A Bayesian Synthesis Approach to Data Fusion for Open Source Longitudinal Data

• In Event: 1-005 - Data Integration

Abstract

The exponential growth in the amount of data collected on individuals has created numerous opportunities for methodological developments and examining new theories. Simultaneously, the number of different sources from which this information is collected continues to grow, creating potential obstacles for effectively combining publicly available data before it can be explored. Data fusion is a method for combining information from separate but related datasets-permitting more flexible (and potentially more complex) analyses to be conducted than separate analyses of each individual dataset. Determining appropriate methods for pooling these independent samples from different sources can be challenging, particularly with longitudinal data that varies in the number and timing of assessments.

While most data fusion methods employ a frequentist framework (e.g., Piccinin & Hofer, 2008), Marcoulides (2017) proposed updating information obtained from datasets via a new approach called Bayesian Synthesis. In brief, instead of combining datasets at once as traditionally done in data fusion, Bayesian Synthesis uses information obtained from one dataset to serve as prior information for the analysis of the next data set. This activity continues sequentially until a single posterior distribution is created incorporating all available datasets.

This study utilizes the new Bayesian Synthesis approach to fuse five publicly available cohorts of the Head Start Family and Child Experiences Survey (FACES). Each FACES cohort consists of a nationally representative sample of about three thousand 3 to 4 year-old children enrolled in Head Start programs (referred to as FACES 1997, FACES 2000, FACES 2003, FACES 2006, and FACES 2009). For comparison, obtained parameter estimates are examined and contrasted to those from traditional data fusion. Results indicate that Bayesian Synthesis provides parameter estimates that may lead to insightful data interpretations. To date, data integration methods have proven to be valuable in helping researchers address important questions with more power and accuracy than any single study. Bayesian Synthesis is another beneficial approach that can effectively be used to conduct such activities and enhance the validity of conclusions obtained from the merging of data from different studies. Implications of the findings, limitations, and suggestions for further research are discussed. For example, Bayesian Synthesis has the added benefit of providing an alternative to requiring raw data (essential for traditional data fusion) through the incorporation of prior information from published studies. The summary statistics from a published study can be converted into prior information for the subsequent analysis of the next data set of interest.

Author

• Katerina M. Marcoulides, University of Florida
  Presenting Author