Characterizing reliability for a Faculty Climate Survey: Estimation model dependencies and reliability generalization

Abstract

Methods. Four reliability estimation models were employed to obtain estimates for faculty appointment and gender group measures derived from four questionnaire scales of a Faculty Climate Survey. Faculty responses were analyzed via (a) coefficient alpha, (b) IRT-Rasch, (c) IRT-Unfolding, and (d) CFA methods. Estimates and their components were compared across-groups within scale and within-group across scales to determine differences among estimation models and to uniquely characterize those differences. Scale dimensionality was assessed per-scale per-group using CFA. Secondary analyses included: (a) independent and dependent-group tests to determine the statistical significance of coefficient alpha differences; (b) bootstrapping simulation to determine the effect of sample size on estimates; and (c) analysis of variance to determine whether attitudinal differences existed between appointment, gender, or appointment-by-gender groups. Results. (1) Reliability estimation models identified important differences between appointment and gender group estimates for scale measures and among scale estimates for each group's set of scale measures. (2) Models were not equally sensitive to detecting differences, either between groups or among scales per group. (3) Alpha and CFA estimates did not always function as lower- and upper-bounds of an expected estimate range: 30% of alpha-CFA range "endpoints" were underestimates of observed ranges. (4) IRT-based estimates were generally located between alpha and CFA estimates, closer to alpha than to CFA estimates. (5) IRT-Unfolding estimates were frequently but not always greater than IRT-Rasch estimates: 30% were less. (6) Alpha and CFA estimation components did not provide comparable item-level information; thus, alpha and CFA plans for characterizing and improving scales differed. (7) IRT-Rasch and IRT-Unfolding estimation components did not provide comparable person-measure information, thereby informing observed differences in IRT-based estimates. (8) Sample size had an effect on CFA estimation: samples of N = 50 achieved highest estimates; samples of N = 500 best reproduced original estimates and components. (9) Modeling error via CFA made meaningful contributions to understanding scale functioning. (10) ANOVA findings were potentially modifiable (e.g., effect sizes), considering obtained reliability estimates. Conclusion. Reliability estimates have group, measure, and model-dependencies that influence the size and nature of obtained estimates and must be accounted for when estimates are interpreted.