Equitable Ability Estimation in Neurodivergent Student Populations with Zero-Inflated Learner Models

Niall Twomey, Sarah McMullan, Anat Elhalal, Rafael Poyiadzi, Luis M. Vaquero

Abstract

Standard educational models often mistake neurodivergent traits—such as dyslexia or sensory processing issues—for low academic ability. This paper introduces IRT-ZILM, a Zero-Inflated Learner Model that separates a student's actual knowledge from the contextual barriers of how a test is delivered. By distinguishing between knowledge gaps and accessibility hurdles, the model achieves significantly more equitable and accurate ability estimations for neurodivergent learners.

Introduction

In the field of Educational Data Mining (EDM), Learner Models (LMs) are essential for estimating a student's "latent ability" based on historical performance. However, these models frequently suffer from contextual inequity. For a neurodivergent (ND) student, an incorrect answer (a "zero") may not stem from a lack of subject matter knowledge. For example, a student with dyslexia might fail a mathematics problem not because of a deficit in numeracy, but because the problem was presented as a dense block of text.

Standard models, such as Item Response Theory (IRT), lack the mechanism to distinguish between these two distinct causes of failure. Consequently, they consistently underestimate the potential of ND learners, creating a feedback loop of biased recommendations and diminished learning opportunities. This research addresses this gap by proposing a model that accounts for the "noise" introduced by mismatched delivery methods.

Methods: The IRT-ZILM Framework

The authors propose the IRT-based Zero-Inflated Learner Model (IRT-ZILM). The "Zero-Inflated" component is a statistical innovation designed to handle data with an excess of zeros—in this case, incorrect responses caused by structural barriers rather than lack of knowledge.

In IRT-ZILM, an incorrect response ($y=0$) is modeled through two distinct paths:

  1. Contextual Failure ($\pi$): The probability that the "Learning Quality Factor" (LQF) is low due to a mismatch between the student's condition and the Delivery and Response Type (DRT) (e.g., a text-heavy prompt for a dyslexic learner).
  2. Ability-based Failure ($1-p$): The standard probability that a student does not know the answer, based on their latent ability and the item's difficulty.

Mathematically, the probability of an incorrect answer is expressed as: $$Pr(Y=0) = \pi + (1-\pi) \cdot (1-p)$$

By parameterizing $\pi$ (the contextual barrier) separately from $p$ (the ability), the model can "forgive" failures caused by poor context. This allows the latent ability parameter to remain an accurate reflection of the student's actual knowledge.

Experimental Setup and Datasets

High-quality, labeled datasets for neurodivergent populations are currently scarce in the EDM community. To evaluate the model, the authors developed a sophisticated simulation environment to generate a synthetic dataset where "ground truth" parameters were known.

The simulation assigned students:

  • Intrinsic Ability: Drawn from a standard normal distribution.
  • ND Conditions: Including Dyslexia, Dyscalculia, or Sensory Processing Disorder (SPD).
  • Item Parameters: Difficulty, discrimination, subject (e.g., Math vs. English), and content type.
  • DRTs: Various modalities for information presentation (Read, Listen, Both) and response (Click, Speak, Write).

This controlled environment allowed researchers to test how effectively different models could recover the true latent ability of students despite the presence of ND-related interference.

Key Insights

The "Two Types of Zeros" Philosophy

The core innovation of this work is treating student failure as a mixture of structural barriers and knowledge gaps. This shift in perspective allows the model to be "equitable by design," ensuring that accessibility issues do not mask intellectual competence.

Identifying NDCs through Interaction

The research demonstrates that by comparing a student's performance across different delivery types (e.g., reading a prompt vs. listening to it), the model can help identify potentially undiagnosed neurodivergent conditions. The Learning Quality Factor ($\pi$) serves as a quantifiable metric for how accessible a specific task is for a specific individual.

Active Selection for Learning Lift

Once the model identifies a student's profile, it can be used to actively select the DRT that maximizes the student's opportunity to demonstrate their true ability. This proactive approach ensures that the medium of instruction does not become a barrier to the message.

Results

The IRT-ZILM model significantly outperformed standard baselines, including standard IRT and Knowledge Tracing Machines (KTMs):

  • Ability Estimation: IRT-ZILM achieved a 0.993 Pearson correlation with true ability, whereas standard IRT lagged at 0.839.
  • Parameter Recovery: For item discrimination, IRT-ZILM reached a 0.932 correlation. In contrast, standard IRT achieved only 0.270, as it was heavily biased by the "noisy" zeros generated by ND students.
  • Learning Lift: By utilizing the model to select the most suitable delivery type, the authors observed a 1.4x to 1.9x "lift" in successful learning opportunities for students with one or more ND conditions.

Limitations and Future Work

While promising, the study's primary limitation is its reliance on synthetic data. Although the simulation is grounded in clinical literature, real-world validation with diverse student populations is the essential next step. Additionally, the current model uses subjective weighting for how specific conditions (like SPD) interact with delivery types.

Future work will focus on expanding the model to cover a broader range of neurodivergent traits and complex, non-binary response types (such as partial credit). Furthermore, the authors aim to develop teacher-facing tools that use these insights to identify which curriculum components are least accessible to neurodivergent students, fostering a more inclusive educational environment.

Summary updated

Abstract

At present, the educational data mining community lacks many tools needed for ensuring equitable ability estimation for Neurodivergent (ND) learners. On one hand, most learner models are susceptible to under-estimating ND ability since confounding contexts cannot be held accountable (eg consider dyslexia and text-heavy assessments), and on the other, few (if any) existing datasets are suited for appraising model and data bias in ND contexts. In this paper we attempt to model the relationships between context (delivery and response types) and performance of ND students with zero-inflated learner models. This approach facilitates simulation of several expected ND behavioural traits, provides equitable ability estimates across all student groups from generated datasets, increases interpretability confidence, and can significantly increase the quality of learning opportunities for ND students. Our approach consistently out-performs baselines in our experiments and can also be applied to many other learner modelling frameworks.
BibTeX
@misc{twomey2022equitable,
  author = {Niall Twomey and Sarah McMullan and Anat Elhalal and Rafael Poyiadzi and Luis M. Vaquero},
  title = {{Equitable Ability Estimation in Neurodivergent Student Populations with Zero-Inflated Learner Models}},
  year = {2022},
  eprint = {2203.10170},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2203.10170},
}
Figure