Two alternatives to the problems of conducting planned and post hoc comparisons in tests of concordance and discordance for G groups of judges are examined. The two models are illustrated using existing data. (Author/JKS)

Interpretations of relative variable importance in multivariate analysis of variance are discussed, with attention to (1) latent construct definition; (2) linear discriminant function scores; and (3) grouping variable effects. Two numerical ranking methods are proposed and compared by the bootstrap approach using two real data sets. (SLD)

Issues regarding tolerance and confidence intervals are discussed within the context of educational measurement, and conceptual distinctions are drawn between these two types of intervals. Points are raised about the advantages of tolerance intervals when the focus is on a particular observed score rather than a particular examinee. (Author/BW)

Methods for assessing collapsibility in regression problems are described, including possible extensions to the class of generalized linear models. These procedures, with terminology borrowed from the contingency table field, can be used in experimental settings or nonexperimental settings where two models viewed as alternative explanations are…

This Monte Carlo simulation studied the relative power of 8 tests for sphericity in randomized block designs where sample size was small (10, 15, 20, and 30) and population covariance matrices of dimension-to-sample size ratio approached 1.0. The locally best invariant test demonstrated substantial power to detect departures from sphericity. (SLD)

Recently, C. E. McCulloch (1987) suggested a modification of the Morgan-Pitman test for comparing the variances of two dependent groups. This paper demonstrates that there are situations where the procedure is not robust. A subsample approach, similar to the Box-Scheffe test, and the Sandvik-Olsson procedure are also assessed. (TJH)

Two algorithms that automatically select subsets of variables (PACE algorithm) and reference variables (Fabin estimators), respectively, used for the noniterative estimators are presented. The PACE algorithm is based on a nonsymmetric matrix sweep operator. A Monte Carlo experiment compares the relative performance of these estimators and others.…

Context for the application and specialization of a Bayesian linear model is m-group regression and the application to the prediction of grade point average. Specialization involves the assumption of homogeneity of regression coefficients (but not intercepts) across groups. Model's predictive efficiency is compared with that of the full m-group…

A marginal maximum likelihood estimation algorithm is presented for the ordered partition model of M. Wilson that does not require the set of available responses to be fully ordered. The model and its estimation algorithm are illustrated in a comparison of alternative scoring schemes for open-ended science items. (SLD)

Two issues in analysis of interactions in complex linear models are considered. One is the omnibus test for interaction, and the other concerns follow-up tests when the omnibus test is significant. Reconsidered omnibus and follow-up tests are illustrated on an educational data set analyzed using the Statistical Analysis System. (SLD)

Focuses on the fact that an educational treatment typically involves an intervention in a growth process. By modelling this process, expected growth for various treatment groups under control conditions may be estimated. Actual growth can be compared with projected growth to estimate the value-added by the program. A simple model is developed. (RC)

Two models are presented to compare trainee and expert assessment strategies. The use of these procedures in assessment training programs is illustrated in a study of 5 postgraduate students in an environmental education course who were required to rate the work of 20 students. (SLD)

In this paper a Bayesian procedure is developed for the simultaneous estimation of the reading ability and difficulty parameters which are assumed to be factors in reading errors by the multiplicative Poisson Model. According to several criteria, the Bayesian estimates are better than comparable maximum likelihood estimates. (Author/JAZ)

Type I error rates for Yao's, James' first-order and second-order, and Johansen's tests of equality of mean vectors for two independent samples were estimated for various conditions defined by the degree of heteroscedasticity and nonnormality. Each procedure can be seriously nonrobust with exponential and log-normal distributions. (TJH)

Forming the matching variable for the Mantel-Haenszel differential item functioning (DIF) procedure through use of the total score as the matching variable (thin) and forming the matching variable by pooling total score levels (thick) were compared in a Monte Carlo study. Reasons thick matching is superior are discussed. (SLD)