📊 statistics

Pairwise Target Rotation for Factor Models

This paper proposes a new interpretability index and a corresponding pairwise target rotation method that effectively incorporate a priori semantic information to improve the recovery of latent structures and the meaningfulness of exploratory factor analysis models.

Original authors: Justin Philip Tuazon, Gia Mizrane Abubo, Joemari Olea

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Justin Philip Tuazon, Gia Mizrane Abubo, Joemari Olea

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

Psychologists and social scientists often try to understand invisible qualities of the human mind, such as anxiety, intelligence, or personality traits. They cannot measure these things directly, so they rely on questionnaires filled with many specific questions. When people answer these questions, the answers are called "manifest variables." The goal of a statistical tool called factor analysis is to find the hidden patterns that connect these answers. It groups questions together that tend to be answered in similar ways, assuming that each group reveals a single underlying trait. However, a major problem has long plagued this method: the math allows for many different ways to group the same questions, and all of these ways fit the data equally well. This creates a puzzle where the computer can produce a dozen different maps of the same territory, but only one of them makes sense to a human reader. For decades, researchers have struggled to decide which map is the true one, often relying on subjective guesses about which grouping feels the most logical.

A team of researchers from the University of the Philippines and the University of Texas at Austin has proposed a new way to solve this puzzle by using the actual meaning of the words in the questions. They argue that if two questions ask about similar ideas, they should naturally belong to the same hidden group. To test this, they developed a new method that treats the meaning of the questions as a guide. Instead of just looking at how people answered, their method looks at the words themselves. They use computer programs to measure how similar the sentences are in meaning, creating a "map of meaning." They then compare this map to the statistical map generated by the data. If the two maps agree—if questions that are semantically similar also end up in the same statistical group—the researchers consider the result to be highly interpretable and meaningful.

The researchers introduced a new score to measure this agreement. They call it an interpretability index. This score does not just check if the math works; it checks if the results make sense in the real world. If the statistical groups align with the semantic groups, the score is high. If the groups are mixed up, the score is low. Using this score, they created a new rotation method, which is a mathematical process that rearranges the groups to find the best possible fit. They named this process "pairwise target rotation," or "priorimax." Unlike older methods that try to force questions into neat, isolated boxes, this new method gently nudges the groups until they match the researcher's expectations about which questions should go together based on their wording.

To see if this idea actually works, the team ran thousands of computer simulations. They created fake data with known hidden structures and then tried to recover those structures using their new method and several older, standard methods. The simulations showed that their new method was better at finding the true hidden structure, especially when the questions had complex relationships where one question might relate to more than one hidden trait. The new method was particularly effective when the "map of meaning" was accurate. When the researchers added noise or errors to the meaning map, the method's performance dropped, which confirmed that the method relies on the quality of the semantic information. However, when the information was clean, the new method consistently outperformed the traditional approaches, finding the correct groups more often and with greater stability.

The researchers then tested their method on real-world data from two well-known psychological surveys. The first was a test measuring depression, anxiety, and stress, containing 42 questions. The second was a test measuring the five major personality traits, containing 50 questions. In both cases, they applied their new rotation method and compared the results to the standard methods used by psychologists for years. The new method produced a higher interpretability score in both cases, suggesting that the groups it found were more consistent with the actual meanings of the questions. The results showed a clear positive link: questions that were similar in meaning tended to cluster together in the statistical analysis. This provided strong evidence that the words people read on a page are deeply connected to the psychological traits they are measuring.

The study concludes that this approach offers a more intuitive and flexible way to analyze data. It allows researchers to incorporate their own knowledge about the subject matter directly into the math, without needing to guess specific numbers beforehand. While the method works best when the semantic information is a good reflection of the reality, it offers a powerful tool for making sense of complex data. By bridging the gap between the meaning of words and the patterns of numbers, the researchers have provided a way to ensure that the hidden structures found in data are not just mathematical artifacts, but meaningful representations of the human experience. This work suggests that in the future, the way we design questions and the way we analyze them can be more closely aligned, leading to clearer insights into the invisible parts of the human mind.

Technical Summary: Pairwise Target Rotation for Factor Models

Problem Statement
Exploratory Factor Analysis (EFA) is a standard method for identifying latent constructs from observed variables. However, the utility of EFA relies heavily on the interpretability of the resulting factor model. A major obstacle to interpretability is rotational indeterminacy: mathematically equivalent loading matrices can reproduce the same covariance structure while implying vastly different substantive meanings. While researchers often rely on subjective judgment or "simple structure" criteria (sparsity) to resolve this, these approaches do not guarantee that the resulting factors are substantively coherent or aligned with the theoretical meaning of the items. Existing methods that incorporate prior information, such as target rotation, often require precise numerical target loadings, which can be difficult to specify in exploratory settings where researchers may only have qualitative expectations about item similarities.

