Analysis of cross-over designs via growth curve models
Introduction
The simplest example of a cross-over design is the two-sequence two-period cross-over design for comparing two treatments. That is, if the two treatments involved are denoted and , then half of the subjects receive treatment in the first period followed by treatment in the second period, whereas the remaining subjects receive the treatments in the reverse order. This design is denoted AB:BA, where the colon is used to separate the treatment sequences. Higher-order cross-over designs are obtained by including more than two treatments or two periods. Typically, in a standard cross-over design, a response variable is measured once at the end of each period for each subject. A repeated measures cross-over design is a cross-over design, where a sequence of measurements is collected within each period. For example, to compare two test drugs, the concentration of the drug in the blood might be measured every 30 min for each subject during 3 h after administration of the drug.
The aim of this paper is to show that statistical models for a repeated measures cross-over design in a useful way can be defined through the growth curve model (GCM). The GCM (other names of the model are generalized linear model, GMANOVA, multivariate linear normal model, etc.) was introduced by Potthoff and Roy (1964) as a method for analysis of growth curve experiments. An extended version of the GCM was presented by von Rosen (1989). In general, the model applies to longitudinal data in which subjects are followed for a period of time. To date the GCM has only been applied to cross-over designs by Banken (1984), in a canonical version, to study hypothesis testing problems. Putt and Chinchilli (1999) developed a mixed effects model and gave conditions under which explicit estimators of the variance components exist as well as hypothesis test of treatment differences. However, explicit maximum-likelihood (ML) estimators of the fixed effects are not obtained, unless independent random errors are assumed, in which case the estimators coincide with the ordinary least-squares estimators. The main advantages of using the GCM is that explicit ML estimators of the parameters are obtained.
Two specific cross-over designs, called the AB:BA and ABAB:BABA design, are studied. Models for these designs are described and the ML estimators for the parameters are derived by utilizing the theory for the GCM. Two models are defined for each of the design, with respect to inclusion or exclusion of a so-called carry-over effect. When the carry-over effect is omitted, the extended GCM is applied. Furthermore, a model with a structured dispersion matrix is defined for the AB:BA design, and a specific linear transformation is used to derive estimators in a convenient way. To illustrate numerically, ML estimates are calculated for an ABAB:BABA study published by Ciminera and Wolfe (1953) and later re-analysed by Putt and Chinchilli (1999).
Section snippets
GCMs
For reviews of the model see von Rosen (1991) and Kshirsagar and Smith (1995). We will define two types of the GCM which will be referred to as GCM1 and GCM2. GCM1 is the same model as introduced by Potthoff and Roy, whereas GCM2 is an extended version of GCM1 introduced by von Rosen (1989). In the following, will be a random observable matrix, where each subject or unit is represented by a column in . Let further, for an arbitrary matrix of real numbers, denote the rank of ,
Conclusion and discussion
Models for the repeated measures and cross-over design have been defined in terms of the GCM1, and [ML] estimators of the parameters have been derived by utilizing the theory for multivariate linear normal models. The models differ from the standard situation in that the main parameters, which we call the cross-over parameters, are linear combinations of , for some matrices and , whereas the matrix generally is not of interest. The number of cross-over parameters is
Acknowledgements
The authors are very grateful to three referees for detailed and constructive comments which improved the paper a lot.
References (13)
Maximum likelihood estimators in multivariate linear normal models
J. Multivariate Anal.
(1989)Moments for a multivariate linear model with an application to the growth curve model
J. Multivariate Anal.
(1990)- Åsenblad, N., 2001. Multivariate linear normal models for the analysis of cross-over designs. Licentiat Thesis,...
- Banken, L., 1984. Eine verallgemeinerung des GMANOVA modells. Dissertation, University of Trier, Trier,...
- et al.
An example of the use of extended cross-over design in the comparison of NPH insulin mixtures
Biometrics
(1953) - et al.
Design and Analysis of Cross-Over Trials
(1989)