Churn & Retention Analytics
Zero attrition marked entrapment, not loyalty.
Public dataset; the analysis, the interaction terms and the deployed app are mine.
01 · The finding
Zero attrition marked entrapment, not loyalty.
Departments with no recorded exits across multiple years modelled at 90%+ underlying risk. A logistic regression with interaction terms found the paradox that raw rates concealed — and moved the retention budget away from the departments that looked worst.
The decision this implies: stop ranking departments by observed exit rate. A department with no exits and high modelled risk is not a retention success, and funding it last is the opposite of what the raw table recommends.
02 · The question
I asked which departments to fund, and the raw rates answered wrong.
- Scope
- 1,400+ employees across all departments.
- Data
- Public HR attrition dataset.
- Method
- Logistic regression with interaction terms, fitted on departmental strata.
03 · What the data showed
Three factors moved exit risk, and the department label hid all of them.
New managers carry a 2.4x exit risk in their first year.
A manager transition raised absolute exit probability by 13.7 percentage points — a 2.4x multiplier, and the single largest effect in the model.
Sustained overtime roughly triples the odds of leaving.
Overtime sustained across periods carried an approximate 3x multiplier on exit risk. The effect is in sustained overtime, not occasional overtime.
Under-30s leave at 2.3x the rate of their older colleagues.
The under-30 cohort exits at 2.3x the rate of the rest of the population, which concentrates the cost of every other risk factor in the same group.
04 · What I recommend
Three changes follow directly from the model.
Re-rank departments for retention spend by modelled risk rather than observed exit rate.
Add a check-in at the manager-transition point, targeted at the first year.
Flag sustained overtime as a retention signal rather than a capacity signal.
05 · Method and caveats
The model is a logistic regression, and its limits are the dataset's.
The model is specified as follows.
- Logistic regression with interaction terms between department and the risk factors.
- Fitted on 1,400+ employees.
- Built in R; results presented in Tableau.
These hypotheses were tested and rejected.
- Commute distance was tested as a driver of attrition and rejected — R² = 0.000. This removed a planned intervention from the agenda before it was funded.
These limits apply.
- Public dataset, so the fields are those the publisher chose to record.
- The Simpson's paradox result depends on departmental strata being small; it should be re-checked on any population where they are not.