Purpose:

Blood pressure is used as surrogate endpoint for stroke and mortality, but there is limited evidence on its validity as surrogate endpoint. The purpose of this study was to explore the validity of blood pressure as surrogate for the risk of stroke and mortality according to published validity criteria.

Method:

We used 9 randomized trials from a recent meta-analysis and extracted data from all treatment arms on blood pressure after one year of treatment and the number of strokes and fatalities at the end of the trials. We extended a model by Daniels and Hughes (1997) and performed a Bayesian meta-regression of the following model:

                         .                   

With (1) we estimated the unknown parameters, , , ,  and . These are the intercepts, regression coefficients and between-trial variances. The observed variables are:,  the logarithms of the relative risk of death and stroke in study i. The standard errors of these relative risks are denoted  and .  The impact on the surrogate endpoint in trial i is denoted, and its standard error is. We evaluated different versions of (1) with respect to model specification for best possible fit.   

The analyses were performed in WinBUGS.

Results:

The model with the best fit according to the DIC (deviance information criterion) was the full model (1) without between-trial variance. When intercepts were omitted, the best model was the one in which stroke and death were assumed to be independent, and with between-trial variance present for stroke and absent for death.

The model indicates an association between the surrogate and both clinical endpoints. A linear model had good fit for the mortality, but not for stroke. The problem with prediction of stroke results stems from a large variation in the stroke outcome that can not be explained by blood pressure alone. This variation was expressed by between-trial variance and treatment effect on the clinical endpoint not captured by the measure of blood pressure in our estimated models.

Conclusion:

Blood pressure may yield valid predictions for mortality but not for stroke. The Bayesian approach allows evaluation of published surrogate validity criteria.