Skip to content
Back to search results

Survival Analysis

Introduced by
Edward L. Kaplan and Paul Meier
Year
1958
Type
Method
Family
time-event-modeling
Font size

What does it do?

Survival analysis is a collection of statistical methods for modeling time from a defined starting point until an event of interest occurs (disease, recovery, death, equipment failure).

Kaplan and Meier's nonparametric estimator (1958) and David Cox's proportional hazards model (1972) jointly enabled analysis of censored data—individuals whose event times are unknown because they left the study or were still event-free at follow-up. Indispensable in oncology, cardiology, infectious disease research, engineering reliability, and any field where time-to-event matters.

The intuition behind the method

Many studies follow people over time until they experience an outcome (disease diagnosis, death) or the study ends. Some individuals experience the outcome (events), while others reach study completion without the outcome (censored). Standard statistical methods ignore censored data or make strong assumptions, biasing results. Survival analysis accounts for censoring and follow-up time, allowing proper comparison of event rates between groups. The survival function S(t) = probability of being event-free at time t. Hazard h(t) = instantaneous rate of event occurrence at time t. Hazard ratio compares event rates between groups: HR = 1 (no difference), HR < 1 (reduced hazard), HR > 1 (increased hazard). This framework enables researchers to identify treatment effects, prognostic factors, and risk stratification despite incomplete data.

History

Kaplan and Meier introduced their nonparametric survival estimator in 1958, revolutionizing clinical trial analysis by properly handling censoring. David Cox's proportional hazards model (1972) extended this to multivariable settings, becoming the standard for covariate adjustment. The partial likelihood approach enabled efficient computation without specifying the baseline hazard function. Modern developments include inverse probability weighting for informative censoring (Robins, 1993), accelerated failure time models (Kalbfleisch & Prentice, 1980), and competing risks analysis (Fine & Gray, 1999) for settings where multiple event types preclude single analysis.

When to use it

Use survival analysis when the outcome of interest is time to event and data contain censoring. Oncology: comparing survival between treatment arms (standard randomized trials). Cardiology: analyzing time to cardiovascular events (MI, stroke) in cohort studies. Infectious disease: time from infection to recovery or progression. Medical implants: device failure or replacement time. Industrial engineering: time to equipment failure (reliability analysis). Any longitudinal study with staggered entry and/or early withdrawal can use survival methods. If no censoring exists (all participants followed to completion), simpler methods (t-test, logistic regression) suffice, though survival analysis remains valid.

Limitations and considerations

  • Proportional hazards assumption: Cox model assumes constant treatment effect over time; violations require alternative methods (accelerated failure time models, time-varying coefficients, stratification).
  • Sparse data at long follow-up times: survival estimates become less stable as participants are censored; confidence intervals widen, limiting inference for late-event prediction.
  • Administrative censoring: if censoring is related to prognosis (informative censoring), standard methods are biased; requires joint modeling or inverse probability weighting.
  • Assumes conditional independence given covariates; unmeasured confounding threatens causal inference, particularly in observational studies.
  • Small event counts reduce power and stability of Cox coefficients; generally need ≥10 events per covariate for stable estimation.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Correctly handles censored data without bias: exploits information from censored individuals (follow-up time) rather than discarding them.
  • Flexibly estimates both univariate (Kaplan-Meier) and multivariable (Cox) associations; baseline covariates are adjusted without assuming linear effects.
  • Hazard ratios are intuitive: single metric comparing event rates, easily interpreted by clinicians and communicated in publications.
  • Nonproportional hazards detectable: assumption diagnostics alert researchers when treatment effects change over time; alternative models (AFT, stratification, time-varying coefficients) address violations.
  • Risk stratification: predicted survival curves enable clinicians to quantify prognosis and individualize treatment decisions based on baseline characteristics.

Applications

  • Oncology trials: comparing overall survival and disease-free survival between chemotherapy regimens, immunotherapies, or surgical approaches; establishing standard-of-care practices.
  • Cardiovascular research: modeling time from baseline risk factors (hypertension, diabetes, smoking) to first major adverse cardiac event; developing risk prediction algorithms.
  • Infectious disease: analyzing time from diagnosis to viral suppression (HIV), treatment completion (tuberculosis), or clinical improvement (COVID-19); comparing treatment efficacy.
  • Medical devices and implants: reliability testing of pacemakers, artificial joints, stents; predicting device failure and estimating replacement intervals.
  • Occupational epidemiology: time from occupational exposure (asbestos, silica dust) to disease diagnosis; quantifying hazard and establishing exposure limits.

How it works

How it works

1

Administer survival analysis by defining: (1) the study start date (e.g., diagnosis, treatment initiation, study enrollment), (2) the event of interest (e.g., death, disease recurrence), and (3) the censoring indicator (whether the event occurred or was censored). Follow participants prospectively from baseline to event occurrence or end of follow-up. Record two key variables: time-to-event (days/months/years from start to event or last contact) and event status (1 = event occurred, 0 = censored). Collect baseline characteristics (age, treatment group, biomarkers) as covariates. Ensure event definitions are clearly operationalized and consistently ascertained across participants.

