Educational and Psychological Measurement最新文献

Differential item functioning (DIF) has been a long-standing problem in educational and psychological measurement. In practice, the source from which DIF originates can be complex in the sense that an item can show DIF on multiple background variables of different types simultaneously. Although a variety of non-item response theory-(IRT)-based and IRT-based DIF detection methods have been introduced, they do not sufficiently address the issue of DIF evaluation when its source is complex. The recently proposed least absolute shrinkage and selection operator (LASSO) regularization method has shown promising results of detecting DIF on multiple background variables. To provide more insight, in this study, we compared three DIF detection methods, including the non-IRT-based logistic regression (LR), the IRT-based likelihood ratio test (LRT), and LASSO regularization, through a comprehensive simulation and an empirical data analysis. We found that when multiple background variables were considered, the Type I error and Power rates of the three methods for identifying DIF items on one of the variables depended on not only the sample size and its DIF magnitude but also on the DIF magnitude of the other background variable and the correlation between them. We presented other findings and discussed the limitations and future research directions in this paper.

Observed variable and factor selection are critical components of factor analysis, particularly when the optimal subset of observed variables and the number of factors are unknown and results cannot be replicated across studies. The Replicable Factor Analytic Solutions (RFAS) algorithm was developed to assess the replicability of factor structures-both in terms of the number of factors and the variables retained-while identifying the "best" or most replicable solutions according to predefined criteria. This study evaluated RFAS performance across 54 experimental conditions that varied in model complexity (six-factor models), interfactor correlations (ρ = 0, .30, and .60), and sample sizes (n = 300, 500, and 1000). Under default settings, RFAS generally performed well and demonstrated its utility in producing replicable factor structures. However, performance declined with highly correlated factors, smaller sample sizes, and more complex models. RFAS was also compared to four alternative variable selection methods: Ant Colony Optimization (ACO), Weighted Group Least Absolute Shrinkage and Selection Operator (LASSO), and stepwise procedures based on target Tucker-Lewis Index (TLI) and ΔTLI criteria. Stepwise and LASSO methods were largely ineffective at eliminating problematic variables under the studied conditions. In contrast, both RFAS and ACO successfully removed variables as intended, although the resulting factor structures often differed substantially between the two approaches. As with other variable selection methods, refining algorithmic criteria may be necessary to further enhance model performance.

Fleiss's Kappa is an extension of Cohen's Kappa, developed to assess the degree of interrater agreement among multiple raters or methods classifying subjects using categorical scales. Like Cohen's Kappa, it adjusts the observed proportion of agreement to account for agreement expected by chance. However, over time, several paradoxes and interpretative challenges have been identified, largely stemming from the assumption of random chance agreement and the sensitivity of the coefficient to the number of raters. Interpreting Fleiss's Kappa can be particularly difficult due to its dependence on the distribution of categories and prevalence patterns. This paper argues that a portion of the observed agreement may be better explained by the interaction between category prevalence and inherent category characteristics, such as ambiguity, appeal, or social desirability, rather than by chance alone. By shifting away from the assumption of random rater assignment, the paper introduces a novel agreement coefficient that adjusts for the expected agreement by accounting for category prevalence, providing a more accurate measure of interrater reliability in the presence of imbalanced category distributions. It also examines the theoretical justification for this new measure, its interpretability, its standard error, and the robustness of its estimates in simulation and practical applications.

The Common Persons (CP) equating design offers critical advantages for high-security testing contexts-eliminating anchor item exposure risks while accommodating non-equivalent groups-yet few studies have systematically examined how CP characteristics influence equating accuracy, and the field still lacks clear implementation guidelines. Addressing this gap, this comprehensive Monte Carlo simulation (N = 5,000 examinees per form; 500 replications) evaluates CP equating by manipulating 8 factors: test length, difficulty shift, ability dispersion, correlation between test forms and CP characteristics. Four equating methods (identity, IRT true-score, linear, equipercentile) were compared using normalized RMSE and %Bias. Key findings reveal: (a) when the CP sample size reaches at least 30, CP sample properties exert negligible influence on accuracy, challenging assumptions about distributional representativeness; (b) Test factors dominate outcomes-difficulty shifts ( $\Delta {\delta }_{\mathrm{XY}}$ = 1) degrade IRT precision severely (|%Bias| >22% vs. linear/equipercentile's |%Bias| <1.5%), while longer tests reduce NRMSE and wider ability dispersion ( ${\sigma }_{\theta }$ = 1) enhances precision through improved person-item targeting; (c) Equipercentile and linear methods demonstrate superior robustness under form differences. We establish minimum operational thresholds: ≥30 CPs covering the score range suffice for precise equating. These results provide an evidence-based framework for CP implementation by systematically examining multiple manipulated factors, resolving security-vs-accuracy tradeoffs in high-stakes equating (e.g., credentialing exams) and enabling novel solutions like synthetic respondents.

In applied research across education, the social and behavioral sciences, and medicine, path models frequently incorporate both continuous and ordinal manifest variables to predict binary outcomes. This study employs Monte Carlo simulations to evaluate six estimators: robust maximum likelihood with probit and logit links (MLR-probit, MLR-logit), mean- and variance-adjusted weighted and unweighted least squares (WLSMV, ULSMV), and Bayesian methods with noninformative and weakly informative priors (Bayes-NI, Bayes-WI). Across various sample sizes, variable scales, and effect sizes, results show that WLSMV and Bayes-WI consistently achieve low bias and RMSE, particularly in small samples or when mediators have few categories. By contrast, categorical MLR approaches tended to yield unstable estimates for modest effects. These findings offer practical guidance for selecting estimators in mixed-scale path analyses and underscore their implications for robust inference.

