How to Report Kaplan-Meier Curves, Numbers at Risk, and Cox Regression in Medical Journals
Publishing survival analysis in clinical journals requires synchronized figures and tables: Kaplan-Meier curves with aligned numbers-at-risk tables, reverse Kaplan-Meier median follow-up, and multivariable Cox models with Schoenfeld residual diagnostics. Here is how to format every element to meet ICMJE and journal standards.
HumanPen Team
· 12 min read
Essential reporting standards for survival curves and Cox regression in medical journals
To report survival analysis in medical journals, investigators must present three synchronized components: Kaplan-Meier curves with confidence intervals and time-aligned numbers-at-risk tables, median follow-up calculated via the reverse Kaplan-Meier method, and multivariable Cox proportional-hazards models validated with Schoenfeld residual tests. Omitting numbers at risk, reporting naive follow-up averages, or ignoring non-proportional hazards triggers immediate statistical desk rejection across oncology, cardiology, and surgical journals.
Clinical manuscripts submitted to peer-reviewed medical publications—including JAMA Oncology, The Lancet Oncology, and the Journal of Clinical Oncology—undergo specialized statistical screening. Editors and biostatistical reviewers routinely evaluate whether time-to-event outcomes conform to established epidemiological guidelines, such as the CONSORT clinical trial reporting framework for trial endpoints. When a submission presents survival curves without numbers at risk or reports a hazard ratio without verifying model assumptions, reviewers cannot evaluate whether attrition, small numbers in the tail, or non-proportional hazards have biased the published findings.
Common reporting errors frequently stall revisions. Researchers often mistakenly compute median follow-up as a simple average among observed patients, drop confidence intervals from survival curves, or fail to state median survival time as `Not Reached` when fewer than 50% of participants experience the event. Furthermore, when authors condense time-to-event methods under strict portal word ceilings—as discussed in our guide on shortening structured medical abstracts—they frequently cut essential follow-up parameters. Transparent reporting requires retaining precise definitions of time origin, censoring mechanisms, and model diagnostics.
A Kaplan-Meier curve without an aligned numbers-at-risk table and a Cox model lacking Schoenfeld residual checks fail modern clinical reporting standards and risk immediate editorial rejection.
Time origin, censoring definitions, and median follow-up reporting
Every survival analysis begins by establishing a precise time origin (`t = 0`). In clinical trials, the origin is typically the date of randomization; in observational cohorts, it may represent histological diagnosis, surgical intervention, or initiation of medical therapy. Investigators must explicitly define the event of interest (such as all-cause mortality, disease recurrence, or graft failure) and distinguish right-censoring (`status = 0`) from event occurrence (`status = 1`). Right-censoring occurs when a participant remains event-free at the study cutoff date, withdraws consent, or is lost to follow-up. Setting the time origin ambiguously or assigning treatment status based on events that occur after cohort entry introduces immortal time bias, which spuriously favors the intervention group.
Reporting median follow-up time requires the reverse Kaplan-Meier method (Schemper & Smith, 1996), rather than a naive summary of observed times. Naive calculations—such as taking the median follow-up of all enrolled participants or evaluating only event-free survivors—produce severe downward bias because patients who experience early events contribute artificially truncated durations. In the reverse Kaplan-Meier approach, the event indicator is inverted (`status_rev = 1 - status`): individuals who experience the primary clinical event are treated as censored at their event date, while censored individuals are treated as having experienced the censoring event. This method is implemented in R via `survfit(Surv(time, 1 - status) ~ 1)` and in Python via `lifelines` as `KaplanMeierFitter().fit(durations=time, event_observed=1 - status)`.
