Educational and Psychological Measurement最新文献

Self-report personality tests used in high-stakes assessments hold the risk that test-takers engage in faking. In this article, we demonstrate an extension of the multidimensional nominal response model (MNRM) to account for the response bias of faking. The MNRM is a flexible item response theory (IRT) model that allows modeling response biases whose effect patterns vary between items. In a simulation, we found good parameter recovery of the model accounting for faking under different conditions as well as good performance of model selection criteria. Also, we modeled responses from N = 3,046 job applicants taking a personality test under real high-stakes conditions. We thereby specified item-specific effect patterns of faking by setting scoring weights to appropriate values that we collected in a pilot study. Results indicated that modeling faking significantly increased model fit over and above response styles and improved divergent validity, while the faking dimension exhibited relations to several covariates. Additionally, applying the model to a sample of job incumbents taking the test under low-stakes conditions, we found evidence that the model can effectively capture faking and adjust estimates of substantive trait scores for the assumed influence of faking. We end the article with a discussion of implications for psychological measurement in high-stakes assessment contexts.

A plethora of techniques exist to determine the number of factors to retain in exploratory factor analysis. A recent and promising technique is the Next Eigenvalue Sufficiency Test (NEST), but has not been systematically compared with well-established stopping rules. The present study proposes a simulation with synthetic factor structures to compare NEST, parallel analysis, sequential ${\chi }^2$ test, Hull method, and the empirical Kaiser criterion. The structures were based on 24 variables containing one to eight factors, loadings ranged from .40 to .80, inter-factor correlations ranged from .00 to .30, and three sample sizes were used. In total, 360 scenarios were replicated 1,000 times. Performance was evaluated in terms of accuracy (correct identification of dimensionality) and bias (tendency to over- or underestimate dimensionality). Overall, NEST showed the best overall performances, especially in hard conditions where it had to detect small but meaningful factors. It had a tendency to underextract, but to a lesser extent than other methods. The second best method was parallel analysis by being more liberal in harder cases. The three other stopping rules had pitfalls: sequential ${\chi }^2$ test and Hull method even in some easy conditions; the empirical Kaiser criterion in hard conditions.

An index extending the widely used omega-hierarchical coefficient is discussed, which can be used for evaluating the influence of a second-order factor on the interrelationships among the components of a hierarchical measuring instrument. The index represents a useful and informative complement to the traditional omega-hierarchical measure of explained overall scale score variance by that underlying construct. A point and interval estimation procedure is outlined for the described index, which is based on model reparameterization and is developed within the latent variable modeling framework. The method is readily applicable with popular software and is illustrated with examples.

In computerized adaptive testing (CAT), examinees see items targeted to their ability level. Postoperational data have a high degree of missing information relative to designs where everyone answers all questions. Item responses are observed over a restricted range of abilities, reducing item-total score correlations. However, if the adaptive item selection depends only on observed responses, the data are missing at random (MAR). We simulated data from three different testing designs (common items, randomly selected items, and CAT) and found that it was possible to re-estimate both person and item parameters from postoperational CAT data. In a multidimensional CAT, we show that it is necessary to include all responses from the testing phase to avoid violating missing data assumptions. We also observed that some CAT designs produced "reversals" where item discriminations became negative causing dramatic under and over-estimation of abilities. Our results apply to situations where researchers work with data drawn from adaptive testing or from instructional tools with adaptive delivery. To avoid bias, researchers must make sure they use all the data necessary to meet the MAR assumptions.

Maintaining consistent item difficulty across test forms is crucial for accurately and fairly classifying examinees into pass or fail categories. This article presents a practical procedure for classifying items based on difficulty levels using functional data analysis (FDA). Methodologically, we clustered item characteristic curves (ICCs) into difficulty groups by analyzing their functional principal components (FPCs) and then employed a neural network to predict difficulty for ICCs. Given the degree of similarity between many ICCs, categorizing items by difficulty can be challenging. The strength of this method lies in its ability to provide an empirical and consistent process for item classification, as opposed to relying solely on visual inspection. The findings reveal that most discrepancies between visual classification and FDA results differed by only one adjacent difficulty level. Approximately 67% of these discrepancies involved items in the medium to hard range being categorized into higher difficulty levels by FDA, while the remaining third involved very easy to easy items being classified into lower levels. The neural network, trained on these data, achieved an accuracy of 79.6%, with misclassifications also differing by only one adjacent difficulty level compared to FDA clustering. The method demonstrates an efficient and practical procedure for classifying test items, especially beneficial in testing programs where smaller volumes of examinees tested at various times throughout the year.

