WorksMathematics
Is It Worth Collecting Additional Covariates?
Abstract
In many studies, basic covariates are routinely recorded at negligible cost, whereas acquiring additional covariates requires expensive or time-consuming procedures. Thus, a natural question is whether it is worth to collecting these additional covariates. Using the reduction in prediction interval width as criterion, we evaluate and test the gain brought by the additional covariates and hence answer this question. We consider the gain both marginally and conditionally on the observable characteristics. The conditional formulation captures subpopulation heterogeneity and supports individualized data acquisition decisions. To avoid restrictive parametric assumptions, we adopt a model-free framework that accommodates flexible nonparametric or machine learning methods. We propose an efficient influence function based method, with proven theoretical properties to achieve our purpose under mild nuisance convergence rates. Interestingly, we find the natural conformal prediction based method cannot achieve the same purpose: It fails in the marginal case and imposes much more stringent or even infeasible requirement on the nuisance parameter estimation procedures. We demonstrate the superior finite-sample performance of the proposed procedures through simulations and real data applications.