| Follow-Up Reporting Method | Statistical Formula & Definition | Inherent Bias & Failure Mode | Journal Reviewer Status |
|---|---|---|---|
| Naive Observed Cohort Median | Median of all recorded observation times across all subjects | Severely underestimates follow-up by treating early death or event times as truncated observation periods | Unacceptable; rejected by statistical editors |
| Follow-Up Among Survivors Only | Median observation time calculated exclusively among subjects with `status = 0` | Produces conditional selection bias; overestimates follow-up if late-enrolled patients remain alive without adequate tracking | Discouraged; misrepresents cohort tracking distribution |
| Potential Follow-Up Duration | Calendar interval from patient enrollment date to the study administrative closure date | Assumes zero loss to follow-up; overstates actual observation time when dropouts occur | Acceptable only in prospective trials with near-zero attrition |
| Reverse Kaplan-Meier Estimator | Kaplan-Meier estimation with inverted status indicator (`status_rev = 1 - status`) | Unbiased non-parametric estimator; accurately accounts for differential enrollment times and censoring patterns | Gold standard; universally required by clinical oncology journals |
When analyzing specific disease endpoints, researchers must account for competing risks. If a patient dies of an unrelated cardiovascular event before experiencing cancer recurrence, standard Kaplan-Meier estimation assumes that the individual remained at equal ongoing risk of recurrence after death. Treating competing deaths as uninformative right-censoring violates the independent censoring assumption and inflates cumulative recurrence probabilities. When competing events occur in more than 5% to 10% of the study cohort, authors should compute Cumulative Incidence Functions (`CIF`) using the Aalen-Johansen estimator or fit a Fine-Gray subdistribution hazards model alongside cause-specific Cox models.
Peer reviewers require median follow-up calculated via the reverse Kaplan-Meier method with an interquartile range; simple averages or survivor-only summaries distort trial observation depth.
Constructing publication-ready Kaplan-Meier plots: aligned numbers at risk and RMST
The KMunicate guidelines (Morris et al., 2019, BMJ Open, `PMID 31575572`) established empirical standards for presenting Kaplan-Meier survival curves in medical journals. An audit of major medical publications revealed that many published curves lacked aligned numbers at risk, omitted confidence intervals, or presented isolated curves where readers could not distinguish meaningful differences from noisy tail fluctuations. A publication-grade survival figure must provide readers with immediate visual clarity regarding both survival probability and the remaining sample size supporting the curve at each clinical milestone.
- Aligned numbers-at-risk table: Place the numbers at risk directly below the horizontal time axis, ensuring that every column aligns with the main x-axis time ticks (such as 0, 6, 12, 18, and 24 months) across all study arms.
- Interval event and censor counts: Display the cumulative number of events or the number of censored individuals within each interval alongside or beneath the numbers at risk.
- Pointwise 95% confidence bands: Include pointwise 95% confidence intervals or shaded confidence bands around survival trajectories, showing where estimates are statistically robust.
- Censoring tick marks: Mark individual right-censored patients along the step curves using small vertical tick marks, enabling readers to identify censoring patterns over time.
- Tail truncation threshold: Truncate or dash the curve when the remaining number of patients at risk drops below 10% to 15% of the starting cohort, preventing overinterpretation of flat horizontal tails driven by single surviving outliers.
When survival trajectories cross, the conventional log-rank test loses statistical power because it assumes proportional hazards across the entire follow-up duration. In modern oncology trials—particularly those evaluating immune checkpoint inhibitors with delayed treatment effects or surgical therapies with early perioperative risks—curves frequently overlap or cross. In such settings, relying solely on a non-significant log-rank p-value misleads clinicians. Investigators should report Restricted Mean Survival Time (`RMST`), which calculates the area under the Kaplan-Meier curve up to a prespecified milestone time `t*`. The difference in RMST between study arms provides an intuitive, assumption-free measure of absolute survival time gained (such as an average gain of 3.4 months of life over 36 months).
When survival curves cross, the log-rank test is underpowered and uninterpretable; reporting Restricted Mean Survival Time differences up to a clinically meaningful milestone provides an assumption-free measure of treatment benefit.
Multivariable Cox regression: testing proportional hazards and handling violations
Multivariable Cox proportional-hazards regression models evaluate the association between multiple prognostic covariates and time-to-event outcomes. The central mathematical assumption of the semi-parametric Cox model is that the hazard ratio (`HR`) comparing any two exposure levels remains constant across the entire observation period. Evaluating this assumption requires computing Schoenfeld residuals, which assess whether covariate coefficients vary as a function of time. In R, this is performed using `cox.zph()`, while in Python `lifelines` it is executed via `CoxPHFitter().check_assumptions()`. A statistically significant correlation between scaled Schoenfeld residuals and time (`p < 0.05`) demonstrates that the proportional hazards assumption has been violated.
