Executive Summary

This report analyzes the drivers of employee turnover using the Employee Attrition dataset from Kaggle. We compare three logistic regression models to identify whether financial incentives (Monthly Income) or work conditions (Overtime) better predict employee departure.

Model Predictors
Model 1 Monthly.Income only
Model 2 Monthly.Income + Overtime
Model 3 All variables (full model)

1. Data Preparation

library(caret)
library(dplyr)
library(ggplot2)
library(pROC)
# File browser pops up twice -- select train.csv first, then test.csv
train_set <- read.csv(file.choose())
test_set  <- read.csv(file.choose())

# Convert character columns to factors
# Attrition levels are "Stayed" (reference) and "Left" (positive class)
fix_factors <- function(df) {
  df %>%
    mutate(across(where(is.character), as.factor)) %>%
    mutate(Attrition = factor(Attrition, levels = c("Stayed", "Left")))
}

train_set <- fix_factors(train_set)
test_set  <- fix_factors(test_set)

cat("Training rows:", nrow(train_set), "\n")
## Training rows: 59598
cat("Test rows    :", nrow(test_set),  "\n")
## Test rows    : 14900
cat("Columns      :", ncol(train_set), "\n")
## Columns      : 24
train_set %>%
  count(Attrition) %>%
  mutate(Percent = round(n / sum(n) * 100, 1))
##   Attrition     n Percent
## 1    Stayed 31260    52.5
## 2      Left 28338    47.5

The dataset is class-imbalanced – more employees stayed than left. Accuracy alone is not sufficient; we also examine confusion matrices and ROC curves.

2. Model 1 – Monthly Income Only

model1 <- glm(
  Attrition ~ Monthly.Income,
  data   = train_set,
  family = binomial
)

summary(model1)
## 
## Call:
## glm(formula = Attrition ~ Monthly.Income, family = binomial, 
##     data = train_set)
## 
## Coefficients:
##                  Estimate Std. Error z value Pr(>|z|)   
## (Intercept)    -2.081e-02  2.902e-02  -0.717  0.47334   
## Monthly.Income -1.059e-05  3.813e-06  -2.777  0.00548 **
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## (Dispersion parameter for binomial family taken to be 1)
## 
##     Null deviance: 82477  on 59597  degrees of freedom
## Residual deviance: 82469  on 59596  degrees of freedom
## AIC: 82473
## 
## Number of Fisher Scoring iterations: 3
prob1 <- predict(model1, newdata = test_set, type = "response")
pred1 <- factor(ifelse(prob1 > 0.5, "Left", "Stayed"), levels = c("Stayed", "Left"))

Confusion Matrix – Model 1

cm1 <- confusionMatrix(pred1, test_set$Attrition, positive = "Left")
cm1
## Confusion Matrix and Statistics
## 
##           Reference
## Prediction Stayed Left
##     Stayed   7868 7032
##     Left        0    0
##                                         
##                Accuracy : 0.5281        
##                  95% CI : (0.52, 0.5361)
##     No Information Rate : 0.5281        
##     P-Value [Acc > NIR] : 0.5033        
##                                         
##                   Kappa : 0             
##                                         
##  Mcnemar's Test P-Value : <2e-16        
##                                         
##             Sensitivity : 0.0000        
##             Specificity : 1.0000        
##          Pos Pred Value :    NaN        
##          Neg Pred Value : 0.5281        
##              Prevalence : 0.4719        
##          Detection Rate : 0.0000        
##    Detection Prevalence : 0.0000        
##       Balanced Accuracy : 0.5000        
##                                         
##        'Positive' Class : Left          
## 
cm1_df <- as.data.frame(cm1$table)

ggplot(cm1_df, aes(x = Reference, y = Prediction, fill = Freq)) +
  geom_tile(colour = "white") +
  geom_text(aes(label = Freq), size = 6, fontface = "bold") +
  scale_fill_gradient(low = "#dce9f5", high = "#185FA5") +
  labs(
    title = "Confusion Matrix - Model 1: Monthly.Income",
    x = "Actual", y = "Predicted", fill = "Count"
  ) +
  theme_bw()

