|
Abstract:
|
Recent medical advances have motivated the development of personalized treatment decision rules (TDRs) that incorporate complex data (e.g. imaging). Given that this data is costly and time consuming to collect, we aim to develop optimal TDRs that reduce unnecessary testing while remaining valuable for physicians and patients. For any given TDR, there is an associated degree of uncertainty that we aim to quantify via a confidence measure. We partition a new patient's data, which can be of multiple types (e.g. count, categorical), into observed and unobserved components, based on what patient-specific measures are collected. We propose estimating confidence through repeated imputations of a patient's unobserved covariates, using the data from which the TDR is estimated and the patient's observed data. The TDR is applied to each imputed patient dataset and the probability the patient will be assigned to each treatment is estimated. Using imputation by chained equations, we study this approach through extensive simulation studies. A theoretical reasoning for imputation in a simple case is presented and recommendations for practical use are provided.
|
