Survival Analysis: KM vs Cox vs Parametric

Survival analysis covers any method for modeling time-to-event data — time until death, relapse, equipment failure, grant closeout, or attrition — where some subjects haven’t experienced the event by the end of observation (they’re censored, not missing). Three method families dominate the field: the Kaplan-Meier estimator, the Cox proportional hazards model, and parametric survival models (Weibull, exponential, log-normal, log-logistic). Most confusion about which to use doesn’t come from not understanding any one method — it comes from treating them as interchangeable tools for “analyzing survival data” instead of as answers to three genuinely different questions.

This guide is a decision framework, not a how-to for any single method. For the mechanics of building and reading a specific model, see How to Read a Kaplan-Meier Curve and The Cox Proportional Hazards Model.

Start with the question, not the method

Each of the three approaches answers a structurally different question:

“What does survival look like over time, in this sample or group?” — a purely descriptive question about the shape of the survival curve itself. This is Kaplan-Meier territory.
“How does survival differ across groups or covariates, after adjusting for other variables, without assuming a specific hazard shape?” — a comparative, covariate-adjusted question where you want a hazard ratio, not a curve. This is Cox regression territory.
“I need a predicted survival time or probability at a specific horizon, my hazard ratio isn’t constant over time, or I need to extrapolate beyond my observed follow-up.” — this is where parametric models take over, because Kaplan-Meier and Cox both share a limitation the question above exposes: neither one can say anything about survival past the last observed event time, and Cox’s central assumption (proportional hazards) can fail outright.

Picking a method by habit — “our field always reports Kaplan-Meier curves with a log-rank p-value” — produces a technically valid but incomplete analysis if the real question was a covariate-adjusted comparison or an absolute-time prediction. The rest of this guide works through each branch.

Kaplan-Meier: when the question is descriptive

The Kaplan-Meier estimator (Edward Kaplan and Paul Meier, 1958) is nonparametric: it makes no assumption about the shape of the underlying survival distribution. It builds a step-function estimate of the survival probability at each observed event time, correctly handling right-censored subjects along the way. That’s exactly why it’s the default first look at any time-to-event dataset — it describes what actually happened without imposing a model shape on it.

Kaplan-Meier is the right tool when:

You want to report or visualize the survival experience of a single group, or compare a small number of groups (typically via a log-rank test for the group comparison).
You have no continuous covariate you need to adjust for simultaneously — KM stratifies by a categorical grouping variable but doesn’t handle continuous covariates or multiple adjustment variables at once.
The deliverable is the curve itself: median survival, survival probability at a specific time point, or a visual comparison between arms.

Where it falls short: KM gives you a description, not an adjusted effect estimate. It can’t produce a hazard ratio that controls for age, stage, baseline severity, and treatment simultaneously. And because it’s built entirely from observed event times, it says nothing about survival beyond the last event in the data — there’s no way to extrapolate a KM curve forward.

Cox proportional hazards: when the question is covariate-adjusted comparison

The Cox model (David Cox, 1972) is semi-parametric: it estimates how covariates multiply the hazard, without specifying the shape of the baseline hazard itself. That’s the useful trade-off — you get an adjusted, multivariable hazard ratio (see How to Interpret a Hazard Ratio) without having to correctly guess the underlying distribution of event times.

Cox regression is the right tool when:

You need to compare survival across a variable of interest (treatment, exposure, genotype) while adjusting for multiple other covariates at once.
The deliverable is a hazard ratio with a confidence interval and p-value, not a predicted survival curve for a specific patient.
You’re not confident about the underlying event-time distribution and would rather not assume one — the semi-parametric form sidesteps that choice.

The trade-off is the proportional hazards assumption: the model assumes the hazard ratio between any two covariate values is constant over the entire follow-up period. If it isn’t — a treatment effect that fades over time, or Kaplan-Meier curves that visibly cross rather than staying separated — a standard Cox model is misspecified. The Cox model guide covers how to check this with Schoenfeld residuals and what to do when it fails (stratified Cox models, time-varying covariates, or moving to a parametric model). Like KM, a Cox model also generally isn’t used to extrapolate survival probability past the observed follow-up window with any real precision.

Parametric models: when PH fails, or you need an absolute prediction

Parametric survival models specify a distributional shape for the event-time process directly — exponential (constant hazard), Weibull (monotonically increasing or decreasing hazard), log-normal or log-logistic (hazard that rises then falls, common for post-surgical or infection-related events). Because the whole distribution is specified, these models can do two things KM and Cox structurally cannot:

Produce an absolute predicted survival time or probability for a given covariate profile — not just a relative hazard ratio, but “this patient has a 60% chance of being event-free at 3 years,” or a median predicted survival time in actual time units.
Extrapolate beyond the observed follow-up. KM and Cox are both anchored to the observed event times in the dataset; a fitted parametric distribution can be projected forward, which is exactly why parametric models dominate health-economic modeling and cost-effectiveness analyses that need lifetime survival projections from a trial with only a few years of follow-up.