3. Model 2 – Monthly Income + Overtime

model2 <- glm(
  Attrition ~ Monthly.Income + Overtime,
  data   = train_set,
  family = binomial
)

summary(model2)
## 
## Call:
## glm(formula = Attrition ~ Monthly.Income + Overtime, family = binomial, 
##     data = train_set)
## 
## Coefficients:
##                  Estimate Std. Error z value Pr(>|z|)    
## (Intercept)    -1.001e-01  2.965e-02  -3.375 0.000737 ***
## Monthly.Income -1.038e-05  3.819e-06  -2.717 0.006578 ** 
## OvertimeYes     2.374e-01  1.750e-02  13.566  < 2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## (Dispersion parameter for binomial family taken to be 1)
## 
##     Null deviance: 82477  on 59597  degrees of freedom
## Residual deviance: 82285  on 59595  degrees of freedom
## AIC: 82291
## 
## Number of Fisher Scoring iterations: 3
prob2 <- predict(model2, newdata = test_set, type = "response")
pred2 <- factor(ifelse(prob2 > 0.5, "Left", "Stayed"), levels = c("Stayed", "Left"))

Confusion Matrix – Model 2

cm2 <- confusionMatrix(pred2, test_set$Attrition, positive = "Left")
cm2
## Confusion Matrix and Statistics
## 
##           Reference
## Prediction Stayed Left
##     Stayed   5487 4530
##     Left     2381 2502
##                                           
##                Accuracy : 0.5362          
##                  95% CI : (0.5281, 0.5442)
##     No Information Rate : 0.5281          
##     P-Value [Acc > NIR] : 0.02397         
##                                           
##                   Kappa : 0.0541          
##                                           
##  Mcnemar's Test P-Value : < 2e-16         
##                                           
##             Sensitivity : 0.3558          
##             Specificity : 0.6974          
##          Pos Pred Value : 0.5124          
##          Neg Pred Value : 0.5478          
##              Prevalence : 0.4719          
##          Detection Rate : 0.1679          
##    Detection Prevalence : 0.3277          
##       Balanced Accuracy : 0.5266          
##                                           
##        'Positive' Class : Left            
## 
cm2_df <- as.data.frame(cm2$table)

ggplot(cm2_df, aes(x = Reference, y = Prediction, fill = Freq)) +
  geom_tile(colour = "white") +
  geom_text(aes(label = Freq), size = 6, fontface = "bold") +
  scale_fill_gradient(low = "#dce9f5", high = "#185FA5") +
  labs(
    title = "Confusion Matrix - Model 2: Monthly.Income + Overtime",
    x = "Actual", y = "Predicted", fill = "Count"
  ) +
  theme_bw()

4. Model 3 – Full Model (All Variables)

model3 <- glm(
  Attrition ~ . - Employee.ID,
  data   = train_set,
  family = binomial
)

