1010 — Different Methods, Different Results: Comparing Logistic Regression and QCA Findings From the Same VA Dataset
Lead/Presenter: Edward Miech, COIN - Indianapolis
All Authors: Miech EJ (Indianapolis COIN/PRIS-M QUERI)
Myers LM (Indianapolis COIN/PRIS-M QUERI)
Baye F (Indiana University School of Public Health)
Bravata DM (Indianapolis COIN/PRIS-M QUERI)
This analysis was part of a larger study that conducted the first national assessment of care quality for patients with TIA or minor stroke across the VHA. The objective was to use Qualitative Comparative Analysis (QCA) to identify processes of care associated with reduced risk of death at 365 days among Veterans with a TIA or minor stroke, and then compare those results with the findings from logistic regression analyses conducted independently using the same VA dataset. QCA is a mathematical, inductive and case-oriented method for examining multifactorial causality, when several factors together may be jointly necessary and/or sufficient for an outcome. QCA uses Boolean algebra and set theory to identify configurations of factors that connect directly to an outcome, then systematically removes all redundant factors to determine the crucial difference-makers.
QCA and logistic regression analyses were conducted on the same VA dataset. This dataset included 8107 patients and examined 30 different processes of care as part of a national assessment of care quality for patients with TIA or minor stroke across the VA in FY11. Logistic regression analyses were performed using SAS version 9.2; QCA analyses were performed in R 3.3.1 using the QCApro package. Analyses were conducted at the patient level.
An adjusted multivariate logistic regression model identified eight different processes of care associated with reduced risk of death at 365 days. QCA analyses, using the same dataset, yielded a two-factor solution where Veterans who received two specific processes of care in the VA (carotid artery imaging and antithrombotics at discharge) had a significantly lower death rate at 365 days: 7.6% (326/4299) vs 12.1% (401/3329) (Chi-square statistic = 3.33; p < .0001).
Using an identical dataset, QCA analysis identified a multifactorial solution not found using traditional regression-based methods, with direct implications for implementation.
QCA offers both VA research and operations a set-theoretic approach for evaluating how specific configurations of factors connect to outcomes. Fundamentally different than regression-based methods both in terms of its math (Boolean vs linear algebra) and its search target (necessary and sufficient conditions vs correlation), QCA analyses can yield fresh, evidence-based insights into complex phenomena.