Survival Analysis 101 for HR: Time-to-Event Made Simple
Andras Rusznyak
11/4/20256 min read
Ha magyarul szeretnéd olvasni a cikket, kattints ide
A plain-English overview of survival analysis (time-to-event analytics) and why it beats simple rates.
The core concepts (event, time origin, censoring, hazard, survival curve) without heavy math.
When and how to use Kaplan–Meier and Cox models in HR—plus common pitfalls to avoid.
A tool-agnostic workflow you can run with HRIS/ATS data.
A detailed practical example (new-hire attrition) with data, steps, and decisions.
IMPORTANT NOTE:
We utilized generative AI in the making of this article.
If you manage hiring, retention, onboarding or mobility, you manage timing. Most HR questions are not just “how many?” but “how soon?”
How soon do new hires leave?
How fast do we fill a role?
How long until a promotion or full productivity?
Traditional metrics (averages, quarterly rates) blur timing and censor information. Survival analysis preserves the timeline for each person or requisition and correctly handles people who haven’t had the event yet (still employed, role still open). The payoff is cleaner insight and more targeted action—for the same data you already have.
Why it matters now
Survival analysis in plain English
Event
What you’re measuring the time until: resignation, first promotion, training completion, return from leave, role filled.
Time origin (T0)
When the clock starts: hire date, requisition open date, program start date, leave start date.
Censoring (right-censoring)
Some cases haven’t reached the event by the “analysis date,” or their data ends (e.g., still employed). We keep them and mark them as “censored” rather than dropping them—this is a key advantage of survival methods.
Survival function, S(t)
The probability the event has not happened by time t (e.g., “% of new hires still employed at day 180”).
Hazard rate, h(t)
The instantaneous risk of the event at time t, given it hasn’t occurred yet (e.g., “risk of leaving in week 8 conditional on still being here”). Hazards help you find risk peaks (onboarding weeks, probation end, manager change).
New-hire attrition: time to resignation within first 6–12 months.
Time-to-fill (TTF): days from requisition open to accepted offer.
Onboarding ramp: time to first productivity milestone or certification.
Internal mobility: time to first promotion or lateral move.
Return-to-work: time from leave start to effective return.
Methods that help
1) Kaplan–Meier (KM) curves — descriptive power
Output: a curve per group (e.g., team, location, manager support tier) showing “% remaining” over time.
What you learn: when most events happen (median time), and which groups diverge.
Use the log-rank test to check if curves differ significantly.
2) Cox proportional hazards (Cox PH) — explain the why
Output: hazard ratios (HRs) per covariate (e.g., HR=1.40 means 40% higher risk at any time, all else equal).
What you learn: which factors raise or lower risk after controlling for others.
Assumption: proportional hazards (relative risks stable over time). Check with residual plots or time interaction terms.
3) When to go beyond Cox/KM
Parametric models (Weibull/Exponential) for direct time predictions and small samples.
Time-varying covariates (e.g., hazard change after manager switch, pay adjustment).
Competing risks (e.g., involuntary vs voluntary exits; internal transfer “competes” with resignation).
Practical example: New-hire attrition in Customer Support
Business question
Early attrition (first 180 days) in Customer Support has crept from 9% to 14%. We need to know when risk peaks and which factors drive it—so we can act during onboarding.
Event & window
Event: first voluntary resignation.
T0: employment start date.
Censoring: analysis snapshot at 210 days after each hire or any involuntary exit (treated as competing event, excluded here for clarity).
Data required
Core: employee ID, start_date (T0), voluntary_exit_date (if any), snapshot_date (censor), team, site, shift (day/evening/night), schedule volatility (std dev of weekly hours), manager support score at week 3 (1–5), base pay band, prior internal experience (Y/N).
Optional: time-varying flags (e.g., manager change date, schedule change date).
Step-by-step
Descriptive KM
Plot survival curves for shift and manager support tier (≤2.5 vs ≥3.5).
Insight: median “survival” (still employed) at 180 days is lower for night shift and low support. Log-rank tests show both differences are statistically significant.
Manager message: “Risk spikes between weeks 6–10—onboarding content needs reinforcement then.”
Cox model (drivers, controlled)
Covariates: shift, manager support tier, schedule volatility (continuous), base pay band, prior internal experience, site fixed effects.
Results (illustrative interpretation):
Night shift HR=1.55 (≈55% higher risk than day shift, p<0.01).
Low manager support HR=1.40 (40% higher risk, p<0.05).
Schedule volatility HR=1.20 per SD (20% higher risk, p<0.05).
Prior internal experience HR=0.78 (22% lower risk, p<0.05).
