TL;DR
This paper introduces the Correlation Thresholding (CT) algorithm, a novel graph-theoretic approach to Exploratory Factor Analysis (EFA). CT simultaneously addresses key EFA challenges – determining the number of factors, learning factor loading constraints, and selecting a rotationally unique solution – in a unified framework, offering a more cohesive alternative to traditional methods.
What Problem Does This Paper Address?
Exploratory Factor Analysis (EFA) is a widely used statistical technique to uncover underlying latent structures in data, particularly in social sciences. However, performing EFA traditionally involves tackling three distinct and often complex problems: (1) deciding how many latent factors are present, (2) identifying which observed variables load onto which factors (and which don't), and (3) choosing a stable, interpretable solution from a multitude of mathematically equivalent options. These separate approaches can lead to a fragmented and sometimes inconsistent analysis. While newer penalized EFA methods offer improvements, they often require extensive computational searches and still necessitate pre-specifying the number of factors. This paper aims to provide a unified and more efficient framework for EFA by addressing these issues simultaneously.
Methods
The core of this paper is the Correlation Thresholding (CT) algorithm. It leverages a graph-theoretic perspective, viewing the relationships between observed variables as a graph. The central heuristic is that correlations between variables sharing a common factor (within-factor correlations) are generally stronger than those between variables that don't share a factor (between-factor correlations).
The CT algorithm works by:
- Constructing a correlation graph: Based on a sample correlation matrix.
- Thresholding correlations: Applying a threshold to the correlation matrix to identify potentially meaningful connections.
- Identifying independent maximal cliques: These cliques in the graph are interpreted as the "children" (observed variables) of the latent factors. The number of such cliques directly estimates the number of latent factors.
- Estimating factor loadings: Based on the clique structure, constraints are imposed on the factor loading matrix.
- Model selection: Multiple candidate structures are generated by varying the correlation threshold, and a final model is selected using a criterion like BIC.
This approach aims to solve the three core EFA problems concurrently, unlike traditional methods that address them sequentially.
Datasets
The study uses simulated data for its extensive simulation studies. For a real-world example, the authors analyze data from the Adolescent Brain Cognitive Development (ABCD) Study, specifically focusing on the youth version of the UPPS-P impulsivity scale. This scale is hypothesized to have five latent factors, with four observed variables per factor.
Key Insights and Ideas
- Unified Framework: The CT algorithm's main innovation is its ability to tackle the number of factors, factor loading constraints, and rotational uniqueness simultaneously, offering a more integrated approach to EFA.
- Graph-Theoretic Intuition: The core idea that stronger correlations indicate shared latent factors is elegantly translated into a graph-based method, making the underlying structure more interpretable.
- Robustness: Despite relying on assumptions like "thresholdability" (stronger within-factor correlations) and "unique child condition" (each factor having at least one unique indicator), the CT algorithm demonstrates surprising robustness to violations of these assumptions, performing competitively even when they are not perfectly met.
- Computational Efficiency: Compared to penalized EFA methods (LASSO, MCP), the CT algorithm evaluates significantly fewer candidate models, making it computationally more efficient, especially in high-dimensional settings.
Key Results
- Structural Accuracy: The CT algorithm generally outperformed EFA, EFA-LASSO, and EFA-MCP in terms of structural accuracy (measured by Hamming distance and F₁ score) across various simulation conditions, especially when thresholdability was moderately violated.
- Model Fit: CT also showed strong performance in model fit (using TLI and RMSEA), often matching or exceeding other methods, particularly when thresholdability assumptions were reasonably met.
- High-Dimensional Performance: In high-dimensional settings (n < p), CT maintained good structural accuracy and computational efficiency, where penalized EFA methods become computationally prohibitive.
- Real Data Example: In a real data analysis of impulsivity measures, the CT algorithm achieved the highest F₁ score (0.974) and obtained the hypothesized number of factors, significantly outperforming other methods in structural accuracy and computational efficiency.
Limitations and Future Work
- Assumptions: The CT algorithm's theoretical guarantees rely on the "thresholdability" and "unique child condition" assumptions. While the study shows good robustness, extreme violations can degrade performance.
- Future Work: The authors suggest generalizing the CT algorithm to non-linear factor analysis models by using non-linear correlation metrics (e.g., polychoric correlations). They also propose incorporating subsampling or bootstrap aggregation techniques to handle sample variability when the population structure might be thresholdable but sample estimates are not.
Summary updated