The core idea
Survival analysis studies time until a defined event. In retention work, the event might be a voluntary departure. People still employed at the end of observation are right-censored: their eventual departure time is not yet known. Kaplan–Meier estimation can describe the probability of remaining event-free over time without imposing a particular distribution. The term survival is statistical terminology; it is not a judgement about employees.
Source and attribution [1]Using it in practice
Specify the starting event, outcome, observation end and treatment of other departures. Keep joining and leaving dates accurate. Show the number at risk alongside a curve and interpret the tail cautiously when few people remain observed. Consider whether censoring is plausibly unrelated to future event risk, conditional on the analysis. Different event types may require a competing-risks approach.
An example, not a reported case
Worked example · illustrative
In an invented cohort of ten people, one leaves at month one, giving an estimated retention probability of 9/10. Two others reach the end of their observation at month two and are censored. One of the seven still at risk leaves at month three. The Kaplan–Meier estimate becomes 9/10 × 6/7, approximately 77.1%. Censored people are not counted as having left at their censoring date.
What to watch for
A survival curve does not explain why people leave. Treating every non-voluntary departure as harmless censoring can be inappropriate for the question. Hazard is an instantaneous conditional rate, not the same as a cumulative probability. Formal comparisons and models require attention to assumptions and uncertainty.
Choose the event carefully
All exits and voluntary resignation answer different questions. Define internal transfers, retirement and contract completion explicitly. If another event prevents the event of interest, seek an approach suited to competing outcomes rather than automatically applying a basic curve.
Read the late part cautiously
A curve may look stable simply because few people remain under observation. Report numbers at risk and uncertainty. Avoid comparing a well-observed early period with a sparse late period as if the evidence were equally strong.
Build one clear observation record per spell
Record the start date, end of observation, event indicator and relevant definition of departure. Decide how rehires and internal moves are handled before analysis. Check impossible dates and duplicate spells. Someone still employed at the cut-off contributes observed time without being treated as a departure. Keep the distinction between an event and the end of available observation explicit in the dataset and the explanation.
Ask whether censoring is informative
If observation stops because a person moves into a part of the business with different records, that loss may be related to their future departure risk. A standard approach may then require more careful treatment. Similarly, redundancy and voluntary resignation are different outcomes. State the question and seek appropriate methods for competing events where necessary. Do not automatically classify every non-resignation as innocuous censoring simply because the software accepts that input.
Read the curve with its supporting information
Show the number still under observation and an appropriate expression of uncertainty. Early differences based on many people are not equivalent to late differences based on a few. A median time is not available from the observed curve if estimated retention never falls to one half. Avoid extending the line beyond the observed evidence and describing the extension as a measured result. The curve summarises timing; it does not explain motives.
Use timing to investigate a process
If departures cluster before a particular training milestone, examine the experience leading to it: scheduling, support, expectations and assessment. That timing may help prioritise inquiry but does not prove the milestone caused exits. Compare relevant groups and speak with people where appropriate. A service improvement should be evaluated separately. The statistical term 'survival' is technical terminology, and the report should use respectful plain language about employees remaining in or leaving employment.
Explain why incomplete observation is still informative
An employee who remains at the end of the study has not had a recorded exit, but their full eventual tenure is unknown. Survival methods can use the observed time without pretending the person left on the study end date. Define the event carefully: voluntary resignation, any exit and transfer are different outcomes. Check how data entry, missing dates and changes of employment affect the start and end of observation.
Distinguish remaining probability from instantaneous risk
A survival curve describes the estimated proportion remaining beyond a time point under the analysis assumptions. A hazard concerns the event rate among those still at risk at that point; it is not simply the same as the probability of leaving during the whole period. Compare cohorts with attention to follow-up and context, and seek specialist support for modelling assumptions or competing events. Use the results to identify when support may be useful, not to assign a predicted departure date to an individual.
Take it into your next conversation
Three useful questions.
- What question can this method answer, and what can it not establish?
- Are the comparison, observation period and assumptions defensible?
- What decision follows, and how will its consequences be reviewed?
Related terms
Go to the evidence
Sources & attribution
Further reading: Nadeem Khan and Dave Millner (2023), Introduction to People Analytics: A practical guide to data-driven HR, 2nd edition, PDF pp. 227, 244 and 294–295. Edition & reading notes ↗
[1] NIST: Kaplan–Meier estimation with censored data ↗
The core idea is an original summary of the cited work. Application notes, examples and sketches are our interpretations, not quotations or reproductions of the authors’ figures. Publisher records may require access to read the full original work.
Application guidance expanded on 26 September 2026. These practical suggestions and invented examples are editorial explanations; further-reading sources are not the original authors of the historical models unless expressly identified. Original overview review dates are retained.
Published 2026-09-20 · Reviewed 2026-09-22. Editorial approach