Multidimensional constructs and moderated linear and nonlinear factor analysis

Technical Summary: Multidimensional Constructs and Moderated Linear and Nonlinear Factor Analysis

Problem Statement
Multidimensional factor models incorporating moderation on all model parameters have historically been restricted to single-factor and two-factor configurations. This limitation creates a misalignment with existing psychological measures, which frequently assess three to five dimensions of a latent construct. The primary technical hurdle in extending Moderated Nonlinear Factor Analysis (MNLFA) to three or more factors is ensuring the validity of the latent variable covariance matrix for all subgroups defined by the moderating design matrix.

While Fisher's z-transformation effectively bounds a single correlation parameter to the $[-1, 1]$ interval for two-factor models, applying independent transformations to multiple correlations does not guarantee that the resulting correlation matrix remains positive definite. Consequently, existing methods cannot simultaneously ensure that individual correlations stay within bounds and that the entire matrix is valid, a longstanding challenge in covariance prediction and simulation.

Methodology
The paper proposes a generalized framework for MNLFA that allows for moderation in factor correlation parameters for models with $m \geq 3$ factors. The approach involves decomposing the person-specific factor covariance matrix ($\Phi_n$) into factor variances and a factor correlation matrix ($\Phi^R_n$). While variances are handled via exponential transforms, the paper introduces two distinct methods for inducing moderation in the correlation matrix to ensure positive definiteness:

1. Partial Correlations and Cholesky Factorization: This method maps unconstrained parameters ($\gamma$) to partial correlations using a Fisher's z-transformation (or hyperbolic tangent). These partial correlations are then used to construct a lower-triangular Cholesky factor ($\Phi^L$), which is subsequently multiplied by its transpose to reconstruct the valid correlation matrix ($\Phi^R = \Phi^L (\Phi^L)^T$).

   • Limitation: This approach is sensitive to factor ordering. The interpretation of a partial correlation depends on which factors are "partialed out" first, meaning the impact of a moderator on a specific correlation is conditional on the moderation of preceding correlations in the ordering.

2. Iterative Generalized Fisher Transform (GFT): Based on Archakov and Hansen (2021), this method utilizes the matrix exponential and matrix logarithm. It iteratively updates a symmetric matrix $A$ (initialized with unconstrained off-diagonal elements) by adjusting its diagonal elements until the matrix exponential $e^A$ converges to a matrix with unit diagonals and off-diagonal elements within $(-1, 1)$.

   • Advantage: This approach allows for the interpretation of moderation on specific correlations without the conditional dependencies inherent in the Cholesky approach.
   • Limitation: It is an iterative algorithm that can be computationally sensitive to the magnitude of unconstrained parameters. The author notes that pre-conditioning unconstrained parameters using a scaled hyperbolic tangent transform (e.g., mapping to $(-\pi, \pi)$) can improve convergence speed.

Estimation and Identification
The paper explores two estimation strategies for these multidimensional models:

• Bayesian Approach: Implemented in Stan using a non-centered parameterization to improve sampling efficiency. The author employs spike-and-slab priors (specifically a logit-normal continuous parameterization) to regularize Differential Item Functioning (DIF) and impact parameters. However, the author reports that conditional likelihood approaches in Stan were prohibitively slow and memory-intensive for sample sizes exceeding 5,000 cases, particularly when requiring nuisance parameters for binary/ordinal items.
• Penalized Maximum Likelihood (PML): To address computational bottlenecks, the author derives closed-form analytical gradients for the marginal likelihood function with respect to all model parameters (intercepts, loadings, variances, correlations, and moderation effects). This avoids the numerical gradient approximation required by Bayesian software. The framework incorporates penalization (Ridge, LASSO, and Alignment penalties) to handle model complexity and identify partial invariance.
  • Implementation Note: While the analytical gradients alleviate some computational burden, the author notes that constructing person-specific mean and covariance matrices for the likelihood evaluation remains a bottleneck for very large samples ($N > 10,000$), and the current PML implementation is still slow for item factor analysis.

Key Contributions

1. Generalization of MNLFA: The paper successfully extends the MNLFA framework from two factors to $m$-dimensional models by providing valid parameterizations for moderated factor correlations.
2. Novel Correlation Parameterizations: It introduces and compares two methods for inducing valid, moderated correlation structures: a Cholesky-based partial correlation approach and an iterative Generalized Fisher Transform.
3. Analytical Gradients for PML: The derivation of closed-form gradients for the multidimensional MNLFA likelihood function facilitates a penalized maximum likelihood estimation approach that is theoretically more efficient than simulation-based Bayesian methods for continuous indicators.
4. Regularization Framework: The integration of spike-and-slab priors (in the Bayesian context) and various penalty functions (in the PML context) provides a mechanism for detecting partial invariance and preventing overfitting in complex, high-dimensional models.

Results and Claims
The paper does not present a large-scale simulation study or a definitive empirical application with final conclusions on model performance across all conditions. Instead, it presents:

• Theoretical Validation: Mathematical derivations demonstrating how the proposed transformations (Cholesky and GFT) map unconstrained parameters to valid correlation matrices.
• Illustrative Examples: R code and numerical examples showing how the GFT method converges from unconstrained matrices to valid correlation matrices and how the Cholesky method reconstructs covariance matrices.
• Computational Observations: The author reports that while the Bayesian implementation in Stan was too slow for large datasets, the PML approach with analytical gradients offers a more feasible path, though it still faces scalability challenges with person-specific likelihood evaluations.

Significance
The paper claims to address a critical methodological gap in the study of measurement invariance. By enabling the estimation of multidimensional MNLFA models, the framework allows researchers to simultaneously investigate differences in latent factor parameters and measurement model characteristics across subpopulations. This is particularly significant for fields like psychopathology, where constructs are inherently multidimensional (3–5 factors) and the pattern of non-invariance across groups (e.g., sex, age) is of substantive interest for developing targeted interventions.

The proposed framework offers a "frequentist alternative" to existing Bayesian penalty methods, aiming to balance flexibility with computational feasibility. The author concludes that while challenges regarding scalability and penalty design remain, this work advances the methodological toolkit for evaluating measurement invariance in complex, multidimensional constructs, promoting fairness and validity in psychological assessment.