Methodology
The authors propose a new framework comprising an interpretability index and a corresponding rotation method called pairwise target rotation (or priorimax).

Premise for Interpretability: The method is grounded in the "Deep Lexical Hypothesis," which posits that the semantic content of questionnaire items is intrinsically linked to the latent constructs they measure. The authors hypothesize that manifest variables with similar semantic meanings should exhibit similar loading profiles (i.e., similar patterns of association with latent factors).
Similarity Functions:
- Semantic Similarity ( $D_s$ ): Calculated using natural language processing (specifically the Universal Sentence Encoder) to generate vector embeddings of item questions. The similarity between two items is measured by the angular distance between their embeddings, normalized to $[0, 1]$ .
- Loading Similarity ( $D_f$ ): Calculated based on the squared standardized loadings of manifest variables across all factors. This measures the agreement in loading patterns between pairs of variables.
The Interpretability Index ( $\omega$ ):
- The authors define a multiset of pairs containing the semantic similarity and loading similarity for every pair of manifest variables.
- The index is constructed from two components:
  - $\tau$ (Kendall Tau-b): Measures the monotonic concordance between semantic similarity and loading similarity. It is mapped to $[0, 1]$ .
  - $\theta$ (Slope-based): Derived from the Ordinary Least Squares (OLS) slope of the relationship between semantic and loading similarities. This component measures the "separability" or distinctness of loading patterns for semantically dissimilar items. It is also mapped to $[0, 1]$ .
- The final index, $\omega$ , is the geometric mean of $\tau$ and $\theta$ : $\omega = \sqrt{\tau \theta}$ . Higher values indicate better interpretability.
Priorimax Rotation:
- Instead of specifying a target loading matrix, the researcher provides a prior matrix ( $C$ ), which encodes pairwise similarities (e.g., semantic similarities) between items.
- The rotation method is formulated as an optimization problem: find the orthogonal rotation matrix $R$ that maximizes $\omega$ given the unrotated loadings and the prior matrix.
- The optimization is solved using the COBYQA algorithm with multiple random starts to approximate the global maximum.

Key Contributions

A New Index ( $\omega$ ): A quantitative measure of factor model interpretability that evaluates the agreement between a priori information (specifically item semantics) and the empirical loading structure. Unlike previous indices, it does not rely on estimated factor scores, thereby avoiding factor score indeterminacy.
Pairwise Target Rotation: A rotation method that incorporates prior information in the form of pairwise similarities rather than specific target loadings. This makes it more intuitive for exploratory settings where exact loading values are unknown, but item relationships are understood.
Generalization: The framework allows the prior matrix to be constructed from any source of a priori information (e.g., expert judgment, theoretical groupings), not just semantic embeddings.

Results

Simulation Study:
- The $\omega$ -index correctly indicated better model fit when noise levels (specific variances) were low and when the prior matrix was a good proxy for the true loadings.
- The priorimax rotation outperformed classical orthogonal rotations (Varimax, Equamax, Quartimax) in recovering the true loading matrix specifically when cross-loadings were present (generic loading matrices). However, the method performed worse than standard rotations for simple loading matrices that strictly follow a simple structure, as standard rotations are designed to seek such sparsity while priorimax does not.
- Performance degraded as the noise level in the prior matrix increased, but remained robust provided the perturbation was not excessive.
- Using the sample correlation matrix as the prior matrix was found to be beneficial only when cross-loadings were present; it performed worse than standard rotations for simple structures.
Empirical Demonstration:
- The method was applied to the Depression Anxiety Stress Scales (DASS) and the Big Five Personality Test.
- In both cases, the priorimax rotation yielded the highest $\omega$ -index compared to unrotated, Varimax, Equamax, and Quartimax solutions.
- The results demonstrated a positive monotonic relationship between semantic similarity and loading similarity, empirically supporting the link between item wording and latent constructs.

Significance and Claims
The paper claims that the proposed method offers an "intuitive yet flexible" way to incorporate a priori information into EFA. By shifting the focus from achieving sparsity to achieving semantic coherence, the method addresses the subjectivity inherent in factor interpretation. The authors assert that the $\omega$ -index serves as a useful measure of relative goodness-of-fit for comparing different rotated solutions. Furthermore, they suggest that the method can aid in ex post validation (where a high $\omega$ indicates good validity) and potentially ex ante validation of questionnaires by examining latent structures derived solely from item semantics. The paper concludes that while the method is most effective when the prior information is a strong proxy for the true loadings, it provides a robust alternative to traditional rotation methods, especially in complex models with cross-loadings.

More like this