A Critical Review of Some Qualitative Research Methods Used to Explore Rater Cognition

Internationally, many assessment systems rely predominantly on human raters to score examinations. Arguably, this facilitates the assessment of multiple sophisticated educational constructs, strengthening assessment validity. It can introduce subjectivity into the scoring process, however, engendering threats to accuracy. The present objectives are to examine some key qualitative data collection methods used internationally to research this potential trade‐off, and to consider some theoretical contexts within which the methods are usable. Self‐report methods such as Kelly's Repertory Grid, think aloud, stimulated recall, and the NASA task load index have yielded important insights into the competencies needed for scoring expertise, as well as the sequences of mental activity that scoring typically involves. Examples of new data and of recent studies are used to illustrate these methods’ strengths and weaknesses. This investigation has significance for assessment designers, developers and administrators. It may inform decisions on the methods’ applicability in American and other rater cognition research contexts.     

Rater effects: Ego engagement in rater decision-making

The decision-making behaviors of 8 raters when scoring 39 persuasive and 39 narrative essays written by second language learners were examined, first using Rasch analysis and then, through think aloud protocols. Results based on Rasch analysis and think aloud protocols recorded by raters as they were scoring holistically and analytically suggested that rater background may have contributed to rater expectations that might explain individual differences in the application of the performance criteria of the rubrics when rating essays. The results further suggested that rater ego engagement with the text and/or author may have helped mitigate rater severity and that self-monitoring behaviors by raters may have had a similar mitigating effect.     

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 Effects on Essay Scoring: A Multilevel Analysis of Severity Drift,Central Tendency,and Rater Experience

This study examined rater effects on essay scoring in an operational monitoring system from England's 2008 national curriculum English writing test for 14‐year‐olds. We fitted two multilevel models and analyzed: (1) drift in rater severity effects over time; (2) rater central tendency effects; and (3) differences in rater severity and central tendency effects by raters’ previous rating experience. We found no significant evidence of rater drift and, while raters with less experience appeared more severe than raters with more experience, this result also was not significant. However, we did find that there was a central tendency to raters’ scoring. We also found that rater severity was significantly unstable over time. We discuss the theoretical and practical questions that our findings raise.     

Automated Scoring of Constructed‐Response Science Items: Prospects and Obstacles

Content‐based automated scoring has been applied in a variety of science domains. However, many prior applications involved simplified scoring rubrics without considering rubrics representing multiple levels of understanding. This study tested a concept‐based scoring tool for content‐based scoring, c‐rater?, for four science items with rubrics aiming to differentiate among multiple levels of understanding. The items showed moderate to good agreement with human scores. The findings suggest that automated scoring has the potential to score constructed‐response items with complex scoring rubrics, but in its current design cannot replace human raters. This article discusses sources of disagreement and factors that could potentially improve the accuracy of concept‐based automated scoring.     

Conceptualizing Rater Judgments and Rating Processes for Rater‐Mediated Assessments

Rater‐mediated assessments exhibit scoring challenges due to the involvement of human raters. The quality of human ratings largely determines the reliability, validity, and fairness of the assessment process. Our research recommends that the evaluation of ratings should be based on two aspects: a theoretical model of human judgment and an appropriate measurement model for evaluating these judgments. In rater‐mediated assessments, the underlying constructs and response processes may require the use of different rater judgment models and the application of different measurement models. We describe the use of Brunswik's lens model as an organizing theme for conceptualizing human judgments in rater‐mediated assessments. The constructs vary depending on which distal variables are identified in the lens models for the underlying rater‐mediated assessment. For example, one lens model can be developed to emphasize the measurement of student proficiency, while another lens model can stress the evaluation of rater accuracy. Next, we describe two measurement models that reflect different response processes (cumulative and unfolding) from raters: Rasch and hyperbolic cosine models. Future directions for the development and evaluation of rater‐mediated assessments are suggested.     

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.     

Autoscoring Essays Based on Complex Networks

This article presents a novel method, the Complex Dynamics Essay Scorer (CDES), for automated essay scoring using complex network features. Texts produced by college students in China were represented as scale‐free networks (e.g., a word adjacency model) from which typical network features, such as the in‐/out‐degrees, clustering coefficient (CC), and dynamic networks, were obtained. The CDES integrates the classical concepts of network feature representation and essay score series variation. Several experiments indicated that the network measures different essay qualities and can be clearly demonstrated to develop complex networks for autoscoring tasks. The average agreement of the CDES and human rater scores was 86.5%, and the average Pearson correlation was .77. The results indicate that the CDES produced functional complex systems and autoscored Chinese essays in a method consistent with human raters. Our research suggests potential applications in other areas of educational assessment.     

