Conditional Standard Errors of Measurement for Scale Scores

Standard errors of measurement of scale scores by score level (conditional standard errors of measurement) can be valuable to users of test results. In addition, the Standards for Educational and Psychological Testing (AERA, APA, & NCME, 1985) recommends that conditional standard errors be reported by test developers. Although a variety of procedures are available for estimating conditional standard errors of measurement for raw scores, few procedures exist for estimating conditional standard errors of measurement for scale scores from a single test administration. In this article, a procedure is described for estimating the reliability and conditional standard errors of measurement of scale scores. This method is illustrated using a strong true score model. Practical applications of this methodology are given. These applications include a procedure for constructing score scales that equalize standard errors of measurement along the score scale. Also included are examples of the effects of various nonlinear raw-to-scale score transformations on scale score reliability and conditional standard errors of measurement. These illustrations examine the effects on scale score reliability and conditional standard errors of measurement of (a) the different types of raw-to-scale score transformations (e.g., normalizing scores), (b) the number of scale score points used, and (c) the transformation used to equate alternate forms of a test. All the illustrations use data from the ACT Assessment testing program.     

Conditional Standard Errors of Measurement for Scale Scores Using IRT

An IRT method for estimating conditional standard errors of measurement of scale scores is presented, where scale scores are nonlinear transformations of number-correct scores. The standard errors account for measurement error that is introduced due to rounding scale scores to integers. Procedures for estimating the average conditional standard error of measurement for scale scores and reliability of scale scores are also described. An illustration of the use of the methodology is presented, and the results from the IRT method are compared to the results from a previously developed method that is based on strong true-score theory.     

Reporting Valid and Reliable Overall Scores and Domain Scores

In educational assessment, overall scores obtained by simply averaging a number of domain scores are sometimes reported. However, simply averaging the domain scores ignores the fact that different domains have different score points, that scores from those domains are related, and that at different score points the relationship between overall score and domain score may be different. To report reliable and valid overall scores and domain scores, I investigated the performance of four methods using both real and simulation data: (a) the unidimensional IRT model; (b) the higher-order IRT model, which simultaneously estimates the overall ability and domain abilities; (c) the multidimensional IRT (MIRT) model, which estimates domain abilities and uses the maximum information method to obtain the overall ability; and (d) the bifactor general model. My findings suggest that the MIRT model not only provides reliable domain scores, but also produces reliable overall scores. The overall score from the MIRT maximum information method has the smallest standard error of measurement. In addition, unlike the other models, there is no linear relationship assumed between overall score and domain scores. Recommendations for sizes of correlations between domains and the number of items needed for reporting purposes are provided.     

STANDARD ERRORS OF MEASUREMENT AT DIFFERENT ABILITY LEVELS

Four methods are outlined for estimating or approximating from a single test administration the standard error of measurement of number-right test score at specified ability levels or cutting scores. The methods are illustrated and compared on one set of real test data.     

Conditional Standard Error of Measurement in Prediction

Previous methods for estimating the conditional standard error of measurement (CSEM) at specific score or ability levels are critically discussed, and a brief summary of prior empirical results is given. A new method is developed that avoids theoretical problems inherent in some prior methods, is easy to implement, and estimates not only a quantity analogous to the CSEM at each score but also the conditional standard error of prediction (CSEP) at each score and the conditional true score standard deviation (CTSSD) at each score, The new method differs from previous methods in that previous methods have concentrated on attempting to estimate error variance conditional on a fixed value of true score, whereas the new method considers the variance of observed scores conditional on a fixed value of an observed parallel measurement and decomposes these conditional observed score variances into true and error parts. The new method and several older methods are applied to a variety of tests, and representative results are graphically displayed. The CSEM-Iike estimates produced by the new method are called conditional standard error of measurement in prediction (CSEMP) estimates and are similar to those produced by older methods, but the CSEP estimates produced by the new method offer an alternative interpretation of the accuracy of a test at different scores. Finally, evidence is presented that shows that previous methods can produce dissimilar results and that the shape of the score distribution may influence the way in which the CSEM varies across the score scale.     

