Multimodal Data Fusion to Detect Preknowledge Test-Taking Behavior Using Machine Learning

In various fields, including college admission, medical board certifications, and military recruitment, high-stakes decisions are frequently made based on scores obtained from large-scale assessments. These decisions necessitate precise and reliable scores that enable valid inferences to be drawn about test-takers. However, the ability of such tests to provide reliable, accurate inference on a test-taker's performance could be jeopardized by aberrant test-taking practices, for instance, practicing real items prior to the test. As a result, it is crucial for administrators of such assessments to develop strategies that detect potential aberrant test-takers after data collection. The aim of this study is to explore the implementation of machine learning methods in combination with multimodal data fusion strategies that integrate bio-information technology, such as eye-tracking, and psychometric measures, including response times and item responses, to detect aberrant test-taking behaviors in technology-assisted remote testing settings.

Using item scores and response times in person-fit assessment

The use of joint models for item scores and response times is becoming increasingly popular in educational and psychological testing. In this paper, we propose two new person-fit statistics for such models in order to detect aberrant behaviour. The first statistic is computed by combining two existing person-fit statistics: one for the item scores, and one for the item response times. The second statistic is computed directly using the likelihood function of the joint model. Using detailed simulations, we show that the empirical null distributions of the new statistics are very close to the theoretical null distributions, and that the new statistics tend to be more powerful than several existing statistics for item scores and/or response times. A real data example is also provided using data from a licensure examination.

Detecting Preknowledge Cheating via Innovative Measures: A Mixture Hierarchical Model for Jointly Modeling Item Responses, Response Times, and Visual Fixation Counts

Preknowledge cheating jeopardizes the validity of inferences based on test results. Many methods have been developed to detect preknowledge cheating by jointly analyzing item responses and response times. Gaze fixations, an essential eye-tracker measure, can be utilized to help detect aberrant testing behavior with improved accuracy beyond using product and process data types in isolation. As such, this study proposes a mixture hierarchical model that integrates item responses, response times, and visual fixation counts collected from an eye-tracker (a) to detect aberrant test takers who have different levels of preknowledge and (b) to account for nuances in behavioral patterns between normally-behaved and aberrant examinees. A Bayesian approach to estimating model parameters is carried out via an MCMC algorithm. Finally, the proposed model is applied to experimental data to illustrate how the model can be used to identify test takers having preknowledge on the test items.

R-package LNIRT for joint modeling of response accuracy and times

In computer-based testing it has become standard to collect response accuracy (RA) and response times (RTs) for each test item. IRT models are used to measure a latent variable (e.g., ability, intelligence) using the RA observations. The information in the RTs can help to improve routine operations in (educational) testing, and provide information about speed of working. In modern applications, the joint models are needed to integrate RT information in a test analysis. The R-package LNIRT supports fitting joint models through a user-friendly setup which only requires specifying RA, RT data, and the total number of Gibbs sampling iterations. More detailed specifications of the analysis are optional. The main results can be reported through the summary functions, but output can also be analysed with Markov chain Monte Carlo (MCMC) output tools (i.e., coda, mcmcse). The main functionality of the LNIRT package is illustrated with two real data applications.

Bridging Models of Biometric and Psychometric Assessment: A Three-Way Joint Modeling Approach of Item Responses, Response Times, and Gaze Fixation Counts

Recently, joint models of item response data and response times have been proposed to better assess and understand test takers' learning processes. This article demonstrates how biometric information such as gaze fixation counts obtained from an eye-tracking machine can be integrated into the measurement model. The proposed joint modeling framework accommodates the relations among a test taker's latent ability, working speed and test engagement level via a person-side variance-covariance structure, while simultaneously permitting the modeling of item difficulty, time-intensity, and the engagement intensity through an item-side variance-covariance structure. A Bayesian estimation scheme is used to fit the proposed model to data. Posterior predictive model checking based on three discrepancy measures corresponding to various model components are introduced to assess model-data fit. Findings from a Monte Carlo simulation and results from analyzing experimental data demonstrate the utility of the model.

Joint Modeling of Response Accuracy and Time in Between-Item Multidimensional Tests Based on Bi-Factor Model

With the popularity of computer-based testing (CBT), it is easier to collect item response times (RTs) in psychological and educational assessments. RTs can provide an important source of information for respondents and tests. To make full use of RTs, the researchers have invested substantial effort in developing statistical models of RTs. Most of the proposed models posit a unidimensional latent speed to account for RTs in tests. In psychological and educational tests, many tests are multidimensional, either deliberately or inadvertently. There may be general effects in between-item multidimensional tests. However, currently there exists no RT model that considers the general effects to analyze between-item multidimensional test RT data. Also, there is no joint hierarchical model that integrates RT and response accuracy (RA) for evaluating the general effects of between-item multidimensional tests. Therefore, a bi-factor joint hierarchical model using between-item multidimensional test is proposed in this study. The simulation indicated that the Hamiltonian Monte Carlo (HMC) algorithm works well in parameter recovery. Meanwhile, the information criteria showed that the bi-factor hierarchical model (BFHM) is the best fit model. This means that it is necessary to take into consideration the general effects (general latent trait) and the multidimensionality of the RT in between-item multidimensional tests.

