P1‐555: Comparing Machine Learning to Estimate Propensity Scores and Conventional Approaches for Addressing Confounding in Cognitive Aging Research: an Example With Diabetes and Cognition in the Uk Biobank
Year: 2019
Type: article
Abstract: Random forests are machine learning methods that can provide accurate predictions in the context of severe non-linear associations and interactions. We compared performance of models adjusting for confounding using propensity scores estimated with random forests to models using propensity scores estimated with logistic regression and models using traditional covariate adjustment in data from the UK biobank, where the true confounding structure is unknown, and simulated data, where the true confounding structure is known. Using UK Biobank baseline data (n=86,814), we used random forests to estimate a propensity score (probability of a diabetes) using educational attainment, age, body mass index, and the first 10 genetic principal components as predictors. We then estimated the association of diabetes with verbal reasoning score by adjusting for the random forest-based propensity score, by adjusting for a propensity score obtained using logistic regression incorporating main effects for the same covariates, and by direct adjustment for the same covariates. Confidence intervals and two-sided p-values were obtained using bootstrap. We repeated this process in simulated data generated using a known data generating process. In UK Biobank, conventional propensity and regression adjustment produce consistent coefficient estimates with a mean difference between estimates of the association between diabetes and cognitive change of 0.00, 95% CI (−0.06, 0.06), p=0.94. Propensity adjustment using random forest predictions produced effect estimates further from the null by 0.049, 95% CI (−0.019, 0.12), p=0.14, and 0.050, 95% CI (0.023, 0.077), p<0.01, compared with conventional propensity and regression adjustment, respectively. Using simulated data, propensity adjustment with random forests produces estimates for the effect of diabetes on cognition further from the true null value than conventional approaches estimated with knowledge of the data generating process. However, propensity adjustment using random forests estimation outperformed conventional approaches when non-linear and interaction terms that were part of the data generating process were omitted from the estimation of propensity scores or the set of adjustment covariates in the conventional approaches. Random forests outperformed conventional propensity score modeling only when conventional models are severely mis-specified. If extreme discontinuities and non-linearities are plausible, random forests may be appealing.
Source: Alzheimer s & Dementia
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W2980645015