A Bayesian Analysis With Informative Prior on Disease Prevalence for Predicting Missing Values Due To Verification Bias

AIM: Verification bias is one of the major problems encountered in diagnostic accuracy studies. It occurs when a standard test performed on a non-representative subsample of subjects which have undergone the diagnostic test. In this study we extend a Bayesian model to correct this bias. METHODS: The study population is patients that have undergone at least two repeated failed IVF/ICSI (in vitro fertilization/intra cytoplasmic sperm injection) cycles. Patients were screened using ultrasonography and those with polyps were recommended for hysteroscopy. A Bayesian modeling was applied on mechanism of missing data using an informative prior on disease prevalence. The parameters of the model were estimated through Markov Chain Monte Carlo methods. RESULTS: A total of 238 patients were screened, 47 of which had polyps. Those with polyps were strongly recommended to undergo hysteroscopy, 47/47 decide to have a hysteroscopy and in 37/47 polyps confirmed. None of the 191 patients with no polyps detected in ultrasonography underwent a hysteroscopy. A model using Bayesian approach was applied with informative prior on polyp prevalence. False and true negatives were estimated in the Bayesian framework. The false negative was obtained 14 and 177 true negatives were obtained, so sensitivity and specificity was estimated easily after estimating the missing data. Sensitivity and specificity were equal to 74% and 94% respectively. CONCLUSION: Bayesian analyses with informative prior seem to be powerful tools in the simulation of experimental space.