PRIME-IPD: A systematic method to prepare Individual participant data for systematic reviews and meta-analyses

Article type
Authors
Riddle A1, Dewidar O1, Ghogomu E1, Matthew C1, Hussein A2, Trawin J3, Gaffy M4, Cousens S5, Bhutta ZA6, Tugwell P7, Wells G2, Welch V1
1Bruyere Research Institute, University of Ottawa
2Cardiovascular Research Methods, University of Ottawa Heart Institute
3Research Assistant at Women's Health Research Institute, University of British Columbia
4Hospital for Sick Children, University of Toronto
5Department of Infectious Disease Epidemiology, London School of Hygiene and Tropical Medicine (LSHTM)
6Centre for Global Child Health, The Hospital for Sick Children
7Ottawa Hospital Research Institute, Clinical Epidemiology Program
Abstract
Background: Individual participant data (IPD) meta-analysis (MA) enables nuanced effect modification analyses and may help standardize outcome measures across studies. However, there are multiple challenges to preparing IPD for MA, such as differences in outcome variables, differences in data collection methods, incomplete data dictionaries. These obstacles and others result in the consumption of extensive amounts of time and resources to prepare IPD for MA.
Objectives: The purpose of this tool is to describe a systematic approach to preparing data for IPD-MA analysis.
Methods: We reviewed relevant guidance from the Cochrane Handbook, Get Real-IPD working group, Cochrane Multiple Interventions group and other published literature. We developed a five-step approach to preparing IPD, through iterative consultation with the advisory board for an IPD-MA and systematic review.
Results: The five steps are: Processing, Replication, Imputation and Merging and Estimation (PRIME). The processing step verifies that the variables of interest are available in the original datasets, identifies missing values and relabels all variables of interests across the datasets to have common variable names. The replication step verifies that the processed dataset is consistent with the analyzed dataset in the published papers, using standardized differences. The imputation step involves an algorithm for how to handle datasets with missing values, including multiple imputation and merging of all imputed datasets. The merging step calls for combining all datasets after dealing with missing data. The final estimation step involves calculating any new variables required for analysis, such as categorical variables for gradations of intensity or severity of outcome variables. The outcomes and value of each of these five steps are illustrated using our systematic review and network meta-analysis of deworming for children.
Conclusion: The purpose of this guidance is to standardize the process of preparing data for IPD-MA. This guidance needs to be evaluated by application to other systematic reviews and meta-analyses.
Patient or healthcare consumer involvement: no