Differential Item Functioning Assessment in Cognitive Diagnostic Modeling: Application of the Wald Test to Investigate DIF in the DINA Model

Analyzing examinees’ responses using cognitive diagnostic models (CDMs) has the advantage of providing diagnostic information. To ensure the validity of the results from these models, differential item functioning (DIF) in CDMs needs to be investigated. In this article, the Wald test is proposed to examine DIF in the context of CDMs. This study explored the effectiveness of the Wald test in detecting both uniform and nonuniform DIF in the DINA model through a simulation study. Results of this study suggest that for relatively discriminating items, the Wald test had Type I error rates close to the nominal level. Moreover, its viability was underscored by the medium to high power rates for most investigated DIF types when DIF size was large. Furthermore, the performance of the Wald test in detecting uniform DIF was compared to that of the traditional Mantel‐Haenszel (MH) and SIBTEST procedures. The results of the comparison study showed that the Wald test was comparable to or outperformed the MH and SIBTEST procedures. Finally, the strengths and limitations of the proposed method and suggestions for future studies are discussed.     

Multilevel Cognitive Diagnosis Models for Assessing Changes in Latent Attributes

Cognitive diagnosis models (CDMs) have been developed to evaluate the mastery status of individuals with respect to a set of defined attributes or skills that are measured through testing. When individuals are repeatedly administered a cognitive diagnosis test, a new class of multilevel CDMs is required to assess the changes in their attributes and simultaneously estimate the model parameters from the different measurements. In this study, the most general CDM of the generalized deterministic input, noisy “and” gate (G‐DINA) model was extended to a multilevel higher order CDM by embedding a multilevel structure into higher order latent traits. A series of simulations based on diverse factors was conducted to assess the quality of the parameter estimation. The results demonstrate that the model parameters can be recovered fairly well and attribute mastery can be precisely estimated if the sample size is large and the test is sufficiently long. The range of the location parameters had opposing effects on the recovery of the item and person parameters. Ignoring the multilevel structure in the data by fitting a single‐level G‐DINA model decreased the attribute classification accuracy and the precision of latent trait estimation. The number of measurement occasions had a substantial impact on latent trait estimation. Satisfactory model and person parameter recoveries could be achieved even when assumptions of the measurement invariance of the model parameters over time were violated. A longitudinal basic ability assessment is outlined to demonstrate the application of the new models.     

A Note on the Invariance of the DINA Model Parameters

Cognitive diagnosis models (CDMs), as alternative approaches to unidimensional item response models, have received increasing attention in recent years. CDMs are developed for the purpose of identifying the mastery or nonmastery of multiple fine-grained attributes or skills required for solving problems in a domain. For CDMs to receive wider use, researchers and practitioners need to understand the basic properties of these models. The article focuses on one CDM, the deterministic inputs, noisy "and" gate (DINA) model, and the invariance property of its parameters. Using simulated data involving different attribute distributions, the article demonstrates that the DINA model parameters are absolutely invariant when the model perfectly fits the data. An additional example involving different ability groups illustrates how noise in real data can contribute to the lack of invariance in these parameters. Some practical implications of these findings are discussed .     

Lord's Wald Test for Detecting DIF in Multidimensional IRT Models: A Comparison of Two Estimation Approaches

Lord's Wald test for differential item functioning (DIF) has not been studied extensively in the context of the multidimensional item response theory (MIRT) framework. In this article, Lord's Wald test was implemented using two estimation approaches, marginal maximum likelihood estimation and Bayesian Markov chain Monte Carlo estimation, to detect uniform and nonuniform DIF under MIRT models. The Type I error and power rates for Lord's Wald test were investigated under various simulation conditions, including different DIF types and magnitudes, different means and correlations of two ability parameters, and different sample sizes. Furthermore, English usage data were analyzed to illustrate the use of Lord's Wald test with the two estimation approaches.     

Parameter Invariance and Skill Attribute Continuity in the DINA Model

Cognitive diagnosis models (CDMs) typically assume skill attributes with discrete (often binary) levels of skill mastery, making the existence of skill continuity an anticipated form of model misspecification. In this article, misspecification due to skill continuity is argued to be of particular concern for several CDM applications due to the lack of invariance it yields in CDM skill attribute metrics, or what in this article are viewed as the “thresholds” applied to continuous attributes in distinguishing masters from nonmasters. Using the deterministic input noisy and (DINA) model as an illustration, the effects observed in real data are found to be systematic, with higher thresholds for mastery tending to emerge in higher ability populations. The results are shown to have significant implications for applications of CDMs that rely heavily upon the parameter invariance properties of the models, including, for example, applications toward the measurement of growth and differential item functioning analyses.     

