84 chapter 2 of most studies, the reference test was considered positive only if the histopathological size and location of the mPTC was consistent with pre-operative imaging of the thyroid nodule. All other mPTC were considered incidental findings (negative reference test) [165, 170, 174, 217]. Secondly, some studies applied the intermediate histopathological diagnosis of ‘(follicular) tumour of unknown malignant potential’ (TUMP or FT-UMP). In accordance with most studies, we considered these lesions malignant [109, 202, 215, 309]. Statistical analysis Disease prevalence, index test sensitivity and specificity for individual studies were calculated using the classic equations. Disease prevalence was calculated as the number of histopathologically proven malignancies divided by all indeterminate thyroid nodules included in the study, to prevent overestimation of the prevalence by counting only nodules with available histopathology as the denominator. The percentage of indeterminate nodules with available histopathology was calculated as the number of nodules with available histopathology divided by all indeterminate nodules with a conclusive index test. The pooled sensitivity, specificity, positive and negative likelihood ratio (LR) and their corresponding 95% confidence intervals (95% CI) were estimated using the metandi and midas commands in Stata/MP [347, 348]. The metandi command uses a twolevel mixed logistic regression model with independent binomial distributions for the true positives and true negatives conditional on the sensitivity and specificity in each study, and a bivariate normal model for the logit transforms of sensitivity and specificity between studies [348]. The user-written midas command uses a bivariate mixed-effects binary regression model to estimate pooled test performance parameters [347]. Pooled results are presented in forest plots and summary receiver operating characteristic (SROC) plots, including area under the SROC curve (AUC). Positive predictive value (PPV) and negative predictive value (NPV) are not static test performance parameters, but depend on the pre-test probability (i.e., prevalence) of disease. Therefore, Bayesian conditional probability plots are presented displaying the post-test probabilities versus prior probabilities using the estimated summary likelihood ratios. Based on realistically occurring variance in prevalence as seen in clinical practice, PPV and NPV are featured for populations with a 15%, 25% and 40% prevalence of malignancy. If available data allowed, additional subgroup analyses were performed for Bethesda III and Bethesda IV nodules, including equivalent categories. Heterogeneity across studies was assessed by visual inspection of forests plots and quantified using the I2 statistics. Risk of publication bias was visually assessed using funnel plots and quantified using an Egger’s test. The metandi and midas commands are only applicable to a minimum of four studies. For pooling of fewer studies we used the Stata/MP metaprop command and random effects modelling to estimate pooled sensitivity, specificity, positive and negative LR. Metaprop provides no more than two significant figures. Moreover, several studies with a small sample size technically limited the
RkJQdWJsaXNoZXIy MTk4NDMw