Estimating predictive values for blood pressure measurements from multivariate regression models with covariates.

Predictive values are useful in estimating the probability distribution of a 'true' or underlying measurement, that is, without measurement error or within-person variability. They have been applied to blood pressure data to estimate the true probability that a person is hypertensive currently, or that he/she will become hypertensive based on previous data from childhood. The current work extends these results to situations where covariates are of interest. One can use multivariate regression models to model predictive values for future levels as functions of covariates as well as current measured levels. I compare predictive value estimates obtained from these models to those obtained from ordinary linear regression and from logistic regression with use of data on childhood blood pressure from East Boston, MA. Estimates obtained using the multivariate model are preferable either in terms of bias in the estimates themselves or in terms of their variability. This is particularly true with covariates included in the model. The difference between the multivariate and ordinary regression estimates depends on the conditional reliability of future levels given current blood pressure levels and covariates. I also discuss predictive value estimates for true current level given observed level as well as covariates. These also depend on the reliability of the current measure given values of covariates.