I am conducting a meta-analysis on emotion recognition and theory of mind and am unsure on how to proceed to avoid a unit-of-analysis-error for some of the studies. In these studies, the results of e.g. an emotion recognition task are reported in the following way.

ASD ADHD Control

M (SD) M (SD) M (SD)

Anger

3.66 2.38 4.08

(2.31) (1.84) (1.26)

Fear

4.25 3.02 4.31

(2.01) (1.44) (2.16)

Sadness

2.25 2.07 2.03

(2.61) (1.24) (1.36)

...

In these tasks, various emotional expressions are presented in a single task, and the results for each emotions are presented separately. In most other studies, a total score is reported, which is the sum of all of these subscales. I see two ways to go forward:

Option 1:

Treating each emotion as if it were a separate study, i.e. computing an effect size for each of the emotions, then aggregating them. To do this using the R package metafor, I would have to guess the "value of the correlation of the sampling errors within clusters" (rho).

Option 2:

Computing summary statistics for the total score from the given data, then computing the effect size for this scale. From what I could gather, I can compute the mean of the total score by summing all the means of the subscales. To calculate the variance the total score (the square root of which would be the standard deviation of the total score), I would have to guess the correlation between the subscales as the scales cannot be assumed to be independent.

I am now wondering which option would be preferable and whether what needs to be guessed for both options is mathematically equivalent? I would appreciate any guidance.

Thank you.

More Manuel Burkart's questions See All
Similar questions and discussions