A Shared Parameter Model for Systolic Blood Pressure Accounting for Data Missing Not at Random in the HUNT Study (2203.16602v1)

Abstract: In this work, blood pressure eleven years ahead is modeled using data from a longitudinal population-based health survey, the Trondelag Health (HUNT) Study, while accounting for missing data due to dropout between consecutive surveys (20-50 %). We propose and validate a shared parameter model (SPM) in the Bayesian framework with age, sex, body mass index, and initial blood pressure as explanatory variables. Further, we propose a novel evaluation scheme to assess data missing not at random (MNAR) by comparing the predictive performance of the fitted SPM with and without conditioning on the missing process. The results demonstrate that the SPM is suitable for inference for a dataset of this size (cohort of 64385 participants) and structure. The SPM indicates data MNAR and gives different parameter estimates than a naive model assuming data missing at random. The SPM and naive models are compared based on predictive performance in a validation dataset. The naive model performs slightly better than the SPM for the present participants. This is in accordance with results from a simulation study based on the SPM where we find that the naive model performs better for the present participants, while the SPM performs better for the dropouts.