<?xml version="1.0" encoding="UTF-8"?><?xml-stylesheet type="text/xsl" href="static/style.xsl"?><OAI-PMH xmlns="http://www.openarchives.org/OAI/2.0/" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://www.openarchives.org/OAI/2.0/ http://www.openarchives.org/OAI/2.0/OAI-PMH.xsd"><responseDate>2026-11-21T04:13:44Z</responseDate><request verb="GetRecord" identifier="oai:macsphere.mcmaster.ca:11375/33066" metadataPrefix="dim">https://prod-ms-be.lib.mcmaster.ca/server/oai/request</request><GetRecord><record><header><identifier>oai:macsphere.mcmaster.ca:11375/33066</identifier><datestamp>2026-09-24T22:15:26Z</datestamp><setSpec>com_11375_271</setSpec><setSpec>col_11375_272</setSpec></header><metadata><dim:dim xmlns:dim="http://www.dspace.org/xmlns/dspace/dim" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:doc="http://www.lyncode.com/xoai" xsi:schemaLocation="http://www.dspace.org/xmlns/dspace/dim http://www.dspace.org/schema/dim.xsd">
   <dim:field mdschema="dc" element="contributor" qualifier="advisor">Jeganathan, Pratheepa</dim:field>
   <dim:field mdschema="dc" element="contributor" qualifier="author">Bilkhu, Simranjeet</dim:field>
   <dim:field mdschema="dc" element="contributor" qualifier="department">Statistics</dim:field>
   <dim:field mdschema="dc" element="date" qualifier="accessioned">2026-04-22T18:49:06Z</dim:field>
   <dim:field mdschema="dc" element="date" qualifier="issued">2026</dim:field>
   <dim:field mdschema="dc" element="identifier" qualifier="uri">https://hdl.handle.net/11375/33066</dim:field>
   <dim:field mdschema="dc" element="identifier" qualifier="uri">https://doi.org/10.71548/20964</dim:field>
   <dim:field mdschema="dc" element="description" qualifier="abstract">Factor analytic models provide a framework for explaining high-dimensional data through a reduced set of latent variables. A fundamental challenge in these models is structural non-identification: the likelihood is invariant to orthogonal rotations, sign-switching, and column permutations of the factor loadings. To obtain a unique solution, structural constraints must be imposed. Two approaches based on sparsity-induced identification are the positive lower triangular (PLT) constraint and the unordered generalized lower triangular (UGLT) constraint. The PLT constraint resolves the rotation problem by imposing a fixed triangular geometry on the factor loadings, whereas the UGLT constraint generalizes this to a more flexible sparsity pattern.

In this thesis, we provide a comparative study of Bayesian factor models under these two identification strategies. We review the relevant prior specifications for each, including a column wise spike and slab prior that aids in inferring the number of relevant factors, and examine the corresponding Markov chain Monte Carlo (MCMC) algorithms by implementing them in R. Central to both algorithms is parameter expansion (PX), a technique that improves MCMC convergence for latent variable models by sampling from an expanded model that preserves the observed-data likelihood. Specifically, we demonstrate how a PX scheme based on the generalized inverse Gaussian distribution can be used to accelerate MCMC. For the PLT model, we show that integrating PX into the Gibbs sampler overcomes the slow mixing often inherent in rigid triangular constraints. For the UGLT model, we extend the partially collapsed Gibbs sampler of Frühwirth-Schnatter et. al by incorporating PX and adding a post-processing step to determine the number of relevant factors.

We evaluate the performance of both models through a series of simulation studies across varying numbers of observations, variables, and signal-to-noise ratios, under both PLT and UGLT data-generating structures. Model accuracy is assessed by comparing the amount of variance explained by the factor loadings and idiosyncratic variances against their true values, as well as computing the MSE of the posterior estimates of the factor loadings. The quality of the posterior predictive distribution is evaluated using the continuous ranked probability score (CRPS). We also examine the effect of parameter expansion through MCMC diagnostics, including trace plots, ACF plots, and effective sample size.

The results indicate that the UGLT model shows a clear advantage when the data-generating process follows a UGLT structure. Under a PLT structure with noisy variables, the sparse PLT and UGLT models perform worse than the non-sparse PLT model, suggesting that the spike-and-slab column sparsity prior does not adapt well when irrelevant variables are present. Notably, while the sparse models do identify zero loadings, they do not consistently assign them to the correct rows. We designed another simulation setting to test test how well each model identifies the relevant factors, we fit both models with more factors than the true number. We find that post-processing under the UGLT model better identifies the irrelevant factors compared to the sparse PLT model.</dim:field>
   <dim:field mdschema="dc" element="description" qualifier="degreetype">Thesis</dim:field>
   <dim:field mdschema="dc" element="description" qualifier="degree">Master of Science (MSc)</dim:field>
   <dim:field mdschema="dc" element="language" qualifier="iso">en</dim:field>
   <dim:field mdschema="dc" element="subject">Factor Analysis</dim:field>
   <dim:field mdschema="dc" element="subject">Bayesian Statistics</dim:field>
   <dim:field mdschema="dc" element="subject">Markov Chain Monte Carlo</dim:field>
   <dim:field mdschema="dc" element="title">A Comparative Study of Bayesian Factor Models under Positive Lower Triangular and Unordered Generalized Lower Triangular Identification Constraints</dim:field>
   <dim:field mdschema="dc" element="title" qualifier="alternative">Identification Constraints in Bayesian Factor Models</dim:field>
   <dim:field mdschema="dc" element="type" lang="en">Thesis</dim:field>
   <dim:field mdschema="others" element="access-status">embargo</dim:field>
</dim:dim></metadata></record></GetRecord></OAI-PMH>