Heterogeneous returns to personality: the role of occupational choice

Abstract

We analyze the role of personality in occupational choice and wages using data from Germany for the years 1992 to 2009. Characterizing personality by use of seven complementary measures (Big Five personality traits, locus of control, and a measure of reciprocity), the empirical findings show that personal characteristics are important determinants of occupational choice. Associated with that, identical personality traits are differently rewarded across occupations. Hence, breaking down the analysis on the level of occupational groups provides more detailed results of returns to personality. By evaluating different personality profiles, we additionally estimate the influence of personality as a whole. The estimates establish occupation-specific patterns of significant returns to particular personality profiles. These findings underline the importance to consider the occupational distribution when analyzing returns to personality due to its heterogeneous valuation.

Appendices

Appendix

Construction of measures for non-cognitive skills

Measures of non-cognitive skills are created by extracting information from the questions answered in SOEP, see Gerlitz and Schupp (2005). The Big Five are conscientiousness, extraversion, agreeableness, openness to experience, and neuroticism. Each of the traits is measured with the help of three questions relating to this particular attitude. Reciprocity is assessed with six questions and locus of control with the help of ten question. Questions relating to non-cognitive skills are answered with 7-point Likert-scales ranging from 1 for “does not apply” to 7 for “does apply”.

Before defining the scales, we conducted a principal component analysis in order to find out which items really load onto the facets. Principal component analysis applies an eigenvalue-decomposition of the correlation or covariance matrix. The extracted eigenvectors describe a series of uncorrelated linear combinations of the variables that contain most of the variance. Thereby, the number of underlying factors is analyzed and it is controlled whether variables intending to assess these factors are indeed able to measure those. The items that have been shown to load onto the facets are then used to define our measures of non-cognitive skills.

In case of the Big Five, we assume five underlying traits for the 15 variables. Building averages of three items each to construct a measure for a personality trait is only appropriate if all three items can clearly be attributed to that trait. Applying the Kaiser-Guttman criterion, we consider as many factors as there are eigenvalues larger than one. This procedure reveals that the 15 items describing the Big Five have indeed five underlying factors and all items clearly load onto them. For locus of control, principal component analysis generated two eigenvalues larger than one and another one equal to one, pointing to two or even three underlying factors. However, there were only two items clearly loading onto the second factor. Therefore we extract only one factor which would be perceived to characterize external attitude. As the concept of locus of control is primarily seen as a one-dimensional concept ranging from external to internal (see Rotter 1975, for a discussion of unidimensionality-multidimensionality), we decided to build one scale.³² It includes the two items that load on the second factor (internal locus of control), but discards four items that did not load unambiguously onto both factors. The final scale therefore includes six items and higher values are coded to a more external attitude.

Regarding the six variables aiming at measuring reciprocity, principal component analysis revealed two eigenvalues larger than one. These two underlying factors can be defined as positive and negative reciprocity. However, we decided to extract one single factor capturing the intensity of reciprocity since this is a more general trait which is of interest with respect to social cooperation apparent in professional environments.

Now we began defining the variables measuring personality facets. Some of the variables of the Big Five traits are reversely coded which is regarded in the calculation of the average scores. Personality variables were set to missing in case that one or more of the items defining it was missing. In the next step we build seven measures, one for each of the non-cognitive skills that we extracted. For the Big Five inventory, each measure is the average value of the three variables that load onto the particular trait. For locus of control, we only use six out of ten variables to construct the average as has been explained above.

Finally, scales of all seven non-cognitive skills are standardized to have mean zero and unit variance. For the analysis of occupation-specific returns to non-cognitive skills, the scales are standardized within each group of occupation to have mean zero and unit variance. Thereby, consistent interpretation across occupational groups is possible: returns to non-cognitive skills can be interpreted as the percentage change in wage for a change of the score by one standard deviation.

Further Tables

Robustness check: selection correction

Given the possible selection bias of the occupation-specific sample due to correlated error terms of occupation and wage ($Cov(u_{ij}, \varepsilon _i)\ne 0$), we estimate an augmented wage equation which aims at correcting for the selection process. When conditioning on occupation, we have to account for the fact that the distribution of occupations is the result of sorting due to characteristics and preferences of individuals as well as labor demand factors that have a direct influence on wages, too. We apply a correction method proposed by Bourguignon et al. (2007), who put forward an extension to the approach of Dubin and McFadden (1984). While Dubin and McFadden (1984) assume linearity between $\varepsilon$ (error term of the wage equation) and the Gumbel distributed original error terms from the multinomial logit ($u_j$), Bourguignon et al. (2007) propose a transformation of the $u_j$ to normally distributed error terms $u_j^*$ so that the linearity assumption holds for the $u_j^*$ and $\varepsilon$:

$$\begin{aligned} u_j^*=J(u_j)=\varPhi ^{-1}(G(u_j)), \quad j=1,\ldots ,8 \end{aligned}$$

(9)

with $G(u)=\exp (-e^{-u})$ as the cumulative distribution function of the Gumbel distribution. Then for every $j$, the expected values of $\varepsilon$ and $u_j^*$ are linearly related

$$\begin{aligned} E(\varepsilon | u_1, \ldots , u_8) = \sigma \sum \limits _{j=1,\ldots ,8} r_j^*u_j^* \end{aligned}$$

(10)

where $r_j^*$ is the correlation between $\varepsilon$ and $u_j^*$. Bourguignon et al. (2007) then show that the outcome equation can consistently be estimated. In our approach, this means that the wage equation conditional on choosing occupation $j$ can be written as

$$\begin{aligned}&{\mathrm {ln}}(\hbox {wage}_\mathrm{i}|{\mathrm {occ}}_\mathrm{i}=\hbox {j})=\alpha _i + \mathbf {X'}_i\mathbf {\beta } + \mathbf {n'}_i\mathbf {\gamma } \nonumber \\&\quad +\, \sigma \left[ r_j^* m(P_j) + \sum \limits _{k=1,\ldots ,8, k \ne j} r_k^*\cdot m(P_k) \frac{P_k}{(P_k-1)} \right] + w_j, \end{aligned}$$

(11)

where $j=1,\ldots ,8$ and $i=1,\ldots ,N$. Furthermore, $w_j$ is a residual that is mean-independent of the regressors and $m(P_j)$ is the integral

$$\begin{aligned} m(P_j)=\int J(v-\ln P_j)g(v)dv, \quad \forall j, j=1,\ldots ,8 \end{aligned}$$

(12)

with $g(u)=\exp (-u-e^{-u})$ as the probability distribution function of the Gumbel distribution.³³ Overall, each wage equation has eight correction terms which are consistent estimators of conditional expected values of the residuals for each possible choice derived from the multinomial logit model.