Classical,Generalizability, and Multifaceted Rasch Detection of Interrater Variability in Large,Sparse Data Sets

Classical test theory (CTT), generalizability theory (GT), and multi-faceted Rasch model (MFRM) approaches to detecting and correcting for rater variability were compared. Each of 4,930 students' responses on an English examination was graded on 9 scales by 3 raters drawn from a pool of 70. CTT and MFRM indicated substantial variation among raters; the MFRM analysis identified far more raters as different than the CTT analysis did. In contrast, the GT rater variance component and the Rasch histograms suggested little rater variation. CTT and MFRM correction procedures both produced different scores for more than 50% of the examinees, but 75% of the examinees received identical results after each correction. The demonstrated value of a correction for systems of well-trained multiple graders has implications for all systems in which subjective scoring is used.     

Effects of Assigning Raters to Items

Student responses to a large number of constructed response items in three Math and three Reading tests were scored on two occasions using three ways of assigning raters: single reader scoring, a different reader for each response (item-specific), and three readers each scoring a rater item block (RIB) containing approximately one-third of a student's responses. Multiple group confirmatory factor analyses indicated that the three types of total scores were most frequently tau-equivalent. Factor models fitted on the item responses attributed differences in scores to correlated ratings incurred by the same reader scoring multiple responses. These halo effects contributed to significantly increased single reader mean total scores for three of the tests. The similarity of scores for item-specific and RIB scoring suggests that the effect of rater bias on an examinee's set of responses may be minimized with the use of multiple readers though fewer than the number of items.     

Predicting Operational Rater‐Type Classifications Using Rasch Measurement Theory and Random Forests: A Music Performance Assessment Perspective

The purpose of this study was to build a Random Forest supervised machine learning model in order to predict musical rater‐type classifications based upon a Rasch analysis of raters’ differential severity/leniency related to item use. Raw scores (N = 1,704) from 142 raters across nine high school solo and ensemble festivals (grades 9–12) were collected using a 29‐item Likert‐type rating scale embedded within five domains (tone/intonation, n = 6; balance, n = 5; interpretation, n = 6; rhythm, n = 6; and technical accuracy, n = 6). Data were analyzed using a Many Facets Rasch Partial Credit Model. An a priori k‐means cluster analysis of 29 differential rater functioning indices produced a discrete feature vector that classified raters into one of three distinct rater‐types: (a) syntactical rater‐type, (b) expressive rater‐type, or (c) mental representation rater‐type. Results of the initial Random Forest model resulted in an out‐of‐bag error rate of 5.05%, indicating that approximately 95% of the raters were correctly classified. After tuning a set of three hyperparameters (n_tree, m_try, and node size), the optimized model demonstrated an improved out‐of‐bag error rate of 2.02%. Implications for improvements in assessment, research, and rater training in the field of music education are discussed.     

ESTIMATING THE RELIABILITY, VALIDITY, AND INVALIDITY OF ESSAY RATINGS

In an essay rating study multiple ratings may be obtained by having different raters judge essays or by having the same rater(s) repeat the judging of essays. An important question in the analysis of essay ratings is whether multiple ratings, however obtained, may be assumed to represent the same true scores. When different raters judge the same essays only once, it is impossible to answer this question. In this study 16 raters judged 105 essays on two occasions; hence, it was possible to test assumptions about true scores within the framework of linear structural equation models. It emerged that the ratings of a given rater on the two occasions represented the same true scores. However, the ratings of different raters did not represent the same true scores. The estimated intercorrelations of the true scores of different raters ranged from .415 to .910. Parameters of the best fitting model were used to compute coefficients of reliability, validity, and invalidity. The implications of these coefficients are discussed.     

Evaluation of construct-irrelevant variance yielded by machine and human scoring of a science teacher PCK constructed response assessment

Machine learning has been frequently employed to automatically score constructed response assessments. However, there is a lack of evidence of how this predictive scoring approach might be compromised by construct-irrelevant variance (CIV), which is a threat to test validity. In this study, we evaluated machine scores and human scores with regard to potential CIV. We developed two assessment tasks targeting science teacher pedagogical content knowledge (PCK); each task contains three video-based constructed response questions. 187 in-service science teachers watched the videos with each had a given classroom teaching scenario and then responded to the constructed-response items. Three human experts rated the responses and the human-consent scores were used to develop machine learning algorithms to predict ratings of the responses. Including the machine as another independent rater, along with the three human raters, we employed the many-facet Rasch measurement model to examine CIV due to three sources: variability of scenarios, rater severity, and rater sensitivity of the scenarios. Results indicate that variability of scenarios impacts teachers’ performance, but the impact significantly depends on the construct of interest; for each assessment task, the machine is always the most severe rater, compared to the three human raters. However, the machine is less sensitive than the human raters to the task scenarios. This means the machine scoring is more consistent and stable across scenarios within each of the two tasks.     

