Announcement

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

  • Query about interaction terms with time varying covariates

    Hi all,

    I am studying the effect of a medication (beta-blockers dispensed after the diagnosis of cancer) on a cancer outcome (e..g. breast cancer death and breast cancer recurrence). We have modeled medication use as a time-varying covariate, meaning that user time on the drug begins accumulating from the first medication dispensing, and carries on until the end of follow-up. In my models, I also have variables ascertained at baseline such as age, and I am interested in figuring out if some of these variables modify the effect of the medication. The steps for how I created my time-varying medication variables are detailed below:

    1) First, I created a variable 'fbetablocker' representing the first time a beta-blocker was dispensed after diagnosis.

    2) I then created a variable 'timetofbetablocker' representing the time after diag (in years) to a patient's first beta-blocker.

    3) stset the data.

    4) stsplit the data at 'timetofbetablocker'.

    5) Generate an analysis variable that counts someone as a user from the first time they used a medication of interest, and a non-user before this time.

    6) Run the survival command. My final time-varying variable is named 'timetofbetablockerana', so the cox models read
    Code:
    stcox timetofbetablockerana............etc
    Now, I am essentially wondering if running interactions with this variable is a 'valid' measure of effect modification or not.

    Code:
    stcox timetofbetablockerana##i.newage
    is the code I've been using ('newage' is just age categorised), and then
    Code:
    timetofbetablockerana#i.newage
    to test for global differences.

    Another option (and one that I've run as well) is to simply create a binary variable (not time-varying) indicating whether or not someone in the study took a medication after diagnosis or not. I could then interact that variable with age.

    The two methods mentioned above give different results, so there must be one that has a more sound statistical justification. I have asked several people in my supervisory team etc as well, and none seem to have much of an idea.

    Any help would be much appreciated.

    Thanks, Oliver

  • #2
    Hi, is anyone able to help me with my question? I realise this may be more statistical theory than STATA specific, but I don't really know where else to ask. If I should be asking this sort of question elsewhere, does anyone have recommendations of where I can ask these sort of statistical specific questions? I have asked my biostats team at my university and they were unsure, so I am very unsure who to contact next.

    Thanks, Oliver

    Comment


    • #3
      I am going to ask one more time before I give up haha. Is anyone possibly able to help me with my question? Thanks!

      Comment


      • #4
        Here's a quick answer.

        The approach with the time-varying covariate is more correct.

        The approach where you assume users of beta blockers are exposed from diagnosis is not appropriate. It violates one of the fundamental rules of survival analysis of "don't condition on the future" (or "don't look into the future). I'm assuming you are studying individuals who were first prescribed beta blockers post diagnosis. The two approaches will give different results because you are using different definitions of exposure (i.e., more exposed person-time when you assume users are exposed from diagnosis).

        There's a reason why I said the time-varying covariate is *more* correct, rather than "correct". I probably should have said "probably more correct" since one needs much more information about your study design, data collectiuon, and hypothesis to say anything with certainty. This is a difficult area and I suggest you consult a biostatistician. This is a routine question for a biostatistician, so if your biostat team have said they were "unsure" my guess is that they were unwilling to help you rather than unable.

        Here's a link to a paper that gives an insight into some of the issues:

        Examining Bias in Studies of Statin Treatment and Survival in Patients With Cancer
        https://jamanetwork.com/journals/jam...rticle/2649379
        Last edited by Paul Dickman; 08 Apr 2022, 08:39.

        Comment

        Working...
        X