Análise de regressão logística para predizer a evasão de colaboradores de uma empresa

Abstract

The objective of this study was to find a logistic regression model that could estimate the probability of an employee leaving a company based on the explanatory variables - level of satisfaction, last evaluation score, number of projects, average number of hours worked per month , working time at the company, occurrence of work accident, promotion in the last 5 years, department, salary range (low, medium, high) and type of work (field, office and laboratory). For this, 4 different regression models were studied, Model 1, Model 2, Model 3 and Model 4, which were created with modifications in some variables such as department and interaction factor. All models were split on a training and test basis to assess which model had the best fit and predictability. The models were evaluated in relation to the area under the ROC curve (AUC). The model with the highest AUC was Model 4 (87.09). It was possible to observe that the variables Average number of hours worked and Salary range contribute positively to the employee's departure; the variables Level of satisfaction, Number of projects and Time worked decrease the chance of leaving the company. It is also noted that employees with an accident at work (78.57% lower) and promotion in the last 5 years (76.04% lower) are less likely to leave the company. Those employees who work in the laboratory (34.87% smaller) or office (6.83% smaller) are less likely to leave the company compared to those who work in the field. When time worked in the company is low, the impact of high grades is a decrease in the employee's chance of leaving. When seniority is longer, high grades increase the employee's chance of leaving. When the last evaluation score is low, the increase in the time worked consists of a decrease in the employee's chance of leaving. However, when the last evaluation score is maximum, the increase in the time worked means an increase in the chance of leaving. Finally, even with a good ability to adjust ROC and predictability, there are variables that are not measured and can influence whether or not the employee leaves, such as the psychological factor.