ABSTRACT
Utilizing the perspective of finite mixture modeling, this note considers whether a finding of a plausible one-parameter logistic model could be spurious for a population with substantial unobserved heterogeneity. A theoretically and empirically important setting is discussed involving the mixture of two latent classes, with the less restrictive two-parameter logistic model holding within each of the classes while the more restrictive one-parameter logistic model is plausible overall. A numerical example is presented then of obtaining the one-parameter logistic model as a spurious result in a finite mixture study. Implications for behavioral and social research are discussed in light of these findings.
Acknowledgment
Thanks are due to T. Asparouhov and S. Embretson for valuable discussions on simulation studies and item response model mixtures.
Notes
1. Using the R-package ‘ltm’, the p-value for the overall goodness of fit test of the Rasch model is found to be p = .86 while the p-value associated with the test of the nested 1PL-model against the 2PL model is p = .439, which is in agreement with the pertinent results presented in the current section in the main text (e.g., Rizopoulos, Citation2006, Citation2018; the tests are carried out with the ltm-commands “GoF.rasch” and “anova”, respectively, and the last stated p-value is that associated with the ∆χ2-test reported above in this note). The largest z-value (test statistic value) across the five tests of possible guessing on any of the items within the 1PL-model is thereby 1.113, suggesting no guessing on an individual item; similarly, the overall test statistic for guessing on at least one item is p = .956 and thus non-significant as well. These findings show that formally a guessing (or related chance processes leading to a correct response) explanation cannot act against the reported plausibility of the 1PL-model. Last but not least, the highest two-way marginal residual (ratio of squared difference between observed and expected frequencies under the 1PL-model, to expected frequency) is 1.49 and thus not significant (relative to the pertinent df = 1; we note that these residuals are not independent and so it would be appropriate to de-emphasize this finding if interpreted only as lack of significance). These results provide further evidence in favor of the earlier claim of plausible fit of the 1PL-model (see also main text).