Paul L Hebert1, Leslie T Taylor, Jason J Wang, Margo A Bergman. 1. Department of Veterans Affairs, Northwest Center for Outcomes Research in Older Adults, VA Puget Sound Health Care System, Seattle, WA, USA. Paul.Hebert2@va.gov
Abstract
OBJECTIVE: To describe a method for imputing missing follow-up blood pressure data in a clinical hypertension trial using blood pressures abstracted from medical charts. METHODS: We tested a two-step method. In the first, a longitudinal mixed-effects model was estimated on blood pressures abstracted from medical charts. In the second, the patient-specific fitted values from this model at follow-up were used to impute blood pressures missing at follow-up in the trial. Simulations that imposed alternative missing data mechanisms on observed trial data were used to compare this approach to imputation approaches that do not incorporate data from charts. RESULTS: For data that are missing at random, incorporating the fitted values from chart-based longitudinal models leads to estimates of the trial-based blood pressures that are unbiased and have lower mean squared deviation than do blood pressures imputed without the chart-based data. For data that are missing not at random, incorporating fitted values ameliorates but does not eliminate the inherent missing data bias. CONCLUSIONS: Incorporating chart data into an imputation algorithm via the use of longitudinal mixed-effects model is an efficient way to impute longitudinal data that are missing from a randomized trial.
OBJECTIVE: To describe a method for imputing missing follow-up blood pressure data in a clinical hypertension trial using blood pressures abstracted from medical charts. METHODS: We tested a two-step method. In the first, a longitudinal mixed-effects model was estimated on blood pressures abstracted from medical charts. In the second, the patient-specific fitted values from this model at follow-up were used to impute blood pressures missing at follow-up in the trial. Simulations that imposed alternative missing data mechanisms on observed trial data were used to compare this approach to imputation approaches that do not incorporate data from charts. RESULTS: For data that are missing at random, incorporating the fitted values from chart-based longitudinal models leads to estimates of the trial-based blood pressures that are unbiased and have lower mean squared deviation than do blood pressures imputed without the chart-based data. For data that are missing not at random, incorporating fitted values ameliorates but does not eliminate the inherent missing data bias. CONCLUSIONS: Incorporating chart data into an imputation algorithm via the use of longitudinal mixed-effects model is an efficient way to impute longitudinal data that are missing from a randomized trial.