summary(model3)
## 
## Call:
## glm(formula = Attrition ~ . - Employee.ID, family = binomial, 
##     data = train_set)
## 
## Coefficients:
##                                    Estimate Std. Error z value Pr(>|z|)    
## (Intercept)                       2.563e-01  8.586e-02   2.985 0.002836 ** 
## Age                              -5.998e-03  1.004e-03  -5.974 2.31e-09 ***
## GenderMale                       -6.262e-01  2.078e-02 -30.131  < 2e-16 ***
## Years.at.Company                 -1.360e-02  1.174e-03 -11.588  < 2e-16 ***
## Job.RoleFinance                  -9.223e-02  4.838e-02  -1.906 0.056596 .  
## Job.RoleHealthcare               -7.214e-02  4.238e-02  -1.702 0.088761 .  
## Job.RoleMedia                    -9.867e-02  3.616e-02  -2.729 0.006353 ** 
## Job.RoleTechnology               -8.476e-02  4.847e-02  -1.749 0.080344 .  
## Monthly.Income                   -6.718e-06  8.246e-06  -0.815 0.415242    
## Work.Life.BalanceFair             1.327e+00  3.129e-02  42.410  < 2e-16 ***
## Work.Life.BalanceGood             2.934e-01  2.958e-02   9.916  < 2e-16 ***
## Work.Life.BalancePoor             1.509e+00  3.757e-02  40.167  < 2e-16 ***
## Job.SatisfactionLow               4.880e-01  3.572e-02  13.664  < 2e-16 ***
## Job.SatisfactionMedium            1.133e-02  2.721e-02   0.417 0.676953    
## Job.SatisfactionVery High         4.988e-01  2.701e-02  18.469  < 2e-16 ***
## Performance.RatingBelow Average   3.339e-01  2.948e-02  11.324  < 2e-16 ***
## Performance.RatingHigh            5.078e-03  2.647e-02   0.192 0.847882    
## Performance.RatingLow             5.972e-01  4.847e-02  12.321  < 2e-16 ***
## Number.of.Promotions             -2.492e-01  1.043e-02 -23.900  < 2e-16 ***
## OvertimeYes                       3.512e-01  2.189e-02  16.043  < 2e-16 ***
## Distance.from.Home                9.950e-03  3.636e-04  27.365  < 2e-16 ***
## Education.LevelBachelor’s Degree  4.581e-02  2.757e-02   1.662 0.096584 .  
## Education.LevelHigh School        3.035e-02  3.072e-02   0.988 0.323037    
## Education.LevelMaster’s Degree    3.087e-02  3.048e-02   1.013 0.311008    
## Education.LevelPhD               -1.564e+00  5.447e-02 -28.714  < 2e-16 ***
## Marital.StatusMarried            -2.559e-01  2.957e-02  -8.654  < 2e-16 ***
## Marital.StatusSingle              1.573e+00  3.220e-02  48.863  < 2e-16 ***
## Number.of.Dependents             -1.573e-01  6.660e-03 -23.621  < 2e-16 ***
## Job.LevelMid                     -1.003e+00  2.266e-02 -44.271  < 2e-16 ***
## Job.LevelSenior                  -2.616e+00  3.262e-02 -80.221  < 2e-16 ***
## Company.SizeMedium                6.143e-03  2.713e-02   0.226 0.820849    
## Company.SizeSmall                 2.063e-01  2.958e-02   6.975 3.06e-12 ***
## Company.Tenure                   -2.036e-04  4.501e-04  -0.452 0.651100    
## Remote.WorkYes                   -1.775e+00  2.916e-02 -60.888  < 2e-16 ***
## Leadership.OpportunitiesYes      -1.627e-01  4.753e-02  -3.424 0.000618 ***
## Innovation.OpportunitiesYes      -1.410e-01  2.784e-02  -5.064 4.10e-07 ***
## Company.ReputationFair            4.698e-01  3.962e-02  11.858  < 2e-16 ***
## Company.ReputationGood           -6.027e-02  3.535e-02  -1.705 0.088173 .  
## Company.ReputationPoor            7.564e-01  3.972e-02  19.043  < 2e-16 ***
## Employee.RecognitionLow           3.956e-02  2.620e-02   1.510 0.131020    
## Employee.RecognitionMedium        4.354e-02  2.769e-02   1.572 0.115845    
## Employee.RecognitionVery High    -8.300e-02  5.019e-02  -1.654 0.098177 .  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## (Dispersion parameter for binomial family taken to be 1)
## 
##     Null deviance: 82477  on 59597  degrees of freedom
## Residual deviance: 57833  on 59556  degrees of freedom
## AIC: 57917
## 
## Number of Fisher Scoring iterations: 5
prob3 <- predict(model3, newdata = test_set, type = "response")
pred3 <- factor(ifelse(prob3 > 0.5, "Left", "Stayed"), levels = c("Stayed", "Left"))

Confusion Matrix – Model 3

cm3 <- confusionMatrix(pred3, test_set$Attrition, positive = "Left")
cm3
## Confusion Matrix and Statistics
## 
##           Reference
## Prediction Stayed Left
##     Stayed   6080 1855
##     Left     1788 5177
##                                           
##                Accuracy : 0.7555          
##                  95% CI : (0.7485, 0.7624)
##     No Information Rate : 0.5281          
##     P-Value [Acc > NIR] : <2e-16          
##                                           
##                   Kappa : 0.5092          
##                                           
##  Mcnemar's Test P-Value : 0.2742          
##                                           
##             Sensitivity : 0.7362          
##             Specificity : 0.7728          
##          Pos Pred Value : 0.7433          
##          Neg Pred Value : 0.7662          
##              Prevalence : 0.4719          
##          Detection Rate : 0.3474          
##    Detection Prevalence : 0.4674          
##       Balanced Accuracy : 0.7545          
##                                           
##        'Positive' Class : Left            
## 
cm3_df <- as.data.frame(cm3$table)

ggplot(cm3_df, aes(x = Reference, y = Prediction, fill = Freq)) +
  geom_tile(colour = "white") +
  geom_text(aes(label = Freq), size = 6, fontface = "bold") +
  scale_fill_gradient(low = "#dce9f5", high = "#185FA5") +
  labs(
    title = "Confusion Matrix - Model 3: All Variables",
    x = "Actual", y = "Predicted", fill = "Count"
  ) +
  theme_bw()