Occasions and the Reliability of Classroom Observations: Alternative Conceptualizations and Methods of Analysis

Student–teacher interactions are dynamic relationships that change and evolve over the course of a school year. Measuring classroom quality through observations that focus on these interactions presents challenges when observations are conducted throughout the school year. Variability in observed scores could reflect true changes in the quality of student–teacher interaction or simply reflect measurement error. Classroom observation protocols should be designed to minimize measurement error while allowing measureable changes in the construct of interest. Treating occasions as fixed multivariate outcomes allows true changes to be separated from random measurement error. These outcomes may also be summarized through trend score composites to reflect different types of growth over the school year. We demonstrate the use of multivariate generalizability theory to estimate reliability for trend score composites, and we compare the results to traditional methods of analysis. Reliability estimates computed for average, linear, quadratic, and cubic trend scores from 118 classrooms participating in the MyTeachingPartner study indicate that universe scores account for between 57% and 88% of observed score variance.     

THE CHOICE OF SCALE FOR EDUCATIONAL MEASUREMENT: AN IRT PERSPECTIVE

Two methods of constructing equal-interval scales for educational achievement are discussed: Thurstone's absolute scaling method and Item Response Theory (IRT). Alternative criteria for choosing a scale are contrasted. It is argued that clearer criteria are needed for judging the appropriateness and usefulness of alternative scaling procedures, and more information is needed about the qualities of the different scales that are available. In answer to this second need, some examples are presented of how IRT can be used to examine the properties of scales: It is demonstrated that for observed score scales in common use (i.e., any scores that are influenced by measurement error), (a) systematic errors can be introduced when comparing growth at selected percentiles, and (b) normalizing observed scores will not necessarily produce a scale that is linearly related to an underlying normally distributed true trait.     

Comparing Graphical and Verbal Representations of Measurement Error In Test Score Reports

Research has shown that many educators do not understand the terminology or displays used in test score reports and that measurement error is a particularly challenging concept. We investigated graphical and verbal methods of representing measurement error associated with individual student scores. We created four alternative score reports, each constituting an experimental condition, and randomly assigned them to research participants. We then compared comprehension and preferences across the four conditions. In our main study, we collected data from 148 teachers. For comparison, we studied 98 introductory psychology students. Although we did not detect statistically significant differences across conditions, we found that participants who reported greater comfort with statistics tended to have higher comprehension scores and tended to prefer more informative displays that included variable-width confidence bands for scores. Our data also yielded a wealth of information regarding existing misconceptions about measurement error and about score-reporting conventions.     

A Comparison of Three Scoring Methods for Tests With Selected-Response and Constructed-Response Items

A number of methods for scoring tests with selected-response (SR) and constructed-response (CR) items are available. The selection of a method depends on the requirements of the program, the particular psychometric model and assumptions employed in the analysis of item and score data, and how scores are to be used. This article compares 3 methods: unweighted raw scores, Item Response Theory pattern scores, and weighted raw scores. Student score data from large-scale end-of-course high school tests in Biology and English were used in the comparisons. In the weighted raw score method evaluated in this study, the CR items were weighted so that SR and CR items contributed the same number of points toward the total score. The scoring methods were compared for the total group and for subgroups of students in terms of the resultant scaled score distributions, standard errors of measurement, and proficiency-level classifications. For most of the student ability distribution, the three scoring methods yielded similar results. Some differences in results are noted. Issues to be considered when selecting a scoring method are discussed.     

Supporting Content-Based Feedback in On-Line Writing Evaluation with LSA

This paper describes tests of an automated essay grader and critic that uses Latent Semantic Analysis. Several methods which score the quality of the content in essays are described and tested. These methods are compared against human scores for the essays and the results show that LSA can score as accurately as the humans. Finally, we describe the implementation of the essay grader/critic in an undergraduate course. The outcome showed that students could write and revise their essays on-line, resulting in improved essays. Implications are discussed for the use of the technology in undergraduate courses and how it can provide an effective approach to incorporating more writing both in and outside of the classroom.     

