Improving Implementation and QI Research with Regression Risk Analysis

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): Quality Improvement and Implementation Research (QI Research) is most valuable when its results can be reported using measures that are unbiased, intuitive and unambiguous. Unfortunately logistic regression, the most common method for adjusting for confounding when outcomes are binary (e.g. dead vs. alive, sick vs. healthy, etc) generally yields an adjusted odds ratio, a measure that is neither intuitive nor an unbiased estimator of the more desirable adjusted risk ratio. The adjusted odds ratio overstates the magnitude of impact compared to an adjusted risk ratio, especially when outcomes are not rare. Often the adjusted risk difference would be of independent interest. The study investigators have recently published a paper describing Regression Risk Analysis (RRA), an innovative approach that allows researchers to accurately estimate adjusted risk ratios and adjusted risk differences from nonlinear models like logistic regression. RRA represents a significant advance over current methods. This project will extend RRA to make it more useful for study designs frequently encountered in QI research, such as complex survey samples, clustering, multinomial regression (i.e., outcomes are in more than 2 categories), and interactions between variables. By allowing logistic regression to be reported in terms of the adjusted risk ratio and the adjusted risk difference, RRA will enhance researchers' capacity to report findings that are intuitive, actionable, and interpretable. Consumers of research, such as administrators and policy makers will also benefit. This project will extend regression risk analysis to enhance its relevance for QI research, including to: 1. Refine the method of estimating risk measures (ratios and differences) and their standard errors from logistic regression to account for: Complex sample designs, including stratification, clustering, and equal or disproportionate weighting; Interactions (effect modification) between variables; and Multinomial regression. 2. Validate the estimates using Monte Carlo simulations; and, 3. Develop both SAS and STATA code with teaching examples to make these techniques accessible to typical health services/quality improvement researchers. We will use the internet to make these methods and our computer code widely available. This project will add a significant and powerful method to the tool box of QI research methods. At the conclusion of this research project, QI researchers who use logistic regression will have access to regression risk analysis, improving their capacity to communicate interpretable and actionable results to the consumers of their research. In turn, translation of research will benefit from evidence being described in terms of both relative (risk ratio), and absolute (risk difference) measures that are unbiased and intuitive. PUBLIC HEALTH RELEVANCE: The applicants recently described how to improve logistic regression by estimating two unbiased and intuitive measures, adjusted risk ratio (ARR) and adjusted risk difference (ARD), instead of the less intuitive adjusted odds ratio, which always exaggerates effect size. They intend to extend their method to handle complex survey design, clusters, multinomial outcomes, and interactions, each of which is often encountered in real-world implementation and quality improvement research. They will develop and validate these extensions, and create and disseminate user-friendly software, thus enhancing the capacity of QI researchers to translate their results.