2

For descriptive survival: use Kaplan-Meier method. At each event time t_i, calculate S(t_i) = S(t_{i-1}) × [1 − (d_i / n_i)] where d_i is number of events at t_i and n_i is number at risk (not yet had event or censored). Plot cumulative survival curves by group; compare using log-rank test (χ² test of observed vs expected events under null hypothesis of equal survival). For univariate/multivariate association: use Cox proportional hazards regression. The hazard function h(t|X) = h₀(t) × exp(β₁X₁ + β₂X₂ + ...), where h₀(t) is baseline hazard and β coefficients quantify covariate effects. Estimate coefficients via partial likelihood; calculate hazard ratios (HR) = exp(β) with 95% confidence intervals. Test proportional hazards assumption: Schoenfeld residual plots and time-varying coefficient tests (p > 0.05 supports assumption). Report concordance index (C-index): 0.5 = random, 0.7–0.8 = good discrimination, >0.8 = excellent.

3

Kaplan-Meier curves plot cumulative probability of remaining event-free over time; curves that diverge early suggest strong treatment or group effects. Log-rank p-value tests equality of survival curves; p < 0.05 indicates significant difference. For Cox regression, interpret hazard ratios as: HR = 1.5 for a predictor means 50% higher hazard (event rate) per unit increase, adjusting for other covariates. HR < 1 indicates reduced hazard (protective effect). Confidence intervals that exclude 1 indicate statistical significance. Examine Schoenfeld residuals (plot coefficient over time) to verify proportional hazards assumption; departures indicate the covariate effect changes with time (violation suggests fit a time-varying coefficient model instead). Plot cumulative hazard or cumulative baseline hazard to visualize model fit. Use predicted survival curves (stratified by exposure groups or risk categories) to communicate results clinically.

Common pitfalls

  • Treating censored data as failures: incorrectly coding censored times as events biases hazard estimates downward and hazard ratios toward 1.
  • Failing to check proportional hazards assumption: plot Schoenfeld residuals or test time-varying coefficients; violation invalidates Cox model interpretation.
  • Omitting censoring indicators: must explicitly code event status (1 = occurred, 0 = censored); wrong coding leads to incorrect likelihood and estimates.
  • Comparing groups with unequal follow-up times without adjustment: survival analysis accounts for varying follow-up, but naïve comparison of event counts is biased.
  • Overfitting with too many covariates relative to events: Cox regression with many predictors and few events produces unstable, unreliable estimates; use penalized regression (ridge, lasso) or dimensionality reduction.
  • Using median follow-up as justification for omitting survival analysis: median follow-up describes the data but does not replace proper analysis of individual follow-up times.

Frequently asked questions

What does a hazard ratio of 0.6 mean, and how is it different from relative risk?

A hazard ratio (HR) of 0.6 means the instantaneous event rate (hazard) is 60% that of the reference group at any given time, adjusting for covariates—equivalent to 40% risk reduction. HR is from survival/Cox analysis. Relative risk (RR) is from cohort or trial data: RR = [incidence in exposed] / [incidence in unexposed]. HR approximates RR when events are rare (<10%), but they diverge with common outcomes. HR specifically compares event rates accounting for censoring and follow-up time; RR does not require survival methodology. Always report which measure you use.

What is the proportional hazards assumption, and what do I do if it's violated?

The proportional hazards assumption means the treatment effect (hazard ratio) is constant over time. Test it by plotting Schoenfeld residuals against time; a trend suggests violation. Alternatively, fit a time-varying coefficient model: include an interaction between the covariate and a function of time (e.g., log time). If assumption is violated, options include: (1) stratify analysis by the problematic variable (separate baseline hazards), (2) use accelerated failure time (AFT) models, or (3) report treatment effect at specific time points (e.g., 1-year, 5-year survival) separately. Always check and report on this assumption.

How should I interpret the log-rank test p-value in my Kaplan-Meier analysis?

The log-rank test compares the observed number of events in each group to the number expected under the null hypothesis (equal survival curves). p < 0.05 indicates statistically significant difference in survival between groups. Like all p-values, it tests a point null hypothesis; a large p-value does not prove equal survival, only insufficient evidence to reject equality. Also report median survival time and survival probabilities at key timepoints (1-year, 5-year), confidence intervals, and the number at risk at each interval on the Kaplan-Meier plot.

How do I handle multiple outcomes (e.g., disease recurrence and death)?

When multiple distinct events are possible, use competing risks analysis. If you model disease recurrence with Cox regression while ignoring death, you bias estimates because death censors recurrence follow-up. Fine-Gray subdistribution hazards model separates the marginal probability of each event. Calculate cumulative incidence function (CIF) for each event type (not 1-minus-other); compare using Gray's test. This approach is essential in oncology (recurrence vs non-cancer death) and cardiology (MI vs stroke).

Related methods

Sources

  1. Kaplan, E. L., & Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282), 457–481.

    Open DOI reference
    Copied.
  2. Cox, D. R. (1972). Regression models and life-tables. Journal of the Royal Statistical Society, Series B, 34(2), 187–220.

    Open DOI reference
    Copied.