When Does Scale Anchoring Work? A Case Study

Providing information to test takers and test score users about the abilities of test takers at different score levels has been a persistent problem in educational and psychological measurement. Scale anchoring, a technique which describes what students at different points on a score scale know and can do, is a tool to provide such information. Scale anchoring for a test involves a substantial amount of work, both by the statistical analysts and test developers involved with the test. In addition, scale anchoring involves considerable use of subjective judgment, so its conclusions may be questionable. We describe statistical procedures that can be used to determine if scale anchoring is likely to be successful for a test. If these procedures indicate that scale anchoring is unlikely to be successful, then there is little reason to perform a detailed scale anchoring study. The procedures are applied to several data sets from a teachers’ licensing test.     

The Effects of the Number of Scale Points and Non-Normality on the Generalizability Coefficient: A Monte Carlo Study

Increasingly, assessment practitioners use generalizability coefficients to estimate the reliability of scores from performance tasks. Little research, however, examines the relation between the estimation of generalizability coefficients and the number of rubric scale points and score distributions. The purpose of the present research is to inform assessment practitioners of (a) the optimum number of scale points necessary to achieve the best estimates of generalizability coefficients and (b) the possible biases of generalizability coefficients when the distribution of scores is non-normal. Results from this study indicate that the number of scale points substantially affects the generalizability estimates. Generalizability estimates increase as scale points increase, with little bias after scales reach 12 points. Score distributions had little effect on generalizability estimates.     

Standard Error of Measurement

The standard error of measurement (SEM) is the standard deviation of errors of measurement that are associated with test scores from a particular group of examinees. When used to calculate confidence bands around obtained test scores, it can be helpful in expressing the unreliability of individual test scores in an understandable way. Score bands can also be used to interpret intraindividual and interindividual score differences. Interpreters should be wary of over-interpretation when using approximations for correctly calculated score bands. It is recommended that SEMs at various score levels be used in calculating score bands rather than a single SEM value.     

Examining the Precision of Cut Scores Within a Generalizability Theory Framework: A Closer Look at the Item Effect

An Angoff standard setting study generally yields judgments on a number of items by a number of judges (who may or may not be nested in panels). Variability associated with judges (and possibly panels) contributes error to the resulting cut score. The variability associated with items plays a more complicated role. To the extent that the mean item judgments directly reflect empirical item difficulties, the variability in Angoff judgments over items would not add error to the cut score, but to the extent that the mean item judgments do not correspond to the empirical item difficulties, variability in mean judgments over items would add error to the cut score. In this article, we present two generalizability-theory–based analyses of the proportion of the item variance that contributes to error in the cut score. For one approach, variance components are estimated on the probability (or proportion-correct) scale of the Angoff judgments, and for the other, the judgments are transferred to the theta scale of an item response theory model before estimating the variance components. The two analyses yield somewhat different results but both indicate that it is not appropriate to simply ignore the item variance component in estimating the error variance.     

Condensed Mastery Profile Method for Setting Standards for Diagnostic Assessment Systems

The purpose of this study was to develop a standard‐setting method appropriate for use with a diagnostic assessment that produces profiles of student mastery rather than a single raw or scale score value. The condensed mastery profile method draws from established holistic standard‐setting methods to use rounds of range finding and pinpointing to specify cut points between performance levels. Panelists are convened to review profiles of mastery and specify cut points between performance levels based on the total number of skills mastered. Following panelist specification of cut points, a statistical method is implemented to smooth cut points over grades to decrease between‐grade variability. Procedural evidence, including convergence plots, standard errors of pinpointing ratings, and panelist feedback, suggest the condensed mastery profile method is a useful and technically sound approach for setting performance standards for diagnostic assessment systems.     

Studying Score Stability with a Harmonic Regression Family: A Comparison of Three Approaches to Adjustment of Examinee‐Specific Demographic Data

For assessments that use different forms in different administrations, equating methods are applied to ensure comparability of scores over time. Ideally, a score scale is well maintained throughout the life of a testing program. In reality, instability of a score scale can result from a variety of causes, some are expected while others may be unforeseen. The situation is more challenging for assessments that assemble many different forms and deliver frequent administrations per year. Harmonic regression, a seasonal‐adjustment method, has been found useful in achieving the goal of differentiating between possible known sources of variability and unknown sources so as to study score stability for such assessments. As an extension, this paper presents a family of three approaches that incorporate examinees' demographic data into harmonic regression in different ways. A generic evaluation method based on jackknifing is developed to compare the approaches within the family. The three approaches are compared using real data from an international language assessment. Results suggest that all approaches perform similarly and are effective in meeting the goal. The paper also discusses the properties and limitations of the three approaches, along with inferences about score (in)stability based on the harmonic regression results.     

