Ridderprint

Chapter 6 126 6.4.3 Analysis Event history analysis (Allison, 1984) was conducted in the statistical software R (R Core Team, 2017) using the R package survival (Tharenou, 2015). We estimated the survival and hazard functions (Kaplan & Meier, 1958). The survival function (Equation 1) estimates, at time t , the probability that an employee stays beyond time t . Differences between the organizations’ survival curves were assessed statistically using log-rank tests (Harrington & Fleming, 1982). The hazard function (Equation 2) estimates, at time t , the probability that an employee leaves given survival until time t . (1) ( ) = { ≥ } = 1 − ( ) = ( ) (2) λ( ) = → ≤ < + $ ≥ % We conducted Cox proportional hazard regression models (Cox, 1972; Therneau & Grambsch, 2000) to estimate the association between survival time and multiple predictors simultaneously. Covariates for which employees had multiple time-dependent measurements were encoded as time-varying (Therneau et al., 2017). We used likelihood- ratio tests to assess model fit, where the test-statistic D is twice the difference in log- likelihood between the alternative and the null model (Wilks, 1938). We validate the proportional hazard assumption using scaled Schoenfeld residuals (Grambsch & Therneau, 1994). 12 We confirmed the robustness of our results by running our models using different strategies to handle missing values. 13 6.5 Results We conducted all analyses separately per organization. Table 6.1 displays the correlations among our main variables, but these are difficult to interpret due to the longitudinal nature of the data. At both organizations, voluntary turnover related negatively to performance ratings and tenure. At Consumer Goods, turnover also related negatively to international experience and potential ratings. While turnover related negatively to employees’ age at Shell, turnover correlated positively to the higher age band dummy at Consumer Goods. Finally, at both organizations, positive associations existed between employees’ performance, potential, and international experiences. 12 Proportionality was not violated in the Consumer Goods data. At Shell, two country dummies demonstrated significant violations because eleven graduates in Nigeria turned over early ( t < 580) whereas the only leaver in Gabon left late ( t = 2433). This did not affect the results of our hypothesis tests. 13 Results of hypothesis tests did not change when only complete cases were retained or when Shell’s missing performance values were replaced by a rating of 1.0 instead of 0.9.