| Diagnostic Evaluation | Analytical Test & Visualization | Violation Threshold | Corrective Methodological Strategy |
|---|---|---|---|
| Schoenfeld Residual Correlation | Generalized linear correlation between scaled Schoenfeld residuals and transformed time (`cox.zph`) | Statistically significant slope (`p < 0.05`) for individual covariates or global model | Stratify by non-proportional confounders or incorporate time-varying coefficients |
| Log-Minus-Log Survival Plot | Plot of `log(-log(S(t)))` versus `log(t)` across discrete categorical strata | Parallel curves indicate proportionality; crossing, diverging, or converging curves indicate violation | Fit piecewise time-split models across distinct follow-up intervals |
| Time-by-Covariate Interaction | Addition of an explicit interaction term between covariate `X` and time `t` in the regression formula | Significant interaction parameter estimate (`beta_interaction != 0`, `p < 0.05`) | Report separate hazard ratios for distinct early and late follow-up periods |
| Outlier & Leverage Influence | Inspection of score residuals and dfbeta values (`residuals(fit, type='dfbeta')`) | Individual observation exerting disproportionate leverage on parameter estimates (`abs(dfbeta) > 0.2`) | Conduct sensitivity analyses excluding extreme influential observations |
When proportional hazards violations occur, clinical researchers have three established methodological options:
- Stratified Cox regression: If a baseline covariate (such as geographic study site or age bracket) violates proportional hazards but is not the primary exposure of interest, stratify the baseline hazard function by that variable (`strata(site)` in R or `strata=['site']` in `lifelines`). Stratification allows each stratum to have its own baseline hazard without estimating a single fixed hazard ratio.
- Time-varying coefficients and piecewise Cox models: When the primary treatment effect changes over time (for example, high initial surgical hazard followed by long-term protection), split follow-up time into discrete intervals (such as 0 to 6 months and 6 to 24 months) using `survSplit()` in R. Report distinct hazard ratios for each time window rather than a single distorted average.
- Landmark analysis: Set a clinically grounded landmark time point (such as 6 months post-enrollment) and evaluate subsequent survival exclusively among patients alive and event-free at that time, preventing immortal time bias and isolating late-stage dynamics.
When preparing the multivariable Cox regression table for publication, authors must present complete numerical disclosures. Every row must report the variable name, category levels with an explicit reference group (`Ref = 1.00`), unadjusted hazard ratio, multivariable-adjusted hazard ratio, exact 95% confidence intervals, and Wald test p-values. Furthermore, models must satisfy the Events Per Variable (`EPV >= 10–20`) guideline. In survival analysis, statistical power is determined by the total number of observed events (`status = 1`), not the total number of enrolled patients. Including 8 covariates in a model with only 35 observed events produces severe coefficient instability and inflated effect estimates.
Never publish a Cox regression table without reporting Schoenfeld residual diagnostic p-values; unexplained proportional hazards violations invalidate published hazard ratios.
Automated survival analysis workflow in HumanPen Start Project
Clinical research teams translating cohort spreadsheets or trial databases into journal-ready submissions can streamline their analytical pipeline through HumanPen Start Project. Attach your data and papers, describe what you need, approve the plan before work starts, and get the reviewed report, charts, and code ready to preview online or download.
The platform provides an isolated computational sandbox supporting dataset uploads up to 2 GB in Excel, CSV, Word, PDF, or zip formats. Analyses are executed using transparent open-source Python (`lifelines`, `pandas`, `matplotlib`) or R (`survival`, `survminer`) scripts. If preliminary cohorts were recorded in SPSS (`.sav`) or Stata (`.dta`) formats, they are processed directly through Python and R libraries. Researchers can inspect real-world execution standards in the public International Stroke Trial secondary analysis project on `/projects`, which processed 19,435 randomized patients into a 37-page statistical analysis report (`SAR.docx`), 300 dpi publication-grade Kaplan-Meier curves, and reproducible `run_all` code.
- Data auditing and follow-up validation: Automated screening of time origins, verification of non-negative observation intervals, and calculation of reverse Kaplan-Meier median follow-up with interquartile ranges.
- 300 dpi publication-grade Kaplan-Meier figures: Production of high-resolution vector and bitmap figures featuring KMunicate-compliant numbers-at-risk tables, interval event counts, and 95% confidence bands.
