It is useful to distinguish between repeated measures that are collected at fixed or varying occasions. If the measurements are taken at fixed occasions, all individuals provide measurements at the same set of occasions, usually regularly spaced, such as once every year. Alternatively, the original design is a fixed design, but because of planning problems, the data collection does not take place at the intended moments. For fixed occasion designs, especially when the occasions are regularly spaced and when there are no missing data, repeated measures analysis of variance (ANOVA) is a viable alternative for multilevel analysis. A comparison of the ANOVA approach and multilevel analysis is given in section 5.2.
5.2 EXAMPLE WITH FIXED OCCASIONS:
The data structure for the multilevel analysis of these data is generally different, depending on the specific program that is used. Most multilevel software requires that the data are structured with the measurement occasions defining the lowest level, and student-level variables repeated over the cases. The multilevel regression model for longitudinal data is a straight forward application of the multilevel regression model in chapter 2. It can also be written as a sequence of models for each level. At the lowest, the repeated measures level, we have: Yti = 0i + li + liTti + 2iXti + eti In repeated measures applications, the coefficients at the lowest level are often indicated by the Greek letter . This has the advantage that the subject level coefficients, which in repeated measures are at the second level, can be represented by the usual Greek letter , and so on. In the equation 5.1, Yti is the response variable of individual i measured at measurement occasion t, T is the time variable that indicates the measurement occasion, and Xti is a time varying covariate. For example, Yti could be the GPA of a student at measurement occasion t, Tti indicates the occasion at which the GPA is measured, and Xti the job status of the student at time t. Student characteristics, such as gender, are time-invariant covariates, which enter the equation at the second level: 0i = 00 + 01Zi + u0i 1i = 10 + 11Zi + u1i 2i = 20 + 21Zi + u2i By substitution, we get the single equation model: Yti = 00 + 10Tti + 20Xti + 01Zi + 11TtiZi + 21XtiZi + u1iTti + u2iXti + u0i + eti Using variables instead of letters, the equation for our GPA example becomes: Yti = 00 + 10Occasionti + 20Jobti + 01Sexi + 11OccasiontiSexi + 21JobtiSexi + u1iOccasionti + u2iJobti + u0i + eti In longitudinal research, we sometimes have repeated measurements of individuals, who are all measured together on a small number of fixed occasions. This is typically the case with experimental designs involving repeated measures and panel research. If we simply want to test the null hypothesis that the means are equal for all occasions, we can use repeated measures analysis of variance. If we use repeated measures univariate analysis of variance (Stevens, 2009, p. 420), we must assume sphericity. Spericity means that there are complex restrictions on the variance and co variances between the repeated measures; for details see Stevens (2009, Chapter 13). A specific form of sphericity, which is easily understood, is compound symmetry, sometimes referred to as uniformity. Compound symmetry requires that all population variances of the repeated measures are equal, and that all population covariances of the repeated measures are equal. If sphericity is not met, The F-ratio used in analysis of variance is positively biased, and we reject the null-hypothesis too often. A different approach is to specify the repeated measures as observations on a multivariate response vector and use multivariate analysis of variance (MANOVA). This does not require sphericity, and is considered the preferred approach if analysis of variance is used on repeated measures (OBrien & Kaiser, 2009). However, the multivariate test is more complicated, because it is based on transformation of the repeated measures, and what is tested are actually contrasts among the repeated measures.
5.4 ADVANTAGES OF MULTIELVEL ANALYSIS FOR
LONGITUDINAL DATA: Using multilevel models to analyse repeated measures data has several advantages. Bryk and Raudenbush (1992) mention five key points. First by modelling varying regression coefficients at the measurement occasion level, we have growth curves that are different for each subject. This fits in with the way individual development is generally conceptualized (see Willet, 1988). Second, the number of repeated measures and their spacing may differ across subjects. Third, the covariances between the repeated measures can be modelled as well, by specifying a specific structure for the variances and covariances at either level. This approach will be discussed in section 5.5. Fourth, if we have balanced data and use RML estiation, the usual analysis of variance based F-tests and t-tests can be derived from the multilevel regression results (see Raudenbush, 1993a). This shows that analysis of variance on repeated measures is a special case of the more general multilevel regression model. Fifth, in the multilevel model it is simple to add higher levels, t investigate the effect of family or social groups on individual development. A sixth advantage, not mentioned by Bryk and Raudenbush, is that it is straightforward to include time-varying or time-constant explanatory variable to the model, which allows us to model both the average group development and the development of different individuals over time.
7.2 COUNT DATA:
In the Poisson distribution, the probability of observing y events (y= 0, 1, 2, 3 ) is: Pr(y) = exp(-)y/y!