–job wage growth. Indeed, job turnover decreases with age as workers gain more labor market experience (Jovanovic & Mincer, 1982; Light & McGarry, 1998; Topel & Ward, 1992). Topel and Ward have shown that the probability of a worker leaving their first job increases by 50% with every additional five years of experience. Yet, in later jobs experience has the opposite effect: additional experience is associated with a decrease in the probability of leaving a job (1992). Because of this finding Topel and Ward (1992) argue that job change patterns
stabilize over time as individuals settle on their eventual career path. Jovanovic and Mincer (1982) argued that this result is observed mainly because job tenure increases with age; however, Light and McGarry (1998) have disputed this interdependency by controlling for individual unobservable effects and noting that the effect of experience is robust conditional upon heterogeneous levels of age.
Turning to tenure, Light and McGarry’s data displays an upward trend in job tenure towards final employment: On the left tail of the distribution, for workers who do not switch jobs, the mean [median] tenure of the first job is 8.65 [8.94] years; for workers who switch jobs once during the eight year period, the mean [median] tenure of the first job equals 2.00 [1.05] years and for the second job 6.40 [6.48] years. At the right tail of
2Light and McGarry (1998) have used eight years of data in their sample whilst Hall (1982), and Topel
and Ward (1992) have used ten years of data. Light and McGarry (1998) estimate that workers hold an average of 5.5 jobs, though Hall (1982), and Topel and Ward (1992) estimate 6.5 and 6.1-6.9 jobs respectively.
the distribution, for individuals who hold ten jobs during the observed period, the mean [median] tenure of the first job is 0.58 [0.33] years while mean [median] tenure for the tenth job equals 3.40 [3.00]; tenure in between the first and last job follows a
non-negative growth path. Thus, switching jobs leads to more durable employment (1998). However, Topel and Ward have argued that while job tenure may be negatively
correlated with job separations for short horizons, the relationship is weakened over time. The initial four quarters of employment coincide with a drop from 20% to 11% in the probability that a worker will switch jobs. After this period the observed exit hazard is relatively flat (1992). Lastly, wage (growth) and job mobility are negatively related. At means, a 10% within–career wage increase leads to a reduction in the quarterly
probability of leaving a job by 2% (Topel & Ward, 1992).
According to Light and McGarry (1998), and Topel and Ward (1992) the empirical results have shown that, among several models of job search and matching, mobility data is best approximated by the “experience good” model of Jovanovic (1979). Within this model, the quality of a match between the worker and the firm is not known ex ante but is learned over time as the match is “experienced”. During this process information about productivity becomes available to both the worker and the firm. According to Jovanovic (1979), a worker will switch jobs if the quality of the match turns out to be worse than was initially believed. In this case the wage will be adjusted downward until the worker decides to separate from her job if the wage falls below the available wage in an outside opportunity. Although match quality is considered to be time–invariant, the observed quality of a match varies over time.
However, the “experience good” model assumes that both workers and firms exhibit risk–neutral preferences and are infinitely lived (Jovanovic, 1979, p. 111). In contrast it has been generally recognized that individuals are not risk–neutral in their preferences. Moreover several papers 3 have observed significant amounts of heterogeneity within
individual preferences and risk–attitudes.4
More objections can be found in psychological literature. An argument has been advanced by Loewenstein, Weber, Hsee, and Welch (2001) that decision making under uncertainty is not well explained by the consequentialist orientation: the notion that individuals assess the desirability and likelihood of possible outcomes of choice alternatives and integrate this information into a decision through some form of
expectations–based calculus. Instead they have proposed the “risk as feelings” hypothesis
3See e.g. Dohmen et al. (2011); Eckel, Johnson, and Montmarquette (2005); Harrison, Lau, and
Rutström (2007); Harrison, Lau, Rutström, and Sullivan (2005)
4 At this point is it worth noting that extensions to the model may not necessarily lead to greater
empirical accuracy. According to Bruhin, Fehr-Duda, and Epper (2010) allowing for risk heterogeneity in models is not sufficient as most individuals do not act according to the standard model of preferences, expected utility theory.The results of Bruhin et al. (2010) reveal that the majority of individuals behave according to cumulative prospect theory (Tversky & Kahneman, 1992) but that even within this group heterogeneity is still observed.
which differentiates between the cognitive and emotional level of risk response. People react to risk cognitively and respond to it emotionally. Although there exists an interrelationship between the two response levels, their determinants are different. Cognitive factors are adequately identified by decision theory, namely probabilities and outcome valences. However, emotional reactions are sensitive to the vividness of
associated imagery, proximity in time and several other variables that are not present in decision making theory (Loewenstein et al., 2001). Even though Loewenstein et al. (2001) provide a possible explanation of which factors have an effect on decision making under uncertainty, these factors become relevant at the time in which the decision is actually taken. Applying this framework to decisions of job change, one would have to assume that such decisions are made “in the moment”. However, this need not be the case.
Indeed, Arnold and Feldman (1982) have been first to suggest that the decision to switch jobs is made over longer periods of time. According to them the turnover process begins with the perceived existence of job alternatives and the intention to search for them. The latter is in term affected by individual demographic characteristics, tenure, perceived job security, and cognitive and affective orientation towards the current job position. Consequently, empirical psychological research has also confirmed age and tenure to be significantly correlated with job turnover (e.g. Arnold & Feldman, 1982; Cotton & Tuttle, 1986). Additional factors which have been found to have a statistically significant correlation with job turnover include age, gender, job satisfaction, industry, education, number of codependents and marital status (e.g. Arnold & Feldman, 1982; Cotton & Tuttle, 1986; Royalty, 1998; Tett & Meyer, 1993). Conceptually, it is interesting to note that the economic and psychological research can be viewed as complements to one another. While Jovanovic (1979)’s “experience good” model may describe the matching process within the worker–firm context, the discussed psychological literature has added a dimension that relates job change decisions to individual
characteristics and beliefs.
Finally it is worth noting that there is an overlap between determinants of risk–attitudes and job turnover. Specifically, Dohmen et al. (2011) report that being female, being older, being married or widowed, having three children (i.e. codependents), having less years of education and working in blue collar industries is negatively
correlated with self–reported willingness to take risks in general.
In light of the overlap between the determinants of risk and the determinants of job separation patterns found in both economic and psychological literature, as well as potential limitations of current economic theory to completely explain mobility data it is possible to motivate the exploration of potential correlations between risk and job change patterns. To do this we use the soep data set which contains job mobility and risk data.
Data Description
The soep is a wide–ranging representative panel survey of approximately 11,000 households on a variety of topics such as: household composition, occupational
biographies, employment, earnings, health and satisfaction indicators (see Wagner, Frick, & Schupp, 2007, for a detailed description). The first wave of the survey was conducted in 1984. All household members over the age of 17 are surveyed.
Sample selection
We eliminate respondents if (1) no risk data for an individual is available in 2004; (2) they began employment before 2004 and (3) total job duration over the period is zero,
even though the individual was employed. The deletions associated with each criterion are summarized in table 1.
