A case study of an individual participant data meta-analysis of diagnostic accuracy showed that prediction regions represented heterogeneity well

TY - JOUR

T1 - A case study of an individual participant data meta-analysis of diagnostic accuracy showed that prediction regions represented heterogeneity well

AU - López Malo Vázquez de Lara, Aurelio

AU - Bhandari, Parash Mani

AU - Wu, Yin

AU - Levis, Brooke

AU - Thombs, Brett

AU - Benedetti, Andrea

AU - DEPRESsion Screening Data (DEPRESSD) PHQ-9 Collaboration

AU - Härter, Martin

AU - Löwe, Bernd

N1 - © 2023. The Author(s).

PY - 2023/6/7

Y1 - 2023/6/7

N2 - The diagnostic accuracy of a screening tool is often characterized by its sensitivity and specificity. An analysis of these measures must consider their intrinsic correlation. In the context of an individual participant data meta-analysis, heterogeneity is one of the main components of the analysis. When using a random-effects meta-analytic model, prediction regions provide deeper insight into the effect of heterogeneity on the variability of estimated accuracy measures across the entire studied population, not just the average. This study aimed to investigate heterogeneity via prediction regions in an individual participant data meta-analysis of the sensitivity and specificity of the Patient Health Questionnaire-9 for screening to detect major depression. From the total number of studies in the pool, four dates were selected containing roughly 25%, 50%, 75% and 100% of the total number of participants. A bivariate random-effects model was fitted to studies up to and including each of these dates to jointly estimate sensitivity and specificity. Two-dimensional prediction regions were plotted in ROC-space. Subgroup analyses were carried out on sex and age, regardless of the date of the study. The dataset comprised 17,436 participants from 58 primary studies of which 2322 (13.3%) presented cases of major depression. Point estimates of sensitivity and specificity did not differ importantly as more studies were added to the model. However, correlation of the measures increased. As expected, standard errors of the logit pooled TPR and FPR consistently decreased as more studies were used, while standard deviations of the random-effects did not decrease monotonically. Subgroup analysis by sex did not reveal important contributions for observed heterogeneity; however, the shape of the prediction regions differed. Subgroup analysis by age did not reveal meaningful contributions to the heterogeneity and the prediction regions were similar in shape. Prediction intervals and regions reveal previously unseen trends in a dataset. In the context of a meta-analysis of diagnostic test accuracy, prediction regions can display the range of accuracy measures in different populations and settings.

AB - The diagnostic accuracy of a screening tool is often characterized by its sensitivity and specificity. An analysis of these measures must consider their intrinsic correlation. In the context of an individual participant data meta-analysis, heterogeneity is one of the main components of the analysis. When using a random-effects meta-analytic model, prediction regions provide deeper insight into the effect of heterogeneity on the variability of estimated accuracy measures across the entire studied population, not just the average. This study aimed to investigate heterogeneity via prediction regions in an individual participant data meta-analysis of the sensitivity and specificity of the Patient Health Questionnaire-9 for screening to detect major depression. From the total number of studies in the pool, four dates were selected containing roughly 25%, 50%, 75% and 100% of the total number of participants. A bivariate random-effects model was fitted to studies up to and including each of these dates to jointly estimate sensitivity and specificity. Two-dimensional prediction regions were plotted in ROC-space. Subgroup analyses were carried out on sex and age, regardless of the date of the study. The dataset comprised 17,436 participants from 58 primary studies of which 2322 (13.3%) presented cases of major depression. Point estimates of sensitivity and specificity did not differ importantly as more studies were added to the model. However, correlation of the measures increased. As expected, standard errors of the logit pooled TPR and FPR consistently decreased as more studies were used, while standard deviations of the random-effects did not decrease monotonically. Subgroup analysis by sex did not reveal important contributions for observed heterogeneity; however, the shape of the prediction regions differed. Subgroup analysis by age did not reveal meaningful contributions to the heterogeneity and the prediction regions were similar in shape. Prediction intervals and regions reveal previously unseen trends in a dataset. In the context of a meta-analysis of diagnostic test accuracy, prediction regions can display the range of accuracy measures in different populations and settings.

KW - Humans

KW - Sensitivity and Specificity

KW - Depressive Disorder, Major/diagnosis

KW - Seizures

KW - Data Collection

U2 - 10.1038/s41598-023-36129-w

DO - 10.1038/s41598-023-36129-w

M3 - SCORING: Journal article

C2 - 37286580

VL - 13

SP - 9275

JO - SCI REP-UK

JF - SCI REP-UK

SN - 2045-2322

IS - 1

ER -