2
$\begingroup$

I have what is a very basic question about meta-analysis. If data from individual studies about the relationship of interest are presented in both bivariate and multivariate analyses, which should I use? Would it be appropriate to calculate effect sizes based on multivariate analyses (e.g., adjusted odds ratios), or should only data from bivariate analyses be used?

$\endgroup$
  • $\begingroup$ I should also mention that the covariates in the multivarite analyses are likely to differ across studies to be included in the meta-analysis. $\endgroup$ – Thomas S E Jun 30 '14 at 0:40
2
$\begingroup$

You can use both types of effect sizes, but you should make sure that you run seperate meta-analyses. That is, one synthesis should be based on bivariate effect sizes; a separate meta-analysis should focus on partial effects (e.g., regression coefficients; adjusted odds ratio etc.) sizes. Here is a recommendation from Aloe/Thompson (2013: 400) (you can download the paper free of charge):

"Partial effect sizes should not be combined with bivariate correlations, nor should different types of partial effect sizes be combined. Meta-analysts should present two sets of analyses: one for bivariate correlations and one for partial effects. "

Aloe, A. M., & Thompson, C. G. (2013). The synthesis of partial effect sizes. Journal of the Society for Social Work and Research, 4(4), 390 – 405.

$\endgroup$
  • $\begingroup$ Thanks for your response. If I am understanding things correctly, this means partial effect sizes from analyses with different covariates should not be included in the same meta-analysis. For example, if I wanted to examine the effect of diet on heart disease, it would be important to covary other meaningful predictors of heart disease (e.g., family history, exercise, stress, sex, smoking etc.). It is common for studies to covary a mix of these predictors but very rare for two studies to include precisely the same set of covariates. If I understand things correctly, this means in reality it wi $\endgroup$ – user49332 Jul 1 '14 at 17:40
  • 1
    $\begingroup$ @Thomas-S-E Did you check out the paper that I posted? It adresses many of those issues... On page 396 (section "Synthesizing Partial Effects") it says "[...] when analyzing a collection of partial-effect sizes, the reviewer needs to include predictor variables that reflect differences in model complexity that are likely to influence the sizes of the partial correlation". So, I would run meta-regressions that control for the most important predictors in your field. If you find my answer helpful, please consider upvoting it. $\endgroup$ – Bernd Weiss Jul 2 '14 at 6:11

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.