📊 statistics

Bridging Item Response Theory and Factor Analysis: A Four-Parameter Mixture-Dichotomized Model with Bayesian Estimation

This paper establishes the analytical equivalence between four-parameter Item Response Theory (IRT) and Factor Analysis (FA) models using a hierarchical mixture formulation, develops a Bayesian estimation procedure for this unified framework, and validates its performance through simulations and empirical applications.

Original authors: Ján Pavlech, Patrícia Martinková

Published 2026-06-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ján Pavlech, Patrícia Martinková

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to measure how much a person knows or how they feel using a test. In the world of psychology and education, there are two main "languages" used to build these tests: Item Response Theory (IRT) and Factor Analysis (FA).

Think of these two languages like two different dialects for describing the same landscape. For a long time, we knew they could translate perfectly into each other for simple tests (the "two-parameter" models). But when tests get more complex—accounting for things like guessing (getting an answer right by luck) or inattention (missing an easy question because you were daydreaming)—the translation broke down. The "Factor Analysis" dialect didn't have a way to talk about these messy human behaviors.

This paper is like a translator's manual that finally bridges that gap. Here is what the authors did, explained simply:

1. The Problem: The "Perfect Score" Illusion

Imagine a student takes a 20-question math test. They get 15 right.

The Old Way: The test says, "This student knows 15 questions."
The Real World: Maybe they guessed on 5 of those 15, or maybe they skipped 5 easy ones because they were distracted. The old math couldn't separate the real knowledge from the luck or the distraction.

The authors wanted to build a new mathematical model (a 4-Parameter Model) that could see through the noise. They wanted to distinguish between:

Ability: What you actually know.
Guessing: Getting it right by luck.
Inattention: Getting it wrong by accident.

2. The Solution: A "Two-Layer" Filter

The authors took the Factor Analysis framework (which usually looks at the "big picture" of how questions relate) and added a special hierarchical mixture filter.

Think of this like a security checkpoint with two gates:

Gate 1 (The Ability Gate): Did the person actually know the answer? If their "mental score" was high enough, they pass through.
Gate 2 (The Behavior Gate):
- If they didn't know the answer (failed Gate 1), there's still a chance they guessed correctly.
- If they did know the answer (passed Gate 1), there's still a chance they slipped up (inattention) and got it wrong.

By adding this second layer, the model can mathematically separate a "lucky guess" from a "real answer."

3. The Big Discovery: They Are Actually the Same

The most exciting part of the paper is the proof. The authors showed that this new "Four-Parameter Factor Analysis" model is mathematically identical to the existing "Four-Parameter IRT" model.

Analogy: It's like realizing that a "Soda" and a "Pop" are just different names for the exact same drink.
Why it matters: For decades, researchers using Factor Analysis couldn't easily model guessing and inattention. Now, they can use their existing tools to do exactly what IRT researchers have been doing, and vice versa. The paper proves they are two sides of the same coin.

4. The New Tool: A "Clean Score"

Because this model understands the difference between knowing, guessing, and slipping, it can give you a new kind of score.

Old Score: "You got 15/20 right."
New "NGNI" Score (Non-Guessing, Non-Inattentive): "Based on your pattern of answers, we estimate you actually knew the answers to about 12 items, guessed on 3, and missed 5 due to distraction."

This allows for a much more honest view of a person's true ability, stripping away the "noise" of luck and carelessness.

5. How They Did It (The Engine)

To make this work, the authors built a new computer engine using Bayesian estimation.

The Analogy: Imagine trying to solve a giant jigsaw puzzle where some pieces are missing. Instead of guessing blindly, the computer runs thousands of simulations, constantly adjusting the picture until it finds the most likely arrangement.
They wrote this engine in two popular programming languages (R and Python) so other researchers can use it immediately.

6. Did It Work?

They tested their new engine in two ways:

Simulation: They created fake tests with known answers and saw if their model could find them. It did, often more accurately than older methods, especially when there were fewer test-takers.
Real Data: They applied it to a real admission test and a real anxiety survey. The results matched up perfectly with the old IRT methods, proving the "translation" works.

Summary

This paper is a bridge. It connects two major schools of thought in testing (IRT and Factor Analysis) by showing they can both handle the messy reality of human behavior (guessing and inattention). It provides a new way to calculate scores that are "purified" of luck and distraction, and it gives researchers the free software tools to do it themselves.

Technical Summary: Bridging Item Response Theory and Factor Analysis via a Four-Parameter Mixture-Dichotomized Model

Problem Statement
Item Response Theory (IRT) and Factor Analysis (FA) are foundational frameworks for modeling latent traits using multi-item data. While the equivalence between the two-parameter (2P) IRT model and the dichotomized single-factor FA model is well-established, extending this correspondence to IRT models with additional parameters—specifically the three-parameter (3P) and four-parameter (4P) models—is less common. The 4P IRT model incorporates pseudo-guessing ( $c$ ) and inattention ( $d$ ) parameters to account for response behaviors such as guessing or careless mistakes. However, standard dichotomized FA models cannot naturally accommodate the lower and upper asymptotes introduced by these parameters. Consequently, there is a lack of a formal FA formulation that is analytically equivalent to the 4P IRT model, limiting the ability to utilize FA-specific estimation frameworks or interpretative advantages for higher-parameter models.

