📊 statistics

Logistic Multidimensional Data Analysis for Ordinal Response Variables using a Cumulative Link function

This paper introduces a multidimensional data analysis framework for ordinal response variables based on a continuous latent variable and cumulative logit models, which accommodates both supervised and unsupervised scenarios by distinguishing between dominance and proximity variables, and is estimated using a derived expectation-majorization-minimization algorithm.

Original authors: Mark de Rooij, Ligaya Breemer, Dion Woestenburg, Frank Busing

Published 2026-07-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mark de Rooij, Ligaya Breemer, Dion Woestenburg, Frank Busing

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

The Big Picture: Turning "Maybe" into a Map

Imagine you are trying to understand a group of people based on their answers to a survey. But there's a catch: the answers aren't simple "Yes" or "No." They are on a sliding scale, like a Likert scale: Strongly Disagree, Disagree, Neutral, Agree, Strongly Agree.

Traditionally, statisticians treat these answers like numbers (1, 2, 3, 4, 5) and draw straight lines to find patterns. The authors of this paper argue that this is like trying to measure the temperature with a ruler—it's the wrong tool. Instead, they propose a new way to look at this data that respects the "ordered" nature of the answers without forcing them into a rigid numerical box.

They call their new framework Cumulative Logistic Multidimensional Data Analysis. That's a mouthful, so let's break it down into two main ideas: The Map and The Compass.

1. The Two Types of Questions: "Dominance" vs. "Proximity"

The paper starts by saying that not all survey questions work the same way. They distinguish between two types of questions, using a clever analogy:

Type A: The "Dominance" Questions (The Mountain Climb)

Think of a math test. If you are a math genius, you can solve easy problems and hard problems. If you are a beginner, you might solve the easy ones but fail the hard ones.

The Analogy: Imagine a mountain. The higher you climb (your ability), the more items you can "conquer."
The Paper's Term: Dominance Variables.
How it works: The paper uses a method called Inner Product (like projecting a shadow). If you are "high up" on the map, you are likely to get a high score on the question. The further you go in one direction, the better you do.

Type B: The "Proximity" Questions (The Target Practice)

Now think about a question like: "How often do you eat spicy food?"

The Analogy: Imagine a dartboard. Some people love spicy food (they stand right next to the "Spicy" target). Some people hate it (they stand far away). But there's also a "Just Right" spot. If you are too far away in the "No Spice" direction, you hate it. If you are too far in the "Spicy" direction, you might also hate it (maybe you prefer mild). You are happiest when you are close to a specific point.
The Paper's Term: Proximity Variables.
How it works: The paper uses a method called Distance. The answer depends on how close you are to the "ideal" point for that question. It's not about going "higher"; it's about being "near."

2. The New Tool: Drawing the Map

The authors created a new way to draw a map (called a Biplot) that shows both the people (the dots) and the questions (the lines or circles) on the same piece of paper.

For Dominance Questions (The Mountain): The questions are drawn as arrows (vectors).
- If a person's dot is far out in the direction of the arrow, they likely answered "Strongly Agree."
- If they are in the opposite direction, they likely "Strongly Disagree."
- The paper adds little markers on the arrows (like 1|2, 2|3) to show exactly where the "cut-off" is between categories.
For Proximity Questions (The Target): The questions are drawn as dots surrounded by concentric circles (like a bullseye).
- If a person's dot is inside the inner circle, they are very close to the "ideal" answer (e.g., "Always").
- If they are in the ring between the circles, they are in the middle category (e.g., "Sometimes").
- If they are outside the outer circle, they are far away (e.g., "Never").

3. Adding the "Compass" (Predictor Variables)

Often, researchers want to know why people answer the way they do. They have extra data like Age, Gender, or Education.

The Paper's Innovation: They figured out how to draw these extra factors on the same map.
How it looks:
- Numerical factors (like Age): Drawn as arrows. If the "Age" arrow points in the same direction as a "Pro-Environment" question, it means older people tend to answer that way.
- Categorical factors (like Gender): Drawn as specific points. If the "Female" point is close to a specific question's target, it suggests women in the survey lean toward that answer.

