
Bivariate meta-analysis of diagnostic test accuracy
mars_dta.RdFit a Gaussian or binomial bivariate random-effects model for logit sensitivity and logit specificity from one 2-by-2 table per independent study.
Arguments
- data
Data frame with one row per study for a single index test.
- studyID
Name of the study identifier column.
- tp, fn, fp, tn
Names of the true-positive, false-negative, false-positive, and true-negative count columns, respectively.
- moderators
Optional one-sided formula of study-level predictors. Each endpoint has its own intercept and moderator coefficients.
- estimation_method
Restricted maximum likelihood (
"REML", the Gaussian default) or maximum likelihood ("MLE", required for binomial).- structure
Between-study covariance structure. Use
"UN"for two variances and a covariance, or"DIAG1"for two independent random effects.- zero_cell
For Gaussian fits, use
"add_half"to add 0.5 to both cells of any stratum containing a zero, or"none"to require finite uncorrected logits. Binomial fits use the original counts and require"none"if this argument is supplied.- missing
Use
"fail"to reject missing counts or moderator values, or"omit"to drop entire studies with missing inputs. Study IDs must always be present and unique.- control
Named list of optimizer controls passed to
mars()or tostats::optim()with L-BFGS-B for binomial fits.- family
Sampling model.
"gaussian"(default) fits transformed study estimates."binomial"models the counts directly.- quadrature
Number of adaptive Gauss–Hermite nodes per dimension for binomial fits, an integer from 3 to 35. Increase from the default of 11 to check numerical stability if counts or heterogeneity are extreme.
Value
A mars_dta object containing the underlying mars fit or binomial
optimizer result (fit),
original tables (counts), endpoint-level analysis data (effects),
labeled logit-scale heterogeneity covariance (tau), and omitted study
IDs (omitted_studies). The effects include raw counts and a continuity
correction flag for each endpoint. Gaussian effects also include yi
and vi. For binomial fits, fit$log_likelihood includes the binomial
normalizing constants.
Details
Sensitivity uses TP and FN, and specificity uses TN and FP. Both strata
must have positive totals. For Gaussian fits, compute_es() calculates the
two logit proportions and their approximate variances. Conditional on the
stratum totals, their within-study sampling covariance is zero. For
binomial fits, TP and TN have conditionally independent binomial likelihoods
given the study's random effects. A zero cell is valid as long as both
successes and failures are observed for each endpoint across the included
studies. Both models allow true logit values to covary across studies.
A threshold may be a study-level moderator when each study contributes one table. Multiple tests or thresholds measured within a study require additional modeling of their dependence.
References
Reitsma, J. B., et al. (2005). Bivariate analysis of sensitivity and specificity produces informative summary measures in diagnostic reviews. Journal of Clinical Epidemiology, 58(10), 982–990. doi:10.1016/j.jclinepi.2005.02.022
Chu, H., and Cole, S. R. (2006). Bivariate meta-analysis of sensitivity and specificity with sparse data: a generalized linear mixed model approach. Journal of Clinical Epidemiology, 59(12), 1331–1332. doi:10.1016/j.jclinepi.2006.06.011
Rutter, C. M., and Gatsonis, C. A. (2001). A hierarchical regression approach to meta-analysis of diagnostic test accuracy evaluations. Statistics in Medicine, 20(19), 2865–2884. doi:10.1002/sim.942 .
Harbord, R. M., Deeks, J. J., Egger, M., Whiting, P., and Sterne, J. A. C. (2007). A unification of models for meta-analysis of diagnostic accuracy studies. Biostatistics, 8(2), 239–251. doi:10.1093/biostatistics/kxl004 .
Examples
counts <- data.frame(study = 1:8, TP = c(44, 50, 56, 59, 64, 68, 69, 71),
TN = c(52, 42, 62, 74, 53, 65, 59, 72))
counts <- transform(counts, FN = 80 - TP, FP = 90 - TN)
fit <- mars_dta(counts, studyID = "study")
predict(fit)
#> profile endpoint logit estimate lower upper
#> 1 1 sensitivity 1.1674200 0.7626784 0.6701812 0.8355983
#> 2 1 specificity 0.7124426 0.6709407 0.5825646 0.7486782