Table 1
Number of sample deletions by reason Number of
respondents
Reason for deletion
54,244 Number of respondents in 2004 wave of the soep -30,363 1. Did not answer risk question in 2004
21,881
-21,214 2. Started first job before 2004 667
-8 3. Zero job duration reported 659 Sample used for analysis
Selection rule (1) has been chosen in order to obtain the longest possible
observation period.5 Selection rule (2) is imposed to control for possible endogeneity in the form of reverse causality between employment decisions and risk–attitudes prior to 2004. Selection rule (3) is meant to correct for errors–in–variables. After imposing these sample selection rules we are left with 659 individuals.
Most notably we do not control for individuals who enter and leave the sample between 2004 and 2012 in order to maintain statistical power.6 Survival rates for our
sample are presented in fig. 1. Ten (1.5%) individuals are tracked for a period of one year and 53 (8.0%) individuals are tracked for a period of two years. After this, we see a decline in sample size of roughly 12% to 14% per year. Roughly half (46%) of our sample
5Earliest data on self–reported willingness to take risks in general is available since 2004.
6If we remove individuals who have not been tracked for the full period between 2004 and 2012
our sample drops to 60 individuals. This number is too small given that we use maximum likelihood estimation techniques.
is monitored up to five years. These differences in measurement may bias the results for the number of job separations, because individuals who leave the sample early are represented as having a possibly understated amount of job separations.
Figure 1 . Survival rates for 659 individuals in sample
649 596 513 435 356 264 179 87 0
Number of individuals left in sample
0 1 2 3 4 5 6 7 8 9
Number of years tracked in sample
Notes: Numbers of years tracked in sample since first observed
employ-ment.
Risk Attitudes and mobility
The main variable of interest is the general risk question, which prompts individuals to provide an assessment of their global willingness to take risks. The translation from German, as presented in Dohmen et al. (2011, p. 5), is: “How do you see yourself: are you generally a person who is fully prepared to take risks or do you try to avoid taking risks? Please tick a box on the scale, where the value 0 means: ’not at all willing to take risks’ and the value 10 means: ’very willing to take risks’.” Note that the measure does not involve explicit lotteries over outcomes. Instead participants are requested to provide an assessment of their overall risk attitude, across various types of lotteries. As has been noted by Dohmen et al. (2011, p. 6) the absence of explicitly formulated lotteries could lead to variability in responses as individuals formulate subjective beliefs about the riskiness of the decision environment; however, Dohmen et al. (2011) have validated the risk measure through a field experiment. Furthermore, they have found that answers to the general risk question are stronger predictors of risk attitudes in non–financial contexts as opposed to answers to a hypothetical lottery. The general risk question is available for 2004, 2006, and the 2008-2012 period. During the available periods non–response rates for the general risk question have been less than one percent (see table A1).
Job separation is determined through a categorical variable which describes whether the respondent has changed jobs since the last interview, or for first time respondents, if a job change has occurred since the beginning of the year.7 The possible responses are:
(1) “not employed”, (2) “employed, no change”, (3) “employed, no info if change”, (4) “employed, with change”, and (5) “first time employed”. Respondents can only be “first
time employed” once. If a respondent indicates “first time employed” twice or more, this is altered to “employed with change”. Non-response rates throughout 2004-2012 are near zero.8 A limitation of this variable is that it is not possible to differentiate between
voluntary and involuntary unemployment.
7For an elaborate description see http://panel.gsoep.de/soep-docs/surveypapers/diw_ssp0149
.pdf, pp. 4-5.
8One individual did not respond in 2006 and 2007; two individuals did not respond in 2009-2011
and three individuals did not respond in 2012. Approximately 19 to 22 thousand individuals have been surveyed each year.
Exploratory analysis of risk and job separation patterns
In this section we first present the distribution of risk attitudes as measured by the general risk question. Next, we proceed in a fashion similar to Light and McGarry (1998) and examine our sample with respect to job separations, and job duration after breaking down our sample by job number and number of job separations.
Figure 2 shows the distribution of general risk attitudes within our sample. Each bar indicates the frequency of individuals choosing a given number on the 11 point risk scale. The figure displays considerable heterogeneity in risk attitudes. The modal response is 5 and the mean is 5.40 while the standard deviation is 2.18. Twenty of the 659 (3%) individuals have chosen a value of 0, indicating that they are not willing to take risks at all. In contrast, 15 (2%) individuals have chosen a value of 10, indicating that they are very willing to take risks.
Figure 2 . Willingness to Take Risks in General
0.03 0.03 0.04 0.09 0.09 0.24 0.14 0.17 0.12 0.03 0.02 0 .05 .1 .15 .2 .25 .3 Fraction mean −1 sd +1 sd −2 sd +2 sd 0 1 2 3 4 5 6 7 8 9 10
Responses to General Risk Question ( 0 = not at all willing; 10 = very willing)
Notes: Response to general risk question in 2004 on an 11-point scale by
individuals from our sample.
We continue with an analysis of job change patterns within our sample. The workers in our sample encounter 529 job separations and 1,118 jobs throughout the observed period. The average number of job separations is 0.89 with a standard deviation of 0.96. This is lower than the 4.9 average found by Light and McGarry (1998) and the 6.1 average found by Topel and Ward (1992). This discrepancy in results can be
attributed to differences in sample selection: our sample is not restricted to individuals whose employment began solely in 2004 and ended in 2012.9 Because we allow individuals
9Light and McGarry (1998) include individuals if they have been tracked for eight years after the
school leaving age of 16, while Topel and Ward (1992) begin tracking individuals at 18 years of age and track for 10 years.
to enter and leave the sample during the observational period, the distribution of job separations (fig. 3) within our sample differs from that of Light and McGarry (1998).
As is revealed by fig. 3, within our sample almost half of all workers do not separate jobs at all. Because nobody has been continuously unemployed, this means that these individuals have maintained one employer throughout their presence in the sample. At the other extreme we can see that 0.3% underwent five job separations. Closer inspection of the data revealed that two individuals held six jobs during a nine year observational period. Unlike the samples of Light and McGarry (1998), and Topel and Ward (1992) most of the mass within our sample is concentrated on the left hand side of the distribution.
Figure 3 . Distribution of the number of job separations in 2004-2012
0.480 0.317 0.141 0.049 0.011 0.003 0 .1 .2 .3 .4 .5 .6 Fraction −1 sd mean +1 sd +2 sd +3 sd 0 1 2 3 4 5
Total number of job separations 2004−2012
Notes: Zero indicates that an individual did not separate jobs at all and
was thus continuously employed. Average number of job separations equals 0.89 with standard deviation 0.96.
Next we examine whether average job duration increases with every successive transition. Given the negative correlation between average job duration and the probability of transition (Jovanovic & Mincer, 1982; Light & McGarry, 1998; Topel & Ward, 1992), it would be possible to motivate the use of average job duration as a control variable in subsequent analysis. In table 2 job duration is reported after breaking the sample down by the total number of job separations and job number. For individuals who have underwent zero job separations the average job duration is 2.20 years.10 Looking down the columns for one to three job separations we can see that there is a non-negative trend in job duration. For columns with four and five job separations the data appears to be more sporadic, however final job duration does exceed first job duration.