- Schoenfeld residual diagnostic suite: Systematic execution of `cox.zph` or `check_assumptions` across all candidate covariates, accompanied by diagnostic residual plots and global model tests.
- Multivariable Cox reporting tables: Formatted summary tables providing unadjusted and adjusted hazard ratios, 95% confidence intervals, reference category labels, and model log-likelihood metrics.
All project deliverables serve as a reproducible methodological and learning reference draft. The principal investigator, biostatistician, or postgraduate author must examine the analysis scripts, verify that clinical assumptions match the study protocol, and report all institutional and AI disclosures in accordance with ICMJE and journal submission policies. Data security remains paramount: every project operates in an isolated environment with transit encryption, never shares data across users, does not train models on private files, retains data for 6 months after the last activity with a 3-day deletion reminder, and permits permanent manual deletion at any time.
Automated statistical pipelines generate reproducible code and publication-ready figures, but clinical investigators and study biostatisticians retain sole responsibility for evaluating model assumptions and clinical validity.
Frequently asked questions
What should I report if median survival time is not reached in our cohort? When fewer than 50% of participants experience the primary event during the study period, the Kaplan-Meier curve remains above the 0.50 horizontal survival probability line, making median survival mathematically undefined (`NA`). Authors must never extrapolate the curve beyond the observed data. Instead, report median survival as `Not Reached` (`NR`), state the median follow-up time calculated via the reverse Kaplan-Meier method, and report milestone survival rates (such as 1-year, 3-year, or 5-year survival percentages with 95% confidence intervals).
How should investigators handle Kaplan-Meier curves that cross during follow-up? When survival curves cross, the proportional hazards assumption is violated, causing the conventional log-rank test to lose statistical power and rendering a single overall hazard ratio misleading. Investigators should report Restricted Mean Survival Time (`RMST`) differences up to a clinically meaningful milestone `t*`, which provides an assumption-free quantification of average survival time gained. Alternatively, fit a piecewise Cox proportional-hazards model reporting separate hazard ratios before and after the crossing point.
Can I run survival analysis if my raw clinical data is stored in SPSS or Stata formats? Yes. In HumanPen Start Project, SPSS (`.sav`) and Stata (`.dta`) files are directly imported using Python (`pyreadstat`) or R (`haven`), preserving variable labels, value codes, and censor status definitions. The underlying survival algorithms run in transparent, open-source Python (`lifelines`) or R (`survival`), delivering reproducible scripts that eliminate proprietary software lock-in.
How many events are required per predictor variable in a multivariable Cox model? Multivariable survival models require at least 10 to 20 observed events per predictor variable degree of freedom (`EPV >= 10–20`) to prevent coefficient overfitting and severe estimation bias. In time-to-event analysis, effective sample size is dictated by the total number of observed failure events (`status = 1`), not the total number of enrolled patients (`N`). For example, in a trial of 500 patients where only 30 events occurred, a multivariable Cox model can reliably accommodate only 2 to 3 predictor variables.
Sources
- Morris et al. (2019, BMJ Open, PMID 31575572) — Proposals on Kaplan-Meier Plots in Medical Research (KMunicate) — Empirical evaluation of Kaplan-Meier reporting practices in top medical journals and standardized recommendations for aligned numbers-at-risk tables.
- lifelines — KaplanMeierFitter & At-Risk Counts Documentation — Open-source Python library documentation for estimating univariate Kaplan-Meier survival curves and rendering at-risk count tables.
- lifelines — CoxPHFitter & Proportional Hazards Assumption Checks — Technical reference for fitting multivariable Cox models and evaluating proportional hazards using Schoenfeld residual tests.
- CRAN — survival Package (survfit, coxph, cox.zph) — The foundational R package for survival analysis, providing the core routines for Surv objects, Kaplan-Meier curves, and proportional hazards testing.
Primary source documentation verified via live HTTP retrieval on November 18, 2026. Clinical researchers should verify journal-specific author guidelines and reporting checklists before submitting survival analysis figures and tables.
KEEP READING
Staggered DiD Panel Workflow: TWFE Bias, Bacon Decomposition, and Modern Estimators
9 min read
Academic writingPropensity Score Matching vs IPW: Balance, Love Plots & E-Values
11 min read
Academic writingRegression Discontinuity Design RDD Checklist: rdrobust Bandwidth, Manipulation Tests & RD Plots
11 min read