Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Competing risks analysis with different exit times for different events

    In clinical trials of initial breast cancer treatment, patients ordinarily remain under observation until the earliest of the following events: local tumor recurrence, distant tumor recurrence, breast cancer death, death from other causes. Or, for women who experience no events, their outcomes are all censored at the scheduled end of the study. (Some women may experience no events and withdraw from the study early--these are also censored observations. They are problematical because of potential selection bias, but that is a side-issue at the moment.) Any outcomes that might actually happen after the observed outcome, are not observed. This standard framework lends itself readily to analysis using the -stcrreg- command, which implements competing risk analysis.

    I have some trial data in which some time after the end of the original study, investigators went back and determined who was alive and who dead, and whether the deaths were due to breast cancer. Notably, they did not determine whether (nor when) any recurrences happened between the time the women exited regular study observation and death. So we have a situation where the death outcomes are censored at a later time than the recurrence outcomes are. That is, in this data, we have recurrences whose observation is censored by death, but deaths are never censored by recurrence. I have been unable to figure out how to represent this in -stset- because the exit time with respect to recurrence differs from that with respect to death. For example, knowing that a woman survived until a given date and had no recurrence before some earlier date is uninformative about whether she remained at risk for recurrence throughout that interim period. Neither is it clear to me whether it is possible to analyze this data with -stcrreg- after I clear the -stset- hurdle. It sounds a bit like multiple failures, which as I understand it, -stcrreg- does not accommodate. I'm hoping somebody with more experience and a deeper understanding can clarify my thinking, or perhaps suggest an alternate approach.

  • #2
    This is a very interesting situation. From your description, I assume that the new data includes dates or ages of death for those who died. If this is the case, then I have only one suggestion for utilizing the new data: ignore recurrence and analyze time to breast cancer death, designating death from other causes as a competing risk. This approach can at best be considered an intent-to-treat analysis of the original assigned treatments.
    Steve Samuels
    Statistical Consulting
    [email protected]

    Stata 14.2

    Comment


    • #3
      Thanks, Steve. And your assumptions about the data are correct.

      Comment


      • #4
        Just for the sake of learning further about such a thought-provoking design, I wonder whether "recurrence" as a (binary, count or tvc variable, depending on the data) predictor couldn't be "adjusted" in the model, keeping the structure as described in #2.
        Best regards,

        Marcos

        Comment


        • #5
          Marcos,

          Thanks for the suggestion. I will give it a try and see what it does. The main problem I see is that if recurrence is recorded as 0 in the data, that information is correct only up to the first exit date. It may be incorrect as of the date at which death was ascertained. Moreover, this ascertainment error is biased: if recurrence is recorded as having occurred, then that remains true on the death ascertainment date, but if it is recorded as not having occurred, it may or may not have changed in between.

          Comment


          • #6
            Interesting problem. What is the study cutoff date for the survival analysis: date of last contact or last date for complete death ascertainment? In comparison, SEER cancer registries in the US use date of last contact as study cutoff date if "active follow-up" and last date of complete death ascertainment otherwise (so called "presumed alive").

            Comment


            • #7
              What is the study cutoff date for the survival analysis: date of last contact or last date for complete death ascertainment?
              Well, had they not collected the subsequent vital status information it would have been date of last contact, of course. The additional data on vital status would seem to enhance the data set because it provides a number of additional death events, so I would like to make it the date of complete death ascertainment, but the recurrence outcomes are actually censored at the time of last contact. So the question is whether it is somehow possible to incorporate all the data available into the analysis and still get valid estimates from a competing risks analysis.

              Comment

              Working...
              X