Table 2
Duration of each job held during 2004-2012
Job Number Number of Job separations during 2004-2012
0 1 2 3 4 5 1 Mean 2.20 1.74 1.36 1.25 1.10 0.15 (S.D.) (2.04) (1.31) (1.02) (0.84) (0.74) (0.21) [Median] [1.60] [1.60] [1.10] [1.30] [0.90] [0.15] 2 Mean 2.62 1.49 1.50 1.43 0.70 (S.D.) (2.18) (1.35) (1.30) (1.23) (0.14) [Median] [2.30] [1.00] [1.50] [1.30] [0.70] 3 Mean 2.13 1.50 1.07 0.65 (S.D.) (2.03) (1.53) (1.13) (0.49) [Median] [1.60] [1.05] [0.80] [0.65] 4 Mean 1.59 0.61 0.10 (S.D.) (1.60) (0.59) (0.00) [Median] [0.85] [0.40] [0.10] 5 Mean 1.37 1.30 (S.D.) (1.02) (0.99) [Median] [1.30] [1.30] 6 Mean 1.90 (S.D.) (1.41) [Median] [1.90]
The relatively low mean job duration for individuals who have never switched jobs prompts us to investigate the composition of our sample with respect to the number of job separations and the time tracked in sample since first employment. The results are presented in table 3 overleaf.
The “total” column displays the skew in the sample which was discussed in fig. 3. Table 3 has a visible triangular appearance when separating between the zero and positive elements within the table. This pattern indicates that as individuals are tracked for longer periods of time, they also experience more job separations. While this may not be surprising, it does imply that our sample exhibits a possibly strong correlation between the period within which individuals were tracked and the number of job separations.
Table 3
Number of individuals by job separation and years tracked in sample since first job Number of job
separations in Number of years tracked in sample since first job
2004-2012 1 2 3 4 5 6 7 8 9 Total 0 10 50 69 54 40 35 19 17 22 316 1 0 3 13 20 29 36 43 35 30 209 2 0 0 1 4 9 18 17 23 21 93 3 0 0 0 0 1 3 5 14 9 32 4 0 0 0 0 0 0 1 3 3 7 5 0 0 0 0 0 0 0 0 2 2 Total 10 53 83 78 79 92 85 92 87 659
From this we can also infer that the low duration reported in the first column of table 2 is possibly overstated due to the fraction of individuals who have been tracked for a relatively brief period of time. It is possible that this will pose a problem for our analysis in the next section where we use two regression models and formulate several specifications in order to estimate the relationship between job separation and
self–reported willingness to take risks in general. Because of the potential problems with our current sample, we also provides estimates for an alternative sample that attempts to deal with these difficulties.
Model estimation of job separation and risk
In this section we perform a regression analysis of the relationship between the number of job separations and the self–reported willingness to take risks in general. In the first subsection we analyze the sample of 659 individuals that has been described in the previous section. Due to several methodological restrictions that arise when using the original sample we also conduct an analysis in the second subsection using a different sample and compare our results.
Base sample
Methodology. Denoting the total number of job separations of individual i as Ni, we assume that Ni follows a Poisson distribution conditional upon rate Λi:
P (Ni = k | Λi) = eΛiΛk
i
k! (k = 0, 1, 2, . . .), (1)
where Λi is a nonnegative random variable such that
Λi = E(Ni) = Var(Ni) (2)
Because the response variable consists of count data the appropriate regression model is Poisson regression (Hastie & Tibshirani, 1990).11 However, our response variable
does not satisfy the condition imposed by eq. (2). In our case the variance (0.96) exceeds the mean(0.80).12 Because the response variable is “overdispersed” we also estimate the
parameters for a negative binomial model (Hilbe, 2011). In this case the response variable does not follow a Poisson distribution but instead follows a negative binomial type 1 distribution (Stasinopoulos et al., 2010):
P (Ni = k | Λi, αi) = Γ(k + i/αi) Γ(k + 1)Γ(1/αi) " (Λiαi)k Λiαi+ 1 #k+(1/αi) (3)
where Λi > 0 is the mean, (Λi+ αiΛ2i) is the variance, αi > 0 is the dispersion parameter and Γ(· ) is the gamma function. Unlike the Poisson distribution, here an additional parameter is included for modelling overdispersion. Finally, robust standard errors are used to control for mild violations of underlying assumptions (Cameron & Trivedi, 2009).
We model the parameter Λi using the linear combination in eq. (4) where xi represents self–reported willingness to take risks in general for the 2004 wave per
individual i, and Z is a vector representing a set of control variables which are based on
11We have also estimated the regression using the Ordinary Least Squares (OLS) model. Our results
were qualitatively unchanged with only small differentials in coefficients.
12Additionally, the null hypothesis that the data fits a Poisson distribution is rejected with p = 0.0071
the overlapping determinants of risk and job change patterns discussed in the theoretical framework.
Λi = E(Ni| xi, Zi) = exp [β0+ β1,ixi+ β2Zi] (4) Among the Z’s we control for time tracked in sample, employment effects (Light & McGarry, 1998; Topel & Ward, 1992) and individual characteristics (Arnold & Feldman, 1982; Cotton & Tuttle, 1986; Royalty, 1998; Tett & Meyer, 1993). The time tracked in sample is the same metric as used in table 3; it is calculated by taking the difference between the year in which first employment was observed and the last year that the individual has been reported to be in the sample.
Employment effects are controlled for by including total unemployment during 2004-2012 and average job duration. Total unemployment is calculated by summing the fraction of the number of months unemployed per year over the observational period. This way we attempt to implicitly control for individuals whose job separations were not voluntary. Average job duration is calculated by summing over the final duration
recorded at each job and dividing by the total number of jobs held over the observational period. Average job duration is included based on our analysis in the previous section, where we have shown that job duration tends to increase as individuals switch jobs. Because our sample consists of individuals who have been observed over different periods, at means, the average job duration across these groups may differ.
Turning to individual characteristics we include controls for age, gender and education. While we recognize that age is time–variant, the variance does not differ across individuals conditional upon their time tracked in sample. Gender and education are included based on specifications similar to Arnold and Feldman (1982), and Royalty (1998). For education, we record the highest education attained in 2004 according to the casmin classification (Brauns, Scherer, & Steinmann, 2003). Descriptive statistics for the controls are provided in table A2.
Estimation results. Incidence rate ratios13
(irr) of regressing the number of job separations throughout 2004-2012 on 2004 self–reported willingness to take risks in general are reported in table 4. Columns 1 through 6 correspond to the Poisson model and columns 7 through 11 to the negative binomial model.14 It is immediately visible that
once duration controls are included, the overdispersion parameter (αi) is reduced to near zero values. This is suggestive of our previous findings, which indicate that the tail in our
13Incidence rate ratios are obtained by raising model coefficients to the power of e, thus accounting for
the log-normal relationship between the dependent and independent variables. Incidence rate ratios are multiplicative meaning that a change of R = ∆r in self–reported willingness to take risks corresponds to
a [IRR]R change in the number of job separations.
distribution caused by the sample selection procedure may significantly affect results.15
Indeed, we can see that when duration controls are added in columns two and eight, the coefficient on self–reported willingness to take risks drops by 1.7 percentage points relative to the base specification in columns one and seven. Furthermore, the coefficient is no longer significant. Meanwhile, the coefficient for time tracked in sample is relatively large and significant at the 0.1% level: a one year increase in the time tracked in sample corresponds to a 35.9% increase in the number of job separations. This further confirms our suspicions that our sample selection methods may significantly affect the results of our estimations.
