Analysis of longitudinally observed irregularly timed multivariate outcomes: regression with focus on cross-component correlation.
Components of repeatedly observed multivariate outcomes (for example, the two components of blood pressure measures (SBP(it), DBP(it)), obtained on subject i at arbitrarily spaced times t) are often analysed separately. We present a unified approach to regression analysis of such irregularly timed multivariate longitudinal data, with particular attention to assessment of the magnitude and durability of cross-component correlation. Maximum likelihood estimates are presented for component-specific regression parameters and autocorrelation and cross-correlation functions. The component-specific autocorrelation function has the 'damped exponential' form [see text], which generalizes the AR(1), MA(1) and random intercept models for univariate longitudinal outcomes. The cross-component correlation function (CCCF) has an analogous form, allowing damped-exponential decay of cross-component correlation as time between repeated measures elapses. Finite sample performance is assessed through simulation studies. The methods are illustrated through blood pressure modelling and construction of multivariate prediction regions.