A2074
Title: Learning directed latent variable networks with measurement error
Authors: Siliang Zhang - East China Normal University (China) [presenting]
Yunxiao Chen - London School of Economics and Political Science (United Kingdom)
Irini Moustaki - London School of Economics (United Kingdom)
Abstract: Recovering directed dependence structure among latent constructs is challenging when they are observed only through error-prone indicators. A common two-stage strategy estimates factor scores and applies standard causal discovery algorithms to them; treating these noisy estimates as observed variables distorts the conditional independence relations among the latent constructs. A likelihood-based procedure is proposed for learning sparse directed latent graphs in linear-Gaussian structural equation models with an identified measurement model. Estimation proceeds via the marginal observed-data likelihood, propagating measurement error from the indicators into the structural model. Acyclicity is imposed through a smooth, differentiable DAG constraint, and the resulting non-convex program is solved by an augmented Lagrangian scheme. Because this step is tuned to preserve true adjacencies at the cost of redundant edges, BIC-based backward pruning is then applied to recover a sparse graph. Identifiability of the latent Markov equivalence class is established under standard assumptions. Simulation studies demonstrate that explicitly modeling measurement error yields substantial gains in structural recovery over two-stage procedures based on estimated factor scores. An application to the Policing by Consent module of European Social Survey Round 5 contrasts the inferred latent structure with the conclusions of the official topline report.
