I am attempting to pool results about the relationship between change in body composition and change in metabolic markers (e.g. cholesterol, triglycerides etc.) from a variety of studies. Most of these studies are RCTs (with results at the level of the study, or trial arms), with BMI as the primary outcome and metabolic markers as secondary outcomes. I am not interested in deriving a pooled estimate of the effect of the intervention on primary or secondary outcomes, but rather the pooled effect of change in primary outcome on secondary outcomes.

Results reported include:

  • Primary outcome - baseline (BL), follow-up (FU), change from BL to FU, relevant standard deviations
  • Secondary outcomes - as for primary outcome

No studies reported the relationship between primary and secondary outcomes (either correlation or association).

My options for pooling results from these studies seem to be:

  1. Calculate study-level change in secondary outcomes as a function of change in primary outcome (this broadly seems to be treating them as ecological studies, hence title of question). Meta regression of change in secondary outcomes on change in primary seems to be an option here - but how should I weight the individual studies?
  2. Multivariate meta-analysis (? outside my comfort zone!) - allowing me to derive simultaneous intervention effects on primary and secondary outcomes. Not sure this gains me a lot though.
  3. Borrow methods from multivariate MA (e.g. 1 and 2) to estimate or impute within study correlation, allowing me to estimate the standardised regression coefficient of secondary on primary outcomes. Thereby enabling pooling of unadjusted standardised coefficients.

However, none of these 3 options will lead to an informative meta-analysis (see comments below).

My question is therefore:

Are there any better approaches I can use to derive a pooled estimate of change in primary outcome on secondary outcomes, given no studies reported correlations between these outcomes?

Edit 1: to focus the question

  • 1
    $\begingroup$ I think I would vote for do not do it. Option one is purely ecological and unlikely to be too convincing and options two and three seem to rely on imputing the very thing which is of primary interest. $\endgroup$ – mdewey Jun 2 '17 at 12:49
  • $\begingroup$ To get at your question, you need to correlations between these two outcomes. Solutions 1 and 2 clearly don't answer your question. I'm not sure how you would accomplish 3. My suggestion is to request the data from these studies and compute the correlation yourself and then meta-analyze the correlations. $\endgroup$ – dbwilson Jun 2 '17 at 13:30
  • $\begingroup$ mdewey and @dbwilson - you both seem to be giving me the reasoning I need to justify that, with the current data, MA would be pointless. I'm going to edit the question to clarify the outstanding question - can anything be done without any correlations between the outcomes $\endgroup$ – ChrisP Jun 2 '17 at 13:48

Your option 1 is doable (and I have done this to examine mediation, albeit indirectly). However, what you are testing is whether treatment-related change in the primary outcome is related to treatment-related change in the secondary outcome. This is not the same thing as the correlation between the two outcomes. Also, you have an ecological validity challenge: you don't know that the same people who are changing on the one outcome are the same as those changing on the other. In my own work, I have examined whether change in a proximal outcome (employment resulting from prison-based educational programming) was related to the distal changes in future criminality (i.e., arrest). In terms of the analysis, you can use inverse-variance weighted meta-analytic regression methods. Keep in mind that this does not account for any error in the effect size that is being used as the independent variable. Essentially, you are assessing whether the observed change in the primary outcome is related to change in the secondary outcome. I hope this helps.

  • $\begingroup$ Given that I only have variance for the related outcomes independently, rather than their common variance I'm not clear how IV weighting would be possible? $\endgroup$ – ChrisP Jun 2 '17 at 21:52
  • $\begingroup$ Use the variance for the effect size that is the DV. $\endgroup$ – dbwilson Jun 2 '17 at 22:36

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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