STATISTICS IN MEDICINE Statist. Med. 2001; 20:3789–3805 (DOI: 10.1002/sim.1172) Analysis of ambulatory blood pressure monitor data using a hierarchical model incorporating restricted cubic splines and heterogeneous within-subject variances ‡ Paul C. Lambert 1; * , Keith R. Abrams 1 , David R. Jones 1 , Aidan W. F. Halligan 2 and Andrew Shennan 3 1 Department of Epidemiology and Public Health; University of Leicester; 22-28 Princess Road West; Leicester LE1 6TP UK. 2 Department of Obstetrics and Gynaecology; University of Leicester; U.K. 3 Fetal Health Research Group; St Thomas’ Hospital; GKT Kings College; U.K. SUMMARY Hypertensive disorders of pregnancy are associated with signicant maternal and foetal morbidity. Mea- surement of blood pressure remains the standard way of identifying individuals at risk. There is growing interest in the use of ambulatory blood pressure monitors (ABPM), which can record an individual’s blood pressure many times over a 24-hour period. From a clinical perspective interest lies in the shape of the blood pressure prole over a 24-hour period and any dierences in the prole between groups. We propose a two-level hierarchical linear model incorporating all ABPM data into a single model. We contrast a classical approach with a Bayesian approach using the results of a study of 206 pregnant women who were asked to wear an ABPM for 24 hours after referral to an obstetric day unit with high blood pressure. As the main interest lies in the shape of the prole, we use restricted cubic splines to model the mean proles. The use of restricted cubic splines provides a exible way to model the mean proles and to make comparisons between groups. From examining the data and the t of the model it is apparent that there were heterogeneous within-subject variances in that some women tend to have more variable blood pressure than others. Within the Bayesian framework it is relatively easy to incorporate a random eect to model the between-subject variation in the within-subject variances. Although there is substantial heterogeneity in the within-subject variances, allowing for this in the model has surprisingly little impact on the estimates of the mean proles or their condence/credible intervals. We thus demonstrate a powerful method for analysis of ABPM data and also demonstrate how heterogeneous within-subject variances can be modelled from a Bayesian perspective. Copyright ? 2001 John Wiley & Sons, Ltd. * Correspondence to: Paul C. Lambert, Department of Epidemiology and Public Health, University of Leicester, 22-28 Princess Road West, Leicester LE1 6TP, U.K. ‡ Presented at the International Society for Clinical Biostatistics Twenty-rst International Meeting, Trento, Italy, September 2000. Received January 2001 Copyright ? 2001 John Wiley & Sons, Ltd. Accepted October 2001