|
Abstract:
|
Follow-up observations for prospective cohort studies are nearly always subject to missingness, with probabilities of response possibly associated with (1) characteristics that do not change over time, which may be collected on all individuals at baseline, and/or (2) the outcome measure of interest, which may not be universally observable. When estimating the association between baseline covariates and the outcome, failing to accurately adjust for limited response may lead to bias. Unfortunately, neither the parametric form of the probability of response is known nor the values of outcomes dictating that probability for all individuals. Using nonparametric identifiability results of a flexible class of logit additive models for response, we describe a novel estimation approach of the association between baseline covariates and outcomes using all available data from the study and estimates of functionals of the outcome at each follow-up time, provided by experts and/or external surveys. Using this method, the probability of response does not require full parameterization and may be applied to cases of intermittent missingness.
|
