Survival Analysis: Predicting When, Not Just If

Why it exists

Simple metrics like "average user lifetime" are often wrong. Calculating an average today ignores all the users who are still active and haven't churned. This artificially lowers the average and gives a misleading picture of retention. We need a method that correctly incorporates the information from subjects who have not yet experienced the event.

The mental model

Think of predicting a lightbulb's lifespan. You can't wait for every bulb in a batch to burn out; it could take years. Instead, you test them for 1000 hours. Some fail, and you record when. The rest are noted as having survived at least 1000 hours. Survival analysis is the statistical engine that takes both the failure times and these survival times to build an accurate model of reliability. Your active users are the lightbulbs that are still shining.

How it works

Survival analysis uses techniques like the Kaplan-Meier estimator to generate a survival curve, which plots the probability of surviving (e.g., not churning) over time. Instead of a single, misleading average, you get a function that answers questions like, "What percentage of users are still active after 6 months?" It correctly handles "right-censored" data—observations from subjects who drop out of the study before the event occurs (for instance, your analysis period ends while they are still a customer).

When to use it

Use it whenever you are analyzing time-to-event data, especially with censored observations. Three key places this shows up: first, modeling user churn or feature adoption in product analytics; second, A/B tests where the metric is time-to-conversion; and third, hardware reliability engineering for predictive maintenance.

When not to use it

If you only care about if an event happens (like a simple conversion rate), not when, this is overkill. It's also unnecessary if your dataset has no censoring—meaning you have observed the final outcome for every single subject. In that case, simpler descriptive statistics are sufficient.

One canonical example

A SaaS company wants to understand churn. Instead of calculating an average subscription length (which is skewed low by current users), they use a Kaplan-Meier curve. The curve reveals the median survival time is 8 months, meaning 50% of users churn by then. It also shows a 30% probability that a new user will still be subscribed after 12 months. This is far more actionable for financial forecasting than a single, incorrect average.