Innovative Data Integration Models for Handling Evidence Inconsistency in AD/ADRD Research

PROJECT SUMMARY This R03 proposal, responding to PAR-23-179, aims to advance evidence-based practice by developing innovative methods and tools for integrating data from AD/ADRD studies. Our focus is on addressing challenges posed by inconsistencies among evidence from different studies, specifically for comparative effectiveness research in AD/ADRD. Although network meta-analyses have become popular for synthesizing multiple intervention options, their effectiveness largely depends on the assumption that direct and indirect evidence is consistent. However, studies on AD/ADRD can vary significantly in patient population characteristics such as gender and age, making evidence consistency questionable. Summary statistics of moderators are often available from published articles, and network meta-regression incorporating these covariates can partially explain potential evidence inconsistency. Nonetheless, many AD/ADRD studies have small sample sizes due to recruitment challenges and their longitudinal nature, leading to large uncertainties in study-specific covariate summaries. Such measurement errors could bias conclusions from network meta-regressions. This project has two main aims to address these research gaps. The first aim is to develop novel approaches to modeling evidence inconsistency in network meta-analyses with applications to AD/ADRD research. The proposed method will use mixture distributions to model potential evidence inconsistency while remaining parsimonious to avoid overcomplicating the model with too many nuisance parameters corresponding to clinically ignorable inconsistency. This approach will allow for a detailed examination of whether direct and indirect evidence can be combined. The second aim is to develop novel models for network meta-regressions that account for uncertainties in study-specific summaries of moderators with applications to AD/ADRD research. The new method will account for full uncertainties in summary statistics of moderators (e.g., mean ages, proportions of genders, races) extracted from published AD/ADRD studies, thereby correcting potential dilution biases caused by measurement errors. Additionally, we will create a user-friendly, open-source software package to enable applied scientists to accurately use the new models for systematic reviews comparing multiple AD/ADRD interventions. This software will include comprehensive guides and worked examples. By leveraging the collective expertise of our transdisciplinary team in both biostatistics and clinical research, this project aims to significantly advance analytical frameworks in AD/ADRD research. The success of this project will influence the design and prediction of interventions within AD/ADRD and have broad applicability to various other clinical research domains, contributing substantially to advancing evidence-based comparative medicine.