Assumption check: proportional hazards acceptable (no strong violations in residuals).
Takeaway: focus on shift design, manager touchpoints, and stable scheduling, especially for externally hired rookies.
Action design (prescriptive playbook)
Weeks 5–10 play: mandatory 1:1 every week, buddy shadowing, micro-goals with feedback.
Scheduling: cap weekly hour variance for first 8 weeks; offer shift swap bank for nights.
Manager enablement: quick coaching module on “early signals & conversations,” with a checklist.
Targeting rule: apply to new hires on night shift or manager support ≤2.5 or high schedule volatility in first 2 weeks.
Impact & ROI framing
If these plays reduce early-exit hazard by a conservative 15%, projected 6-month attrition falls from 14% → ~12%.
For a 400-hire cohort, that’s ~8 fewer leavers.
Using a blended replacement + ramp cost of €8,000 per leaver, avoid ~€64,000—likely exceeding the cost of mentoring hours and manager coaching.
Evaluation
Run a stepped rollout (two sites first) with matched control or a simple A/B by start month.
Measure KM curves pre/post and Cox HR changes; track secondary metrics (CSAT/QA, schedule variance, manager check-ins completed).
What this gives you
Where HR should use it (today)
Data you need (minimal set)
Entity ID: employee, requisition.
T0 date: hire/open date.
Event flag + date: resignation/offer accepted/promotion date.
Censor date: analysis snapshot or termination for a different reason.
Covariates (optional, for drivers): team, manager, location, job level, schedule pattern, comp band, engagement score, manager change, prior internal experience, channel, agency vs. direct.
Pro tip: Keep one row per person (or requisition) with fields for T0, event_date (or null), and censor_date (snapshot). For advanced modeling, convert to person-period (spells by months/weeks) only if needed.
Interpreting results (so decisions follow)
Median time: “Half of new hires are still employed at 280 days” is more actionable than “12% quarterly attrition.”
Hazard peaks: “Risk spikes in weeks 5–8” can trigger playbooks (buddy program, shadowing, first project, manager 1:1s).
Hazard ratios (drivers): “Night shift HR=1.6” → focus on schedule control, not generic engagement emails.
Scenario value: turning hazard reduction into € saves (e.g., avoided backfill and ramp cost) makes a case for budget.
Pitfalls (and how to avoid them)
Data leakage: using post-event info (e.g., exit interview scores) as predictors. Only use data known before the event.
Mis-defined T0: mixed start clocks ruin interpretation (e.g., sometimes contract signed, sometimes start date). Standardize.
Immortal time bias: giving “exposure” (e.g., training) that happens after T0 and before the event without time-aware modeling. Treat as time-varying or restrict the window.
Mixing events: voluntary + involuntary exits blur signals. Model them separately or use competing risks.
Small samples / over-segmentation: prefer pooled models with fixed effects over slicing into tiny groups.
PH assumption violations: if hazards cross, add time interactions or use stratified Cox/parametric forms.
Tool-agnostic implementation (fast path)
People & roles
HR Analytics (lead), TA/HR Ops (data owners), HRBP/Managers (context), C&B (bands), Legal/DP (privacy).
Workflow
Frame the decision: which event, which window, what action if risk is high?
Define T0 and event; separate competing events.
Assemble data: HRIS/ATS + covariates; mark censor_date = snapshot.
Describe with KM curves; test group differences.
Explain with Cox; check assumptions.
Translate to action: triggers, playbooks, owners, cadence.
Measure lift: compare against baseline/cohort.
Tools (pick your stack)
Excel/BI: basic KM via add-ins or calculated measures; good for communication.
R: survival, survminer.
Python: lifelines, scikit-survival.
SPSS/Stata/SAS: built-in survival procedures for regulated environments.
Why this works
Survival analysis reveals when risk peaks (timing), quantifies who needs help (drivers), and justifies what to do next (plays tied to risk patterns). It keeps every eligible new hire in the data—even those who haven’t left—so you’re not misled by incomplete outcomes.
Bottom line for leaders
Timing matters: survival analysis turns vague rates into time-aware insights leaders can act on.
Start with KM to see when and for whom; use Cox to identify why and by how much.
Translate results into playbooks with owners, run a measured pilot, and track lift.
Keep it ethical: separate voluntary vs involuntary exits, remove post-event data, and review fairness across groups.
You don’t need new systems—clean data + simple models + disciplined rollout beats complexity every time.
Up next: Thank you for reading this article. We will be posting short snippets on HR Analytics while we are working on season 2 of the larger series. Stay tuned.
Have you read our other articles? Go to Motioo Insights
Do you have any questions or comments? Contact us
