1
$\begingroup$

I have a cohort comprising of individuals who participate in a weight-loss program and their weight measurements are collected longitudinally for 12 months (not necessarily at the same time points). Their characteristics at baseline are also collected.

Over the 12 months individuals can have multiple visits as part of the program, and the research question is whether the total number of visits after 12 months is related to achieving a meaningful outcome. The dichotomous outcome is defined as whether an individual's 12 month weight is at least 5% lower than the baseline weight.

My question is whether this study design lends itself to a logistic regression, because the exposure - i.e. total number of visits - is time-varying within that 12 months. Some could spread out their number of visits over the 12 months or complete them within a few weeks. In a case-control study, where the selection is based on outcome and information is collected retrospectively to look at exposure, there is no explicit need to look at when in that look-back period the exposure takes place. This leads me to believe that the study design above can be appropriately addressed with a logistic regression.

Looking for additional thoughts in case I missed anything.

$\endgroup$

1 Answer 1

0
$\begingroup$

There is very likely nothing special about 5% weight loss. I mean, if the participants lost 4.9%, do we really want to say they didn't get the outcome? Additionally, it is easier for very light people to lose 5% as opposed to very heavy people. 5% of 300 is 15 whereas 5% of 150 is 7.5. Lighter people are more likely to achieve the outcome than heavier people simply because they have less weight to lose.

Why not just take the difference between pre and post weight? Or do an ANCOVA so you can adjust for other lifestyle factors associated with weightloss?

$\endgroup$
5
  • $\begingroup$ It's true that the 5% is not a magical number, but it is based on what is clinically meaningful for the study. A % weight change is used specifically for the reason you outlined - heavier individuals are more likely to lose more weight than lighter individuals. Control variables can also be added to a logistic regression... $\endgroup$ Sep 9, 2022 at 17:33
  • 1
    $\begingroup$ This answer also does not address the intended question of whether a logistic regression is appropriate for the above study... $\endgroup$ Sep 9, 2022 at 17:37
  • $\begingroup$ @stat134711 Then let me be clear: While logistic regression can be used in this case, you risk completely miss-characterizing the effect of your intervention due to the issues I mention above. You can do it, but I would advise you to instead compute the average weight loss across your subjects perhaps adjusting for pre-exposure variables important for weight-loss like baseline weight, physcial activity, etc. $\endgroup$ Sep 9, 2022 at 17:55
  • $\begingroup$ What you're doing is dichotmoizing a continuous measure (namely weight loss). I'm not above measuring the percent difference in weight loss and using that as the outcome in say a Beta regression, but to dichotomize completely and say patients who have not lost exactly 5% of their baseline weight or more -- even if that ends up being 4.9% -- seems to me as a very important limitation. $\endgroup$ Sep 9, 2022 at 17:58
  • $\begingroup$ I'm fully aware of the information lost by dichotomizing a continuous variable. I didn't fully lay out an SAP in the question because that's too much text. Yes, there will be adjustment for baseline characteristics. The 5% is after discussion with the clinician. The analysis is very secondary and hypothesis generating. Things would be different it were a primary analysis, and I would push towards a more efficient use of the data. Thanks for the discussion $\endgroup$ Sep 9, 2022 at 18:15

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.