Journal of Educational and Behavioral Statistics

I consider recent attempts to establish standards, principles, and goals for artificial intelligence (AI) through the lens of educational measurement. Distinctions are made between generative AI and AI-adjacent methods and applications of AI in formative versus summative assessment contexts. While expressing optimism about its possibilities, I caution that the examples of truly generative AI in educational testing have the potential to be overexaggerated, that efforts to establish standards for AI should complement the Standards for Educational and Psychological Testing and focus attention on the issues of fairness and social responsibility, and that scientific advance and transparency in the development and application of AI in educational assessment may be incompatible with the competitive marketplace that is funding this development.

The surge of AI in education raises concerns about measurement downsides. Calls for clear standards are warranted. Fortunately, the psychometrics field has a long history of developing relevant standards—like sample invariance and item bias avoidance—crucial for reliable, valid, and interpretable assessments. This established body of knowledge, not unlike traffic laws for self-driving cars, should guide AI assessment development. Measuring new constructs necessitates stronger construct validity research. Instead of rewriting the rulebook, our focus should be on educating AI developers about these standards. This commentary specifically addresses the concern of empowering instructors not with high-stakes testing but with effective item writing through AI. We explore the potential of AI to transform item development, a key area highlighted by researchers. While AI tools offer exciting possibilities for tackling educational challenges, equipping instructors to leverage them effectively remains paramount.

If psychometrics has long concerned itself with validity, reliability, and fairness, then what could psychometrics learn from the cybernetic theories of AI? Through engagement with Burstein’s (2023) Responsible AI Standards, this paper unpacks some paradigmatic differences between psychometrics and cybernetics, points to how recursivity and contingency are both a challenge and opportunity for psychometrics, and how this matters epistemologically, ethically and politically. Following these epistemological differences, the paper raises ethico-political concerns with the promise of the “human-in-the-loop”.

I review opportunities and threats that widely accessible Artificial Intelligence (AI)-powered services present for educational statistics and measurement. Algorithmic and computational advances continue to improve approaches to item generation, scale maintenance, test security, test scoring, and score reporting. Predictable misuses of AI for these purposes will result in biased scores, construct underrepresentation, and differential impact over time. Recent efforts to develop standards for AI use in testing like those of Burstein are promising. I argue that similar efforts to develop AI standards for educational measurement will benefit from increased attention to the context of test use and explicit commitment to ongoing monitoring of bias and scale drift over time.

This commentary examines the Duolingo English Test Responsible AI standards and provides some thoughts on specific ways we can evaluate the use of AI for automated scoring.

A test battery with two different levels of adaptation is presented: a within-subtest level for the selection of the items in the subtests and a between-subtest level to move from one subtest to the next. The battery runs on a two-level model consisting of a regular response model for each of the subtests extended with a second level for the joint distribution of their abilities. The presentation of the model is followed by an optimized MCMC algorithm to update the posterior distribution of each of its ability parameters, select the items to Bayesian optimality, and adaptively move from one subtest to the next. Thanks to extremely rapid convergence of the Markov chain and simple posterior calculations, the algorithm can be used in real-world applications without any noticeable latency. Finally, an empirical study with a battery of short diagnostic subtests is shown to yield score accuracies close to traditional one-level adaptive testing with subtests of double lengths.

Item response theory (IRT) models the relationship between the possible scores on a test item against a test taker’s attainment of the latent trait that the item is intended to measure. In this study, we compare two models for tests with polytomously scored items: the optimal scoring (OS) model, a nonparametric IRT model based on the principles of information theory, and the generalized partial credit (GPC) model, a widely used parametric alternative. We evaluate these models using both simulated and real test data. In the real data examples, the OS model demonstrates superior model fit compared to the GPC model across all analyzed datasets. In our simulation study, the OS model outperforms the GPC model in terms of bias, but at the cost of larger standard errors for the probabilities along the estimated item response functions. Furthermore, we illustrate how surprisal arc length, an IRT scale invariant measure of ability with metric properties, can be used to put scores from vastly different types of IRT models on a common scale. We also demonstrate how arc length can be a viable alternative to sum scores for scoring test takers.

In a randomized trial that collects text as an outcome, traditional approaches for assessing treatment impact require that each document first be manually coded for constructs of interest by human raters. An impact analysis can then be conducted to compare treatment and control groups, using the hand-coded scores as a measured outcome. This process is both time and labor-intensive, which creates a persistent barrier for large-scale assessments of text. Furthermore, enriching one’s understanding of a found impact on text outcomes via secondary analyses can be difficult without additional scoring efforts. The purpose of this article is to provide a pipeline for using machine-based text analytic and data mining tools to augment traditional text-based impact analysis by analyzing impacts across an array of automatically generated text features. In this way, we can explore what an overall impact signifies in terms of how the text has evolved due to treatment. Through a case study based on a recent field trial in education, we show that machine learning can indeed enrich experimental evaluations of text by providing a more comprehensive and fine-grained picture of the mechanisms that lead to stronger argumentative writing in a first- and second-grade content literacy intervention. Relying exclusively on human scoring, by contrast, is a lost opportunity. Overall, the workflow and analytical strategy we describe can serve as a template for researchers interested in performing their own experimental evaluations of text.

Normative studies are needed to obtain norms for comparing individuals with the reference population on relevant clinical or educational measures. Norms can be obtained in an efficient way by regressing the test score on relevant predictors, such as age and sex. When several measures are normed with the same sample, a multivariate regression-based approach must be adopted for at least two reasons: (1) to take into account the correlations between the measures of the same subject, in order to test certain scientific hypotheses and to reduce misclassification of subjects in clinical practice, and (2) to reduce the number of significance tests involved in selecting predictors for the purpose of norming, thus preventing the inflation of the type I error rate. A new multivariate regression-based approach is proposed that combines all measures for an individual through the Mahalanobis distance, thus providing an indicator of the individual’s overall performance. Furthermore, optimal designs for the normative study are derived under five multivariate polynomial regression models, assuming multivariate normality and homoscedasticity of the residuals, and efficient robust designs are presented in case of uncertainty about the correct model for the analysis of the normative sample. Sample size calculation formulas are provided for the new Mahalanobis distance-based approach. The results are illustrated with data from the Maastricht Aging Study (MAAS).

Topic models are mathematical and statistical models used to analyze textual data. The objective of topic models is to gain information about the latent semantic space of a set of related textual data. The semantic space of a set of textual data contains the relationship between documents and words and how they are used. Topic models are becoming more common in educational measurement research as a method for analyzing students’ responses to constructed-response items. Two popular topic models are latent semantic analysis (LSA) and latent Dirichlet allocation (LDA). LSA uses linear algebra techniques, whereas LDA uses an assumed statistical model and generative process. In educational measurement, LSA is often used in algorithmic scoring of essays due to its high reliability and agreement with human raters. LDA is often used as a supplemental analysis to gain additional information about students, such as their thinking and reasoning. This article reviews and compares the LSA and LDA topic models. This article also introduces a methodology for comparing the semantic spaces obtained by the two models and uses a simulation study to investigate their similarities.