|
Abstract:
|
When validating a risk model in an external cohort, often not all predictor variables in the model are available on all cohort members. Missingness can be random or by design (e.g. are available only in case-cohort and nested case-control samples). Weighting methods and imputation can address the missingness problem. We propose methods to improve weighting approaches. We first create a pseudorisk model using information from variables available for the entire cohort. Then we modify known sampling weights by survey calibration so that the weighted sum of the pseudo-risk in the complete data equals its cohort sum. We also study poststratification of the weights, based on pseudo-risk. We compared observed (O) to expected (E) counts in simulations to assess calibration and relative efficiencies of the methods, compared to analyses assuming complete exposure information. We also compare them to using multiple imputation by chained equations (MICE). In our simulations, all weighting and imputation procedures for O/E were unbiased. Survey calibration and poststratification were more efficient than standard weighting, and often more efficient than imputation.
|
