Choosing a survival measure beyond the hazard ratio
Choose a survival summary by the clinical question: time alive within a horizon, survival at a fixed time, or a modeled time ratio.
Fit
When it matters, and when it does not
It matters when
- The survival curves cross, come together or separate late, so the hazard ratio changes over follow-up.
- You want to say how much event-free time a treatment adds, in months, up to a set time.
- You expect a delayed or waning effect, such as with immunotherapy, and must plan for it in advance.
It matters less when
- Another event, such as death from another cause, can stop the outcome from happening. Use competing-risks methods. Read: Choosing an analysis for competing risks
- The hazards are close to proportional and a ratio answers your question. Use Cox regression, and add risks at set times. Read: Cox proportional hazards regression
- The event can happen many times to one patient, such as falls. Use a method for recurrent events. Read: Choosing an analysis for recurrent events
Before you start
What to check in your study
- The time horizon for an absolute survival summarySuch as two years. Fix it in the protocol, before seeing the curves, within the follow-up most patients reach.
- The event, and what ends follow-up without itoutcomeTypecompetingEventsLeaving the study is censoring. An event that makes the outcome impossible is a competing event.
- The measure you will reportsummaryMeasureA difference in restricted mean survival time, survival at set times, a time ratio, or hazard ratios by period.
- What you will do if the hazards are not proportionalName it in the analysis plan. A weighted log-rank test, which suits a delayed effect, must be chosen in advance too.
- Censoring is uninformative: patients who leave have the same outlook as similar patients who stay.How to check: Record why each patient left, and compare those censored early with those followed longer.
- Enough patients are still followed at the horizon.How to check: Show the numbers at risk. A horizon past most patients' follow-up rests on very few.
- Proportional hazards is checked, not assumed.How to check: Look at the log-minus-log plot, the scaled Schoenfeld residuals and whether the curves cross.
- A Weibull model's shape fits the data.How to check: Choose Weibull only when its shape is defensible. Check the fit; other accelerated failure time distributions have different assumptions.
Plain language
How it works
Start with what a patient needs to know. Restricted mean survival time compares average time alive or event-free within a prespecified horizon. It needs no proportional hazards assumption. Survival at a fixed time answers a different absolute-risk question. Choose these summaries for their meaning, rather than after testing whether curves cross.
Choose a Weibull accelerated failure time model when a time ratio answers the question and its common-shape, monotone-hazard assumptions fit. Weibull also has a proportional hazards representation. It is not a general cure for crossing curves. For changing effects, consider flexible parametric models or hazard ratios by period. With proportional hazards, Cox regression may remain useful. Extrapolation beyond observed follow-up needs additional justification.
Worked example
A worked example
- Setting
- A randomized trial of 600 adults with advanced lung cancer, comparing a new immunotherapy with chemotherapy.
- Question
- Does the new drug give patients more time alive, when its benefit is expected to come late?
- Approach
- The protocol fixed a 24-month horizon and chose restricted mean survival time for its clinical meaning. The curves cross.
- Result
- Up to 24 months, patients lived an average of 14 months with chemotherapy and 17 months with the new drug.
- What it means
- The new drug added an average of 3 months of life within the first 24 months. That is a figure a patient can weigh, and it says nothing about time after 24 months.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
A single hazard ratio assumes the effect is constant over the whole follow-up, and when the curves cross or converge that one number can describe no period of the study. Worse, it remains estimable and looks perfectly reasonable when the assumption fails, so nothing in the output announces the problem. And a hazard ratio is a relative quantity on a scale most clinicians do not think in: it cannot tell a patient how much longer they might live.
- Choosing the horizon after seeing the curves.Instead: Fix it in the protocol. Picking the time where the curves differ most is choosing the answer.
- Testing proportional hazards, and changing method only when the test is significant.Instead: Plan for hazards that are not proportional in advance. A test with few events can miss a real change.
- Reading one minus a Kaplan-Meier curve as the risk of the outcome when other deaths compete.Instead: Use the cumulative incidence function. Kaplan-Meier treats competing deaths as censored and overstates the risk when those deaths precede later outcomes.
Inside Mallard
How Mallard uses this
Learn search lists Choosing a survival measure beyond the hazard ratio. Its related methods explain restricted mean survival time and Weibull time ratios.
concept.choosing_survival_measure, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Flexible parametric proportional-hazards and proportional-odds models for censored survival data, with application to prognostic modelling and estimation of treatment effects (opens in a new tab) · Stat Med · doi 10.1002/sim.1203
- Restricted mean survival time: an alternative to the hazard ratio for the design and analysis of randomized trials with a time-to-event outcome (opens in a new tab) · BMC Med Res Methodol · doi 10.1186/1471-2288-13-152
- The accelerated failure time model: a useful alternative to the Cox regression model in survival analysis (opens in a new tab) · Stat Med · doi 10.1002/sim.4780111409
- Moving beyond the hazard ratio in quantifying the between-group difference in survival analysis (opens in a new tab) · J Clin Oncol · doi 10.1200/JCO.2014.55.2208