Abstract:
|
In cohort studies, working with certain data elements may cost much money, time or material. Using all available data may not be feasible. Cox proportional hazards (CPH) and nested case control (NCC) models may be used on cohort incidence data, and share the same partial loglikelihood. CPH models use the full cohort; NCC studies use fewer subjects, but are less efficient. For main effects, NCC's relative efficiency is (m-1)/m when each case has m matched controls. For interactions and/or unequal m, efficiency drops, but by how much is not known. Additionally, selected CPH and NCC models may differ. Here, we explore how efficiency and model selection differ in an atrial fibrillation incidence study among atomic bomb survivors from the Adult Health Study followed for 40+ years. Reviewing paper records of 10,000+ subjects cost much time. However, electronically available data was used for matching on key factors. A NCC study would have saved 40% of costs. We compare efficiency of NCC to CPH for estimating main effects of radiation and interactions of radiation with other key factors. We also examine how model selection differs between the two possible analyses.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.