Rater strategies for reaching agreement on pupil text quality

Novice members of a Norwegian national rater panel tasked with assessing Year 8 pupils’ written texts were studied during three successive preparation sessions (2011–2012). The purpose was to investigate how the raters successfully make use of different decision-making strategies in an assessment situation where pre-set criteria and standards give a rather strict framework. The data sources were the raters’ pair assessment dialogues. The analysis shows that the raters use a ‘shared standards strategy’, but when reaching agreement on text quality they also seem to make very good use of assessment strategies related to their work as writing teachers. Moreover, asymmetries in knowledge and participation among raters contribute to creating an image of writing assessment as a challenging hermeneutic practice. It is suggested that future rater preparation would gain from being attentive to the internalised assessment practices teachers bring to the fore when working as raters.     

The impact of the scoring system of a large-scale standardized EFL writing assessment on its score variability and reliability: Implications for assessment policy makers

Using generalizability (G-) theory and rater interviews as research methods, this study examined the impact of the current scoring system of the CET-4 (College English Test Band 4, a high-stakes national standardized EFL assessment in China) writing on its score variability and reliability. One hundred and twenty CET-4 essays written by 60 non-English major undergraduate students at one Chinese university were scored holistically by 35 experienced CET-4 raters using the authentic CET-4 scoring rubric. Ten purposively selected raters were further interviewed for their views on how the current scoring system could impact its score variability and reliability. The G-theory results indicated that the current single-task and single-rater holistic scoring system would not be able to yield acceptable generalizability and dependability coefficients. The rater interview results supported the quantitative findings. Important implications for the CET-4 writing assessment policy in China are discussed.     

Modeling Rater Response Processes in Evaluating Score Meaning

Rater‐mediated assessments require the evaluation of the accuracy and consistency of the inferences made by the raters to ensure the validity of score interpretations and uses. Modeling rater response processes allows for a better understanding of how raters map their representations of the examinee performance to their representation of the scoring criteria. Validity of score meaning is affected by the accuracy of raters' representations of examinee performance and the scoring criteria, and the accuracy of the mapping process. Methodological advances and applications that model rater response processes, rater accuracy, and rater consistency inform the design, scoring, interpretations, and uses of rater‐mediated assessments.     

Scoring Stability in a Large‐Scale Assessment Program: A Longitudinal Analysis of Leniency/Severity Effects

Although much attention has been given to rater effects in rater‐mediated assessment contexts, little research has examined the overall stability of leniency and severity effects over time. This study examined longitudinal scoring data collected during three consecutive administrations of a large‐scale, multi‐state summative assessment program. Multilevel models were used to assess the overall extent of rater leniency/severity during scoring and examine the extent to which leniency/severity effects were stable across the three administrations. Model results were then applied to scaled scores to estimate the impact of the stability of leniency/severity effects on students’ scores. Results showed relative scoring stability across administrations in mathematics. In English language arts, short constructed response items showed evidence of slightly increasing severity across administrations, while essays showed mixed results: evidence of both slightly increasing severity and moderately increasing leniency over time, depending on trait. However, when model results were applied to scaled scores, results revealed rater effects had minimal impact on students’ scores.     

Rater Cognition: Implications for Validity

The scoring process is critical in the validation of tests that rely on constructed responses. Documenting that readers carry out the scoring in ways consistent with the construct and measurement goals is an important aspect of score validity. In this article, rater cognition is approached as a source of support for a validity argument for scores based on constructed responses, whether such scores are to be used on their own or as the basis for other scoring processes, for example, automated scoring.     

大学英语写作评分方法对评分者严厉程度的影响——整体评分法和分析评分法的对比分析

评分标准在写作测试中非常重要,使用不同的评分方法会影响评卷者的评分行为。研究显示,虽然整体法和分析法两种英语写作评分方法都可靠,但是在两种评分中,评卷者的严厉程度以及考生的写作成绩发生很大变化。总体上,整体法评分中,评卷者的严厉程度趋于一致,接近理想值;分析法评分中,考生的写作成绩更高,同时评卷者的严厉程度也存在显著差异。因而,在决定考生前途命运的重大考试中,整体评分法更受推崇。     

Using Multigroup Confirmatory Factor Analysis to Test Measurement Invariance in Raters: A Clinical Skills Examination Application

Psychometric models based on structural equation modeling framework are commonly used in many multiple-choice test settings to assess measurement invariance of test items across examinee subpopulations. The premise of the current article is that they may also be useful in the context of performance assessment tests to test measurement invariance of raters. The modeling approach and how it can be used for performance tests with less than optimal rater designs are illustrated using a data set from a performance test designed to measure medical students’ patient management skills. The results suggest that group-specific rater statistics can help spot differences in rater performance that might be due to rater bias, identify specific weaknesses and strengths of individual raters, and enhance decisions related to future task development, rater training, and test scoring processes.     

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.     

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.     

A Latent Class Signal Detection Model for Rater Scoring with Ordered Perceptual Distributions

In signal detection rater models for constructed response (CR) scoring, it is assumed that raters discriminate equally well between different latent classes defined by the scoring rubric. An extended model that relaxes this assumption is introduced; the model recognizes that a rater may not discriminate equally well between some of the scoring classes. The extension recognizes a different type of rater effect and is shown to offer useful tests and diagnostic plots of the equal discrimination assumption, along with ways to assess rater accuracy and various rater effects. The approach is illustrated with an application to a large‐scale language test.     

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.