The classic strategy to meta-analysis of diagnostic accuracy data is to pool a univariate measure of accuracy like the diagnostic odds ratio, the positive likelihood ratio or the negative likelihood ratio. For fixed effect estimation a Mantel-Haenszel estimator is implemented and for random effect estimation a DerSimonian-Laird estimator is available.
any object that can be converted to a data frame with integer variables
character, what effect size should be pooled? Either
character, method of estimation. Either
logical, should warnings produced by the internal call to
further arguments to be passed on to
First note that the function
madad is used to calculate effect measures. You can pass on arguments to this function via the
... arguments. This is especially useful for the
correction arguments, see the example.
The Mantel-Haenszel method performs fixed effect estimation of the effect sizes. For the DOR the variance of this estimator is calculated according to Robins et al. (1986) and for the likelihood ratios the variance is based on Greenland et al. (1985).
The DerSimonian-Laird method performs a random effects meta-analysis. For this tau^2, the variance of the log-transformed effect size (DOR, positive or negative likelihood ratio) is calculated by the DerSimonian and Laird (1986) method. The confidence interval for tau^2 is derived by inverting the Q-Test of Viechtbauer (2007).
Zwindermann and Bossuyt (2008) argue, that univariate summary points like the likelihood ratios should be derived from the bivariate model of Reitsma et al (2005). The function
SummaryPts, using output of
reitsma supports this approach.
An object of class
madauni, for which some standard methods are available, see
Performing univariate meta-analysis of diagnostic studies can not be recommended anymore now that bivariate methods are available, at least not if a reasonable number of primary studies is availabel. The package
mada provides this functionality for exploratory purposes and for meta-analysis of a small number of studies. The prefered way is to use
reitsma in conjunction with
The default value of
madad (and hence in the calculation of the effect sizes for
"all", i.e. the continuity correction is added to all studies if any has a zero cell. This is a different default value than the
metafor package uses. Set
"single" to arrive at the same values.
Philipp Doebler <email@example.com>
DerSimonian, R. and Laird, N. (1986). “Meta-analysis in clinical trials.” Controlled clinical trials, 7, 177–188.
Greenland, S. and Robins, J.M. (1985). “Estimation of a Common Effect Parameter from Sparse Follow-Up Data.” Biometrics, 41, 55–68.
Reitsma, J., Glas, A., Rutjes, A., Scholten, R., Bossuyt, P., & Zwinderman, A. (2005). “Bivariate analysis of sensitivity and specificity produces informative summary measures in diagnostic reviews.” Journal of Clinical Epidemiology, 58, 982–990.
Robins, J. and Greenland, S. and Breslow, N.E. (1986). “A general estimator for the variance of the Mantel-Haenszel odds ratio.” American Journal of Epidemiology, 124, 719–723.
Viechtbauer, W. (2007). “Confidence intervals for the amount of heterogeneity in meta-analysis.” Statistics in Medicine, 26, 37–52.
Zwinderman, A., & Bossuyt, P. (2008). “We should not pool diagnostic likelihood ratios in systematic reviews.”Statistics in Medicine, 27, 687–697.
1 2 3 4 5 6 7 8 9 10 11 12 13 14
data(AuditC) ## First example: DOR meta-analysis AuditC.uni <- madauni(AuditC) summary(AuditC.uni) ## Second example: sensitivity analysis ## Do continuity corrections make a difference? AuditC.uni_low <- madauni(AuditC, correction = 0.1) AuditC.uni_single <- madauni(AuditC, correction.control = "single") ## default is "all" confint(AuditC.uni) confint(AuditC.uni_low) confint(AuditC.uni_single)