In columns three and nine we include additional controls for education and observe that these can be significantly negatively correlated with job separation. Relative to the omitted base category which represents individuals who are still in school we can see that individuals who possess any form of schooling will incur a smaller number of job
separations over the span of the observational period. This is remarkable, because we would expect that mobility picks up when individuals finish school and enter the job search process. Even though a clear trend cannot be established, it does appear that individuals who have obtained relatively higher educational attainment in the form of vocational maturity certificates or lower tertiary education will go through less job separations (-37.1% and -86.6% respectively) than those whose highest obtained degrees consist of completing general elementary school (-30.4%) or an intermediate general qualification (-27.5%). These results stand in stark contrast with Royalty (1998) and our sample composition provides us with reasons to doubt the outcomes. It is possible that due to the earlier described skew in our sample, there is an overrepresentation of
individuals who have not switched jobs or switched jobs at most once and who are still in school.
The inclusion of gender and age controls in columns four and ten lead to the effect of risk to increase by 0.9% percentage points relative to the base specification, however the effect is not significant at the five percent level. The effects of education on job separation remain similar, while the effect of time tracked in sample increases to 40.4%. An increase of 4.5 percentage points relative to specifications two and seven. According to these specifications males undergo 17.4% less job separations than females. This is interesting to consider because Dohmen et al. (2011) have found that men consider themselves to be more willing to take risks. If our hypothesis about risk and job separation is correct, then that would imply that men separate from jobs more often. Finally the coefficient for age is slightly lower than unity, but is not significant.
15For completeness, we also provide estimates without duration controls in table A3. While excluding
duration controls does not cause coefficients for self–reported willingness to take risks in general to vary by more than about 4.1 percentage points from the results presented in table 4, it does influence the significance of the results.
T able 4 Incidenc e rate ratios of job sep ar ation on 2004 self–r ep orte d wil lingness to take risks in gener al P oisson estimates Negativ e Binomial estimates (1) (2) (3) (4) (5) (6) (7) (8) (9 ) (10) (11) Self–rep orted willingness to tak e risks in general 1.046 ∗ 1.029 1.030 1.037 1.037 ∗ 1.043 ∗ 1.046 ∗ 1.029 1.030 1.037 1.037 ∗ (0.022) (0.020) (0.0 21) (0.022) (0.019) (0.021) (0.022) (0.020) (0.021) (0.022) (0.019) Time trac k ed in sample 1.359 ∗∗∗ 1.403 ∗∗∗ 1.404 ∗∗∗ 1.476 ∗∗∗ 1.513 ∗∗∗ 1.359 ∗∗∗ 1.403 ∗∗∗ 1.404 ∗∗∗ 1.476 ∗∗∗ (0.027) (0.030) (0.029) (0.030) (0.033) (0.027) (0.030) (0.029) (0.030) Education (casmin )† 1. Inadequately Completed 1.45 3 1.556 ∗ 1.153 1.453 1.557 ∗ (0.291) (0.323) (0.264) (0.291) (0.323) 2. General Elemen tary Sc ho ol 0.696 ∗ 0.727 ∗ 0.644 ∗∗ 0.696 ∗ 0.727 ∗ (0.107) (0.113) (0.099) (0.107) (0.113) 4. In ter mediate General Qualification 0.72 5 ∗∗ 0.721 ∗∗ 0.762 ∗ 0.725 ∗∗ 0.721 ∗∗ (0.089) (0.089) (0.087) (0.089) (0.089) 5. In ter mediate V o cational 0.57 3 0.600 0.570 0.573 0.600 (0.232) (0.259) (0.169) (0.232) (0.259) 6. General Maturit y Certificate 0.877 0.908 0.940 0.877 0.908 (0.094) (0.132) (0.111) (0.094) (0.132) 7. V o cational Maturit y Certificat e 0.62 9 ∗ 0.662 0.608 ∗ 0.629 ∗ 0.662 (0.146) (0.182) (0.144) (0.146) (0.182) 8. Lo w er T ertiary Education 0.13 4 ∗ 0.153 ∗ 0.254 0.134 ∗ 0.153 ∗ (0.109) (0.127) (0.187) (0.109) (0.127) 9. Higher T ertiary Education 0.67 1 0.780 0.924 0.671 0.780 (0.213) (0.295) (0.223) (0.213) (0.295) Male 0.826 ∗ 0.899 0.826 ∗ (0.071) (0.071) (0.07 1) Age 0.987 1.005 0.987 (0.019) (0.012) (0.01 9) A v erage job duration 0.719 ∗∗∗ 0.725 ∗∗∗ 0.719 ∗∗∗ (0.021) (0.022) (0.021) T otal unemplo ymen t in y ears during 2004-2012 0.885 ∗∗ 0.919 0.885 ∗∗ (0.041) (0.042) (0.041) In tercept 0.628 ∗∗∗ 0.0962 ∗∗∗ 0.0912 ∗∗∗ 0.122 ∗∗∗ 0.108 ∗∗∗ 0.0946 ∗∗∗ 0.628 ∗∗∗ 0.0962 ∗∗∗ 0.0912 ∗∗∗ 0.122 ∗∗∗ 0.108 ∗∗∗ (0.080) (0.017) (0.0 18) (0.051) (0.018) (0.029) (0.080) (0.017) (0.018) (0.051) (0.018) Log pseudolik eliho o d -796 -690 -599 -598 -634 -554 -793 -690 -704 -702 -783 αi 0.185 ∗∗∗ 2.55e-07 ∗∗∗ 1.18e-7 ∗∗∗ 6.36e-08 ∗∗∗ 7.66e-08 ∗∗∗ (0.082) (0.000) (0.000) (0.000) (0.000) N 659 659 595 595 659 595 659 659 595 595 659 Notes: Robust Standard errors in paren theses. ∗ p < 0 .05 , ∗∗ p < 0 .01 , ∗∗∗ p < 0 .001 . †Omitted base category is casmin classification 0: “In sc ho ol” . casmin classification 3. “Basic v o cational qualification” has b een dropp ed due to the lo w n um b er of resp onden ts (2).
The specifications in columns 5 and 11 include controls for employment effects. A unit increase in self–reported willingness to take risks in general is now significantly associated with a 3.7% increase in the number of job separations at the five percent level. The effect of time tracked in sample remains significant at the 0.1% level and has now increased to 47.6%. Both controls are significant: a unit increase in average job duration is associated with a decrease of 28.1% in the number of job separations. Average job duration increases if the total job duration increases while keeping the number of jobs fixed. In this case, according the results of Topel and Ward (1992), and Light and McGarry (1998) the number of job separations should decrease as workers move towards more stable employment. Another mechanism by which average job tenure increases is if the number of jobs held decreases, while tenure is held fixed. By definition if the number of jobs held decreases then the number of job separations also decreases. Given that our sample contains a relatively large proportion of individuals who have had relatively few job separations and thus have held relatively few jobs, it is most likely the driver behind the denominator which is responsible for the observed coefficient. It is not as likely that we observe longer job durations because the majority of our workers are tracked for relatively short periods of time. Indeed, if we look at column four of table A3 we can see that the effect of average job duration is closer to unity if we do not include controls for time tracked in sample.
