Using Imputed Untreated Cholesterol in a Propensity Score Model to Reduce Confounding by Indication
Neal Jorgensen, STAT-DD-MS Candidate, Colorado State University
Tuesday, November 9, 2010
11:00 a.m., room 008, Statistics Bldg
Studying the effects of medications on endpoints in an observational setting can lead to challenging statistical problems due to confounding by indication. Participants taking a medication (e.g. statins) will often have a worse risk factor profile, and have worse underlying values of the biomarker which the medication is designed to improve (e.g. lipid levels). This is known as confounding by indication. Traditional methods to account for this rely on the assumption of no unmeasured confounders. In observational studies at baseline we may have many participants taking medication already, and have only their on-treatment value of such biomarkers. We propose a two step approach to reduce confounding by indication: The first step is to consider the untreated value as missing data and to impute this value as a function of the observed treated value, dose and type of medication, and other participant characteristics. The second step constructs a propensity-score which involves modeling the probability of medication use as a function of measured covariates, and the estimated underlying untreated biomarker value. The resulting score can then be used to adjust for the imbalances in the covariates and to reduce indication bias. We illustrate these techniques in models relating statin use to the risk of incident coronary heart disease.
Dr. Rui Song, Advisor, Statistics Department, CSU
Dr. Jennifer Hoeting, Committee Member, Statistics Department, CSU
Dr. Annette Bachand, Outside Member, Env and Rad Health Sciences, CSU