Parametric models come in two parameterizations that answer subtly different questions: the proportional hazards (PH) form (a covariate multiplies the hazard, same interpretation as Cox) and the accelerated failure time (AFT) form (a covariate multiplies survival time itself — an “acceleration factor” that speeds up or slows down the event, which many non-statisticians find more directly interpretable: “this exposure shortens expected survival time by 20%,” rather than a hazard-scale statement). The Weibull distribution is the one family that can be fit in either parameterization with the same underlying model; exponential, log-normal, and log-logistic are AFT-native.

Reach for a parametric model when:

The proportional hazards assumption is checked and violated, and stratifying or adding time-varying covariates to a Cox model doesn’t adequately fix it.
The deliverable genuinely requires an absolute time or probability estimate — a prognosis tool, a sample-size or trial-duration projection, or a cost-effectiveness model — rather than a relative comparison.
You have reasonable grounds (prior literature, visual diagnostics, or subject-matter knowledge of the failure mechanism) to specify a hazard shape, and you’re willing to defend that choice — a wrong distributional assumption is a real risk parametric models carry that KM and Cox don’t.

Decision summary

Your actual question	Method	What you get	What it requires
What does survival look like in this group?	Kaplan-Meier	A survival curve; median survival; log-rank comparison across groups	Nothing distributional — nonparametric
How does survival differ by covariates, adjusted?	Cox proportional hazards	An adjusted hazard ratio with CI/p-value	The proportional hazards assumption holds
PH fails, or I need an absolute time/probability, or an extrapolation	Parametric (Weibull, exponential, log-normal, log-logistic)	A predicted survival time/probability; a projectable curve	A defensible distributional shape

Checking your assumptions before you commit

Before finalizing a method, run the checks that actually validate the choice rather than assuming it fits:

For Cox: test proportional hazards with Schoenfeld residuals (a significant trend over time indicates a violated assumption) and visually inspect log(−log(survival)) plots by group — parallel curves support PH, converging or crossing curves don’t.
For parametric model selection: compare candidate distributions with AIC/BIC, and check the fit visually against a nonparametric (KM-based) cumulative hazard plot before trusting extrapolated predictions — a model that fits the observed window well can still extrapolate badly if the tail behavior of the chosen distribution doesn’t match reality.
Across all three: confirm censoring is handled correctly and is plausibly non-informative (subjects aren’t censored for a reason related to their event risk) — every method here assumes that, and violating it biases all three equally.

Frequently asked questions

Can I use more than one of these methods in the same paper?

Yes, and it’s the common pattern in clinical and epidemiological reporting: lead with a Kaplan-Meier curve and log-rank test for the descriptive picture, then report an adjusted Cox hazard ratio as the primary comparative estimate. Parametric models are added on top of that when the analysis specifically needs an absolute prediction or extrapolation the first two can’t provide — they’re not usually a replacement for reporting KM and Cox, more a supplement for a specific downstream use.

My proportional hazards assumption is violated but I still want a hazard-ratio-style comparison — what are my options before jumping to a parametric model?

Two common fixes inside the Cox framework: stratify the model by the variable that violates PH (it no longer produces a hazard ratio for that variable specifically, but the rest of the model stays semi-parametric), or add a time-varying coefficient / time-by-covariate interaction term to let the hazard ratio change over follow-up. A parametric model is the next step when neither adequately captures the pattern, or when the real need is an absolute prediction rather than a fixed-form hazard ratio.

Do I need a parametric model for a typical clinical trial primary endpoint analysis?

Usually not. Most trial reporting for overall survival or progression-free survival leads with Kaplan-Meier curves and a Cox-based hazard ratio; that’s the standard regulatory and journal expectation. Parametric models become necessary in that same trial’s downstream work — a health-economic model extrapolating survival benefit over a patient’s lifetime from a few years of trial follow-up, for example — not in the primary efficacy analysis itself.

How do I choose which parametric distribution to use?

Start from the shape of the hazard implied by the mechanism you’re modeling: a genuinely constant risk over time suggests exponential; a risk that steadily rises or falls suggests Weibull; a risk that rises then falls (common after surgery, or during an acute infection window) suggests log-normal or log-logistic. Confirm the choice empirically — compare AIC/BIC across candidate distributions and check the fitted survival curve against the nonparametric Kaplan-Meier curve for the same data before relying on any extrapolation from it.

Survival Analysis: Choosing Between Kaplan-Meier, Cox, and Parametric Models

Ask about Survival Analysis: Choosing Between Kaplan-Meier, Cox, and Parametric Models