Survival analysis and clinical outcomes

Survival analysis is needed when what matters is not only whether an event occurs but how long it takes to occur, and when not all patients are followed for the same period. It applies to observational studies, clinical registries and trials in which the outcome is a time to event: survival, time to recurrence, a clinical endpoint or another time-based outcome. Comparing event proportions ignores two things that survival analysis does handle: censoring, meaning the patients who finish follow-up without having had the event, and the moment at which each event happens. Reducing a time-based outcome to a percentage at a fixed date wastes information and can produce comparisons that are hard to defend if follow-up is unequal or censoring is not handled correctly. Time to event is the question, not a detail.

For each patient, you need a well-defined time origin, an event date or date of last observation, and an indicator of whether the event occurred or follow-up was censored, plus the clinical covariates of interest. The client collects and provides these data, whether from a registry or a trial; the service prepares, models and interprets them. A key check before any analysis is the definition of the event and of the time origin: an ambiguous event or a poorly set origin compromises everything that follows, however correct the model. Covariates also need to be dated, or clearly placed relative to the start of follow-up, so that no information from after the start of the risk period is used. The censoring structure and the possible presence of competing risks are examined as well. The database does not need to arrive clean, but the event, its date, the time origin and the censoring rule must be defined without ambiguity, because they are the backbone of the analysis.

We do not apply Cox regression by default: we first assess whether the data meet its assumptions and, if they do not, use another method. This applies to any time-to-event analysis, because the Cox model rests on the proportional hazards assumption, that is, that a variable's effect on the hazard stays constant over time, and that assumption does not always hold. The work includes checking it and, when it fails, turning to alternative models: models with time-dependent effects, stratification or other approaches, depending on the case. When proportional hazards do not hold, restricted mean survival time (RMST) can also be considered if it fits the clinical question. Assuming proportional hazards without checking is one of the most common mistakes in survival analysis, and it produces hazard ratios that are hard to interpret as a constant effect over time. The choice of method adapts to the structure of the data, not the other way round.

When a patient can experience several mutually exclusive types of event, competing risks methods are used, not a standard survival analysis. This applies, for example, when the event of interest is device failure or a recurrence, but the patient may die of another cause first: that death is not harmless censoring, because it prevents the event of interest from ever occurring. If the aim is to estimate the cumulative incidence of the event of interest, treating the competing event as censoring can overestimate it. Depending on the aim, you may want to model the cumulative incidence with Fine-Gray (subdistribution hazard) or the cause-specific hazard with cause-specific models, alongside comparisons using Gray's test. The choice between the two approaches depends on the question, whether risk prediction or an aetiological effect; it is not automatic. Ignoring competing risks when they exist is a methodological error with consequences for the estimate, not a minor nuance.

In an observational study, treatments are compared by controlling for confounding, because the groups were not formed at random and differ in characteristics that affect the outcome. This applies when you want to estimate the effectiveness of treatments in real-world practice, without the random allocation that balances the groups in a trial. Techniques such as propensity score matching, weighting or multivariable adjustment are used, depending on the data structure and the question, together with sensitivity analyses for unmeasured confounding. For the comparison to be valid, the variables that explain why each treatment was indicated must be measured with sufficient quality. The scope needs to be stated precisely: these techniques estimate effectiveness adjusted for the observed factors and reduce bias, but they are not equivalent to the evidence from a randomised trial, nor do they prove causality on their own. What you receive is a more defensible comparison with its assumptions made explicit, not causal proof.

Yes. The service analyses time to event whether the outcome is a trial's primary endpoint or a clinical event in a registry, adapting the method to the design. It applies to the two profiles that share this service: the team that needs to analyse an efficacy or safety endpoint from an academic or pragmatic trial, and the team working with the events recorded in a registry or observational cohort. In both cases the deliverables can include Kaplan-Meier curves, hazard ratios with their clinical interpretation, adjusted models and sensitivity analyses, with the methods section written to the reporting guidelines that fit the design, such as STROBE for observational studies or the relevant guideline for academic trials. There is a limit to state clearly: the service covers the statistical analysis of endpoints in academic and pragmatic trials, not the formal regulatory documentation for a submission to a regulatory agency, which is a separate field. The survival methods are the same; the scope stays within academic research and publication.

Survival analysis focuses on the time until a specific event occurs, such as death, recurrence or failure, and on the factors that speed it up or delay it, while accounting for censoring. Longitudinal data analysis focuses on the trajectory: how a variable measured repeatedly on the same units changes over time, such as a biomarker, a scale or a clinical parameter. They do not answer the same question. Survival analysis answers "when, and with what risk, does this event occur?"; longitudinal analysis answers "how does this repeated measure change over time?". They are often combined, since longitudinal follow-up usually includes survival among its methods and recurrent events sit on the boundary between the two, but one models the time to an outcome and the other how a measure changes. The data source (registry, cohort, trial) does not decide which one is yours: what decides it is whether your question is about when an event happens or about the trajectory of a variable.

We deliver the results of the analysis, interpreted and ready to go into the manuscript: Kaplan-Meier curves, hazard ratio forest plots, regression tables, sensitivity analyses and the methods section written to the applicable reporting guidelines. This applies if you are going to publish in high-impact journals or present results to a committee, and need the analysis in a publishable format, not as software output that has to be reinterpreted. The service includes preparing the technical argument and the additional sensitivity analyses that reviewers may ask for when they question a methodological choice. If the analysis comes from a trial, the reporting guideline or methodological structure is adapted to the design; STROBE applies to observational studies, not to every context. The research team authors and leads the publication; the service provides the analysis, its interpretation and the methodological defence. What you receive is material ready for the manuscript or the report, with its methodological rationale documented.