Multidimensional Item Response Theory (MIRT) is applied routinely in developing educational and psychological assessment tools, for instance, for exploring multidimensional structures of items using exploratory MIRT. A critical decision in exploratory MIRT analyses is the number of factors to retain. Unfortunately, the comparative properties of statistical methods and innovative Machine Learning (ML) methods for factor retention in exploratory MIRT analyses are still not clear. This study aims to fill this gap by comparing a selection of statistical and ML methods, including Kaiser Criterion (KC), Empirical Kaiser Criterion (EKC), Parallel Analysis (PA), scree plot (OC and AF), Very Simple Structure (VSS; C1 and C2), Minimum Average Partial (MAP), Exploratory Graph Analysis (EGA), Random Forest (RF), Histogram-based Gradient Boosted Decision Trees (HistGBDT), eXtreme Gradient Boosting (XGBoost), and Artificial Neural Network (ANN). The comparison was performed using 720,000 dichotomous response data sets simulated by the MIRT, for various between-item and within-item structures and considering characteristics of large-scale assessments. The results show that MAP, RF, HistGBDT, XGBoost, and ANN tremendously outperform other methods. Among them, HistGBDT generally performs better than other methods. Furthermore, including statistical methods' results as training features improves ML methods' performance. The methods' correct-factoring proportions decrease with an increase in missingness or a decrease in sample size. KC, PA, EKC, and scree plot (OC) are over-factoring, while EGA, scree plot (AF), and VSS (C1) are under-factoring. We recommend that practitioners use both MAP and HistGBDT to determine the number of factors when applying exploratory MIRT.

This study examines the performance of ChatGPT, developed by OpenAI and widely used as an AI-based conversational tool, as a data analysis tool through exploratory factor analysis (EFA). To this end, simulated data were generated under various data conditions, including normal distribution, response category, sample size, test length, factor loading, and measurement models. The generated data were analyzed using ChatGPT-4o twice with a 1-week interval under the same prompt, and the results were compared with those obtained using R code. In data analysis, the Kaiser-Meyer-Olkin (KMO) value, total variance explained, and the number of factors estimated using the empirical Kaiser criterion, Hull method, and Kaiser-Guttman criterion, as well as factor loadings, were calculated. The findings obtained from ChatGPT at two different times were found to be consistent with those obtained using R. Overall, ChatGPT demonstrated good performance for steps that require only computational decisions without involving researcher judgment or theoretical evaluation (such as KMO, total variance explained, and factor loadings). However, for multidimensional structures, although the estimated number of factors was consistent across analyses, biases were observed, suggesting that researchers should exercise caution in such decisions.

This study explores the performance of the item response tree (IRTree) approach in modeling missing data, comparing its performance to the expectation-maximization (EM) algorithm and multiple imputation (MI) methods. Both simulation and empirical data were used to evaluate these methods across different missing data mechanisms, test lengths, sample sizes, and missing data proportions. Expected a posteriori was used for ability estimation, and bias and root mean square error (RMSE) were calculated. The findings indicate that IRTree provides more accurate ability estimates with lower RMSE than both EM and MI methods. Its overall performance was particularly strong under missing completely at random and missing not at random, especially with longer tests and lower proportions of missing data. However, IRTree was most effective with moderate levels of omitted responses and medium-ability test takers, though its accuracy decreased in cases of extreme omissions and abilities. The study highlights that IRTree is particularly well suited for low-stakes tests and has strong potential for providing deeper insights into the underlying missing data mechanisms within a data set.

This study investigates the treatment of rapid-guess (RG) responses as missing data within the context of the effort-moderated model. Through a series of illustrations, this study demonstrates that the effort-moderated model assumes missing at random (MAR) rather than missing completely at random (MCAR), explaining the conditions necessary for MAR. These examples show that RG responses, when treated as missing under the effort-moderated model, do not introduce bias into ability estimates if the missingness mechanism is properly accounted for. Conversely, using a standard item response theory (IRT) model (scoring RG responses as if they were valid) instead of the effort-moderated model leads to considerable biases, underestimating group means and overestimating standard deviations when the item parameters are known, or overestimating item difficulty if the item parameters are estimated.

Evaluating differential item functioning (DIF) in assessments plays an important role in achieving measurement fairness across different subgroups, such as gender and native language. However, relying solely on the item response scores among traditional DIF techniques poses challenges for researchers and practitioners in interpreting DIF. Recently, response process data, which carry valuable information about examinees' response behaviors, offer an opportunity to further interpret DIF items by examining differences in response processes. This study aims to investigate the potential of response process data features in improving the interpretability of DIF items, with a focus on gender DIF using data from the Programme for International Assessment of Adult Competencies (PIAAC) 2012 computer-based numeracy assessment. We applied random forest and logistic regression with ridge regularization to investigate the association between process data features and DIF items, evaluating the important features to interpret DIF. In addition, we evaluated model performance across varying percentages of DIF items to reflect practical scenarios with different percentages of DIF items. The results demonstrate that the combination of timing features and action-sequence features is informative to reveal the response process differences between groups, thereby enhancing DIF item interpretability. Overall, this study introduces a feasible procedure to leverage response process data to understand and interpret DIF items, shedding light on potential reasons for the low agreement between DIF statistics and expert reviews and revealing potential irrelevant factors to enhance measurement equity.