Niels G. Waller’s research while affiliated with University of Minnesota and other places

Researchers simulating covariance structure models sometimes add model error to their data to produce model misfit. Presently, the most popular methods for generating error-perturbed data are those by Tucker, Koopman, and Linn (TKL), Cudeck and Browne (CB), and Wu and Browne (WB). Although all of these methods include parameters that control the degree of model misfit, none can generate data that reproduce multiple fit indices. To address this issue, we describe a multiple-target TKL method that can generate error-perturbed data that will reproduce target RMSEA and CFI values either individually or together. To evaluate this method, we simulated error-perturbed correlation matrices for an array of factor analysis models using the multiple-target TKL method, the CB method, and the WB method. Our results indicated that the multiple-target TKL method produced solutions with RMSEA and CFI values that were closer to their target values than those of the alternative methods. Thus, the multiple-target TKL method should be a useful tool for researchers who wish to generate error-perturbed correlation matrices with a known degree of model error. All functions that are described in this work are available in the fungible R library. Additional materials (e.g., R code, supplemental results) are available at https://osf.io/vxr8d/.

We conducted an extensive Monte Carlo study of factor-rotation local solutions (LS) in multidimensional, two-parameter logistic (M2PL) item response models. In this study, we simulated more than 19,200 data sets that were drawn from 96 model conditions and performed more than 7.6 million rotations to examine the influence of (a) slope parameter sizes, (b) number of indicators per factor (trait), (c) probabilities of cross-loadings, (d) factor correlation sizes, (e) model approximation error, and (f) sample sizes on the local solution rates of the oblimin and (oblique) geomin rotation algorithms. To accommodate these design variables, we extended the standard M2PL model to include correlated major factors and uncorrelated minor factors (to represent model error). Our results showed that both rotation methods converged to LS under some conditions with geomin producing the highest local solution rates across many models. Our results also showed that, for identical item response patterns, rotation LS can produce different latent trait estimates with different levels of measurement precision (as indexed by the conditional standard error of measurement). Follow-up analyses revealed that when rotation algorithms converged to multiple solutions, quantitative indices of structural fit, such as numerical measures of simple structure, will often misidentify the rotation that is closest in mean-squared error to the factor pattern (or item-slope pattern) of the data-generating model.

Coefficient α, although ubiquitous in the research literature, is frequently criticized for being a poor estimate of test reliability. In this note, we consider the range of α and prove that it has no lower bound (i.e., α ∈ ( − ∞, 1]). While outlining our proofs, we present algorithms for generating data sets that will yield any fixed value of α in its range. We also prove that for some data sets—even those with appreciable item correlations—α is undefined. Although α is a putative estimate of the correlation between parallel forms, it is not a correlation as α can assume any value below − 1 (and α values below 0 are nonsensical reliability estimates). In the online supplemental materials, we provide R code for replicating our empirical findings and for generating data sets with user-defined α values. We hope that researchers will use this code to better understand the limitations of α as an index of scale reliability.

Although many textbooks on multivariate statistics discuss the common factor analysis model, few of these books mention the problem of factor score indeterminacy (FSI). Thus, many students and contemporary researchers are unaware of an important fact. Namely, for any common factor model with known (or estimated) model parameters, infinite sets of factor scores can be constructed to fit the model. Because all sets are mathematically exchangeable, factor scores are indeterminate. Our professional silence on this topic is difficult to explain given that FSI was first noted almost 100 years ago by E. B. Wilson, the 24th president (1929) of the American Statistical Association. To help disseminate Wilson’s insights, we demonstrate the underlying mathematics of FSI using the language of finite-dimensional vector spaces and well-known ideas of regression theory. We then illustrate the numerical implications of FSI by describing new and easily implemented methods for transforming factor scores into alternative sets of factor scores. An online supplement (and the fungible R library) includes R functions for illustrating FSI.