Rater accuracy and training group effects in Expert- and Supervisor-based monitoring systems

This study evaluated rater accuracy with rater-monitoring data from high stakes examinations in England. Rater accuracy was estimated with cross-classified multilevel modelling. The data included face-to-face training and monitoring of 567 raters in 110 teams, across 22 examinations, giving a total of 5500 data points. Two rater-monitoring systems (Expert consensus scores and Supervisor judgement of correct scores) were utilised for all raters. Results showed significant group training (table leader) effects upon rater accuracy and these were greater in the expert consensus score monitoring system. When supervisor judgement methods of monitoring were used, differences between training teams (table leader effects) were underestimated. Supervisor-based judgements of raters’ accuracies were more widely dispersed than in the Expert consensus monitoring system. Supervisors not only influenced their teams’ scoring accuracies, they overestimated differences between raters’ accuracies, compared with the Expert consensus system. Systems using supervisor judgements of correct scores and face-to-face rater training are, therefore, likely to underestimate table leader effects and overestimate rater effects.     

The Effects of Incomplete Rating Designs in Combination With Rater Effects

Researchers have explored a variety of topics related to identifying and distinguishing among specific types of rater effects, as well as the implications of different types of incomplete data collection designs for rater‐mediated assessments. In this study, we used simulated data to examine the sensitivity of latent trait model indicators of three rater effects (leniency, central tendency, and severity) in combination with different types of incomplete rating designs (systematic links, anchor performances, and spiral). We used the rating scale model and the partial credit model to calculate rater location estimates, standard errors of rater estimates, model–data fit statistics, and the standard deviation of rating scale category thresholds as indicators of rater effects and we explored the sensitivity of these indicators to rater effects under different conditions. Our results suggest that it is possible to detect rater effects when each of the three types of rating designs is used. However, there are differences in the sensitivity of each indicator related to type of rater effect, type of rating design, and the overall proportion of effect raters. We discuss implications for research and practice related to rater‐mediated assessments.     

Writing evaluation: rater and task effects on the reliability of writing scores for children in Grades 3 and 4

We examined how raters and tasks influence measurement error in writing evaluation and how many raters and tasks are needed to reach a desirable level of .90 and .80 reliabilities for children in Grades 3 and 4. A total of 211 children (102 boys) were administered three tasks in narrative and expository genres, respectively, and their written compositions were evaluated in widely used evaluation methods for developing writers: holistic scoring, productivity, and curriculum-based writing scores. Results showed that 54 and 52% of variance in narrative and expository compositions were attributable to true individual differences in writing. Students’ scores varied largely by tasks (30.44 and 28.61% of variance), but not by raters. To reach the reliability of .90, multiple tasks and raters were needed, and for the reliability of .80, a single rater and multiple tasks were needed. These findings offer important implications about reliably evaluating children’s writing skills, given that writing is typically evaluated by a single task and a single rater in classrooms and even in some state accountability systems.     

An Investigation of Rater Cognition in the Assessment of Projects

In the United Kingdom, the majority of national assessments involve human raters. The processes by which raters determine the scores to award are central to the assessment process and affect the extent to which valid inferences can be made from assessment outcomes. Thus, understanding rater cognition has become a growing area of research in the United Kingdom. This study investigated rater cognition in the context of the assessment of school‐based project work for high‐stakes purposes. Thirteen teachers across three subjects were asked to “think aloud” whilst scoring example projects. Teachers also completed an internal standardization exercise. Nine professional raters across the same three subjects standardized a set of project scores whilst thinking aloud. The behaviors and features attended to were coded. The data provided insights into aspects of rater cognition such as reading strategies, emotional and social influences, evaluations of features of student work (which aligned with scoring criteria), and how overall judgments are reached. The findings can be related to existing theories of judgment. Based on the evidence collected, the cognition of teacher raters did not appear to be substantially different from that of professional raters.     

Statistically Comparing the Performance of Multiple Automated Raters Across Multiple Items

Automated scoring systems are typically evaluated by comparing the performance of a single automated rater item-by-item to human raters. This presents a challenge when the performance of multiple raters needs to be compared across multiple items. Rankings could depend on specifics of the ranking procedure; observed differences could be due to random sampling of items and/or responses in the validation sets. Any statistical hypothesis test of the differences in rankings needs to be appropriate for use with rater statistics and adjust for multiple comparisons. This study considered different statistical methods to evaluate differences in performance across multiple raters and items. These methods are illustrated leveraging data from the 2012 Automated Scoring Assessment Prize competitions. Using average rankings to test for significant differences in performance between automated and human raters, findings show that most automated raters did not perform statistically significantly different from human-to-human inter-rater agreement for essays but they did perform differently on short-answer items. Differences in average rankings between most automated raters were not statistically significant, even when their observed performance differed substantially.     

Estimating Variance Components from Sparse Data Matrices in Large-Scale Educational Assessments

In generalizability theory studies in large-scale testing contexts, sometimes a facet is very sparsely crossed with the object of measurement. For example, when assessments are scored by human raters, it may not be practical to have every rater score all students. Sometimes the scoring is systematically designed such that the raters are consistently grouped throughout the scoring, so that the data can be analyzed as raters nested within teams. Other times, rater pairs are randomly assigned for each student, such that each rater is paired with many other raters at different times. One possibility for this scenario is to treat the data as if raters were nested within students. Because the raters are not truly independent across all students, the resulting variance components could be somewhat biased. This study illustrates how the bias will tend to be small in large-scale studies.     

