Risk-adjusted Kaplan-Meier Analysis — Weighted Survival Curves
Risk-adjusted Kaplan-Meier Survival Analysis · Also known as: weighted Kaplan-Meier, IPTW-adjusted Kaplan-Meier, propensity-score-weighted survival curves, adjusted survival curves
Risk-adjusted Kaplan-Meier analysis combines the non-parametric Kaplan-Meier estimator with inverse probability of treatment weighting (IPTW) or similar risk-adjustment procedures to produce survival curves that are comparable across groups as if the groups had identical distributions of baseline confounders. It is the observational-study analogue of plotting survival curves from a randomised trial.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use risk-adjusted Kaplan-Meier when you have time-to-event data from an observational study or a non-randomised database and wish to compare survival between groups while controlling for measured confounders, without imposing the proportional-hazards assumption required by Cox regression. It is particularly appropriate when you want marginal (population-averaged) survival curves rather than conditional (within-covariate-pattern) curves, and when the distribution of covariates differs substantially between groups. Do not use it if the positivity assumption is violated — that is, if certain covariate combinations appear in only one treatment group — because the required weights become extreme or infinite. It is also inappropriate as the primary analysis when many covariates are unmeasured, since the method corrects only for observed confounding. For hypothesis-generating or small samples (fewer than roughly 30 events per group after weighting) the weighted curves may be unstable and a Cox model may be preferable.
Strengths & limitations
- Provides intuitive, visually interpretable survival curves that are directly comparable across groups after confounding adjustment.
- Does not assume proportional hazards, making it valid even when the hazard ratio changes over time.
- Produces marginal structural estimates — what would happen at the population level — rather than conditional estimates that apply only to a specific covariate pattern.
- Can be combined with robust variance estimation or bootstrapping to give honest confidence intervals that account for weight variability.
- Flexible: weights can be derived from logistic regression, boosting, random forests, or targeted maximum likelihood estimation.
- Relies on the assumption of no unmeasured confounding; unmeasured variables that simultaneously drive treatment selection and the outcome cannot be adjusted away by weighting.
- Extreme or highly variable weights (caused by near-violations of the positivity assumption) inflate variance and can make curves unreliable; weight trimming or truncation introduces its own bias.
- Standard Greenwood confidence intervals are invalid; robust or bootstrap intervals must be used, adding computational complexity.
- The method adjusts only for baseline confounders; time-varying confounding between baseline and the event requires more complex marginal structural models with time-varying weights.
Frequently asked
What is the difference between risk-adjusted Kaplan-Meier and a stratified Kaplan-Meier?
Stratified KM plots separate curves by discrete covariate levels (e.g., sex, age group) but cannot adjust for many variables simultaneously and do not produce a single summary curve for each treatment arm. Risk-adjusted KM uses IPTW to account for the full covariate distribution in a single weighted estimator, yielding one marginal curve per treatment group that reflects the overall population.
Can I use propensity score matching instead of IPTW for adjusted KM curves?
Yes. Propensity score matching creates a matched subset and then standard KM curves are applied to the matched sample; the resulting curves estimate the average treatment effect in the matched (rather than full) population. IPTW estimates the effect in the original population. Both are valid approaches; the choice depends on the estimand of interest and whether the full sample or only matched patients are relevant.
How do I know if my weights are too extreme?
Examine the distribution of stabilised weights. Values exceeding 10 in many settings suggest near-violations of positivity and may severely inflate variance. Effective sample size reduction (sum of weights squared divided by sum of squared weights) should not drop below roughly 20–30% of the original sample before concern is warranted. Weight trimming at the 1st and 99th percentiles is a common pragmatic remedy, but it introduces bias and should be reported as a sensitivity analysis.
Is a weighted log-rank test valid for comparing the adjusted curves?
A weighted log-rank test can be used but requires careful implementation with a robust variance estimator to account for the variability in the estimated weights. Standard log-rank test software that ignores weighting will give incorrect p-values. Alternatively, a weighted Cox model provides a formal test of the marginal hazard ratio while accounting for weight uncertainty through robust standard errors.
Does risk-adjusted Kaplan-Meier handle time-varying confounders?
Standard risk-adjusted KM adjusts only for baseline confounders measured before treatment initiation. When confounders change over time and are affected by prior treatment (e.g., CD4 count in HIV research), time-varying IPTW within the marginal structural model framework is needed. This is a substantially more complex extension beyond the standard baseline-adjusted KM approach.
Sources
- Cole, S. R., & Hernan, M. A. (2004). Adjusted survival curves with inverse probability weights. Computer Methods and Programs in Biomedicine, 75(1), 45–49. DOI: 10.1016/j.cmpb.2003.10.004 ↗
- Hernan, M. A., Brumback, B., & Robins, J. M. (2001). Marginal structural models to estimate the joint causal effect of nonrandomized treatments. Journal of the American Statistical Association, 96(454), 440–448. DOI: 10.1198/016214501753168154 ↗
How to cite this page
ScholarGate. (2026, June 3). Risk-adjusted Kaplan-Meier Survival Analysis. ScholarGate. https://scholargate.app/en/epidemiology/risk-adjusted-kaplan-meier-analysis
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Cox proportional hazardsEpidemiology↔ compare
- Inverse Probability WeightingCausal inference↔ compare
- Kaplan-Meier AnalysisEpidemiology↔ compare
- Propensity Score MatchingResearch Statistics↔ compare
- Survival AnalysisResearch Statistics↔ compare