Finally we observe a significant negative correlation between unemployment and the number of job separations. This is remarkable because it would be expected that higher levels of unemployment would correlate positively with job separation. However, if we look at column four of table A3 again, we can see that the coefficient is positive. As such we again seek to explain this effect by accounting for differences within the time tracked in sample. It is possible that individuals who are tracked for longer periods of time also go through longer spells of continuous employment and thus display less job separations. However, these are ad hoc explanations rather than deductions which are rooted firmly in theory. We end with the specification in column six, which includes all controls. There are no significant differences relative to the previous specifications.
Overall our results are indeterminate. Although the coefficient of self–reported willingness to take risks in general is small and positive, the vicissitudes of the risk coefficient’s significance level withhold us from drawing a clear conclusion about the effects of self–reported willingness to take risks in general on the number of job
separations. In addition to this the consistently high and significant coefficient for time tracked in sample and the fact that we’ve had to provide “creative” explanations for the effects of various controls force us to state that our approach suffers from the way that the sample has been constructed. This impels us to repeat our analysis in the next section with a larger sample in which these concerns have been addressed.
Extended sample
Methodology. The methodology for estimating the relationship between the number of job separations and the self–reported willingness to take risks in general using the extended sample of 2,519 individuals does not differ much from the previous section. The 2,519 individuals have been accounted for by lifting the constraint in table 1 which excludes individuals who began employment before 2004. A full description of the sample and the selection procedure can be found in appendix B.
We realize that by including individuals who began employment before 2004, we allow for reverse causality between our dependent and independent variables. We also realize that by including individuals who began employment before 2004 we are able to model the number of job separations in two ways: either by counting the number of job separations from 2004 onwards or by beginning the count at first employment, which can be as early as the first wave of the soep in 1984. The main disadvantage of the first approach is that we overstate the low number of job separations by ignoring the job separations before 2004. The disadvantage of the second approach is that we now have constructed a sample which contains mobility data from different stages of the job search process. After all, as has been argued by Topel and Ward (1992), and Light and McGarry (1998) most job separations occur in the first 4-5 years of an individual’s career. In an
attempt to mitigate this, we use age as our main control variable. For purposes of completeness we provide estimations for both dependent variables and show that the results are qualitatively unchanged.
The parameter Λi is again modelled according to eq. (4). The difference now is that we exclude employment effects consisting of average job duration and unemployment from our controls vector Z. Average job duration is excluded because this variable would only capture tenure effects during 2004-2012. Hence, it is no longer accurate. Unemployment spells are excluded because we do not possess unemployment data that covers the entire 1984-2012 period. Summary statistics for the control variables are provided in table A4.
Estimation results. The irr’s for regressing the number of job separations
during 2004–2012 (fig. B4a) on 2004 self–reported willingness to take risks in general using the extended sample are reported in table 5. Because the coefficients of the Poisson model and the negative binomial model were nearly identical and because non-zero overdispersion is present we have not included the results of the Poisson model. For completeness we have also regressed the number of job separations during 1984-2012 (fig. B4b) on 2004 self–reported willingness to take risks in general. We have found these
results to be within half a percentage point of the results discussed below and thus we do not consider them to differ significantly.
T able 5 Incidenc e R ate R atios of job sep ar ation on 2004 self–r ep orte d wil lingness to take risks in gener al using an extende d sample of 2,519 individuals (1) (2) (3) (4) (5) Self–rep orted Willingness to tak e risks in general 1.028 ∗∗ 1.032 ∗∗ 1.021 ∗ 1.027 ∗∗ 1.033 ∗∗ (0.010) (0.010) (0.010) (0.010) (0.011) Time trac k ed in sample during 2004-2012 1.264 ∗∗∗ 1.285 ∗∗∗ 1.285 ∗∗∗ (0.016) (0.0 17) (0.017) Age 0.983 ∗∗∗ 0.968 ∗∗∗ 0.968 ∗∗∗ (0.003) (0.005) (0.004) Education ( casmin )† 1. Inadequately Completed 1.663 ∗ 1.695 ∗∗ (0.336) (0.339) 2. General Elemen tary Sc ho ol 1.940 ∗∗∗ 1.977 ∗∗∗ (0.187) (0.190) 4. In termediate General Qualification 1.642 ∗∗∗ 1.638 ∗∗∗ (0.142) (0.141) 5. In termediate V o cational 1.761 ∗∗∗ 1.744 ∗∗∗ (0.172) (0.171) 6. General Maturit y Ce rtificate 1.797 ∗∗∗ 1.775 ∗∗∗ (0.164) (0.161) 7. V o cational Maturit y Certificate 1.986 ∗∗∗ 1.956 ∗∗∗ (0.203) (0.200) 8. Lo w er T ertiary Education 1.732 ∗∗∗ 1.725 ∗∗∗ (0.258) (0.257) 9. Higher T ertiary Education 1.778 ∗∗∗ 1.766 ∗∗∗ (0.217) (0.214) Gender 0.875 ∗∗ (0.040) Constan t 0.888 ∗ 0.137 ∗∗∗ 1.436 ∗∗∗ 0.172 ∗∗∗ 0.176 ∗∗∗ (0.051) (0.016) (0.151) (0.029) (0.030) Log pseudolik eliho o d − 3264 − 3089 − 3251 − 2806 − 2802 α i 0.259 ∗∗∗ 0.0952 ∗∗∗ 0.244 ∗∗∗ 0.0454 ∗∗∗ 0.0423 ∗∗∗ (0.039) (0.031) (0.038) (0.030) (0.030) N 2,519 2,519 2,519 2,350 2,350 Notes: Robust Standard errors in paren theses. ∗ p < 0 .05 , ∗∗ p < 0 .01 , ∗∗∗ p < 0 .001 . † Omitted base category is casmin classification 0: “In sc ho ol”; casmin classification 3. “Basic v o cation al qualification” has b een dropp ed due to the lo w n um b er of resp onden ts.
Generalizing across all specifications, at means, a unit increase in self–reported willingness to take risks in general is correlated with a 2.1%-3.3% increase in the number of job separations. Compared to our base sample analysis, the effect using the extended sample is smaller yet significant across all specifications. The lower coefficients are in line with our expectations as discussed in appendix B, given that the extended sample overrepresents the number of low job separations (see fig. B4a) to an even higher degree compared to the base sample (see fig. 3).
