Portland State University. Systems Science Ph. D. Program
Date of Award
Doctor of Philosophy (Ph.D.) in Systems Science
1 online resource (xi, 312 pages)
This project explores two statistical methods that infer social network structures and statistically test those structures for change over time: regression-based differential network analysis (R-DNA) and information theory-based differential analysis (I-DNA). R-DNA is adapted from bioinformatics and I-DNA employs reconstructability analysis.
This project applies both R-DNA and I-DNA to analyze Medicaid claims data from one-year periods before (May 2011- Apr 2012) and after (Jan 2013-Dec 2013) the formation of the Health Share of Oregon Coordinated Care Organization (CCO). The formation of CCOs was legislated by the state of Oregon in 2012 with the triple aim of improving health outcomes, reducing cost, and increasing patient satisfaction. We hypothesize that Health Share's CCO formation would be followed by several changes in the healthcare delivery network.
Our primary aim is to contribute methodologically to the field of social network analysis by demonstrating and comparing these methods' capacity for network inference and statistical testing. Our secondary aim is to contribute substantively to the field of health policy by identifying changes in the healthcare delivery network that followed Health Share's CCO formation.
Application of R-DNA and I-DNA to claims data involves three steps: (a) the inference of billing provider networks, (b) measurement of a "distance" between networks before and after Health Share's CCO formation, and (c) statistical testing of this distance by resampling. Both methods afford what is akin to a network-level t test for significant network difference between two time periods.
We use R-DNA and I-DNA to analyze three different billing networks: (1) the full network, including all 1,298 billing providers who operated within Health Share's region and were actively billing during both time periods, (2) the network of the top 30 billing providers by patient volume, and (3) a care sector network defined in terms of primary, specialty, ancillary, mental/behavioral, facility, and "other" billing provider types.
To address our primary aim, we demonstrate how four standard methods of data analysis can be used for network inference -- correlation, multiple regression, partial least squares regression, and reconstructability analysis -- and how three methods of data resampling can be used for statistical testing -- permutation, one-sample bootstrapping, and two-sample bootstrapping. We conduct comparisons between these variants of R-DNA and I-DNA and make recommendations for how one might select from them in future social network analysis projects.
In terms of our secondary aim, applications of R-DNA and I-DNA reveal complementary insights and three specific patterns, all of which differ from our hypotheses. First, network connectivity consolidated after Health Share's formation, producing stronger connections within the network's core and weaker ones throughout its periphery. Changes in connectivity between primary and specialty care sectors reveal an increase in patients who received neither type of care, which may indicate increased obstacles to access. Changes between primary and mental/behavioral care sectors reveal a decrease in the number of patients receiving both types of care, which may suggest that referrals did not increase as expected.
We conclude that both I-DNA and R-DNA are useful for inferring social network structures, for descriptively exploring the types of change that occur in them over time, and for testing whether those changes are statistically significant.
Schmidt, Teresa Danielle, "Statistical Analysis of Social Network Change" (2019). Dissertations and Theses. Paper 5415.