Several students and colleagues have asked me recently about an issue that comes up in multivariate meta-analysis when some of the studies include multiple treatment groups and multiple outcome measures. In this situation, one might want to include effect size estimates for each treatment group and each outcome measure. In order to do so in fully multivariate meta-analysis, estimates of the covariances among all of these efffect sizes are needed. The covariance among effect sizes arises for several reasons:

- For a single outcome measure, effect sizes based on different treatment groups compared to a common control group will be correlated because the same control group data is used to calculate both effect sizes;
- Effect sizes based on a single treatment group and a single control group, but for different outcome measures, will be correlated because the outcomes are measured on the same set of units (in both the treatment group and the control group).
- Effect sizes based on different treatment groups and for different outcome measures will be correlated because the outcomes are measured on the same set of units in the control group (though not in the treatment group).

For standardized mean difference (SMD) measures of effect size, formulas for the covariance are readily available for the first two cases (see e.g., Gleser & Olkin, 2009), but not for the third case. Below I review the formulas for the covariance between SMDs in the first two cases and provide a formula for the third case.

# Notation and Model

Suppose that the experiment has a control group that includes \(n_0\) units and \(T\) treatment groups that include \(n_1,...,n_T\) units, respectively. Also suppose that \(J\) outcome measures are made on each unit in each group. The formulas below assume that the data follow a one-way MANOVA model. Let \(y_{ijt}\) denote the score for unit \(i\) on outcome \(j\) in group \(t\). Then I assume that

\[y_{ijt} = \mu_{jt} + \epsilon_{ijt},\]

where the errors are multi-variate normally distributed with mean zero, variance that can differ across outcome but not across treatment group, and correlation that is constant across treatment groups, i.e. \(\text{Var}\left(\epsilon_{ijt}\right) = \sigma^2_j\), \(\text{Cov}\left(\epsilon_{ijt}, \epsilon_{ikt} \right) = \rho_{jk}\).

Denote the mean score on outcome \(j\) in group \(t\) as \(\bar{y}_{jt}\) and the standard deviation of the scores on outcome \(j\) in group \(t\) as \(s_{jt}\), both for \(j = 1,...,J\) and \(t = 0,...,T\) (with \(t = 0\) corresponding to the control group). Also required are estimates of the correlations among outcome measures 1 through \(J\), after partialling out differences between treatment groups. Let \(r_{jk}\) denote the partial correlation between measure \(j\) and measure \(k\), for \(j = 1,...,J - 1\) and \(k = j + 1,...,J\).

With multiple treatment groups, one might wonder how best to compute the standard deviation for purposes of scaling the treatment effect estimates. In their discussion of SMDs from multiple treatment studies, Gleser and Olkin (2009) assume (though they don’t actually state outright) that the standard deviation will be pooled across all \(T + 1\) groups. The pooled standard deviation for outcome \(m\) is calculated as the square root of the pooled variance,

\[s_{jP}^2 = \frac{1}{N - T - 1} \sum_{t=0}^T (n_t - 1)s_{jt}^2,\]

where \(N = \sum_{t=0}^T n_t\). The standardized mean difference for treatment \(t\) on outcome \(j\) is then estimated as

\[d_{jt} = \frac{\bar{y}_{jt} - \bar{y}_{j0}}{s_{jP}}\]

for \(j = 1,...,J\) and \(t = 1,...,T\). The conventional estimate of the large-sample variance of \(d_{jt}\) is

\[\text{Var}(d_{jt}) \approx \frac{1}{n_0} + \frac{1}{n_t} + \frac{d_{jt}^2}{2 (N - T - 1)}.\]

# Covariances

For SMDs based on a common outcome measure and a common control group, but different treatment groups, the large-sample covariance between the effect size estimates can be estimated as

\[\text{Cov}(d_{jt},d_{ju}) \approx \frac{1}{n_0} + \frac{d_{jt} d_{ju}}{2 (N - T - 1)}.\]

The above differs slightly from Gleser and Olkin (2009, Formula 19.19) because it uses the degrees of freedom \(N - T - 1\) in the denominator of the second term, rather than the total sample size. If the total sample size is larger relative to the number of treatment groups, the discrepancy should be minor.

SMDs based on a single treatment group but for different outcome measures follow a structure that is essentially equivalent to what Gleser and Olkin (2009) call a “multiple-endpoint” study. The large-sample covariance between the effect size estimates can be estimated as

\[\text{Cov}(d_{jt},d_{kt}) \approx r_{jk} \left(\frac{1}{n_0} + \frac{1}{n_t}\right) + \frac{r_{jk}^2 d_{jt} d_{kt}}{2 (N - T - 1)}\]

(cf. Gleser & Olkin, 2009, Formula 19.19). Note that if the degrees of freedom are large relative to \(d_{jt}\) and \(d_{kt}\), then the correlation between the effect sizes will be approximately equal to \(\text{Cor}(d_{jt},d_{kt}) \approx r_{jk}\).

Finally, the large-sample covariance between SMDs based on different treatment groups and different outcome measures can be estimated as

\[\text{Cov}(d_{jt},d_{ku}) \approx \frac{r_{jk}}{n_0} + \frac{r_{jk}^2 d_{jt} d_{ku}}{2 (N - T - 1)}.\]

This is similar to the previous formula, but does not include the term corresponding to the covariance between different outcome measures in a common treatment group.

If \(r_{jj} = 1\) is used for the correlation of an outcome measure with itself, all of the above formulas (including the variance of \(d_{jt}\)) can be expressed compactly as

\[\text{Cov}(d_{jt},d_{ku}) \approx r_{jk} \left(\frac{1}{n_0} + \frac{I(t = u)}{n_t}\right) + \frac{r_{jk}^2 d_{jt} d_{ku}}{2 (N - T - 1)},\]

where \(I(A)\) is equal to one if \(A\) is true and equal to zero otherwise.

# References

- Gleser, L. J., & Olkin, I. (2009). Stochastically dependent effect sizes. In H. Cooper, L. V. Hedges, & J. C. Valentine (Eds.), The Handbook of Research Synthesis and Meta-Analysis (2nd ed., pp. 357-376). New York, NY: Russell Sage Foundation.