Decisions in Single Group Repeated Measures Analysis: Statistical Tests and Three Computer Packages

Abstract

Repeated measures analysis is a widely used tool of applied statisticians. However, its users are confronted with what seems to be a myriad of decisions, even in its simple application. Should a univariate or a multivariate mode be used? Should an adjusted univariate test be used, and if so, which one? Should a pooled or partitioned error term be used? How important is the sphericity test? Given answers to these questions, which of three major statistical packages, and which programs within a package, could be used? These questions are discussed and resolved in this presentation.

Self‐efficacy training for patients with end‐stage renal disease

“…Hospitalized individuals and those with acute illness, psychological or cognitive disorders, or physical limitations in self-care were excluded. The sample size needed was calculated using tabulation scheme for sample estimation with repeated measures design, as recommended by Barcikowski and Robey (1984) and Stevens (2002). According to the table (Stevens 2002, p. 511), for a power of 0AE80, average correlation of 0AE50, a medium effect size of 0AE35, four repeated measures and a of 0AE05, 27 subjects were needed for each group.…”

Analysis of Repeated Measures Experiments

“…Repeated sampling of the subjects is usually an attempt not only to reduce underlying variation and thus achieve a more precise result, but also to examine trends over time. The purpose of this paper is to provide an algorithm for the analysis of repeated measures experiments that uses the appropriate test statistics for multivariate analyses (Morrison 1976:207-208, and see 178; Barcikowski and Robey 1984), while allowing the straightforward testing oftime trends or other contrasts of interest. The algorithm functions by reducing each set of measurements on a subject to a single weighted sum, and then performing a univariate analysis on the weighted sums as if they were the original measurements.…”

On the Choice of Appropriate Error Terms in Profile Analysis

“…In the context of the univariate mixed model ANOVA considered in this paper, note that the choice U* = U(UTU) -1/2 in equation (3.3) is a sufficient condition for the first part of proposition 3. Therefore, the approximate e-adjusted F -tests proposed by Greenhouse and Geisser (1959) or Huynh and Feldt (1976) may be employed in the general context under investigation here, and not only for the special case of omnibus interaction tests, as suggested by Rouanet and Lepine (1970), Boik (1981) and Barcikowski and Robey (1984). Alternatively, we may choose U* = V(y Ty)-1/2 where Y corresponds to some comparison embedding U.…”

“…Unfortunately, such flexibility is coupled with some potential difficulties associated with the choice of the appropriate error terms, as pointed out by Barcikowski and Robey (1984), and applied statisticians still lack guidelines to help them to choose between the available alternatives.…”