Specifications two and three serve as comparisons for two different controls: time tracked in sample and age. Both of these controls have been used to account for the variance in the data caused by the difference in tracking individuals for different amounts of time (fig. B1) and the inclusion of additional individuals who began employment before 2004 and are thus most likely older than those who began employment during or after 2004. Under both controls the effect of self–reported willingness to take risk maintains its significance; however, it is 0.7 percentage points lower as compared to the base specification when controlling for age and 0.4 percentage points higher than the base specification when controlling for time tracked in sample. In either case the differences are low and most likely not significant. Age is shown to be negatively correlated with the number of job separations, which is consistent with theory (Dohmen et al., 2011) and our previous regression analysis (table 4). The effect of time tracked in sample is highly significant but lower than in our previous analysis, which is a consequence of the fact that in the extended sample a larger proportion of individuals have been tracked for the full nine year period (fig. B1).
Thereafter, we add controls for education in specification four and also gender in specification five. The effect of self–reported willingness to take risks in general on the number of job separations remains significant and does not deviate from the base specification by more than half a percentage point. The controls for time tracked in sample and age also do not change by any significant amount. As expected, the number of job separations for individuals who have completed any education is higher than for the individuals who are still in school. From our analysis we cannot infer whether individuals who have obtained a relatively higher degree separate jobs more or less often than those who have obtained a relatively lower degree. With respect to gender, we see that men undergo 13.5% less job separations than women. As in our previous analysis this is an interesting observation because previous research has indicated that men report that they are more willing to take risks (Dohmen et al., 2011). This would imply that men undergo more job separations.
Our results using the extended sample serve as an extension to our previous analysis by showing that significance can be achieved while maintaining a similar
magnitude of the effect of self–reported willingness to take risks in general on the number of job separations.
Conclusion
In this thesis the relationship between the number of job separations and
self–reported willingness to take risks has been examined. Our estimation results suggest that, at means, there is evidence of a small, positive, and statistically significant
correlation at the five to one percent level between the number of job separations and the self–reported willingness to take risks in general. We have arrived at this conclusion in two stages. First we analyzed the relation using a base sample of 659 individuals. Though self–reported willingness to take risks in general did appear to have a small positive effect on the number of job separations our estimates lacked statistical
significance. Therefore we conducted another regression analysis on a larger sample of 2,519 individuals which was obtained by including persons who began employment before 2004. The resulting estimated coefficients were slightly lower on average than those of the base sample; however, they were significant in all cases. While some of the results may be biased due to sample composition, the similarity in coefficients across various controls and sample sizes provides evidence that the qualitative results may hold.
Our results contain implications for existing research as well as potential future models of job separations. Most notably our results suggest that the specifications used in Light and McGarry (1998), and Topel and Ward (1992) do not fully account for the variability in mobility data because these specifications are based on the “experience good” model of Jovanovic (1979), which assumes that individuals are risk–neutral. Specifically it assumes that an agent will choose to separate jobs if 1) such an opportunity is available and 2) the wage offered by her current employer falls beneath the wage offered by the outside opportunity. However, the findings presented in this thesis cast doubts upon the justification of the risk–neutrality assumption and the simplified nature of individual decision making. Instead we propose that (extensions to) models of job separation should focus more on the behavioral aspects concerning the choice of an individual to terminate employment. While we have not investigated the exact mechanism that underlies the decision to separate from a job, we can consider a general framework which could be used to arrive at a theoretical result for modelling job separation adequately.
Perusing current models of job separation such as the “experience good” model, it becomes evident that individual agents are modelled as though they are only concerned with current wage levels. Indeed, wage is determined solely by the firm, which bases its decision on the observed output of the agent. In fact, according to Jovanovic (1979) firms are able to screen potential employees and determine an ex-ante wage based on the expected productivity. This modelling approach leads to a “one-way” interpretation of the job separation process, where firms are able to form expectations of employee output (and wages) but employees are not able to form expectations about their employment
context, output levels and wages.16
Instead an adequate model of job separation would allow individuals to form beliefs about their future wage paths. To remain close to the “experience good” model we may even consider that as information about employee productivity becomes available to both the firm and the worker over time, the worker uses this information to update her beliefs regarding her potential future wage path. Consequently the uncertainty about her future wage at the current employer is reduced over time.
When an outside job offer becomes available the worker is potentially able to obtain some information about the conditions under which she will be employed.17 She then
forms a belief about her expected wage at the outside position. However, because the quality of the match with the outside employer has not yet been experienced, the
uncertainty governing the expected wage path is most likely greater than the uncertainty governing the expected wage path at the current employer. The risk component can then be considered as the amount of uncertainty that she is willing to tolerate in order to separate jobs. Or perhaps in alternative terms: the amount of insurance premium that she is willing to pay to be insured from a potential wage decrease.
While the above discussion has been mostly based on speculation, it does motivate future research to take into account risk–attitudes18 when modelling decisions with
respect to job separation. These risk–attitudes may then be combined with beliefs that have been constructed conditional upon the available information at that time.
There are two additional components which may be worthy of closer examination. The first is that currently differences in the utility of various decision outcomes are attributed to differences in wage outcomes. That is to say that utility is a function of wage. However, this view may be myopic. Instead future studies could attempt to systematically explore the components of job satisfaction as whole. In such a case individuals would be assumed to form beliefs about future job satisfaction, rather than just future wage. These other components, such as potential career growth, social connections, intrinsic satisfaction and so forth may be potentially significant factors when modelling a mechanism that underlies the decision to separate from jobs.
The second component is due to a methodological restriction of our research and it concerns potential feedback effects between job separation and risk. Note that in our extended sample we do not control for reverse causality.19 The implication is that it may be possible for job separation patterns to influence risk–attitudes and vice versa. Future
16This is at odds with the reviewed literature, which states that one of the main factors influencing job
separation is job satisfaction. Consequently job satisfaction has a direct effect on output levels and thus wages.
17This is analogous to the ability of firms to “screen” potential employees ex ante (Jovanovic, 1979).
18And potentially other heterogeneous individual characteristics.
19Although from the similarity of the coefficients between the base sample (where we do control for
reverse causality) and the extended sample, we may derive that feedback effects do not play a significant role.
studies should attempt to control for this in an adequate fashion.
Another limitation of our research is that our sample selection method does not exclude individuals who enter and leave the sample within the observational period. The consequence is that possible job separations which would have occurred for these
individuals, are not taken into account. This leads to a potentially excessive amount of lower numbers of job separations, which in term may affect our estimates. Because these individuals may have experienced more job separations beyond the period in which they were observed, our coefficients may underestimate the true effect of self–reported
willingness to take risks in general. While we have mitigated this variability to a certain extent when using the extended sample (see fig. B1), the bias may still persist.
Furthermore, when using the extended sample without accounting for job changes prior to 2004, we have an even greater overrepresentation of low numbers of job separations. However, the estimates have been compared against estimates where we did include job changes prior to 2004 and no significant differences have been found.
Next, we do not differentiate between voluntary and involuntary job separation. Job separation that occurs due to firm–side decisions or due to macroeconomic effects are not necessarily related to individual risk attitudes. What is more, our observational period can be said to consist of roughly two parts: the 2004-2007 period, where
macroeconomic conditions have been relatively stable and the 2008-2012 period, which has been characterised by distress within the financial sector and, more importantly, a period of economic stagnation and overall economic decline. In combination with our restrictions due to sample selection, this may have caused for additional uncontrolled heterogeneity between individuals who have been tracked within the first and second parts of the observational period. While an attempt to control these effects within the base sample has been made by including controls for job duration and unemployment, omitted variable bias may still be present.
