Methods for a longitudinal quantitative outcome with a multivariate Gaussian distribution multi-dimensionally censored by therapeutic intervention

Methods to estimate underlying blood pressure: The Atherosclerosis Risk in Communities (ARIC) Study

Antihypertensive medications complicate studies of blood pressure (BP) natural history; BP if untreated (“underlying BP”) needs to be estimated. Our objectives were to compare validity of five missing data imputation methods to estimate underlying BP and longitudinal associations of underlying BP and age. We simulated BP treatment in untreated hypertensive participants from Atherosclerosis Risk in Communities (ARIC) in visits 1–5 (1987–2013) using matched treated hypertensive participants. The underlying BP was imputed: #1, set as missing; #2, add 10 mmHg for systolic, 5 mmHg for diastolic; #3, add medication class-specific constant; #4, truncated normal regression; and #5, truncated normal regression including prior visit data. Longitudinal associations were estimated using linear mixed models of imputed underlying BP for simulated treated and measured BP for untreated participants. Method 3 was the best-performing for systolic BP; lowest relative bias (5.3% for intercept at age 50, 0% for age coefficient) and average deviation from expected (0.04 to -1.79). Method 2 performed best for diastolic BP; lowest relative bias (0.6% intercept at age 50, 33.3% age <60, 9.1% age 60+) and average deviation (-1.25 to -1.68). Methods 4 and 5 were comparable or slightly inferior. In conclusion, constant addition methods yielded valid and precise underlying BP and longitudinal associations.