Attrition Prediction Case Study

Problem Statement

A large company named XYZ, employs, at any given point of time, around 4000 employees. However, every year, around 15% of its employees leave the company and need to be replaced with the talent pool available in the job market. The management believes that this level of attrition (employees leaving, either on their own or because they got fired) is bad for the company, because of the following reasons -

1. The former employees’ projects get delayed, which makes it difficult to meet timelines, resulting in a reputation loss among consumers and partners
2. A sizeable department has to be maintained, for the purposes of recruiting new talent
3. More often than not, the new employees have to be trained for the job and/or given time to acclimatise themselves to the company

Hence, the management wants to understand what factors they should focus on, in order to curb attrition. In other words, they want to know what changes they should make to their workplace, in order to get most of their employees to stay. Also, they want to know which of these variables is most important and needs to be addressed right away.

Abstract

Aim: To model the probability of attrition of a Company using logistic regression

Approach:

• The primary approach is to use build a Logistic Regression model in order to understand the important factors affecting the company.
• We make use of EDA (Exploratory Data Analysis) operations like univariate and bivariate analysis to identify problems and tackle it building a model and evaluating it.
• Using the data sets available, we make use of R, a statistical computing program to perform various operations and hence come up with the results.

Conclusion: Come up with factors causing attrition and thus help the Company curb it.

Problem Solving Methodology

Analysis — EDA

• EDA was carried out on all available variables of the dataset.
• Continuous and categorical variables were segregated.
• Accordingly, outlier treatment and scaling were done.

Model Building

1. We make use of glm() function to build a logistic regression model. Attrition is dependent variable.

2. We use stepAIC() to remove insignificant categorical variables.

3. We run vif() function and look for values with higher vif and higher P-value.

4. Accordingly, in every iteration, we remove insignificant variables to build the final model.

5. The final model had 16 significant variables: TotalWorkingYears, YearsSinceLastPromotion, YearsWithCurrManager, Avg.hrs, JobSatisfaction4, EnvironmentSatisfaction2, EnvironmentSatisfaction3, EnvironmentSatisfaction4, BusinessTravelTravel_Frequently, DepartmentResearch & Development, DepartmentSales, JobRoleManufacturing Director, MaritalStatusSingle, NumCompaniesWorked5, NumCompaniesWorked7.

Model Evaluation

We start model evaluation assuming the probability cut-off of 0.5.

And Found: Accuracy = 85.81%, Sensitivity = 22.00% and Specificity = 98.14%.

Then, to find optimal probability cut-off , we used confusionMatrix and found the cut-off to be 0.17.

1.For the optimal probability cut-off of 0.17,

• Accuracy = 70.85%
• Sensitivity = 70.33% and
• Specificity = 70. 95%.

2. KS — Statistic

The KS Statistic was found to be 0.4128775 i.e 41.3%

• A good model is one for which the KS statistic:
• is equal to 40% or more
• lies in the top deciles, i.e. 1st, 2nd, 3rd or 4th

3. We further plot the Gain and Lift Charts.

Model Analysis

The gain for the given model is 76.55% by the 4th decile. This means, that if you sort all employees according to probability, then among the top 40% of the employees, 76.55% are more likely to switch the company.

The model’s lift is equal to 1.9 by the 4th decile. It means, the model catches 1.9 times more attrition than a random model would have caught at the 4th decile.

Conclusion

1.The Logistic Regression model built for the given data set.

2.16 Important Variables affecting attrition were found.

For the optimal probability cut-off of 0.17

• Accuracy = 70.85%
• Sensitivity = 70.33% and
• Specificity = 70. 95%.

3. The Model was found to be a good model based on the model evaluation that was done using KS Statistic along with Gain and Lift charts.

• The KS Statistic was found to be 41.3%
• Gain% by the 4th decile was 76.55%
• Lift by the 4th decile was 1.9.

4. This Model with significant variables should help the Company to understand what changes they should make to their workplace, in order to get most of their employees to stay and cut attrition.

The R Code