Statistical methods for the meta-analysis of diagnostic tests must take into account the use of surrogate standards

Evaluating the performance of a new diagnostic test presents a challenge if the conventional "gold" standard is invasive, hazardous, or expensive, especially if that test has been supplanted in usual clinical practice by a "silver" standard test that is more acceptable and perhaps only slightly suboptimal. In such a case, a systematic literature review will typically uncover a mix of study types, some using the gold and some the silver.

We sought to develop and compare statistical methods to account for this kind of heterogeneity in performing a meta-analysis.

We compared the performance of estimation methods based on generalized mixed models which incorporate heterogeneity, especially choice of reference test, and random between-study variation in sensitivity and specificity with more conventional methods which neglect the differences in reference tests. Computer simulations were conducted to assess bias and root mean square error of point estimates and coverage of interval estimates.

Methods ignoring the difference in reference tests severely underestimated sensitivity and specificity under the assumption of conditional independence. Bias was substantial even for references with small departure from the standard and persisted with increasing sample size. Coverage of interval estimates was far from nominal level.

In the presence of varying reference tests, avoidance of bias and invalid confidence intervals for diagnostic performance requires applying a model that accounts for differences in reference test and heterogeneity among studies.