The R Book

Time Series Analysis in Mixed-Effects Models

It is common to have repeated measures on subjects in observational studies, where we would expect that the observation on an individual at time t + 1 would be quite strongly correlated with the observation on the same individual at time t. This contravenes one of the central assumptions of mixed-effects models (p. 627), that the within-group errors are independent. However, we often observe significant serial correlation in data such as these.

The following example comes from Pinheiro and Bates (2000) and forms part of the nlme library. The data refer to the numbers of ovaries observed in repeated measures on 11 mares (their oestrus cycles have been scaled such that ovulation occurred at time 0 and at time 1). The issue is how best to model the correlation structure of the data. We know from previous work that the fixed effect can be modelled as a three-parameter sine–cosine function

and we want to assess different structures for modelling the within-class correlation structure. The dataframe is of class groupedData which makes the plotting and error checking much simpler.

library(nlme)
library(lattice)
data(Ovary)
attach(Ovary)
names(Ovary)

[1]  "Mare"  "Time"  "follicles"
plot(Ovary)

The panel plot has ranked the horses from bottom left to top right on the basis of their mean number of ovules (mare 4 with the lowest number, mare 8 with ...