Nonparametric Preprocessing, Systematic Sensitivity, and Minimum Uncertainty

• Activity Number: 374
• Type: Invited
• Date/Time: Tuesday, July 31, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Risk Analysis
• Author(s): Robert L. Obenchain, PhD, FASA*+
• Companies: Risk Benefit Statistics LLC
• Keywords: comparative effectiveness research ; heterogeneous effect size estimates ; nonparametric preprocessing ; systematic sensitivity ; minimum uncertainty ; local control

Abstract

Global, parametric models typically make strong assumptions that can be wrong or misleading. Due to numerous sources of bias in observational data, such models tend to grossly underestimate the true uncertainty in Comparative Effectiveness Research using massive datasets of medical records or administrative claims. Local Control (LC) is a form of nonparametric preprocessing that makes robust head-to-head treatment comparisons. LC starts by dividing patients into many subgroups (blocks) in any way that assures patients are well-matched on pre-treatment covariates within subgroups. A Local Treatment Difference, LTD = (Avg. Outcome on Treatment) minus (Avg. Outcome on Control), is then computed within each subgroup that contains both treated and control patients. The resulting LTD distribution quantifies patient heterogeneous response to treatment and provides an objective basis for individualized medicine. Sensitivity analyses can vary choice of patient covariates, number of subgroups formed and method of patient matching/clustering. Dynamic displays of LTD histograms then reveal stability or instability in effect-size distributions and realistically quantify minimal uncertainty.