Survival Analysis in Stata: A Guide for Epidemiologists

Survival analysis is the backbone of epidemiological research. Whether you’re estimating time to disease onset, time to hospital readmission, or time to death after a clinical intervention, the analytical framework is the same: you have time-to-event data with censoring, and you need an estimator that handles both.

Stata’s survival analysis suite is among the most mature in any statistical software. But it requires a specific workflow — starting with stset — that is unlike anything in the rest of Stata. If you skip or misconfigure this step, every downstream analysis is wrong. And this is exactly where ChatGPT falls apart.

Step 1: stset — Declaring Survival Data

Before any survival analysis, you must tell Stata your data is survival data. The stset command is not optional — it defines the time variable, the failure event, and any entry time or censoring structure.

Key decisions at this stage:

• What is failure? — The failure() option defines what counts as an event. If died = 1 means the event occurred and died = 0 means censored, specify failure(died). If the failure variable has multiple values (e.g., 1 = disease, 2 = death), specify which value: failure(event == 1).
• Scale matters. — Is time in days, months, or years? The scale affects hazard ratio interpretation. If time is in days and hazard ratios are near 1.0001, consider rescaling to months.
• Check your stset: After running stset, always run stsum to see the summary statistics and verify that the number of subjects, failures, and time at risk look correct.

Step 2: Descriptive Survival — Kaplan-Meier

The Kaplan-Meier curve is the standard visualization for survival data. Always include confidence intervals (ci). Always run the log-rank test. And always report median survival time — it’s more interpretable than the hazard ratio for non-technical audiences.

Step 3: Cox Proportional Hazards

The Cox model estimates hazard ratios: HR = 1.5 means the treatment group has a 50% higher instantaneous rate of the event at any given time, conditional on covariates. An HR < 1 means the treatment is protective.

Critical: the proportional hazards assumption.

The Cox model assumes that hazard ratios are constant over time. If the effect of treatment changes as time passes (e.g., a drug works well initially but wears off), the PH assumption is violated and the hazard ratio is misleading.

A significant p-value on estat phtest means the PH assumption is violated for that variable. Solutions: (1) stratify on the offending variable (stcox ..., strata(variable)), (2) include a time interaction, or (3) use a different model (parametric or accelerated failure time).

ChatGPT never runs this test. It generates stcox and stops. But a Cox model without a PH test is like a regression without checking residuals — you’re publishing results from a model whose key assumption may be violated.

Step 4: Competing Risks

In many studies, there are multiple ways to “fail.” A cancer patient might die from cancer, die from cardiovascular disease, or die from other causes. If you treat non-cancer deaths as censoring, you overestimate the cancer-specific hazard — because you’re assuming that patients who die from other causes would eventually have died from cancer given enough time.

The Fine-Gray model estimates subdistribution hazard ratios (SHR), which account for the competing risk. An SHR > 1 means the treatment group has a higher cumulative incidence of the event of interest, accounting for the fact that some patients experience the competing event instead.

Step 5: Parametric Models

When you have a theoretical reason to believe the hazard follows a specific functional form, parametric models can be more efficient than Cox:

How Sytra Handles Survival Analysis

Sytra understands the full survival analysis pipeline. When you say “run a Cox regression of treatment on time to readmission, adjusting for age, sex, and comorbidity,” it generates:

1. stset with the correct time and failure variables
2. stsum and stci for descriptive statistics
3. sts graph for Kaplan-Meier curves
4. stcox with vce(robust)
5. estat phtest to check the PH assumption
6. If PH is violated, it suggests stratification or a parametric alternative

If your data has competing risks, Sytra detects multiple failure types and suggests the Fine-Gray model. Because it understands the epidemiological methodology, not just the Stata syntax.