On the Speed Sensitivity Parameter in the Lognormal Model for Response Times and Implications for High-Stakes Measurement Practice

In high-stakes testing, often multiple test forms are used and a common time limit is enforced. Test fairness requires that ability estimates must not depend on the administration of a specific test form. Such a requirement may be violated if speededness differs between test forms. The impact of not taking speed sensitivity into account on the comparability of test forms regarding speededness and ability estimation was investigated. The lognormal measurement model for response times by van der Linden was compared with its extension by Klein Entink, van der Linden, and Fox, which includes a speed sensitivity parameter. An empirical data example was used to show that the extended model can fit the data better than the model without speed sensitivity parameters. A simulation was conducted, which showed that test forms with different average speed sensitivity yielded substantial different ability estimates for slow test takers, especially for test takers with high ability. Therefore, the use of the extended lognormal model for response times is recommended for the calibration of item pools in high-stakes testing situations. Limitations to the proposed approach and further research questions are discussed.

The use of item scores and response times to detect examinees who may have benefited from item preknowledge

According to Wollack and Schoenig (2018, The Sage encyclopedia of educational research, measurement, and evaluation. Thousand Oaks, CA: Sage, 260), benefiting from item preknowledge is one of the three broad types of test fraud that occur in educational assessments. We use tools from constrained statistical inference to suggest a new statistic that is based on item scores and response times and can be used to detect examinees who may have benefited from item preknowledge for the case when the set of compromised items is known. The asymptotic distribution of the new statistic under no preknowledge is proved to be a simple mixture of two χ² distributions. We perform a detailed simulation study to show that the Type I error rate of the new statistic is very close to the nominal level and that the power of the new statistic is satisfactory in comparison to that of the existing statistics for detecting item preknowledge based on both item scores and response times. We also include a real data example to demonstrate the usefulness of the suggested statistic.

Detection of Item Preknowledge Using Response Times

Benefiting from item preknowledge is a major type of fraudulent behavior during educational assessments. This article suggests a new statistic that can be used for detecting the examinees who may have benefited from item preknowledge using their response times. The statistic quantifies the difference in speed between the compromised items and the non-compromised items of the examinees. The distribution of the statistic under the null hypothesis of no preknowledge is proved to be the standard normal distribution. A simulation study is used to evaluate the Type I error rate and power of the suggested statistic. A real data example demonstrates the usefulness of the new statistic that is found to provide information that is not provided by statistics based only on item scores.

Cognitive diagnosis modelling incorporating item response times

To provide more refined diagnostic feedback with collateral information in item response times (RTs), this study proposed joint modelling of attributes and response speed using item responses and RTs simultaneously for cognitive diagnosis. For illustration, an extended deterministic input, noisy 'and' gate (DINA) model was proposed for joint modelling of responses and RTs. Model parameter estimation was explored using the Bayesian Markov chain Monte Carlo (MCMC) method. The PISA 2012 computer-based mathematics data were analysed first. These real data estimates were treated as true values in a subsequent simulation study. A follow-up simulation study with ideal testing conditions was conducted as well to further evaluate model parameter recovery. The results indicated that model parameters could be well recovered using the MCMC approach. Further, incorporating RTs into the DINA model would improve attribute and profile correct classification rates and result in more accurate and precise estimation of the model parameters.

Joint Testlet Cognitive Diagnosis Modeling for Paired Local Item Dependence in Response Times and Response Accuracy

In joint models for item response times (RTs) and response accuracy (RA), local item dependence is composed of local RA dependence and local RT dependence. The two components are usually caused by the same common stimulus and emerge as pairs. Thus, the violation of local item independence in the joint models is called paired local item dependence. To address the issue of paired local item dependence while applying the joint cognitive diagnosis models (CDMs), this study proposed a joint testlet cognitive diagnosis modeling approach. The proposed approach is an extension of Zhan et al. (2017) and it incorporates two types of random testlet effect parameters (one for RA and the other for RTs) to account for paired local item dependence. The model parameters were estimated using the full Bayesian Markov chain Monte Carlo (MCMC) method. The 2015 PISA computer-based mathematics data were analyzed to demonstrate the application of the proposed model. Further, a brief simulation study was conducted to demonstrate the acceptable parameter recovery and the consequence of ignoring paired local item dependence.