Methodology
The authors propose a Four-Parameter Factor-Analytic (4P FA) model that extends the traditional dichotomized FA framework through a hierarchical mixture formulation.

Model Construction:
- The model begins with a standard linear FA structure where a continuous latent item response $Y^*_i$ is defined as $Y^*_i = \alpha_i \theta + \epsilon_i$ , with $\theta$ representing the latent trait.
- A first-level dichotomization occurs at a threshold $\tau_i$ , creating a latent binary indicator $Z_i$ . If $Y^*_i \geq \tau_i$ , the respondent is considered to have "solved" the item ( $Z_i=1$ ); otherwise, they have not ( $Z_i=0$ ).
- A second-level mixture model generates the observed binary response $Y_i$ $Y_{i}$ based on $Z_i$ $Z_{i}$ :
  - If $Z_i = 0$ (latent performance below threshold), $Y_i$ follows a Bernoulli distribution with success probability $c_i$ (guessing parameter).
  - If $Z_i = 1$ (latent performance above threshold), $Y_i$ follows a Bernoulli distribution with success probability $d_i$ (inattention parameter, where $d_i < 1$ accounts for lapses).
- This structure yields the 4P FA model, which the authors prove is analytically equivalent to the 4P IRT model. The FA parameters ( $\alpha_i, \tau_i$ ) map to IRT discrimination ( $a_i$ ) and difficulty ( $b_i$ ) via standard transformations, while $c_i$ and $d_i$ are shared directly.
Estimation Procedure:
- A Bayesian estimation procedure is developed to estimate the four item parameters ( $a, b, c, d$ ), latent trait scores ( $\theta$ ), and adjusted ability scores.
- The model is implemented using Directed Acyclic Graphs (DAGs) in R (via the runjags package interfacing with JAGS) and Python (via pymc).
- The JAGS implementation utilizes Gibbs sampling, while the Python implementation employs the No-U-Turn Sampler (NUTS), though the latter requires handling discrete latent variables ( $Z_i$ ) via alternative samplers, impacting computational speed.
- The framework also allows for the estimation of "NGNI" (Non-Guessing and Not-Inattentive) scores, which adjust observed scores by removing the effects of guessing and inattention.

Key Contributions

Theoretical Extension: The paper formally establishes the analytical equivalence between the 4P FA model and the 4P IRT model, extending the FA–IRT correspondence beyond the 2P case.
Hierarchical Mixture Formulation: It introduces a specific hierarchical mixture structure within the FA framework to account for guessing and inattention, providing a latent-variable formulation that standard FA models lack.
Software Implementation: The authors provide open-source implementations in both R and Python, facilitating the application of this complex model.
Score Adjustment: The methodology enables the calculation of ability scores adjusted for specific response distortions (guessing and inattention), offering a diagnostic tool beyond standard IRT ability estimates.

Results

Simulation Study: A simulation study compared the proposed 4P FA Bayesian estimation (4FA-B) against a typical 4P IRT Bayesian estimation (4IRT-B) and a frequentist 4P IRT estimation using the EM algorithm (4IRT-EM).
- Convergence: The EM algorithm (4IRT-EM) frequently failed to converge or diverged, particularly with smaller sample sizes (e.g., $n=50$ ) or larger numbers of items. The Bayesian approaches (both FA and IRT parameterizations) demonstrated superior stability.
- Precision: While the EM algorithm was computationally faster (seconds vs. minutes/hours for Bayesian methods), the Bayesian estimates generally exhibited lower Mean Squared Error (MSE), particularly in small sample scenarios.
- Parameterization: The study found that the FA parameterization (4FA-B) often yielded discrimination and difficulty estimates more aligned with the true simulated values compared to the typical IRT Bayesian parameterization (4IRT-B), suggesting potential practical advantages in estimation performance.
Empirical Examples: The model was applied to two datasets: an admission test (MSATB, 1407 respondents, 20 items) and an anxiety inventory (766 respondents, 29 items).
- Results confirmed strong agreement between the 4P FA and 4P IRT parameter estimates when transformed into a common metric.
- The application demonstrated the calculation of NGNI, NG (non-guessing), and NI (non-inattention) scores, showing how these adjusted scores differ from observed totals based on response patterns and item difficulty.

Significance and Claims
The paper claims to bridge a significant gap in psychometric modeling by providing a rigorous FA counterpart to the 4P IRT model. The authors posit that this equivalence allows researchers to leverage FA-specific estimation frameworks for models with guessing and inattention parameters. They suggest that the FA parameterization may offer practical advantages in estimation performance, particularly regarding stability and precision in Bayesian contexts. Furthermore, the ability to derive "purified" ability scores (NGNI) that account for specific error sources (guessing and inattention) is highlighted as a novel application of the methodology, offering deeper diagnostic insights into respondent behavior than standard IRT models provide. The work does not claim to replace existing IRT methods but rather to extend the FA tradition to accommodate complex response behaviors, thereby unifying two major psychometric traditions under a single, flexible framework.