Systematic reviews and meta-analyses of diagnostic test accuracy usually pool sensitivity and specificity estimates. However, for clinical/patient-care purposes, positive and negative predictive values (PPV and NPV) are arguably more useful.
Is including PPV and NPV as meta-analytic pooling outcomes sensible? One potential argument against this that I can think of is that these measures are influenced by disease prevalence in the studies that report them (unlike sensitivity and specificity). However, a potential counter-argument to that is some sort of stratification by pre-specified prevalence ranges can be performed.
You are right that the prevalence of a condition in the tested population will have an important influence on the +PPV and –PPV. However, you can calculate another informative statistic – the clinical likelihood ratio (positive and negative). This is the effect of the test on the likelihood of the person having the condition. It simply compares the odds of the condition pre and post test. A positive test with a clinical likelihood ratio of 10, for example, means that the condition is ten times more likely to be present if the test is positive. And a negative likelihood ratio of 0·1 means that the test is only a tenth as likely to be present if the test is negative.
Likelihood ratios are based on odds, so they generalise across prevalences in a way that predictive values don't. This might help you deal with the prevalence effects.
Interesting question. The direct answer is, it would make sense if the pre-test probability (prevalence) is similar among the studies, which is usually not the case. Otherwise, such explained heterogeneity would preclude meaningful meta-analysis.
- Badges
- Science topic
- Similar topics
- Social Psychology
- Attitude