Survival Workflow

How to run a survival analysis in Licklider, from data preparation to Kaplan-Meier curves, log-rank tests, and Cox regression.

Survival analysis handles time-to-event data, where the outcome is how long until something happens — death, relapse, treatment failure, or any other defined endpoint — and some observations are censored because the event had not occurred by the end of follow-up.

Licklider supports the two most common survival analyses: Kaplan-Meier curves with log-rank testing for group comparisons, and Cox proportional hazards regression for covariate-adjusted analysis.

Data format

Your dataset needs at minimum:

  • A time column — the duration from the start of follow-up to the event or censoring (for example: survival_time, days_to_event, os_months, pfs)
  • An event indicator column — a binary variable where 1 (or True) means the event occurred and 0 (or False) means the observation was censored

An optional group column allows survival to be compared between groups.

Licklider detects survival data structure automatically from column names and values. If the detection is incorrect, confirm the column roles in the Chat.

Licklider cannot determine automatically whether the chosen time origin, event definition, censoring process, or covariate set is scientifically appropriate for your study. In particular, it cannot infer from the table alone whether censoring is informative, whether competing risks should be modeled separately, whether delayed entry or immortal time bias changes the interpretation, or whether a predictor should be treated as time-varying.

These limits matter because a survival workflow can still produce a clean-looking curve, p-value, or hazard ratio even when the survival question has been posed incorrectly. If the endpoint, time origin, or censoring logic is wrong, the reported median survival, log-rank result, and covariate effects can all look technically complete while supporting the wrong scientific claim.

Kaplan-Meier analysis

Request a Kaplan-Meier curve in the Chat:

  • "Show a Kaplan-Meier curve"
  • "Compare survival between the two groups"
  • "Plot overall survival by treatment arm"

The figure shows:

  • Step-function survival curves, one per group, in distinct colors
  • Censoring tick marks on each curve at the times when observations were censored
  • Median survival per group in the Inspector (shown as not reached when fewer than half of subjects experienced the event)

Log-rank test

When two or more groups are present, the log-rank test runs automatically. The p-value is shown in the Inspector alongside the curves.

For three or more groups, pairwise log-rank comparisons are also available with adjustment for multiple testing.

The log-rank test is surfaced here because a visual difference between curves is not enough to judge whether the group comparison is likely to reflect more than follow-up noise. Multiple-testing adjustment is included for multi-group follow-up comparisons so that screening several pairwise contrasts does not make the evidence look stronger than it is.

Number at risk

The number of subjects remaining at risk at selected time points is shown in the Inspector. This table is required for publication.

At-risk counts are shown because later portions of the curve may be supported by only a small number of remaining participants. Without those counts, a tail separation can look more stable than the follow-up support actually allows.

For more detail → see Kaplan-Meier Analysis and Kaplan-Meier Curve.

Cox proportional hazards regression

When you want to model the effect of one or more predictors on survival while controlling for other variables, use Cox regression:

  • "Run a Cox regression with treatment and age as covariates"
  • "Fit a proportional hazards model"

The results include hazard ratios with 95% confidence intervals, p-values, concordance index, and a check of the proportional hazards assumption.

This workflow pairs Cox regression with Kaplan- Meier because the two outputs answer different questions: KM makes the group-level survival pattern visible, while Cox regression estimates a covariate-adjusted hazard ratio without requiring the baseline hazard function to be specified in advance.

The proportional-hazards check is included because a single hazard ratio is easiest to interpret when it is approximately stable over follow-up. The concordance index is shown to give a quick summary of how well the fitted model orders earlier versus later events.

For more detail → see Cox Proportional Hazards Regression.

Required disclosures

Before a survival result can be used in a claim-bearing context, Licklider confirms:

  • Censoring has been acknowledged and will be reported
  • At-risk information is available
  • Median survival is not claimed if fewer than half of subjects experienced the event

These requirements are checked automatically. Unresolved items appear in the Inspector. For more detail → see Survival Data Detection Guard.

Choosing between KM and Cox

Use Kaplan-Meier when you want to describe and compare survival curves between groups without adjusting for other variables.

Use Cox regression when you want to estimate the effect of a predictor on survival while controlling for other covariates, or when you need to report hazard ratios.

Both can be run on the same dataset. A typical workflow runs KM curves first to visualize the comparison, then Cox regression to produce the covariate-adjusted effect estimate.

Limits Licklider cannot detect automatically

Licklider can help detect survival-shaped data and enforce key disclosures, but it does not guarantee that every survival-design problem has been caught.

It cannot determine automatically:

  • whether censoring is non-informative or whether dropout is related to prognosis
  • whether a competing-risk process should replace a standard KM or Cox interpretation
  • whether left truncation, delayed entry, or immortal time bias changes the target estimand
  • whether an omitted confounder or a poorly chosen covariate set makes the hazard ratio misleading
  • whether a time-varying effect should be modeled instead of reported as one constant HR

These limits matter because they can create a false sense that a technically valid survival output is also scientifically valid. Licklider can surface survival structure, run the requested method, and check a subset of reporting conditions, but it cannot recover study-design intent that is not represented clearly in the data.

Design Rationale & References

This workflow follows a simple rule: a survival result should make both the outcome pattern and the claim limits visible before it is treated as decision-ready. That is why the workflow combines survival-aware outputs with disclosure checks instead of showing only a curve or only a p-value.

Kaplan-Meier and Cox regression are paired because they serve complementary purposes in a standard survival workflow. KM provides a censoring-aware description of how survival evolves over time, while Cox regression supports covariate-adjusted effect estimation when you need to report hazard ratios [1][2].

At-risk counts and censoring disclosure are kept visible because the visual separation between curves can overstate certainty when few subjects remain under observation or when censoring is heavy [3]. Median survival is withheld when fewer than half of subjects experienced the event because a specific median claim would imply more follow-up support than the data contain.

The workflow also keeps the proportional-hazards check visible because a single HR can be misleading when the effect changes over time [4]. Licklider therefore treats "the model ran" and "the model is easy to interpret" as different questions.

  1. Kaplan, E. L., & Meier, P. (1958). Nonparametric Estimation from Incomplete Observations. Journal of the American Statistical Association, 53(282), 457-481.
  2. Cox, D. R. (1972). Regression Models and Life-Tables. Journal of the Royal Statistical Society: Series B (Methodological), 34(2), 187-220.
  3. Pocock, S. J., Clayton, T. C., & Altman, D. G. (2002). Survival plots of time-to-event outcomes in clinical trials: good practice and pitfalls. The Lancet, 359(9318), 1686-1689.
  4. Grambsch, P. M., & Therneau, T. M. (1994). Proportional Hazards Tests and Diagnostics Based on Weighted Residuals. Biometrika, 81(3), 515-526.

What this page does not cover