Elsevier

The Lancet

Volume 359, Issue 9318, 11 May 2002, Pages 1686-1689
The Lancet

Department of Medical Statistics
Survival plots of time-to-event outcomes in clinical trials: good practice and pitfalls

https://doi.org/10.1016/S0140-6736(02)08594-XGet rights and content

Summary

Survival plots of time-to-event data are a key component for reporting results of many clinical trials (and cohort studies). However, mistakes and distortions often arise in the display and interpretation of survival plots. This article aims to highlight such pitfalls and provide recommendations for future practice. Findings are illustrated by topical examples and also based on a survey of recent clinical trial publications in four major journals. Specific issues are: should plots go up or down (we recommend up), how far in time to extend the plot, showing the extent of follow-up, displaying statistical uncertainty by including SEs or CIS, and exercising caution when interpreting the shape of plots and the time-pattern of treatment difference.

Section snippets

Should plots go up or down?

A survival plot going down displays the proportion of patients free of the event (which of course declines over time), whereas a plot going up shows the cumulative proportion experiencing the event by time. In principle, both contain the same information, but the visual perceptions with regard to comparison of treatment groups can be quite different.

For instance, figure 1 shows three ways of displaying the same data on time to non-fatal myocardial infarction or death in the RITA-2 trial.5 The

How far in time to extend the plot?

Follow-up times in any one trial can vary substantially because patients are usually recruited over a long period, and some patients can be lost to follow-up. Length of follow-up is taken into account in the Kaplan-Meier life-table method1, 2, 3 for estimating the proportion of patients who experience an event by time since randomisation. Technically, any survival plot can be extended right through to the longest follow-up time, and five trials we surveyed did just that. However, this extension

Showing the extent of follow-up

So, readers need to be informed about the extent of follow-up, and stating the median follow-up time is often useful. Another helpful device is to display the numbers of patients event-free and still in follow-up in each treatment group at relevant time points, as shown in Figure 1, Figure 3. These numbers at risk of the event convey to the reader the increasing unreliability of estimates as time gets further from randomisation; most trials we surveyed included this information. The numbers on

Displaying statistical uncertainty

Most outcome results of clinical trials include measures of statistical uncertainty—eg, either SEs or CIs—for each treatment group, or a CI for the comparison of groups. However, survival plots often fail to include such measures. Hence the visual impression of any treatment differences, and how they vary over time, can look much more convincing than is really the case, especially if the clinical trial has few outcome events.

For any time since randomisation, the SE (or 95% CI) for the estimated

Interpreting the shape of survival plots

The easiest patterns to interpret are those that show no apparent difference between treatments or when there is a steady divergence between treatments over time. However, in many instances, more complex patterns seem to exist: the treatment difference might look greater early on (figure 1), the divergence between treatments might start later on (Figure 2, Figure 3), or the survival curves might cross. Such putative treatment-time interactions need cautious interpretation since there are rarely

References (15)

There are more references available in the full text version of this article.

Cited by (393)

  • A Guide to Selecting Flexible Survival Models to Inform Economic Evaluations of Cancer Immunotherapies

    2023, Value in Health
    Citation Excerpt :

    Limited guidance is offered to help inform the time point at which the plots are curtailed. Pocock et al35 suggested a point at which 10% to 20% of the patients are still at risk. More recently, Gebski et al36 suggested a less conservative approach, using the size of the decrease of the percent survival estimate at time point t if 1 extra event should occur immediately after t.

View all citing articles on Scopus
View full text