Analyzing compositional outcome data

Hi all,

I have been scratching my head for some time when it comes to analyzing trial data where the outcome variable is a component of another important outcome variable that is affected by the intervention.

In my example, we have conducted a double-blind weight loss trial with two groups. The intervention leads to a significantly lower total body weight vs. controls.

What I am now stuck thinking about is how to model various aspects of body composition (i.e., fat mass and fat free mass) which are components of total body weight.

In the same way that fat mass and fat free mass makes up total body weight, total body weight loss will be composed of fat mass and fat free mass loss. I am therefore thinking that when modeling the effect of the intervention on e.g., fat mass, these models should be adjusted for total body weight.

An option would be to use body fat percentage as the outcome variable. However, the range of body fat percentage is quite narrow as 100 percent body fat is physiologically impossible).

Happy to hear your thoughts about this. Should outcome models with body composition data, e.g., body fat, be adjusted for total body weight when we know that there is an effect of the intervention on total body weight?

are you suggesting adjusting for post-baseline body weight?

Yes. That was my thought. Just to clarify, we have 3 repeated measurements in this study and models are generally adjusted for outcome values at baseline.

i think it suffices to adjust for outcome values at baseline. Because post-baseline body weight is affected by treatment, adjusting for it may hide the treatment effect on the outcome analysed

1 Like

Yes, that is a good point that I have also been thinking about.

I think of body weight as a very strong mediator of the treatment effect on body composition compartments. Adjusting for body weight in this case will then let us say something about the treatment effect on e.g., fat mass independently of the effect on body weight, which is a clinically relevant parameter (body re-composition).

You can possibly omit one of the dimensions and do a multivariate T test (Hotelling test) to simultaneously infer about all components. A totally different approach would use variable clustering followed by principal components. Find out which components are moving together, then test then (2 group comparison).

1 Like

It’s two years later, but if you’re still working with body composition data, I would strongly caution against controlling for outcome body weight for two reasons.

First, it introduces a tautology because outcome body weight includes outcome body fat. So you’re likely to find a significant relationship simply because outcome body fat strongly predicts outcome body fat, which is not surprising.

Second, including any outcome variable in the model changes the study from an RCT to an observational study. This is doable but requires causal modeling assumptions and likely g-methods.

I don’t have an answer yet for your question, but it’s something I’m actively working on and will hopefully have an answer for soon.

1 Like