Rater calibration when observational assessment occurs at large scale: Degree of calibration and characteristics of raters associated with calibration

Observational assessment is used to study program and teacher effectiveness across large numbers of classrooms, but training a workforce of raters that can assign reliable scores when observations are used in large-scale contexts can be challenging and expensive. Limited data are available to speak to the feasibility of training large numbers of raters to calibrate to an observation tool, or the characteristics of raters associated with calibration. This study reports on the success of rater calibration across 2093 raters trained by the Office of Head Start (OHS) in 2008–2009 on the Classroom Assessment Scoring System (CLASS), and for a subsample of 704 raters, characteristics that predict their calibration. Findings indicate that it is possible to train large numbers of raters to calibrate to an observation tool, and rater beliefs about teachers and children predicted the degree of calibration. Implications for large-scale observational assessments are discussed.     

Real-Time Feedback on Rater Drift in Constructed-Response Items: An Example From the Golden State Examination

In this study, patterns of variation in severities of a group of raters over time or so-called "rater drift" was examined when raters scored an essay written under examination conditions. At the same time feedback was given to rater leaders (called "table leaders") who then interpreted the feedback and reported to the raters. Rater severities in five successive periods were estimated using a modified linear logistic test model (LLTM, Fischer, 1973) approach. It was found that the raters did indeed drift towards the mean, but a planned comparision of the feedback with a control condition was not successful; it was believed that this was due to contamination at the table leader level. A series of models was also estimated designed to detect other types of rater effects beyond severity: a tendency to use extreme scores, and tendency to prefer certain categories. The models for these effects were found to be showing significant improvement in fit, implying that these effects were indeed present, although they were difficult to detect in relatively short time periods.     

A Hierarchical Rater Model for Constructed Responses,with a Signal Detection Rater Model

The hierarchical rater model (HRM) re‐cognizes the hierarchical structure of data that arises when raters score constructed response items. In this approach, raters’ scores are not viewed as being direct indicators of examinee proficiency but rather as indicators of essay quality; the (latent categorical) quality of an examinee's essay in turn serves as an indicator of the examinee's proficiency, thus yielding a hierarchical structure. Here it is shown that a latent class model motivated by signal detection theory (SDT) is a natural candidate for the first level of the HRM, the rater model. The latent class SDT model provides measures of rater precision and various rater effects, above and beyond simply severity or leniency. The HRM‐SDT model is applied to data from a large‐scale assessment and is shown to provide a useful summary of various aspects of the raters’ performance.     

Captions, consistency, creativity, and the consensual assessment technique: New evidence of reliability

The consensual assessment technique (CAT) is a measurement tool for creativity research in which appropriate experts evaluate creative products [Amabile, T. M. (1996). Creativity in context: Update to the social psychology of creativity. Boulder, CO: Westview]. However, the CAT is hampered by the time-consuming nature of the products (asking participants to write stories or draw pictures) and the ratings (getting appropriate experts). This study examined the reliability of ratings of sentence captions. Specifically, four raters evaluated 12 captions written by 81 undergraduates. The purpose of the study was to see whether the CAT could provide reliable ratings of captions across raters and across multiple captions and, if so, how many such captions would be required to generate reliable scores, and how many judges would be needed? Using generalizability theory, we found that captions appear to be a useful way of measuring creativity with a reasonable level of reliability in the frame of CAT.     

The Utility of Multiple Raters and Tasks in Science Performance Assessments

Concern about the education system has increasingly focused on achievement outcomes and the role of assessment in school performance. Our research with fifth and eighth graders in California explored several issues regarding student performance and rater reliability on hands-on tasks that were administered as part of a field test of a statewide assessment program in science. This research found that raters can produce reliable scores for hands-on tests of science performance. However, the reliability of performance test scores per hour of testing time is quite low relative to multiple-choice tests. Reliability can be improved substantially by adding more tasks (and testing time). Using more than one rater per task produces only a very small improvement in the reliability of a student's total score across tasks. These results were consistent across both grade levels, and they echo the findings of past research.     

Accounting for Rater Effects With the Hierarchical Rater Model Framework When Scoring Simple Structured Constructed Response Tests

Many large‐scale assessments are designed to yield two or more scores for an individual by administering multiple sections measuring different but related skills. Multidimensional tests, or more specifically, simple structured tests, such as these rely on multiple multiple‐choice and/or constructed responses sections of items to generate multiple scores. In the current article, we propose an extension of the hierarchical rater model (HRM) to be applied with simple structured tests with constructed response items. In addition to modeling the appropriate trait structure, the multidimensional HRM (M‐HRM) presented here also accounts for rater severity bias and rater variability or inconsistency. We introduce the model formulation, test parameter recovery with a focus on latent traits, and compare the M‐HRM to other scoring approaches (unidimensional HRMs and a traditional multidimensional item response theory model) using simulated and empirical data. Results show more precise scores under the M‐HRM, with a major improvement in scores when incorporating rater effects versus ignoring them in the traditional multidimensional item response theory model.