Finally, a distinction has not been made between job–to–job separation and
job–to–unemployment separation. The context in which the decision to separate from the current job is different if a worker is guaranteed a new job upon leaving their current place of employment as opposed to entering unemployment. This change of context may possibly influence the decision making process.
References
Arnold, H. J., & Feldman, D. C. (1982). A multivariate analysis of the determinants of job turnover. Journal of Applied Psychology, 67 (3), 350.
Barsky, R. B., Juster, F. T., Kimball, M. S., & Shapiro, M. D. (1997). Preference parameters and behavioral heterogeneity: An experimental approach in the health and retirement study. The Quarterly Journal of Economics, 112 (2), 537–579. Bartel, A. P., & Borjas, G. J. (1981). Wage growth and job turnover: An empirical
analysis. In Studies in labor markets (pp. 65–90). University of Chicago Press. Blumen, I. (1955). The industrial mobility of labor as a probability process (No. 6).
Cornell University.
Bonin, H., Dohmen, T., Falk, A., Huffman, D., & Sunde, U. (2007). Cross-sectional earnings risk and occupational sorting: The role of risk attitudes. Labour Economics, 14 (6), 926–937.
Brauns, H., Scherer, S., & Steinmann, S. (2003). The casmin educational classification in international comparative research. Springer.
Bruhin, A., Fehr-Duda, H., & Epper, T. (2010). Risk and rationality: Uncovering heterogeneity in probability distortion. Econometrica, 78 (4), 1375–1412.
Cameron, A. C., & Trivedi, P. K. (2009). Microeconometrics using stata (Vol. 5). Stata Press College Station, TX.
Cotton, J. L., & Tuttle, J. M. (1986). Employee turnover: A meta-analysis and review with implications for research. Academy of management Review, 11 (1), 55–70. Dohmen, T., Falk, A., Huffman, D., Sunde, U., Schupp, J., & Wagner, G. G. (2011).
Individual risk attitudes: Measurement, determinants, and behavioral consequences. Journal of the European Economic Association, 9 (3), 522-550.
Eckel, C., Johnson, C., & Montmarquette, C. (2005). Saving decisions of the working poor: Short-and long-term horizons (Vol. 10). Bradford, UK: Emerald Group Publishing Limited.
Feinberg, R. M. (1977). Risk aversion, risk, and the duration of unemployment. The Review of Economics and Statistics, 59 (3), 264-271.
Hall, R. E. (1982). The importance of lifetime jobs in the u.s. economy. The American Economic Review, 72 (4), 716-274.
Harrison, G. W., Lau, M. I., & Rutström, E. E. (2007). Estimating risk attitudes in denmark: A field experiment. The Scandinavian Journal of Economics, 109 (2), 341–368.
Harrison, G. W., Lau, M. I., Rutström, E. E., & Sullivan, M. B. (2005). Eliciting risk and time preferences using field experiments: Some methodological issues. Research in experimental economics, 10 , 125–218.
Press.
Hilbe, J. M. (2011). Negative binomial regression. Cambridge, UK: Cambridge University Press.
Jaeger, D. A., Dohmen, T., Falk, A., Huffman, D., Sunde, U., & Bonin, H. (2010). Direct evidence on risk attitudes and migration. The Review of Economics and Statistics, 92 (3), 684–689.
Jovanovic, B. (1979). Job matching and the theory of turnover. The Journal of Political Economy, 972–990.
Jovanovic, B., & Mincer, J. (1982). Labor mobility and wages. National Bureau of Economic Research.
Light, A., & McGarry, K. (1998). Job change patterns and the wages of young men. Review of Economics and Statistics, 80 (2), 276–286.
Loewenstein, G. F., Weber, E. U., Hsee, C. K., & Welch, N. (2001). Risk as feelings. Psychological bulletin, 127 (2), 267.
March, J. G., & Simon, H. A. (1958). Organizations.
Pratt, J. W. (1964). Risk aversion in the small and in the large. Econometrica: Journal of the Econometric Society, 32 (1/2), 122-136.
Ross, S. A. (1981). Some stronger measures of risk aversion in the small and the large with applications. Econometrica, 49 (3), 621-638.
Royalty, A. B. (1998). Job-to-job and job-to-nonemployment turnover by gender and education level. Journal of Labor Economics, 16 (2), 392–433.
Stasinopoulos, D., Rigby, B., Akantziliotou, C., Heller, G., Ospina, R., & Motpan, N. (2010). gamlss. dist: Distributions to be used for gamlss modelling. R package version. Retrieved from
ttp://cran.r-project.org/web/packages/gamlss.dist/index.html
Tett, R. P., & Meyer, J. P. (1993). Job satisfaction, organizational commitment, turnover intention, and turnover: path analyses based on meta-analytic findings. Personnel psychology, 46 (2), 259–293.
Topel, R. H., & Ward, M. P. (1992). Job mobility and the careers of young men. The Quarterly Journal of Economics, 107 (2), 439–479.
