A Comparison of Pseudo-Maximum Likelihood and Asymptotically Distribution-Free Dynamic Factor Analysis Parameter...

Area of Science:

• Psychology
• Statistics
• Quantitative Psychology

Background:

• Intraindividual variability is crucial across psychology subdisciplines.
• Analyzing multivariate time-series data requires specialized statistical techniques.
• The dynamic factor model (DFM) offers a method for latent variable modeling of time-series data.

Purpose of the Study:

• To compare pseudo-Maximum Likelihood (p-ML) and Asymptotically Distribution Free (ADF) estimation methods for DFM parameters.
• To evaluate the consistency and precision of parameter estimates, standard errors, and chi-square statistics from both methods.
• To assess the performance of p-ML and ADF under varying conditions, particularly with shorter time-series.

Main Methods:

• Monte Carlo simulation was employed to compare estimation techniques.
• DFM parameters were estimated within a covariance-structure framework using block-Toeplitz matrices.
• The study focused on comparing p-ML and ADF estimation methods.

Main Results:

• Both p-ML and ADF yielded consistent model parameter estimates with comparable precision.
• Only the ADF method produced consistent standard errors and chi-square statistics.
• Estimation results were similar across methods, especially for shorter manifest time-series.

Conclusions:

• ADF estimation is recommended for DFM analysis due to its provision of consistent inferential statistics (standard errors, chi-square).
• Both methods perform comparably when dealing with short time-series data.
• The findings contribute to the robust statistical analysis of intraindividual variability using dynamic factor models.