Learn R Programming

metaSEM (version 2.0)

SEC: Summary Estimate Correlation (SEC) of Meta-Analyses with Overlapping Primary Studies

Description

It computes the summary estimate correlation (SEC) matrix of several meta-analyses from their citation (incidence) matrix, together with the corrected covered area (CCA). When the variances of the summary estimates are given, it also returns their approximate sampling covariance matrix, which can be used in a second-order meta-analysis.

Usage

SEC(X, method = c("count", "size", "weighted"), n, q, weights, V)

Value

An object of class SEC, a list with the components:

R

The SEC matrix.

Sigma

The approximate sampling covariance matrix of the summary estimates, or NULL if V is not given.

k

The numbers of studies in the meta-analyses.

n.studies

The number of distinct studies in X.

overlap

The matrix \(X^T X\) with the numbers of studies on the diagonal and the numbers of shared studies off the diagonal.

CCA

The corrected covered area of all meta-analyses. See CCA.

CCA.pairwise

The matrix of the pairwise corrected covered areas.

eigenvalues

The eigenvalues of R.

method

The method used.

Arguments

X

A numeric or logical matrix or a data frame of 0 and 1, with the primary studies in the rows and the meta-analyses in the columns. Missing values are treated as 0 (not included). Rows without any inclusion are removed. Associated n, q, and weight values for removed rows are ignored; values for included rows must be finite.

method

Either "count" (the default), "size", or "weighted". See Details.

n

A vector of the sample sizes of the primary studies, one for each row of X. It is required when method="size".

q

A vector of the variances of the primary-study estimates, one for each row of X. It is required when method="weighted".

weights

Unnormalized weights of the primary studies when method="weighted". It is either a vector, with one weight for each row of X, used in all meta-analyses, or a matrix of the same dimensions as X. If it is missing, the inverse-variance weights 1/q are used.

V

An optional vector of the sampling variances of the summary estimates, one for each column of X.

Author

Mike W.-L. Cheung <[email protected]>

Details

Let \(X\) be the \(S \times m\) citation matrix, where \(X_{si}=1\) when primary study \(s\) is included in meta-analysis \(i\). SEC is the correlation between two meta-analytic summary estimators under a specified covariance model. Three versions are available:

"count"

The count-based approximation \(R_{ij}=k_{ij}/\sqrt{k_i k_j}\), where \(k_i\) is the number of studies in meta-analysis \(i\) and \(k_{ij}\) is the number of studies shared by meta-analyses \(i\) and \(j\). It requires only X. It is exact for unweighted means of independent study estimates with equal variances and approximate otherwise.

"size"

The size-weighted approximation \(R_{ij}=\sum_{s \in i \cap j} n_s/\sqrt{\sum_{s \in i} n_s \sum_{s \in j} n_s}\), where \(n_s\) is the sample size of primary study \(s\). It assumes that the sampling variances are inversely proportional to the sample sizes and that both meta-analyses use inverse-variance weights. It ignores the heterogeneity variance and therefore approximates the conditional target.

"weighted"

The exact weighted SEC for fixed coefficients. Let \(C\) be the matrix of the normalized coefficients, with \(C_{si}=X_{si}w_{si}/\sum_t X_{ti}w_{ti}\), and \(Q=\mathrm{Diag}(q_1,\ldots,q_S)\). The covariance matrix of the summary estimates is \(C^T Q C\), and \(R\) is its correlation matrix. Use the sampling variances \(v_s\) as \(q_s\) for the conditional target and \(v_s+\tau^2\) for the marginal target. When the weights are estimated from the same data, e.g., with a REML estimate of \(\tau^2\), the result is a plug-in approximation.

Each row of X must be a distinct and independent study-level unit that contributes the same estimand to every meta-analysis that includes it. A paper cited by two meta-analyses that use different outcomes, time points, subgroups, or contrasts is not a shared study.

The CCA and the pairwise CCA are computed with CCA. CCA describes the overlap of the inclusion sets; it is not the correlation between the summary estimates.

If V is given, the approximate sampling covariance matrix of the summary estimates is \(\Sigma=D_V R D_V\), where \(D_V\) is a diagonal matrix of \(\sqrt{V_i}\). V and \(R\) should refer to the same target, e.g., the variances of random-effects summaries with the marginal target. The SEC matrix is singular when the effective study coefficient vectors are linearly dependent; with column-specific weights this need not coincide with dependence of the columns of X. A warning is given for a singular or nearly singular matrix; a redundant meta-analysis may be removed before the second-order analysis.

References

Cheung, M. W.-L. (2026). Summary estimate correlation (SEC): Quantifying overlap-induced dependence in meta-analyses. Manuscript in preparation.

Pieper, D., Antoine, S.-L., Mathes, T., Neugebauer, E. A. M., & Eikermann, M. (2014). Systematic review finds overlapping reviews were not mentioned in every other overview. Journal of Clinical Epidemiology, 67(4), 368--375. tools:::Rd_expr_doi("10.1016/j.jclinepi.2013.11.007")

See Also

CCA, Turk23

Examples

Run this code
## A citation matrix of 4 primary studies and 3 meta-analyses
X <- matrix(c(1, 1, 0,
              1, 1, 0,
              1, 0, 1,
              0, 0, 1), ncol=3, byrow=TRUE,
            dimnames=list(paste0("Study", 1:4), c("MA1", "MA2", "MA3")))
SEC(X)

## Size-weighted SEC with the sample sizes of the primary studies
SEC(X, method="size", n=c(50, 100, 200, 80))

## Weighted SEC with the sampling variances of the primary studies.
## Since q = 4/n, it equals the size-weighted SEC.
SEC(X, method="weighted", q=c(0.08, 0.04, 0.02, 0.05))

## Second-order meta-analysis of the data in Turk23
## Standard errors of the summary estimates from their 95% CIs
se <- with(Turk23$data, (upper-lower)/(2*qnorm(0.975)))
SEC(Turk23$X, V=se^2)

Run the code above in your browser using DataLab