Cognitive diagnostic model of best choice: a study of reading comprehension

Abstract

The present study compared the performance of six cognitive diagnostic models (CDMs) to explore inter skill relationship in a reading comprehension test. To this end, item responses of about 21,642 test-takers to a high-stakes reading comprehension test were analyzed. The models were compared in terms of model fit at both test and item levels, classification consistency and accuracy, and proportion of skill mastery profiles. The results showed that the G-DINA performed the best and the C-RUM, NC-RUM, and ACDM showed the closest affinity to the G-DINA. In terms of some criteria, the DINA showed comparable performance to the G-DINA. The test-level results were corroborated by the item-level model comparison, where DINA, DINO, and ACDM variously fit some of the items. The results of the study suggested that relationships among the subskills of reading comprehension might be a combination of compensatory and non-compensatory. Therefore, it is suggested that the choice of the CDM be carried out at item level rather than test level.     

Examining Chi-Square Test Statistics Under Conditions of Large Model Size and Ordinal Data

This study examined the effect of model size on the chi-square test statistics obtained from ordinal factor analysis models. The performance of six robust chi-square test statistics were compared across various conditions, including number of observed variables (p), number of factors, sample size, model (mis)specification, number of categories, and threshold distribution. Results showed that the unweighted least squares (ULS) robust chi-square statistics generally outperform the diagonally weighted least squares (DWLS) robust chi-square statistics. The ULSM estimator performed the best overall. However, when fitting ordinal factor analysis models with a large number of observed variables and small sample size, the ULSM-based chi-square tests may yield empirical variances that are noticeably larger than the theoretical values and inflated Type I error rates. On the other hand, when the number of observed variables is very large, the mean- and variance-corrected chi-square test statistics (e.g., based on ULSMV and WLSMV) could produce empirical variances conspicuously smaller than the theoretical values and Type I error rates lower than the nominal level, and demonstrate lower power rates to reject misspecified models. Recommendations for applied researchers and future empirical studies involving large models are provided.     

Four New Corrected Statistics for SEM With Small Samples and Nonnormally Distributed Data

When the assumption of multivariate normality is violated and the sample sizes are relatively small, existing test statistics such as the likelihood ratio statistic and Satorra–Bentler’s rescaled and adjusted statistics often fail to provide reliable assessment of overall model fit. This article proposes four new corrected statistics, aiming for better model evaluation with nonnormally distributed data at small sample sizes. A Monte Carlo study is conducted to compare the performances of the four corrected statistics against those of existing statistics regarding Type I error rate. Results show that the performances of the four new statistics are relatively stable compared with those of existing statistics. In particular, Type I error rates of a new statistic are close to the nominal level across all sample sizes under a condition of asymptotic robustness. Other new statistics also exhibit improved Type I error control, especially with nonnormally distributed data at small sample sizes.     

Mediated Effects with the Parallel Process Latent Growth Model: An Evaluation of Methods for Testing Mediation in the Presence of Nonnormal Data

This Monte Carlo simulation study investigated the impact of nonnormality on estimating and testing mediated effects with the parallel process latent growth model and 3 popular methods for testing the mediated effect (i.e., Sobel’s test, the asymmetric confidence limits, and the bias-corrected bootstrap). It was found that nonnormality had little effect on the estimates of the mediated effect, standard errors, empirical Type I error, and power rates in most conditions. In terms of empirical Type I error and power rates, the bias-corrected bootstrap performed best. Sobel’s test produced very conservative Type I error rates when the estimated mediated effect and standard error had a relationship, but when the relationship was weak or did not exist, the Type I error was closer to the nominal .05 value.     

Impact of Diagnosticity on the Adequacy of Models for Cognitive Diagnosis under a Linear Attribute Structure: A Simulation Study

Compared to unidimensional item response models (IRMs), cognitive diagnostic models (CDMs) based on latent classes represent examinees' knowledge and item requirements using discrete structures. This study systematically examines the viability of retrofitting CDMs to IRM‐based data with a linear attribute structure. The study utilizes a procedure to make the IRM and CDM frameworks comparable and investigates how estimation accuracy is affected by test diagnosticity and the match between the true and fitted models. The study shows that comparable results can be obtained when highly diagnostic IRM data are retrofitted with CDM, and vice versa, retrofitting CDMs to IRM‐based data in some conditions can result in considerable examinee misclassification, and model fit indices provide limited indication of the accuracy of item parameter estimation and attribute classification.     