4. The Engine: The "EMM" Algorithm

To build these maps, the authors had to invent a new mathematical engine. They call it an Expectation-Majorization-Minimization (EMM) algorithm.

The Analogy: Imagine you are trying to find the lowest point in a foggy valley (the best map). You can't see the bottom.
1. Guess: You take a step.
2. Check: You look around to see if you are going down.
3. Adjust: If you are stuck on a small hill, the algorithm has a special trick to "smooth out" the terrain so you can slide down to the true bottom.
The paper proves this engine works well and doesn't get stuck in fake "lowest points" (local optima) too often, especially when they run it multiple times from different starting spots.

5. Real-World Tests (The Examples)

The authors tested their new map-making tool on three real datasets to prove it works:

Student Exam Scores: They looked at how students answered statistics questions. They found that "Math Knowledge" and "Engagement" were strong arrows pointing toward correct answers. They could see that some questions were "easy" (markers were far left) and some were "hard" (markers far right).
Environmental Habits (Thailand): They asked people about recycling, eating meat, and outdoor activities.
- The Surprise: Most questions (Recycling, Avoiding bad products) acted like Dominance (arrows). The more you care, the more you do it.
- The Twist: One question ("How often do you go outside?") acted like Proximity (circles). It wasn't just about caring; it was about being close to a specific lifestyle. The new map showed this difference clearly, whereas old methods would have missed it.
Global Attitudes (12 Countries): They compared how people in different countries feel about the environment. The map showed that countries like Germany and Austria were very close together (similar answers), while Asian countries formed a different cluster. It also showed that Education was a huge arrow pushing people toward different answers.

Summary

The paper says: "Stop treating 'Maybe' like a number. Treat it like a position on a map."

They built a new system that:

Knows the difference between questions where "more is better" (Dominance) and questions where "closer is better" (Proximity).
Draws a visual map showing people, questions, and reasons (like age or gender) all in one picture.
Uses a smart mathematical engine to ensure the map is accurate.

This allows researchers to see patterns in complex, ordered data that they might have missed with traditional methods.

Technical Summary: Logistic Multidimensional Data Analysis for Ordinal Response Variables

Problem Statement
The paper addresses a gap in multidimensional data analysis (MDA) for ordinal response variables. While standard techniques like Principal Component Analysis (PCA) and Multidimensional Unfolding (MDU) are well-established for continuous data, and logistic regression is standard for categorical data, there is a lack of exponential family generalizations for MDA specifically tailored to ordinal responses. Existing methods often treat ordinal data as continuous (leading to distorted effect sizes and inflated Type-I errors) or focus exclusively on binary outcomes. Furthermore, current frameworks fail to simultaneously handle the two distinct response processes identified by Coombs (1964): dominance (monotonic relationship, e.g., ability tests) and proximity (non-monotonic, single-peaked relationship, e.g., attitudes), and do not adequately integrate predictor variables (covariates) into these geometric models for ordinal data.

Methodology
The authors propose a unified framework of four models based on the cumulative logistic regression (proportional odds) model, assuming a continuous latent variable underlying the observed ordinal scores. The framework distinguishes between response types and the presence of predictors:

Model Types:
- Dominance Variables (Inner Product): Modeled using Cumulative Logistic PCA (CLPCA) (unsupervised) and Cumulative Logistic Reduced Rank Regression (CLRRR) (supervised, where subject scores are linear functions of predictors).
- Proximity Variables (Distance): Modeled using Cumulative Logistic MDU (CLMDU) (unsupervised) and Cumulative Logistic Restricted MDU (CLRMDU) (supervised).
- In all models, the structural part ( $\theta_{ir}$ ) is defined either as an inner product $\langle u_i, v_r \rangle$ for dominance or a negative distance $-d(u_i, v_r)$ for proximity.
Estimation Algorithm:
- The authors derive an Expectation-Majorization-Minimization (EMM) algorithm for Maximum Likelihood Estimation (MLE).
- E-step: Computes the conditional expectation of the complete data negative log-likelihood.
- M-step: Uses a majorization function that reduces the problem to a weighted least squares minimization.
- Updates:
  - For inner product models (PCA/RRR), updates involve Singular Value Decomposition (SVD) or Generalized SVD.
  - For distance models (MDU/RMDU), updates utilize a modified SMACOF algorithm capable of handling negative dissimilarities (arising from the logistic link).
- Threshold parameters are updated separately using standard MLE with the structural part as an offset.
Visualization (Biplots):
- The paper extends biplot methodology to visualize these models.
- CLPCA/CLRRR: Variables are represented as axes with markers indicating estimated thresholds (transition points between categories) based on the underlying latent variable.
- CLMDU/CLRMDU: Variables are represented as points surrounded by concentric circles (decision boundaries) corresponding to category thresholds. The space is partitioned into regions where specific response categories are most probable.

Key Contributions

Theoretical Unification: The paper fills the literature gap by providing the first exponential family generalizations of PCA, RRR, MDU, and RMDU specifically for ordinal response variables, covering both dominance and proximity processes.
Algorithmic Development: It introduces a unified EMM algorithm that handles the non-linearity of the logistic link and the specific geometric constraints of both inner-product and distance-based models.
Visualization: It proposes novel biplot techniques for ordinal data, specifically the use of threshold-based markers for dominance variables and concentric decision circles for proximity variables, allowing for the interpretation of complex response profiles in low-dimensional space.
Software Implementation: The algorithms are implemented in the R package logistic_mapping (functions clpca and clmdu).

Results

Empirical Applications:
1. Student Performance (Dominance): Analysis of 138 students' exam responses (10 items) and psychological predictors. The CLRRR model successfully identified that mathematical knowledge and affective engagement predicted correct answers, while academic procrastination predicted errors. The biplot revealed distinct clusters of items with different response patterns.
2. Pro-Environmental Behavior (Mixed): Analysis of ISSP data (Thailand) regarding environmental attitudes and behaviors. The study contrasted CLRRR and CLRMDU. The proximity model (CLRMDU) provided a better fit, particularly for the "OUT" (outdoor activity) variable, which exhibited a non-monotonic (proximity) response pattern, whereas other behavioral variables followed a dominance pattern.
3. Environmental Efficacy (Proximity): Analysis of ISSP data across 12 countries. The CLRMDU model (3D) revealed distinct country clusters (Asian vs. European) and showed how education, age, and gender influenced attitudes toward environmental efficacy.
Simulation Study: A simulation study varying sample size (250–1000), number of response variables (4, 8), and categories (3, 5) demonstrated that the algorithms converge monotonically. Parameter recovery (measured by Stress-1) improved with larger sample sizes, more response variables, and more categories.

Significance and Claims
The authors claim that this framework provides a versatile tool for analyzing complex datasets with ordinal responses where dimension reduction and the simultaneous modeling of correlated variables are required.

Distinction of Processes: The framework explicitly acknowledges and models the difference between dominance and proximity variables, which standard linear MDA methods fail to do.
Handling Predictors: It extends these capabilities to supervised settings (RRR and RMDU), allowing for the inclusion of external covariates to explain subject positions.
Visual Interpretability: The proposed biplots offer a holistic view of the data structure, dependencies among response variables, and the effects of predictors, complementing numerical coefficient tables.
Assumption Handling: The paper notes that while the proportional odds assumption is central to the model, it may not always hold. However, citing Harrell (2001), the authors suggest that the bias-variance trade-off might favor the simpler proportional odds model over more complex alternatives that introduce excessive variance. They acknowledge that testing this assumption within the MDA framework requires further development (e.g., comparing against Multinomial Multiple Correspondence Analysis).

The paper concludes that these models are distinct from Item Response Theory (IRT) models like the Graded Response Model (GRM) or Generalized Graded Unfolding Model (GGUM), as the primary goal here is dimension reduction and prediction of correlated responses rather than latent trait estimation for test construction.