Conventional methods for assessing the validity and reliability of situational judgment test (SJT) scores have proven to be inadequate. For example, factor analysis techniques typically lead to nonsensical solutions, and assumptions underlying Cronbach’s alpha coefficient are violated due to the multidimensional nature of SJTs. In the current article, we describe how cognitive diagnosis models (CDMs) provide a new approach that not only overcomes these limitations but that also offers extra advantages for scoring and better understanding SJTs. The analysis of the Q-matrix specification, model fit, and model parameter estimates provide a greater wealth of information than traditional procedures do. Our proposal is illustrated using data taken from a 23-item SJT that presents situations about student-related issues. Results show that CDMs are useful tools for scoring tests, like SJTs, in which multiple knowledge, skills, abilities, and other characteristics are required to correctly answer the items. SJT classifications were reliable and significantly related to theoretically relevant variables. We conclude that CDM might help toward the exploration of the nature of the constructs underlying SJT, one of the principal challenges in SJT research.

The most commonly employed item selection rule in a computerized adaptive test (CAT) is that of selecting the item with the maximum Fisher information for the estimated trait level. This means a highly unbalanced distribution of item-exposure rates, a high overlap rate among examinees and, for item bank management, strong pressure to replace items with a high discrimination parameter in the bank. An alternative for mitigating these problems involves, at the beginning of the test, basing item selection mainly on randomness. As the test progresses, the weight of information in the selection increases. In the present work we study, for two selection rules, the progressive methods (Revuelta & Ponsoda, 1998) and the proportional method (Segall, 2004a), different functions that define the weight of the random component according to the position in the test of the item to be administered. The functions were tested in simulated item banks and in an operative bank. We found that both the progressive and the proportional methods tolerate a high weight of the random component with minimal or zero loss of accuracy, while bank security and maintenance are improved.

Research related to the fit evaluation at the item level involving cognitive diagnosis models (CDMs) has been scarce. According to the parsimony principle, balancing goodness of fit against model complexity is necessary. General CDMs require a larger sample size to be estimated reliably, and can lead to worse attribute classification accuracy than the appropriate reduced models when the sample size is small and the item quality is poor, which is typically the case in many empirical applications. The main purpose of this study was to systematically examine the statistical properties of four inferential item-fit statistics: , the likelihood ratio (LR) test, the Wald (W) test, and the Lagrange multiplier (LM) test. To evaluate the performance of the statistics, a comprehensive set of factors, namely, sample size, correlational structure, test length, item quality, and generating model, is systematically manipulated using Monte Carlo methods. Results show that the statistic has unacceptable power. Type I error and power comparisons favor LR and W tests over the LM test. However, all the statistics are highly affected by the item quality. With a few exceptions, their performance is only acceptable when the item quality is high. In some cases, this effect can be ameliorated by an increase in sample size and test length. This implies that using the above statistics to assess item fit in practical settings when the item quality is low remains a challenge.

In a typical study comparing the relative efficiency of two item selection rules in computerized adaptive testing, the common result is that they simultaneously differ in accuracy and security, making it difficult to reach a conclusion on which is the more appropriate rule. This study proposes a strategy to conduct a global comparison of two or more selection rules. A plot showing the performance of each selection rule for several maximum exposure rates is obtained and the whole plot is compared with other rule plots. The strategy was applied in a simulation study with fixed-length CATs for the comparison of six item selection rules: the point Fisher information, Fisher information weighted by likelihood, Kullback-Leibler weighted by likelihood, maximum information stratification with blocking, progressive and proportional methods. Our results show that there is no optimal rule for any overlap value or root mean square error (RMSE). The fact that a rule, for a given level of overlap, has lower RMSE than another does not imply that this pattern holds for another overlap rate. A fair comparison of the rules requires extensive manipulation of the maximum exposure rates. The best methods were the Kullback-Leibler weighted by likelihood, the proportional method, and the maximum information stratification method with blocking.

There has been an increase of interest in psychometric models referred to as cognitive diagnosis models (CDMs). A critical concern is in selecting the most appropriate model at the item level. Several tests for model comparison have been employed, which include the likelihood ratio (LR) and the Wald (W) tests. Although the LR test is relatively more robust than the W test, the current implementation of the LR test is very time consuming, given that it requires calibrating many different models and comparing them to the general model. In this article, we introduce the two-step LR test (2LR), an approximation to the LR test based on a two-step estimation procedure under the generalized deterministic inputs, noisy, "and" gate (G-DINA) model framework, the two-step LR test (2LR). The 2LR test is shown to have similar performance as the LR test. This approximation only requires calibration of the more general model, so that this statistic may be easily applied in empirical research.