Modeling the determinants of turnover intentions: a Bayesian approach

Document Type



employee turnover; research methodology; human resource management; model averaging

Identifier Data



Emerald Insight

Publication Source

Evidence-vased HRM, 6(1), 2-24



Much of what we learn from empirical research is based on a specific empirical model(s) presented in the literature. However, the range of plausible models given the data is potentially larger, thus creating an additional source of uncertainty termed: model uncertainty. The purpose of this paper is to examine the effect of model uncertainty on empirical research in HRM and suggest potential solutions to deal with the same.


Using a sample of call center employees from India, the authors test the robustness of predictors of intention to leave based on the unfolding model proposed by Harman et.al. (2007). Methodologically, the authors use Bayesian Model Averaging (BMA) to identify the specific variables within the unfolding model that have a robust relationship with turnover intentions after accounting for model uncertainty.


The findings show that indeed model uncertainty can impact what we learn from empirical studies. More specifically, in the context of the sample, using four plausible model specifications, the authors show that the conclusions can vary depending on which model the authors choose to interpret. Furthermore, using BMA, the authors find that only two variables, job satisfaction and perceived organizational support, are model specification independent robust predictors of intention to leave.

Practical implications

The research has specific implications for the development of HR analytics and informs managers on which are the most robust elements affecting attrition.


While empirical research typically acknowledges and corrects for the presence of sampling uncertainty through p-values, rarely does it acknowledge the presence of model uncertainty (which variables to include in a model). To the best of the authors’ knowledge, it is the first study to show the effect and offer a solution to studying total uncertainty (sampling uncertainty + model uncertainty) on empirical research in HRM. The work should open more doors toward more studies evaluating the robustness of key HRM constructs in explaining important work-related outcomes.