Researchers conducting systematic reviews and meta-analyses often encounter studies in which the research design is a well conducted cluster randomized trial, but the statistical analysis does not take clustering into account. For example, the study might assign treatments by clusters but the analysis may not take into account the clustered treatment assignment. Alternatively, the analysis of the primary outcome of the study might take clustering into account, but the reviewer might be interested in another outcome for which only summary data are available in a form that does not take clustering into account. This article provides expressions for the approximate variance of risk differences, log risk ratios, and log odds ratios computed from clustered binary data, using the intraclass correlations. An example illustrates the calculations. References to empirical estimates of intraclass correlations are provided.

This study compares two network-based approaches for analyzing binary psychological assessment data: network psychometrics and latent space item response modeling (LSIRM). Network psychometrics, a well-established method, infers relationships among items or symptoms based on pairwise conditional dependencies. In contrast, LSIRM is a more recent framework that represents item responses as a bipartite network of respondents and items embedded in a latent metric space, where the likelihood of a response decreases with increasing distance between the respondent and item. We evaluate the performance of both methods through simulation studies under varying data-generating conditions. In addition, we demonstrate their applications to real assessment data, showcasing the distinct insights each method offers to researchers and practitioners.

The three-parameter logistic (3PL) model in item-response theory (IRT) has long been used to account for guessing in multiple-choice assessments through a fixed item-level parameter. However, this approach treats guessing as a property of the test item rather than the individual, potentially misrepresenting the cognitive processes underlying the examinee's behavior. This study evaluates a novel alternative, the Two-Parameter Logistic Extension (2PLE) model, which re-conceptualizes guessing as a function of a person's ability rather than as an item-specific constant. Using Monte Carlo simulation and empirical data from the PIRLS 2021 reading comprehension assessment, we compared the 3PL and 2PLE models on the recovery of latent ability, predictive fit (Leave-One-Out Information Criterion [LOOIC]), and theoretical alignment with test-taking behavior. The simulation results demonstrated that although both models performed similarly in terms of root-mean-squared error (RMSE) for ability estimates, the 2PLE model consistently achieved superior LOOIC values across conditions, particularly with longer tests and larger sample sizes. In an empirical analysis involving the reading achievement of 131 fourth-grade students from Saudi Arabia, model comparison again favored 2PLE, with a statistically significant LOOIC difference (ΔLOOIC = 0.482, z = 2.54). Importantly, person-level guessing estimates derived from the 2PLE model were significantly associated with established person-fit statistics (C*, U3), supporting their criterion validity. These findings suggest that the 2PLE model provides a more cognitively plausible and statistically robust representation of examinee behavior by embedding an ability-dependent guessing function.

In item response theory modeling, item fit analysis using posterior expectations, otherwise known as pseudocounts, has many advantages. They are readily obtained from the E-step output of the Bock-Aitkin Expectation-Maximization (EM) algorithm and continue to function as a basis of evaluating model fit, even when missing data are present. This paper aimed to improve the interpretability of the root mean squared deviation (RMSD) index based on posterior expectations. In Study 1, we assessed its performance using two approaches. First, we employed the poor person's posterior predictive model checking (PP-PPMC) to compute their significance levels. The resulting Type I error was generally controlled below the nominal level, but power noticeably declined with smaller sample sizes and shorter test lengths. Second, we used receiver operating characteristic (ROC) curve analysis (±) to empirically determine the reference values (cutoff thresholds) that achieve an optimal balance between false-positive and true-positive rates. Importantly, we identified optimal reference values for each combination of sample size and test length in the simulation conditions. The cutoff threshold approach outperformed the PP-PPMC approach with greater gains in true-positive rates than losses from the inflated false-positive rates. In Study 2, we extended the cutoff threshold approach to conditions with larger sample sizes and longer test lengths. Moreover, we evaluated the performance of the optimized cutoff thresholds under varying levels of data missingness. Finally, we employed response surface analysis (±) to develop a prediction model that generalizes the way the reference values vary with sample size and test length. Overall, this study demonstrates the application of the PP-PPMC for item fit diagnostics and implements a practical frequentist approach to empirically derive reference values. Using our prediction model, practitioners can compute the reference values of RMSD that are tailored to their dataset's sample size and test length.

Ensuring fairness in educational and psychological assessments is critical, particularly in detecting differential item functioning (DIF), where items perform differently across subgroups. The Rasch tree method, a model-based recursive partitioning approach, is an innovative and flexible DIF detection tool that does not require the pre-specification of focal and reference groups. However, research systematically examining its performance under realistic measurement conditions, such as when multiple DIF items do not consistently favor one subgroup, is limited. This study builds on prior research, evaluating the Rasch tree method's ability to detect DIF by investigating the impact of DIF balance, along with other key factors such as DIF magnitude, sample size, test length, and contamination levels. Additionally, we incorporate the Educational Testing Service effect size heuristic as a criterion to compare the DIF detection rate performance with only statistical significance. Results indicate that the Rasch tree has better true DIF detection rates under balanced DIF conditions and large DIF magnitudes. However, its accuracy declines when DIF is unbalanced and the percentage of DIF contamination increases. The use of an effect size reduces the detection of negligible DIF. Caution is recommended with smaller samples, where detection rates are the lowest, especially for larger DIF magnitudes and increased DIF contamination percentages in unbalanced conditions. The study highlights the strengths and limitations of the Rasch tree method under a variety of conditions, underscores the importance of the impact of DIF group imbalance, and provides recommendations for optimizing DIF detection in practical assessment scenarios.