I'm conducting a meta-analysis on estimating a gold standard measure from other simpler measures. To determine the error of estimation, I derived the standard deviations of the differences between the two measures. Therefore, as effect sizes, I'm using the natural logarithm of the standard deviations of the differences $(lnSD)$ and their sampling variance $(1/2[n-1])$, as described on Nakagawa et al.

However, many studies included in my analysis, the same subject produces multiple results (multiple estimates [stages in the data.frame below]). Thus, I should determine the variance-covariance matrix of the effects. However, I am not sure if it is appropriate to determine the variance-covariance matrix using any correlation value for this type of effect, because effects variances are based only on the sample size $(1/2[n-1])$.

Is there any alternative to determine the variance-covariance matrix for this type of data, or I should just include the study as a random effect?

    study    yi     vi exercise stage
1      A 2.351 0.0250      row     3
2      A 2.351 0.0250      row     4
3      A 2.261 0.0250      row     5
4      A 1.922 0.0250      row     7
5      A 1.887 0.0250      row    10
6      B 1.884 0.0313    cycle     3
7      B 1.851 0.0313    cycle     4
8      B 1.774 0.0313    cycle     7
9      B 1.740 0.0313    cycle    10
10     C 1.831 0.0278    cycle     3
11     D 1.732 0.0294      run     5

Professor Wolfgang Viechtbauer answered this question on the R help listserv for meta analysis. The covariance between two $\ln(SD)$ is calculated as $r^2/(2(n-1))$; where $r$ is the correlation between the two variables.


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.