3
$\begingroup$

Suppose I want to conduct a trial with two treatment arms in each of which there are 50 patients. The patients in both arms are iid sample from sick population. I would like to compare treatment outcome of the two. Assume that there is only 1 person carrying out the experiments for both arms. Since patients are not recruited simultaneously, the study is carried out in sequential manner. Thus it is entirely possible that the experimenter will learn from the the procedure. I am assuming that I do not have extra time to make kick-in period in study to eliminate learning effect by saturating it.

Q: How does one account learning effect in study which cannot be conducted simultaneously in general study designs? For subjects measured repeatedly with cross over design, there are methods to account learning effect. I do not see how to account this in general design. I assume randomization might help, but it could be the case that the experimenter has super good memory and this helps him or her learning during the study.

$\endgroup$

1 Answer 1

4
$\begingroup$

One approach is to include time, flexibly modeled for example with a regression spline, as a predictor in the model. That allows you to correct for overall changes due to investigator "learning" over time, without having to pre-specify the functional form.

If the nature of the study might lead to differences in the "learning" pattern over time between the treatment groups, you could include an interaction term between the time function and treatment.

You should make sure that the treatments and controls are interlaced appropriately over time. You don't want to have a preponderance of one treatment at early times and the other at late times. That might require some modification of your randomization scheme.

$\endgroup$
3
  • $\begingroup$ A follow up question is how one can diagnose addition of spline necessary here. I would say time is a factor which may influence outcome due to variation of treatment. The treatment at beginning might be different from treatment at the end due to improved training. Should I say time becomes a confounder which is the reason of adding spline part? $\endgroup$
    – user45765
    Aug 8, 2022 at 18:28
  • 1
    $\begingroup$ @user45765 that seems like a good way to put it. You can then examine the magnitude of the time-variation modeled by the spline to evaluate its practical as well as its statistical significance. $\endgroup$
    – EdM
    Aug 8, 2022 at 18:54
  • $\begingroup$ I see. Thanks a lot for the suggestions and clarifications. $\endgroup$
    – user45765
    Aug 8, 2022 at 19:00

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.