Effects of Scale Transformation and Test-Termination Rule on the Precision of Ability Estimation in Computerized Adaptive Testing

Error indices (bias, standard error of estimation, and root mean squared error) obtained on different measurement scales under different test-termination rules in computerized adaptive testing (CAT) were examined. Four ability estimation methods (maximum likelihood estimation, weighted likelihood estimation, expected a posterior, and maximum a posterior), three measurement scales (θ, number-correct score, and ACT score), and three test-termination rules (fixed length, fixed standard error, and target information) were studied for a real and a generated item pool. The findings indicated that the amount and direction of bias, standard error of estimation, and root mean squared error obtained under different ability estimation methods were influenced both by scale transformations and by test-termination rules in a CAT environment. The implications of these effects for testing programs are discussed.     

Quantifying Error and Uncertainty Reductions in Scaling Functions: An ITEMS Module

This module describes and extends X‐to‐Y regression measures that have been proposed for use in the assessment of X‐to‐Y scaling and equating results. Measures are developed that are similar to those based on prediction error in regression analyses but that are directly suited to interests in scaling and equating evaluations. The regression and scaling function measures are compared in terms of their uncertainty reductions, error variances, and the contribution of true score and measurement error variances to the total error variances. The measures are also demonstrated as applied to an assessment of scaling results for a math test and a reading test. The results of these analyses illustrate the similarity of the regression and scaling measures for scaling situations when the tests have a correlation of at least .80, and also show the extent to which the measures can be adequate summaries of nonlinear regression and nonlinear scaling functions, and of heteroskedastic errors. After reading this module, readers will have a comprehensive understanding of the purposes, uses, and differences of regression and scaling functions.     

A Criterion-Referenced Approach to Student Ratings of Instruction

We developed a criterion-referenced student rating of instruction (SRI) to facilitate formative assessment of teaching. It involves four dimensions of teaching quality that are grounded in current instructional design principles: Organization and structure, Assessment and feedback, Personal interactions, and Academic rigor. Using item response theory and Wright mapping methods, we describe teaching characteristics at various points along the latent continuum for each scale. These maps enable criterion-referenced score interpretation by making an explicit connection between test performance and the theoretical framework. We explain the way our Wright maps can be used to enhance an instructor’s ability to interpret scores and identify ways to refine teaching. Although our work is aimed at improving score interpretation, a criterion-referenced test is not immune to factors that may bias test scores. The literature on SRIs is filled with research on factors unrelated to teaching that may bias scores. Therefore, we also used multilevel models to evaluate the extent to which student and course characteristic may affect scores and compromise score interpretation. Results indicated that student anger and the interaction between student gender and instructor gender are significant effects that account for a small amount of variance in SRI scores. All things considered, our criterion-referenced approach to SRIs is a viable way to describe teaching quality and help instructors refine pedagogy and facilitate course development.     

Multiple choice and true/false tests: reliability measures and some implications of negative marking

The standard error of measurement usefully provides confidence limits for scores in a given test, but is it possible to quantify the reliability of a test with just a single number that allows comparison of tests of different format? Reliability coefficients do not do this, being dependent on the spread of examinee attainment. Better in this regard is a measure produced by dividing the standard error of measurement by the test's ‘reliability length’, the latter defined as the maximum possible score minus the most probable score obtainable by blind guessing alone. This, however, can be unsatisfactory with negative marking (formula scoring), as shown by data on 13 negatively marked true/false tests. In these the examinees displayed considerable misinformation, which correlated negatively with correct knowledge. Negative marking can improve test reliability by penalizing such misinformation as well as by discouraging guessing. Reliability measures can be based on idealized theoretical models instead of on test data. These do not reflect the qualities of the test items, but can be focused on specific test objectives (e.g. in relation to cut‐off scores) and can be expressed as easily communicated statements even before tests are written.