2016-10-21

Elly Kaizar, Ohio State University

"Multiple imputation to construct a data bank from a collection of studies"

Meta-analysis is commonly used to combine individual patient data from multiple studies in a single analysis with a single inferential goal. However, collections of similar studies provide rich sources of information. Constructing a data bank opens the door to suites of inquiries that would otherwise be impossible due to the sample size and relatively narrow focus of each individual study. One notable challenge in constructing a data bank is overcoming systematic missingness where different studies use different scales to measure similar variables. Addressing this type of missingness, especially for explanatory variables, is atypical in single-study analysis and there are many open questions. We propose using multiple imputation to address systematic missingness and create a user-friendly data bank, and study two common approaches to multiple imputation: joint models and chained equations. While single-study versions of these methods for normally distributed data may be equivalent, meta-analytic versions diverge. We highlight settings where data bank creators must weigh algorithmic ease against proper imputation and demonstrate the methods for a collection of studies of pediatric traumatic brain injury.

Stay connected TwitterFacebookLinkedInYouTubeInstagram