Assessment of Differential Item Functioning Under Cognitive Diagnosis Models: The DINA Model Example

The assessment of differential item functioning (DIF) is routinely conducted to ensure test fairness and validity. Although many DIF assessment methods have been developed in the context of classical test theory and item response theory, they are not applicable for cognitive diagnosis models (CDMs), as the underlying latent attributes of CDMs are multidimensional and binary. This study proposes a very general DIF assessment method in the CDM framework which is applicable for various CDMs, more than two groups of examinees, and multiple grouping variables that are categorical, continuous, observed, or latent. The parameters can be estimated with Markov chain Monte Carlo algorithms implemented in the freeware WinBUGS. Simulation results demonstrated a good parameter recovery and advantages in DIF assessment for the new method over the Wald method.     

The Place of the Fable in the Character Training of Children

Type I error rate and power for the t test, Wilcoxon-Mann-Whitney (U) test, van der Waerden Normal Scores (NS) test, and Welch-Aspin-Satterthwaite (W) test were compared for two independent random samples drawn from nonnormal distributions. Data with varying degrees of skewness (S) and kurtosis (K) were generated using Fleishman's (1978) power function. Five sample size combinations were used with both equal and unequal variances. For nonnormal data with equal variances, the power of the U test exceeded the power of the t test regardless of sample size. When the sample sizes were equal but the variances were unequal, the t test proved to be the most powerful test. When variances and sample sizes were unequal, the W test became the test of choice because it was the only test that maintained its nominal Type I error rate.     

Using Monte Carlo Normal Distributions to Evaluate Structural Models With Nonnormal Data

Statistical theories of goodness-of-fit tests in structural equation modeling are based on asymptotic distributions of test statistics. When the model includes a large number of variables or the population is not from a multivariate normal distribution, the asymptotic distributions do not approximate the distribution of the test statistics very well at small sample sizes. A variety of methods have been developed to improve the accuracy of hypothesis testing at small sample sizes. However, all these methods have their limitations, specially for nonnormal distributed data. We propose a Monte Carlo test that is able to control Type I error with more accuracy compared to existing approaches in both normal and nonnormally distributed data at small sample sizes. Extensive simulation studies show that the suggested Monte Carlo test has a more accurate observed significance level as compared to other tests with a reasonable power to reject misspecified models.     

Detecting Multidimensional DIF in Polytomous Items with IRT Methods and Estimation Approaches

The purpose of this study was to investigate multidimensional DIF with a simple and nonsimple structure in the context of multidimensional Graded Response Model (MGRM). This study examined and compared the performance of the IRT-LR and Wald test using MML-EM and MHRM estimation approaches with different test factors and test structures in simulation studies and applying real data sets. When the test structure included two dimensions, the IRT-LR (MML-EM) generally performed better than the Wald test and provided higher power rates. If the test included three dimensions, the methods provided similar performance in DIF detection. In contrast to these results, when the number of dimensions in the test was four, MML-EM estimation completely lost precision in estimating the nonuniform DIF, even with large sample sizes. The Wald with MHRM estimation approaches outperformed the Wald test (MML-EM) and IRT-LR (MML-EM). The Wald test had higher power rate and acceptable type I error rates for nonuniform DIF with the MHRM estimation approach.The small and/or unbalanced sample sizes, small DIF magnitudes, unequal ability distributions between groups, number of dimensions, estimation methods and test structure were evaluated as important test factors for detecting multidimensional DIF.     

