Most attrition analysis stops at a turnover percentage. A six-step method to segment, sequence and test causes — and the traps that produce confident wrong answers.
An attrition analysis is a structured investigation into who is leaving, when in their lifecycle they leave, and what distinguishes them from the people who stay. It is not a turnover percentage. A percentage tells you the size of the problem; an attrition analysis tells you where it is concentrated and what to do about it.
The reason most attrition work fails is that it stops one step too early. A team reports 18% turnover, notices it is up from 15%, commissions an engagement survey, and launches a company-wide retention initiative. Meanwhile 60% of the increase sits in two teams with a common cause that a segmentation would have found in an afternoon.
Never analyse total turnover. It combines populations with unrelated causes.
The first cut is always the same:
Then split voluntary into regretted and non-regretted. Two organisations with identical 18% turnover are in completely different health depending on whether the leavers were people they wanted to keep.
If your data cannot support this split, that is the finding. Fixing exit classification is a prerequisite, not a detour.
Attrition is almost never evenly distributed. The goal of segmentation is to find where it concentrates, and there are five cuts that repay the effort:
| Segment | What it reveals | |---|---| | Tenure band | Whether this is a hiring problem or a retention problem | | Manager | The single strongest concentrator of voluntary turnover | | Function / team | Local causes — workload, leadership change, market poaching | | Level | Whether progression has stalled at a specific rung | | Location | Local labour market pressure, return-to-office effects |
Run tenure band first. It is the fastest way to establish what kind of problem you have:
These three findings lead to entirely different interventions. An organisation that responds to a first-year attrition problem with a compensation review has spent money on the wrong thing.
A single-point analysis tells you who left. Sequencing tells you what happened before they left, which is where causes live.
For each leaver, look backwards 12 months and record:
Then compare that profile against a matched group of stayers. This is the step that separates real analysis from storytelling. Without a comparison group, every leaver characteristic looks meaningful — until you discover the stayers had it too.
Three techniques, in increasing order of effort. Most organisations should stop after the second.
Cross-tabulation with a comparison group. For each candidate driver, compare its prevalence among leavers against matched stayers. This alone identifies most actionable causes and requires no modelling. Where you want a significance test on a categorical relationship, a chi-square test is usually sufficient.
Survival analysis. Instead of asking "did they leave?", ask "how long did they stay, and what shortened it?" This handles the fact that current employees have not left yet, which ordinary comparisons mishandle badly. It is the right technique for tenure-cliff questions and for first-90-day attrition.
Regression modelling. A logistic regression estimating leave probability from multiple drivers simultaneously, which is what you need when drivers are correlated — and they always are. Compensation, tenure and level move together, so a univariate analysis will attribute the same effect three times.
Whichever you use, hold out the temptation to jump to a flight risk score before you have understood the drivers. A model that predicts well but explains nothing gives managers a list and no conversation to have.
Quantitative analysis identifies the pattern. Exit data explains the mechanism — imperfectly, and with a known bias.
Departing employees give the safest true reason, not the whole one. "Career opportunity" is the most common exit reason in almost every organisation, and it is frequently a polite version of "my manager" or "I was not going to be promoted here". Treat exit reasons as a hypothesis generator rather than a conclusion, and read exit survey completion rate alongside them — a 40% completion rate means you have heard from the comfortable minority.
Better triangulation sources:
An attrition analysis that ends with a percentage competes badly for budget. One that ends with a cost figure does not.
Multiply regretted departures by role-family cost of turnover. Even a conservative estimate — a share of annual salary, including vacancy cost and time to productivity — usually produces a number large enough to fund the intervention several times over.
Then state the counterfactual: what a three-point reduction in regretted attrition would be worth. That is the sentence that gets the programme approved.
Analysing total turnover. Guarantees a muddled conclusion, because you are averaging across populations with opposite causes.
No comparison group. Leaver characteristics are meaningless in isolation. If 70% of leavers had a manager change, you need to know what share of stayers did too.
Survivorship bias in tenure analysis. Comparing average tenure of leavers against average tenure of current employees is invalid — current employees have not finished accruing tenure. This is precisely what survival analysis exists to handle.
Correlated drivers treated separately. Pay, level and tenure move together. Analysing them one at a time attributes the same effect repeatedly and inflates every one of them.
Trusting exit reasons at face value. Systematically biased toward reasons that preserve the reference.
Small-sample escalation. A four-person team with one departure is at 25% turnover. Set a minimum group size before anyone acts on a rate.
Confusing a spike with a trend. Look at rolling twelve-month figures. Monthly people data is mostly noise.
One page. Not a deck.
If the analysis cannot be reduced to that page, it has not finished.
How much data do I need to run an attrition analysis? Segmentation and cross-tabulation work with a few hundred employees and one to two years of exit history. Survival analysis and regression modelling want at least two to three years and a few hundred separations. Below that, stick to segmentation — it identifies most actionable causes anyway.
What is the difference between attrition analysis and an attrition model? Analysis explains why people have left, using historical data and comparison groups. A model predicts who is likely to leave next. Analysis should come first: a predictive model built before the drivers are understood produces a ranked list that managers cannot act on, because nobody can say what the score means.
Why do exit interviews say "career opportunity" so often? Because it is the safest true answer. It preserves the reference and avoids naming a manager. It is rarely a lie, but it is usually incomplete, which is why exit reasons should be triangulated against engagement data from six to twelve months earlier rather than treated as the finding.
Which segment should I analyse first? Tenure band. It is the fastest way to distinguish a hiring and onboarding problem from a progression or management problem, and those two require completely different responses.
Run this analysis without exporting anything. PeoplePilot Analytics segments attrition by tenure, manager, function and level from live HR data, sequences pre-departure events, and costs regretted attrition automatically. Start a free trial or browse the HR Metrics Library.