5. ROC Curves – All Three Models

roc1 <- roc(test_set$Attrition, prob1, levels = c("Stayed", "Left"), quiet = TRUE)
roc2 <- roc(test_set$Attrition, prob2, levels = c("Stayed", "Left"), quiet = TRUE)
roc3 <- roc(test_set$Attrition, prob3, levels = c("Stayed", "Left"), quiet = TRUE)

lab1 <- paste0("Model 1: Monthly.Income (AUC = ", round(auc(roc1), 3), ")")
lab2 <- paste0("Model 2: Monthly.Income + Overtime (AUC = ", round(auc(roc2), 3), ")")
lab3 <- paste0("Model 3: All Variables (AUC = ", round(auc(roc3), 3), ")")

roc_df <- bind_rows(
  data.frame(fpr = 1 - roc1$specificities, tpr = roc1$sensitivities, model = lab1),
  data.frame(fpr = 1 - roc2$specificities, tpr = roc2$sensitivities, model = lab2),
  data.frame(fpr = 1 - roc3$specificities, tpr = roc3$sensitivities, model = lab3)
)

colour_map <- c("#185FA5", "#D85A30", "#1D9E75")
names(colour_map) <- c(lab1, lab2, lab3)
ggplot(roc_df, aes(x = fpr, y = tpr, colour = model)) +
  geom_line(linewidth = 1) +
  geom_abline(slope = 1, intercept = 0,
              linetype = "dashed", colour = "grey60") +
  scale_colour_manual(values = colour_map) +
  labs(
    title    = "ROC Curves: Employee Attrition Models",
    subtitle = "Comparing three logistic regression specifications",
    x        = "False Positive Rate (1 - Specificity)",
    y        = "True Positive Rate (Sensitivity)",
    colour   = "Model"
  ) +
  theme_bw() +
  theme(
    legend.position  = "bottom",
    legend.direction = "vertical"
  )

6. Performance Summary

perf <- data.frame(
  Model = c(
    "Model 1: Monthly.Income",
    "Model 2: Monthly.Income + Overtime",
    "Model 3: All Variables"
  ),
  Accuracy = round(c(
    cm1$overall["Accuracy"],
    cm2$overall["Accuracy"],
    cm3$overall["Accuracy"]
  ), 3),
  Sensitivity = round(c(
    cm1$byClass["Sensitivity"],
    cm2$byClass["Sensitivity"],
    cm3$byClass["Sensitivity"]
  ), 3),
  Specificity = round(c(
    cm1$byClass["Specificity"],
    cm2$byClass["Specificity"],
    cm3$byClass["Specificity"]
  ), 3),
  AUC = round(c(auc(roc1), auc(roc2), auc(roc3)), 3)
)

knitr::kable(perf, align = "lcccc",
             caption = "Performance metrics across all three models")
Performance metrics across all three models
Model Accuracy Sensitivity Specificity AUC
Model 1: Monthly.Income 0.528 0.000 1.000 0.505
Model 2: Monthly.Income + Overtime 0.536 0.356 0.697 0.530
Model 3: All Variables 0.756 0.736 0.773 0.845

7. Conclusion

  • Model 1 (Monthly.Income only) provides a weak baseline – income alone has limited predictive power for attrition.
  • Model 2 (Monthly.Income + Overtime) improves sensitivity noticeably. Employees working overtime are substantially more likely to leave, regardless of income level.
  • Model 3 (all variables) achieves the best AUC and overall discrimination. Key significant predictors typically include Overtime, Job.Satisfaction, Years.at.Company, and Marital.Status.

The results suggest that work conditions matter more than pay as a driver of attrition – a finding with clear HR policy implications.