配对设计下相对危险度的Wald检验

基于相对危险度提出了一个假设检验问题,在配对设计下用delta方法构建了Wald检验统计量,并对Wald检验统计量进行连续性修正。通过Monte Carlo方法模拟,表明Wald检验统计量控制第一类错误的概率较差,而修正后的Wald检验统计量能较好地控制第一类错误率,且第一类错误率接近于给定的显著性水平,是一个理想检验。     

Effect of Sample Size Ratio and Model Misfit When Using the Difficulty Parameter Differences Procedure to Detect DIF

This study examined the effect of sample size ratio and model misfit on the Type I error rates and power of the Difficulty Parameter Differences procedure using Winsteps. A unidimensional 30-item test with responses from 130,000 examinees was simulated and four independent variables were manipulated: sample size ratio (20/100/250/500/1000); model fit/misfit (1 PL and 3PLc =. 15 models); impact (no difference/mean differences/variance differences/mean and variance differences); and percentage of items with uniform and nonuniform DIF (0%/10%/20%). In general, the results indicate the importance of ensuring model fit to achieve greater control of Type I error and adequate statistical power. The manipulated variables produced inflated Type I error rates, which were well controlled when a measure of DIF magnitude was applied. Sample size ratio also had an effect on the power of the procedure. The paper discusses the practical implications of these results.     

Specification Searches in Multilevel Structural Equation Modeling: A Monte Carlo Investigation

Cluster sampling results in response variable variation both among respondents (i.e., within-cluster or Level 1) and among clusters (i.e., between-cluster or Level 2). Properly modeling within- and between-cluster variation could be of substantive interest in numerous settings, but applied researchers typically test only within-cluster (i.e., individual difference) theories. Specifying a between-cluster model in the absence of theory requires a specification search in multilevel structural equation modeling. This study examined a variety of within-cluster and between-cluster sample sizes, intraclass correlation coefficients, start models, parameter addition and deletion methods, and Type I error control techniques to identify which combination of start model, parameter addition or deletion method, and Type I error control technique best recovered the population of the between-cluster model. Results indicated that a “saturated” start model, univariate parameter deletion technique, and no Type I error control performed best, but recovered the population between-cluster model in less than 1 in 5 attempts at the largest sample sizes. The accuracy of specification search methods, suggestions for applied researchers, and future research directions are discussed.     

Improving Person-Fit Assessment by Correcting the Ability Estimate and Its Reference Distribution

The standardized log-likelihood of a response vector (l_z) is a popular IRT-based person-fit test statistic for identifying model-misfitting response patterns. Traditional use of l_z is overly conservative in detecting aberrance due to its incorrect assumption regarding its theoretical null distribution. This study proposes a method for improving the accuracy of person-fit analysis using l_z which takes into account test unreliability when estimating the ability and constructs the distribution for each l_z through resampling methods. The Type I error and power (or detection rate) of the proposed method were examined at different test lengths, ability levels, and nominal α levels along with other methods, and power to detect three types of aberrance—cheating, lack of motivation, and speeding—was considered. Results indicate that the proposed method is a viable and promising approach. It has Type I error rates close to the nominal value for most ability levels and reasonably good power.     

Resampling and Distribution of the Product Methods for Testing Indirect Effects in Complex Models

Recent advances in testing mediation have found that certain resampling methods and tests based on the mathematical distribution of 2 normal random variables substantially outperform the traditional z test. However, these studies have primarily focused only on models with a single mediator and 2 component paths. To address this limitation, a simulation was conducted to evaluate these alternative methods in a more complex path model with multiple mediators and indirect paths with 2 and 3 paths. Methods for testing contrasts of 2 effects were evaluated also. The simulation included 1 exogenous independent variable, 3 mediators and 2 outcomes and varied sample size, number of paths in the mediated effects, test used to evaluate effects, effect sizes for each path, and the value of the contrast. Confidence intervals were used to evaluate the power and Type I error rate of each method, and were examined for coverage and bias. The bias-corrected bootstrap had the least biased confidence intervals, greatest power to detect nonzero effects and contrasts, and the most accurate overall Type I error. All tests had less power to detect 3-path effects and more inaccurate Type I error compared to 2-path effects. Confidence intervals were biased for mediated effects, as found in previous studies. Results for contrasts did not vary greatly by test, although resampling approaches had somewhat greater power and might be preferable because of ease of use and flexibility.     

Multiplicity Control in Structural Equation Modeling

Abstract

Researchers conducting structural equation modeling analyses rarely, if ever, control for the inflated probability of Type I errors when evaluating the statistical significance of multiple parameters in a model. In this study, the Type I error control, power and true model rates of famsilywise and false discovery rate controlling procedures were compared with rates when no multiplicity control was imposed. The results indicate that Type I error rates become severely inflated with no multiplicity control, but also that familywise error controlling procedures were extremely conservative and had very little power for detecting true relations. False discovery rate controlling procedures provided a compromise between no multiplicity control and strict familywise error control and with large sample sizes provided a high probability of making correct inferences regarding all the parameters in the model.