Tversky, A., & Kahneman, D. (1992). Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and uncertainty, 5 (4), 297–323. Wagner, G., Frick, J., & Schupp, J. (2007). The german socio-economic panel study
Appendix A
Table (A1) Non-response rates for the general risk question
Year individuals surveyed Non–response relative frequency
2004 21,881 138 0.65% 2006 22,210 148 0.67% 2008 19,639 45 0.23% 2009 20,707 85 0.41% 2010 18,848 65 0.34% 2011 21,011 58 0.28% 2012 20,730 76 0.37%
Table (A2) Summary statistics of control variables used in base sample
interval variables
N Mean S.D. Min Max
Age 659 21.873 3.616 18.000 45.000
Average job duration 659 2.068 1.686 0.100 8.900
Total years unemployed 659 0.324 0.833 0.000 6.402
categorical variables N Fraction Gender: 1. Female 351 0.533 2. Male 308 0.467 Total 659 1.000 Education (casmin): 0. In school 248 0.417 1. Inadequately completed 12 0.020
2. General elementary school 51 0.086
3. Basic vocational qualification 2 0.003
4. Intermediate general qualification 107 0.180
5. Intermediate vocational 12 0.020
6. General maturity certificate 130 0.218
7. Vocational maturity certificate 7 0.012
8. Lower tertiary education 5 0.008
9. Higher tertiary education 21 0.035
T able A3 Incidenc e rate ratios of job sep ar ation on 2004 sel f–r ep orte d wil lingness to take risks in gener al excluding dur ation contr ols P oisson estimates Negativ e Binomial estimates (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Self-rep orted willingness to tak e risks in general, 1.046 ∗ 1.055 ∗ 1.062 ∗ 1.053 ∗ 1.070 ∗∗ 1.046 ∗ 1.055 ∗ 1.062 ∗ 1.053 ∗ 1.070 ∗∗ 2004 w a v e (0.022) (0.025) (0.026) (0.022) (0.026) (0.022) (0.025) (0.026) (0.022) (0.026) Education (casmin )† 1. Inadequately completed 1.794 1.815 ∗ 1.482 1.806 ∗ 1.847 ∗ 1.500 (0.538) (0 .477) (0.403) (0.54 3) (0.492) (0.420) 2. General Elemen tary sc ho ol 1.129 1.158 1.059 1.126 1.160 1.061 (0.210) (0 .218) (0.208) (0.20 9) (0.218) (0.208) 4. In termediate general qualification 0.962 0.948 0.977 0.963 0.95 1 0.980 (0.133) (0 .137) (0.138) (0.13 3) (0.136) (0.138) 5. In termediate v o cational 0.767 0.768 0.769 0.764 0.780 0.778 (0.362) (0 .398) (0.384) (0.35 8) (0.396) (0.383) 6. General maturit y certificate 0.959 0.962 0.987 0.958 0.973 0.999 (0.121) (0 .200) (0.180) (0.12 0) (0.189) (0.175) 7. V o cational maturit y certificate 0.927 0.905 0.908 0.925 0.926 0.932 (0.343) (0 .432) (0.418) (0.34 2) (0.420) (0.416) 8. Lo w er tertiary education 0.237 0.244 0.285 0.237 0.25 0 0.292 (0.215) (0 .233) (0.270) (0.21 5) (0.237) (0.276) 9. Higher tertiary education 0.840 0.865 0.899 0.840 0.89 2 0.930 (0.278) (0 .450) (0.417) (0.27 9) (0.431) (0.412) Male 0.852 0.849 0.852 0.849 (0.086) (0.086) (0.0 86) (0.0 86) Age 0.997 0.996 0.9 94 0.994 (0.034) (0.028) (0.0 31) (0.0 26) A v erage job duration 0.894 ∗∗∗ 0.911 ∗∗ 0.891 ∗∗∗ 0.909 ∗∗ (0.023) (0.026) (0.024) (0.027) T otal unemplo ymen t in y ears during 2004-2012 1.105 ∗ 1.111 ∗ 1.108 ∗ 1.114 ∗ (0.046) (0.047) (0.049) (0.050) In tercept 0.628 ∗∗∗ 0.591 ∗∗∗ 0.650 0.721 ∗ 0.728 0.628 ∗∗∗ 0.590 ∗∗∗ 0.686 0.724 ∗ 0.768 (0.080) (0.088) (0.441) (0.101) (0.409) (0.080) (0.088) (0.419) (0.102) (0.400) Log pseudolik eliho o d -796 -705 -703 -785 -697 -793 -704 -70 2 -783 -695 αi 0.185 ∗∗∗ 0.152 ∗∗∗ 0.147 ∗∗∗ 0.145 ∗∗∗ 0.118 ∗∗∗ (0.082) (0.082) (0.080) (0 .074) (0.075) N 659 595 595 6 59 595 659 595 595 659 595 Notes: Robust Standard errors in paren theses. ∗ p < 0 .05 , ∗∗ p < 0 .01 , ∗∗∗ p < 0 .001 . †Omitted base category is CASMIN classification 0: “In sc ho ol” . CASMIN cl assification 3. “Basic v o cational qualification” has b een dropp ed due to the lo w n um b er of resp onden ts (2).
Table (A4) Summary statistics of control variables used in the extended sample
interval variables
N Mean S.D. Min Max
Age 2,519 30.278 5.415 18.000 65.000 categorical variables N Fraction Gender: 1. Female 1,287 0.511 2. Male 1,232 0.489 Total 2,519 1.000 Education (casmin): 0. In school 327 0.139 1. Inadequately completed 62 0.026
2. General elementary school 167 0.071
3. Basic vocational qualification 21 0.008
4. Intermediate general qualification 423 0.180
5. Intermediate vocational 165 0.070
6. General maturity certificate 587 0.250
7. Vocational maturity certificate 328 0.014
8. Lower tertiary education 131 0.056
9. Higher tertiary education 139 0.059
Appendix B
Sample selection
This appendix describes the selection procedure and exploratory analysis for the sample which we use to conduct a second set of estimations. The purpose of introducing this sample is to relax selection rule (2) imposed in table 1 and thus allow for a larger sample size with more statistical power.20
Sample selection is described in table B1. First we remove individuals who have not answered the risk question. Next we remove individuals for whom first employment cannot be determined and thus an accurate job count cannot be calculated. Not imposing selection rule 2 would result in additional overstatement of low job separation counts as we would include individuals who underwent the majority of their job separations prior to the first wave of the soep in 1984. It is reasonable to assume that as such the majority of their job separations have not been accounted for by the data and thus their job
separation count does not accurately reflect their actual mobility (Light & McGarry, 1998; Topel & Ward, 1992). The resulting sample consists of 2,519 individuals, which makes it 3.82 times larger than our original sample of 659 individuals.21
Table (B1) Number of sample deletions by reason
Number of respondents
Reason for deletion
77,934 Number of respondents in 1984-2012 waves of the soep -56,063 1. Did not answer risk question in 2004
21,881
-19,362 2. Started first job before 1984 2,519 Sample used for analysis
Several potential trade–offs have been made by selecting the sample as in table B1. By including job separation before 2004 we allow for potential reverse causality between employment and self–reported willingness to take risks in general. While these effects have not been documented within the context of employment, some existing literature has argued that potential feedback loops may exist between risk–attitudes and decision making (see e.g. Barsky, Juster, Kimball, & Shapiro, 1997). The second trade-off is that we now allow for even more variance in terms of age and potential job mobility patterns:
20This is especially relevant given that both the Poisson and negative binomial model rely on maximum
likelihood estimation, which necessitates the use of a relatively large sample size.
as has been described by Topel and Ward (1992), job mobility patterns tend to “even out” with age as individuals settle into a career. While this may harm the comparability of our research with that of previous works, we note that unlike research by Jovanovic and Mincer (1982), Light and McGarry (1998), and Topel and Ward (1992) we are examining the effects of risk-attitudes on job separation overall, and not for any particular
subsample.22
Survival rates for the extended sample are displayed in fig. B1. Because individuals have been included whose first employment has been observed before 2004, we have slightly modified the calculation of time tracked in sample: we now do not calculate the time tracked since first observed employment, but instead we simply look at the portion of 2004-2012 in which these individuals have been observed. The consequence is that the statistic is somewhat overstated compared to the one in fig. 1 which can be noticed by observing that no individuals drop out after one year as was the case in the original sample.
Figure B1 . Survival rates for 2,519 individuals in sample
0 250 500 750 1000 1250 1500 1750 2000 2250 2500
Number of individuals left in sample
1 2 3 4 5 6 7 8 9
Number of years tracked in sample
Notes: Survival rates are calculated by observing total time tracked in
soep over 2004-2012; in contrast to fig. 1 where total time tracked in sample was calculated since first observed employment.
However, the distortion created by the change in calculation procedure cannot account for the fact that 1,368 (54%) individuals have been observed for the full nine years compared to our original sample, where 54% of the sample remained after only five years. Furthermore, in the original sample annual declines in sample size were around 12% to 14%, whereas in fig. B1 annual decline in sample numbers is around 7% to 8%.
22Light and McGarry (1998), and Topel and Ward (1992) examine job change patterns specific to
