A comparison of univariate and bivariate models in meta-analysis of diagnostic accuracy studies.

AIM

: An implicit diagnostic threshold has been thought to be the cause of between-study variation in meta-analyses of diagnostic accuracy studies. Bivariate models have been used to account for implicit diagnostic thresholds. However, little difference in estimates of test performance has been reported between univariate and bivariate models. This study aims to undertake another comparison of these two models in order to determine if spectrum effects could better explain the variation across studies.

METHODS

Studies were selected from those provided in Ohle et al.'s meta-analysis and quality scored using QUADAS 2. Univariate analyses of sensitivity and specificity were computed using two models: one bias-adjusted and the other not. The univariate sensitivity and specificity results were compared with the bivariate logit-normal summary ROC method.

RESULTS

Similar results were obtained when using summary ROC and univariate pooling methods for sensitivity and specificity. Differences in study characteristics were found for outlier studies in univariate analyses, suggesting spectrum effects.

CONCLUSION

Univariate pooling methods provide an estimate of test performance for an average disease spectrum which is possibly why results concur with the bivariate models. A better appreciation of such spectrum effects can be demonstrated through univariate analyses, especially when the forest plots are examined in either bias-adjusted or non-bias-adjusted univariate models.