The Interaction of Job Satisfaction, Job Search, and Job Changes. An Empirical Investigation with German Panel Data

Abstract

Using the rich data set of the German Socio-Economic Panel (GSOEP) this article analyzes the effects of job characteristics on job satisfaction as well as the conditions under which low job satisfaction leads to job search, and under which job search leads to job changes. Individual fixed effects are included into the analysis in order to hold unobserved heterogeneity constant. According to the empirical results, the strongest determinants of job satisfaction are relations with colleagues and supervisors, task diversity and job security. Furthermore, job satisfaction is an important determinant of the self-reported probability of job search, which in turn effectively predicts actual job changes. The effect of job search on the probability of changing jobs varies with job satisfaction and is strongest at low levels of job satisfaction. The effects of job dissatisfaction on job search and of job search on quits are stronger for workers with lower tenure, better educated workers, workers in the private sector and when the economy and labor market are in a good condition.

Appendix A1: Details of the Estimation

Equations (1)–(3) form a recursive system of equations that can be estimated by treating each of the equations separately. The inclusion of fixed effects is not so straightforward in models for binary or more general ordinal dependent variables such as the (ordered) probit and the (ordered) logit model. The fixed-effects probit model leads to inconsistent parameter estimates (see for example Baltagi 2001, p. 206; Hsiao 2003, p. 194), and the fixed-effects logit model can only be estimated on the subsample of individuals that have longitudinal variation in the dependent variable, which leads to small sample sizes and selective samples.⁹ To circumvent these problems, I applied linear fixed-effects models to the binary and to the ordinal dependent variables. In the case of multinomial ordered variables with more than two classes (job satisfaction and job search), I rescaled the dependent variable before applying the linear regression model as proposed by van Praag and Ferrer-i-Carbonel (2004). The rescaling makes the coefficients of the linear model comparable with the coefficients of the ordered probit model. Van Praag and Ferrer-i-Carbonel (2004) call this probit-adapted OLS (POLS). The rescaling consists of deriving those Z-values of a standard normal distribution that correspond to the cumulative frequencies of the different categories of the ordinal dependent variable. Suppose an ordinal variable x coded from 1 to 4 has the following distribution: P(x = 1) = 0.1, P(x = 2) = 0.3, P(x = 3) = 0.5, and P(x = 4) = 0.1. The cumulated frequencies are then P(x ≤ 1) = 0.1, P(x ≤ 2) = 0.4, P(x ≤ 3) = 0.9, and P(x ≤ 4) = 1, and the corresponding Z-values of the standard normal distribution are: Z_0.1 = −1.28, Z_0.4 = −.25, Z_0.9 = 1.28, and Z₁ = ∞. For a given value of the original ordinal variable, the value of the “cardinalized” dependent variable is constructed by considering the expectation of a standard normally distributed variable under the condition that it is in the interval between those two Z-values that correspond to the class of the value of the original variable. In the above example, this means that cardinalized variable x _c takes on the values:

$$ x_{c} = \left\{ {\begin{array}{*{20}l} {{E(Z|Z < - 1.28) = - \phi ( - 1.28)/\Upphi ( - 1.28)} \hfill} & {{{\text{if}}} \hfill} & {{{\text{x}} = 1} \hfill} \\ {{E(Z| - 1.28 < Z < - 0.25) = [\phi ( - 1.28) - \phi ( - 0.25)]/[\Upphi ( - 0.25) - \Upphi ( - 1.28)]} \hfill} & {{{\text{if}}} \hfill} & {{{\text{x}} = 2} \hfill} \\ {{E(Z| - 0.25 < Z < 1.28) = [\phi ( - 0.25) - \phi (1.28)]/[\Upphi (1.28) - \Upphi ( - 0.25)]} \hfill} & {{{\text{if}}} \hfill} & {{{\text{x}} = 3} \hfill} \\ {{E(Z|1.28 < Z) = \phi (1.28)/[1 - \Upphi (1.28)]} \hfill} & {{{\text{if}}} \hfill} & {{{\text{x}} = 4} \hfill} \\ \end{array} } \right., $$

where Z is a standard normal random variable, φ being the standard normal probability density function, and Φ being the standard normal cumulative density function, which leads to:

$$ x_{c} = \left\{ {\begin{array}{*{20}l} {{ - {\text{1}}{\text{.75}}} \hfill} & {{{\text{if}}} \hfill} & {{{\text{x = 1}}} \hfill} \\ {{ - {\text{.70}}} \hfill} & {{{\text{if}}} \hfill} & {{{\text{x = 2}}} \hfill} \\ {{{\text{.42}}} \hfill} & {{{\text{if}}} \hfill} & {{{\text{x = 3}}} \hfill} \\ {{{\text{1}}{\text{.75}}} \hfill} & {{{\text{if}}} \hfill} & {{{\text{x = 4}}} \hfill} \\ \end{array} } \right. $$

In principle, I follow this approach but I replace the Z-values from the standard normal distribution by the cutoff points from the ordered probit regression instead. I prefer this approach because it uses the information of the whole model and not only the frequency distribution of the dependent variable for the re-scaling.