Edward W. Wolfe

Rater training is essential in ensuring the quality of constructed response scoring. Most of the current knowledge about rater training comes from experimental contexts with an emphasis on short-term effects. Few sources are available for empirical evidence on whether and how raters become more accurate as they gain scoring experiences or what long-term effects training can have. In this study, we addressed this research gap by tracking how the accuracies of new raters change through experience and by examining the impact of an additional training session on their accuracies in scoring calibration and monitoring essays. We found that, on average, raters’ accuracy improved with scoring experience and that individual raters differed in their accuracy trajectories. The estimated average effect of the training was an approximately six percent increase in the calibration essay accuracy. On the other hand, we observed a smaller impact on the monitoring essay accuracy. Our follow-up analysis showed that this differential impact of the additional training on the calibration and monitoring essay accuracy could be accounted for by successful gatekeeping through calibration.

This introductory article describes how constructed response scoring is carried out, particularly the rater monitoring processes and illustrates three potential designs for conducting rater monitoring in an operational scoring project. The introduction also presents a framework for interpreting research conducted by those who study the constructed response scoring process. That framework identifies three classifications of inputs (rater characteristics, response content, and rating context) which typically serve as independent variables in constructed response scoring research as well as three primary outcomes (rating quality, rating speed, and rater attitude) which serve as the dependent variables in those studies. Finally, we explain how each of the articles in this issue can be classified according to that framework.

(2020). Why Should We Care about Human Raters? Applied Measurement in Education: Vol. 33, Using Human Raters in Constructed Response Scoring: Understanding, Predicting, and Modifying Performance, pp. 189-190.

Engelhard (1996) proposed a rater accuracy model (RAM) as a means of evaluating rater accuracy in rating data, but very little research exists to determine the efficacy of that model. The RAM requires a transformation of the raw score data to accuracy measures by comparing rater-assigned scores to true scores. Indices computed based on raw scores also exist for measuring rater effects, but these indices ignore deviations of rater-assigned scores from true scores. This paper demonstrates the efficacy of two versions of the RAM (based on dichotomized and polytomized deviations of rater-assigned scores from true scores) to two versions of raw score rater effect models (i.e., a Rasch partial credit model, PCM, and a Rasch rating scale model, RSM). Simulated data are used to demonstrate the efficacy with which these four models detect and differentiate three rater effects: severity, centrality, and inaccuracy. Results indicate that the RAMs are able to detect, but not differentiate, rater severity and inaccuracy, but not rater centrality. The PCM and RSM, on the other hand, are able to both detect and differentiate all three of these rater effects. However, the RSM and PCM do not take into account true scores and may, therefore, be misleading when pervasive trends exist in the rater-assigned data.

To date, much of the research concerning rater effects has focused on rater severity/leniency. Consequently, other potentially important rater effects have largely ignored by those conducting operational scoring projects. This simulation study compares four rater centrality indices (rater fit, residual-expected correlations, rater slope, and rater threshold variance) in terms of their Type I and Type II error rates under varying levels of centrality magnitude, centrality pervasiveness, and rating scale construction when each of four latent trait models is fitted to the simulated data (Rasch rating scale and partial credit models and the generalized rating scale and partial credit models). Results indicate that the residual-expected correlation may be most appropriately sensitive to rater centrality under most conditions.

A large body of literature exists describing how rater effects may be detected in rating data. In this study, we compared the flag and agreement rates for several rater effects based on calibration of a real data under two psychometric models-the Rasch rating scale model (RSM) and the Rasch testlet-based rater bundle model (RBM). The results show that the RBM provided more accurate diagnoses of rater severity and leniency than do the RSM which is based on the local independence assumption. However, the statistical indicators associated with rater centrality and inaccuracy remain consistent between these two models.

Previous research has investigated the influence of sample size, model misspecification, test length, ability distribution offset, and generating model on the likelihood ratio difference test in applications of item response models. This study extended that research to the evaluation of dimensionality using the multidimensional random coefficients multinomial logit model (MRCMLM). Logistic regression analysis of simulated data reveal that sample size and test length have a large effect on the capacity of the LR difference test to correctly identify unidimensionality, with shorter tests and smaller sample sizes leading to smaller Type I error rates. Higher levels of simulated misfit resulted in fewer incorrect decisions than data with no or little misfit. However, Type I error rates indicate that the likelihood ratio difference test is not suitable under any of the simulated conditions for evaluating dimensionality in applications of the MRCMLM.

Historically, rule-of-thumb critical values have been employed for interpreting fit statistics that depict anomalous person and item response patterns in applications of the Rasch model. Unfortunately, prior research has shown that these values are not appropriate in many contexts. This article introduces a bootstrap procedure for identifying reasonable critical values for Rasch fit statistics and compares the results of that procedure to applications of rule-of-thumb critical values for three example datasets. The results indicate that rule-of-thumb values may over- or under-identify the number of misfitting items or persons.

Cognitive radios (CRs) are recent technological developments that rely on artificial intelligence to adapt a radio's performance to suit environmental demands, such as sharing radio frequencies with other radios. Measuring the performance of the cognitive engines (CEs) that underlie a CR's performance is a challenge for those developing CR technology. This simulation study illustrates how the Rasch model can be applied to the evaluation of CRs. We simulated the responses of 50 CEs to 35 performance tasks and applied the Random Coefficients Multidimensional Multinomial Logit Model (MRCMLM) to those data. Our results indicate that CEs based on different algorithms may exhibit differential performance across manipulated performance task parameters. We found that a multidimensional mixture model may provide the best fit to the simulated data and that the two algorithms simulated may respond to tasks that emphasize achieving high levels of data throughput coupled with lower emphasis on power conservation differently than they do to other combinations of performance task characteristics.