Embedded Files
Let's use statistics and machine learning to understand why employees are leaving an organization
Let's use statistics and machine learning to understand why employees are leaving an organization
Software Used:
Software Used:
R
Tableau
Analyses Used:
Analyses Used:
Correlation Matrix
Logisitic Regression
Confusion Matrix
AUC Curve
What is the purpose of this project?
Begin as a Beginner: I am a beginner to the world of people analytics. As someone exploring the capabilities of R and Tableau, I hope this marks the beginning of a journey where I can share my abilities and learn from others. Feel free to write to me if you have suggestions or would like to see something added!
Descriptive and Inferential Data Analysis: I am passionate about using the story telling of data to better explain HR metrics. Although there might be trends or eye-catching numbers at first glance, my goal is to dig deeper and understand what our data is really showing. You cannot improve what you do not measure.
Model with Machine Learning: The ultimate goal is to create a model to predict whether an employee will quit or stay. Logistic regression is a statistical technique that can differentiate which factors are more impactful towards a dichotomous outcome (e.g., to leave or to stay). Adding machine learning techniques then allows us to understand the strength of our predictive model.
Recommendations: How can the organization reduce turnover? This step is arguably the most important since it connects data to organizational policy.
About this data set: IBM HR Analytics Attrition Data set is an excellent dataset to demonstrate logistic regression and better understand attrition. Attrition is simply the turnover rate of employees. This data is publicly available here.
Exploratory Analysis
What does our data show at a glance? We can make some interesting observations without using advanced methods. Although Excel and R can be used for this process, Tableau produces dynamic data visualizations. These analyses will also help us determine how we should work with our data throughout the project.
Overall, about 84% of the workforce stayed while about 16% left. We can visually see this is a moderately imbalanced dataset. Knowing our dataset is skewed (not normally distributed) helps us determine what further analyses we can conduct.
Is there a wage gap? It seems both overall and among departments that there are no major differences of the monthly average income between genders. This is good news for both our employees and the organization!
Here’s where things get interesting.
Age cohorts: Gen-Z workers show the highest frequency of turnover rate at around 36%- most likely to pursue other career opportunities. Boomers have the next highest proportion of employees leaving at 17%- most likely to retire.
Income: It seems that employees who left were paid less on average than those who stayed. These differences were most significant within the HR and Research and Development departments. (HR should know better!)
What factors significantly relate to each other?
A correlation matrix displays correlations between 2 variables. Although this analysis can be conducted in Excel, R can sort the significant correlations together! This feature can help us quickly visualize the strongest relationships, rather than having to look around the entire visual.
What we see:
o The higher the job level, the higher the monthly income r = .95
o The higher the job level, the higher the total working years r = .78
o The higher the percent salary hike, the higher the performance rating r = .77
o The higher the total working years, the higher the monthly income r = .77
Logistic Regression
Logistic regression is a specialized linear model where regression weights (our input variables) are interpreted on a log odds scale for a dichotomous outcome. We can use this technique to calculate fit statistics and model evaluations.
VIF: After categorizing and splitting our data, we should check if multicollinearity is occurring between our input variables. The Variance Inflation Factor is a technique which creates a linear model with 1 dependent variable against multiple independent variables.
We can see that “Job Level” produces the highest VIF score of 8.19, so we should eliminate it. Normally a VIF score of 10 or higher serves as the cutoff. However, this removal will increase the precision of our other independent variables and increases the overall statistical power of our model.
After removing “Job Level” from the data set, we can now see that none of our VIF scores are close to approaching 10. Let’s create our final model!
Look at
all of
those
significant values!
Interpreting the Model
Let’s use machine learning to further evaluate our logistic regression model. A Confusion Matrix is a classification model, using supervised learning, that allows us to evaluate our linear model’s performance. The number of predicted and actual values are displayed in the chart below. The top left corner shows the number of true positives, the top right corner shows the number of false positives (Type 1 error) the bottom left corner shows the number of false negatives (Type 2 Error), and the bottom right corner shows the number of true negatives.
Overall, our model shows that from 294 example cases, 247 were correctly predicted to stay and 18 were correctly predicted to leave. 23 were falsely predicted to stay and 6 were falsely predicted to leave.
Compared to the dataset, the linear model holds an impressive 90% accuracy rating!
Our AUC is above the standard cutoff of .80!
Since the data set is moderately imbalanced, relying on a model’s accuracy measure is insufficient. Receiver Operating Characteristic Curve (ROC curves) plot true positive rates [TP/(TP + FN)], which is the percentage of terms who were accurately classified as such, and we want to maximize (i.e., sensitivity), against false positive rates [FP/(TN + FP)], which is the percentage of non-terms who were classified as terms, and we want to minimize. We can determine its strength by looking at the are under the ROC curve or AUC. The higher our AUC, the better our model is classifying turnover as staying (1 as 1) and leaving (0 as 0). (Speer et al., 2019).
Recommendations:
Top reasons people leave the organization:
Monthly income- unsurprisingly, those with lower salaries left the organization.
Overtime- those who work longer hours during a standard 40-hour work week might feel burned out or undercompensated.
Tenure- the number of years one works at a company was detected as a more significant predictor than age. While the rate of turnover differed between age cohorts, the amount of time they were eligible to stay with the organization also differed, so understanding individuals’ tenure is a more effective way to predict their behavior.
What should be done:
Incomes can be increased to increase retention. Although we don’t know this organization’s overtime policy, either limiting the amount of overtime or increasing its incentive could decrease turnover.
If dates were provided for when employees started and left the organization, a survival analysis could have been conducted to predict not only what causes an employee to leave, but also when.
While understanding why employees leave an organization is a great start, it is equally important to understand why employees stay. Deploying an engagement survey can provide additional insight for which factors keep employees at the organization.
Why does this matter?
HR professionals, managers, and C-Suite members generally believed seeking a higher income was the key motivator, for employees quitting, during the Great Resignation. Our exploratory visualizations seemed to support this belief, but analyzing our data tells a deeper story. In this organization’s case, working overtime was the most impactful reason! Increasing salaries could decrease the number of employees working overtime, but what if the organization is financially constrained? Addressing the other significant variables could provide a more cost-effective and personalized approach. Overall, using advanced, statistical models created a better judgment of employees’ behavior to predict attrition.
Speer, A. B., Dutta, S., Chen, M., & Trussell, G. (2019). Here to stay or go? Connecting turnover research to applied attrition modeling. Industrial and Organizational Psychology, 12(3), 277-301.
Page updated
Google Sites
Report abuse