1
$\begingroup$

I would like to investigate the effect of a drug on survival with Cox regression. The drug can be administered either once or twice. In the univariate Cox regression, I see that a shorter time interval between the two drug administrations has an effect. I would also like to examine this effect in the multivariable Cox regression analysis. However, since only 50% of the patients were given the drug twice, I do not have the same number of patients as for the other variables. In the strict sense, this is not "missing data", as I know that this variable (time interval between drug administration) has no value for the patients with one dose. How can I deal with this? Or can I simply enter 0 as the time interval? I'm sure I'm not the first person to have this problem or question. In my search, however, I have only ever come across threads with the question of how to deal with "missing data". I look forward to any answers from you. Thank you very much!

$\endgroup$
4
  • $\begingroup$ Welcome to CrossValidated! First of all, do you know what software will you be using? Second, do you know what is the reason the other 50% of subjects were not given the drug twice? $\endgroup$
    – jmarkov
    Commented Nov 16, 2023 at 11:33
  • $\begingroup$ I want to use R for analysis and wanted to ask the question in general. In detail: The immunity (survival) of a vaccine is to be investigated over time. This was administered either once or twice. Thank you in advance! @jmarkov $\endgroup$ Commented Nov 16, 2023 at 12:07
  • $\begingroup$ Right, so the number of times it was administered it is random? I mean, a subject could have been administered twice just by chance? $\endgroup$
    – jmarkov
    Commented Nov 16, 2023 at 12:22
  • $\begingroup$ This is correct. $\endgroup$ Commented Nov 16, 2023 at 12:42

1 Answer 1

1
$\begingroup$

This situation is essentially what's described on this page with respect to loan amounts for individuals who didn't apply for loans: in your case, data on the interval between administrations can't exist for those who only got one administration.

The solution is similar: create a "dummy" predictor variable with a value of 0 for those who only got 1 administration, and a value of 1 for those who got 2 administrations (like the no-loan/loan dummy variable in the answer to the related question). Set the time interval between administrations to 0 for those with only 1 administration, and to the actual value for the others.

The Cox coefficient for that variable will be the log-hazard between those with only 1 dose and those with a (hypothetical) 0 interval between 2 doses. The coefficient(s)* for the time interval will only apply to those who received 2 doses.

After you fit the model, it will be most useful to compare predictions from representative scenarios rather than just relying on the values of the coefficients.


*You probably want to model that time interval flexibly, for example with a regression spline.

$\endgroup$
1
  • $\begingroup$ @CoxRegressionNewbie If this answer meets your needs, could you accept it by clicking the check mark? Thanks $\endgroup$
    – Peter Flom
    Commented Nov 30, 2023 at 20:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.