HR Analytics Case Study: Predicting Employee Attrition and Cutting Turnover

A 1,400-person technology company was bleeding talent — annual voluntary turnover had climbed to 16%, and each departure of a mid-level engineer was costing roughly 1.5× their salary to replace. HR had exit-interview transcripts and gut feelings, but no model. We were asked one question: which employees are most likely to leave, and what actually drives it? Here is the full analysis.

The Data

HR exported an anonymized snapshot of 1,400 employees with these fields:

• Attrition (target): 1 = left voluntarily in the past year, 0 = stayed
• MonthlyIncome, YearsAtCompany, Age
• OverTime: yes / no
• JobSatisfaction: 1 (low) to 4 (very high)
• DistanceFromHome: kilometers
• Department: Sales, R&D, HR

Overall attrition was 16.1%. The job was to move from that single number to who and why.

Phase 1: Exploratory Comparisons

Before modeling, we compared leavers and stayers one variable at a time. This builds intuition and tells you which variables are even worth putting in a model.

Income — Independent Samples T-Test

Do leavers earn less? In SPSS:

Analyze > Compare Means > Independent-Samples T Test

• Test Variable: MonthlyIncome
• Grouping Variable: Attrition (1, 0)

We first read Levene's Test for equal variances (p = .02, so we used the "equal variances not assumed" row). The result:

• Stayers: M = $6,832, SD = $4,818
• Leavers: M = $4,787, SD = $3,640
• t(290.4) = 6.38, p < .001

Leavers earned significantly less. Effect size (Cohen's d) was 0.47 — a meaningful, moderate gap.

Overtime — Chi-Square Test

Is overtime linked to leaving? Using Analyze > Descriptive Statistics > Crosstabs with Chi-square:

• Among employees working overtime, 30.5% left.
• Among those who did not, only 10.4% left.
• χ²(1) = 89.0, p < .001

This was the most striking pattern in the whole dataset — overtime nearly tripled the attrition rate.

Phase 2: The Logistic Regression Model

The single-variable tests are suggestive but confounded — maybe overtime workers also earn less. A logistic regression lets every variable compete while holding the others constant.

Analyze > Regression > Binary Logistic

• Dependent: Attrition
• Covariates: MonthlyIncome, YearsAtCompany, Age, OverTime, JobSatisfaction, DistanceFromHome
• Mark OverTime and Department as Categorical
• Under Options: Hosmer-Lemeshow, CI for exp(B)

Model Fit

• Omnibus test: χ²(7) = 198.4, p < .001 — the model is significant.
• Nagelkerke R² = 0.27 — explains about 27% of attrition variance.
• Hosmer-Lemeshow: χ²(8) = 6.1, p = .64 — good fit (non-significant is what you want).

Variables in the Equation

Predictor	Exp(B) (Odds Ratio)	p
OverTime (yes)	3.41	<.001
JobSatisfaction (per point)	0.66	<.001
MonthlyIncome (per $1k)	0.92	.002
YearsAtCompany (per year)	0.93	.004
Age (per year)	0.97	.03
DistanceFromHome	1.02	.18 (n.s.)

Reading the results:

• OverTime, Exp(B) = 3.41: even after controlling for pay, satisfaction, and tenure, overtime workers had 3.4× the odds of leaving. The earlier chi-square wasn't a confound — overtime is a genuine driver.
• JobSatisfaction, Exp(B) = 0.66: each one-point rise in satisfaction cut the odds of leaving by 34%.
• DistanceFromHome was not significant once other factors were accounted for — a good reminder that a variable everyone assumes matters sometimes doesn't.

Phase 3: Scoring the Workforce

We saved the predicted probabilities (Save > Probabilities in the logistic dialog) so every current employee got a churn-risk score from 0 to 1. Sorting the workforce by that score produced a watchlist: the top 8% of employees by risk accounted for nearly a third of expected departures.

What HR Actually Changed

The analysis pointed at two concrete levers, not vague culture talk:

1. Overtime management. R&D, where overtime was heaviest, got hiring backfill and a workload audit. Within two quarters, sustained overtime dropped 22%.
2. Targeted retention for the watchlist. Managers had structured stay-conversations with high-risk, high-performing employees, focused on pay banding and satisfaction blockers.

Twelve months later, voluntary turnover fell from 16.1% to 11.4%. The estimated saving in replacement costs was over $2.1M.

Pitfalls in People Analytics

• Survivorship bias. Your data only describes people still in the system plus recent leavers. Long-gone patterns are invisible — interpret cautiously.
• Ethics and fairness. We deliberately excluded gender, ethnicity, and marital status from the model. Predicting attrition is fine; building a model that could enable discrimination is not. Decide this before you analyze.
• Acting on correlation as if it were cause. The model tells you overtime is associated with leaving. The workload audit tested whether reducing it caused lower attrition — and it did, which is why we trust it.
• Black-box scores with no story. A risk score nobody understands gets ignored. We always pair the number with the two or three reasons behind it.

Reporting It

A binary logistic regression predicted voluntary attrition from income, tenure, age, overtime, job satisfaction, and commute distance, χ²(7) = 198.4, p < .001, Nagelkerke R² = .27. Overtime was the strongest predictor; employees working overtime were 3.41 times more likely to leave (95% CI [2.4, 4.8], p < .001), while each additional point of job satisfaction reduced the odds of leaving by 34%.

---

Sitting on HR data and unsure what it's telling you? Insighter Digital builds attrition and engagement models in SPSS and Python — ethically scoped, clearly explained, and tied to actions your managers can take. Reach out.