“How long does a typical subscriber stay with us?” sounds like a question you could answer with an average. Take everyone who cancelled, average how long they stayed, done. That answer is badly wrong, and the reason it is wrong is the starting point for a whole branch of statistics. Survival analysis was developed in medicine to study time until death or relapse, and the same tools apply to time until churn, time until a second purchase, time until a loan defaults, or time until a machine fails.

The problem: people who have not finished their story

Consider ten subscribers to an illustrative meal-kit service, observed up to today:

  • Cancelled after 2, 4, 5, 7 and 9 months.
  • Still active after 3, 5, 8, 10 and 12 months (they joined at different times, so we have watched them for different lengths).

Averaging the five cancellers gives 5.4 months. But that ignores the five who are still subscribed, including two who have already stayed 10 and 12 months. Averaging all ten observed durations gives 6.5 months, which is also wrong: for active subscribers we know only that their tenure is at least what we have seen. Both shortcuts underestimate true tenure.

Observations where the event has not yet happened are called censored (more precisely, right-censored). Throwing them away biases results towards short durations, because long-lived customers are exactly the ones most likely to still be around. Treating their current tenure as final has the same effect. Survival methods use censored observations correctly: as information that the customer survived at least this long.

The Kaplan-Meier estimate, step by step

The Kaplan-Meier method (Kaplan and Meier, 1958) estimates the survival curve: the probability of still being a customer after each length of time. At every time when someone churns, it asks: of the people still at risk (subscribed and still under observation just before this point), what fraction left? Survival is then multiplied by one minus that fraction.

MonthAt riskChurnedSurvival so far
21011 × 9/10 = 0.900
4810.900 × 7/8 = 0.788
5710.788 × 6/7 = 0.675
7510.675 × 4/5 = 0.540
9310.540 × 2/3 = 0.360

Watch the “at risk” column. At month 4, only 8 remain at risk: one churned at month 2, and one was censored at month 3, so they simply leave the denominator without counting as churn. The subscriber still active at 5 months counts as at risk at month 5 (they were there) but not afterwards. That is how censored customers contribute: fully, for as long as we observed them.

The median survival time is the first point where the curve drops to 50% or below: here, month 9. Half of subscribers are estimated to last at least nine months, far more than the naive 5.4. With ten customers the estimate is extremely uncertain, of course; real analyses use thousands and show confidence bands.

Hazard: the risk at each moment

The survival curve has a companion concept, the hazard: the chance of churning in a given period among those who have survived to it. Hazard shapes are often more useful for business decisions than the survival curve itself.

  • High early hazard that falls: many customers leave in the first months, then the survivors settle in. This is typical for consumer subscriptions and argues for investing in onboarding.
  • Spikes at specific points: churn concentrated at month 12 points to annual renewals, or at the end of an introductory discount. Retention effort should arrive just before those spikes.
  • Rising hazard over time: common for physical equipment wearing out, and sometimes for products that customers outgrow.

Comparing groups and adding explanations

Kaplan-Meier curves can be drawn separately for segments, such as customers acquired through referral versus paid ads, or monthly versus annual plans, and compared with a log-rank test. To study several factors at once, the Cox proportional hazards model (David Cox, 1972) estimates how each factor multiplies the hazard. A result might read: customers who skipped onboarding have 1.8 times the churn hazard of those who completed it, holding plan type and city constant. The model assumes these multipliers stay constant over tenure; when they clearly do not (for example, a discount that protects only in the first three months), that assumption needs checking and the model adjusting.

Business uses beyond churn

Anything framed as “time until something happens, where some cases have not happened yet” fits. Time from first to second purchase, where most new customers have not yet bought again. Time to employee attrition, with current staff censored. Time to loan default, with performing loans censored. Time to resolve a support ticket, with open tickets censored. In each case, simple averages over completed cases are biased in the same direction as our meal-kit example.

A few cautions. Censoring is assumed to be unrelated to the risk of the event; if customers stop being observed because, say, they switched to a different account type in a way that relates to their likelihood of leaving, estimates are biased. Definitions matter too: is a paused subscription churned? Is a customer who returns after three months a survivor or a new start? Settle those rules before drawing curves, and you have one of the most honest ways of measuring how long relationships last.

Written by the RP Strativa Team. The meal-kit service, the 1.8 hazard ratio and other figures are illustrative.

Leave a Reply

Trending