Journal of Educational and Behavioral Statistics

Formulas for estimating sample sizes are presented to provide specified levels of power for tests of significance from a longitudinal design allowing for subject attrition. These formulas are derived for a comparison of two groups in terms of single degree-of-freedom contrasts of population means across the study timepoints. Contrasts of this type can often capture the main and interaction effects in a two-group repeated measures design. For example, a two-group comparison of either an average across time or a specific trend across time (e.g., linear or quadratic) can be considered. Since longitudinal data with attrition are often analyzed using an unbalanced repeated measures model (with a structured variance-covariance matrix for the repeated measures) or a random-effects model for incomplete longitudinal data, the variance-covariance matrix of the repeated measures is allowed to assume a variety of forms. Tables are presented listing sample size determinations assuming compound symmetry, a first-order autoregressive structure, and a non-stationary random-effects structure. Examples are provided to illustrate use of the formulas, and a computer program implementing the procedure is available from the first author.

For the comparison of more than two independent samples the Kruskal-Wallis H test is a preferred procedure in many situations. However, the exact null and alternative hypotheses, as well as the assumptions of this test, do not seem to be very clear among behavioral scientists. This article attempts to bring some order to the inconsistent, sometimes controversial treatments of the Kruskal-Wallis test. First we clarify that the H test cannot detect with consistently increasing power any alternative hypothesis other than exceptions to stochastic homogeneity. It is then shown by a mathematical derivation that stochastic homogeneity is equivalent to the equality of the expected values of the rank sample means. This finding implies that the null hypothesis of stochastic homogeneity can be tested by an ANOVA performed on the rank transforms, which is essentially equivalent to doing a Kruskal-Wallis H test. If the variance homogeneity condition does not hold then it is suggested that robust ANOVA alternatives performed on ranks be used for testing stochastic homogeneity. Generalizations are also made with respect to Friedman’s G test.

The tetrachoric correlation describes the linear relation between two continuous variables that have each been measured on a dichotomous scale. The treatment of the point estimate, standard error, interval estimate, and sample size requirement for the tetrachoric correlation is cursory and incomplete in modern psychometric and behavioral statistics texts. A new and simple method of accurately approximating the tetrachoric correlation is introduced. The tetrachoric approximation is then used to derive a simple standard error, confidence interval, and sample size planning formula. The new confidence interval is shown to perform far better than the confidence interval computed by SAS. A method to improve the SAS confidence interval is proposed. All of the new results are computationally simple and are ideally suited for textbook and classroom presentations.

Imputation methods are popular for the handling of missing data in psychology. The methods generally consist of predicting missing data based on observed data, yielding a complete data set that is amiable to standard statistical analyses. In the context of Bayesian factor analysis, this article compares imputation under an unrestricted multivariate normal model (Multiple Imputation [MI]) to imputation under the statistical model of interest (Data Augmentation [DA]). The former method is popular in applied research, but the latter method is more straightforward from a Bayesian perspective. Simulations demonstrate that DA yields less-biased parameter estimates for moderate sample sizes and high missingness proportions. MI, however, yields less-biased parameter estimates for large sample sizes with misspecified models. The incorporation of auxiliary variables in DA is also addressed, and BUGS code is provided.