[Paper Review] Graphical Markov models, unifying results and their interpretation

Nanny Wermuth|arXiv (Cornell University)|May 10, 2015

Bayesian Modeling and Causal Inference76 references3 citations

TL;DR

This paper unifies graphical Markov models, particularly regression graph models, to model causal pathways in longitudinal and intervention studies. By integrating conditional independence constraints with directed acyclic and undirected graphs, it enables tracing structural effects of variable omission, subpopulation selection, and order changes in data generation—offering a framework for predicting outcomes and integrating knowledge across studies.

ABSTRACT

Graphical Markov models combine conditional independence constraints with graphical representations of stepwise data generating processes.The models started to be formulated about 40 years ago and vigorous development is ongoing. Longitudinal observational studies as well as intervention studies are best modeled via a subclass called regression graph models and, especially traceable regressions. Regression graphs include two types of undirected graph and directed acyclic graphs in ordered sequences of joint responses. Response components may correspond to discrete or continuous random variables and may depend exclusively on variables which have been generated earlier. These aspects are essential when causal hypothesis are the motivation for the planning of empirical studies. To turn the graphs into useful tools for tracing developmental pathways and for predicting structure in alternative models, the generated distributions have to mimic some properties of joint Gaussian distributions. Here, relevant results concerning these aspects are spelled out and illustrated by examples. With regression graph models, it becomes feasible, for the first time, to derive structural effects of (1) ignoring some of the variables, of (2) selecting subpopulations via fixed levels of some other variables or of (3) changing the order in which the variables might get generated. Thus, the most important future applications of these models will aim at the best possible integration of knowledge from related studies.

Motivation & Objective

• To unify and systematize graphical Markov models, especially regression graph models, for modeling causal processes in longitudinal and intervention studies.
• To address the challenge of modeling how structural effects emerge when variables are ignored, subpopulations are selected, or variable generation order is altered.
• To ensure that the generated distributions reflect key properties of joint Gaussian distributions for reliable inference and prediction.
• To provide a framework for integrating knowledge across related empirical studies by modeling developmental pathways and structural changes.

Proposed method

• Combines conditional independence constraints with graphical representations using ordered sequences of joint responses in regression graphs.
• Uses mixed graphs with both directed acyclic edges (for causal dependencies) and undirected edges (for conditional independence) to represent data-generating processes.
• Applies the properties of joint Gaussian distributions to ensure that the generated distributions support valid inference and prediction.
• Models structural effects of variable omission by analyzing conditional distributions under marginalization.
• Analyzes subpopulation selection by conditioning on fixed levels of observed variables to assess downstream impacts.
• Investigates changes in variable generation order by reordering the sequence of responses in the graphical model to predict altered dependencies.

Experimental results

Research questions

• RQ1 How can graphical Markov models be unified to represent causal processes in longitudinal and intervention studies?
• RQ2 What are the structural effects of ignoring certain variables in a regression graph model?
• RQ3 How does selecting a subpopulation via fixed levels of some variables alter the conditional independence structure?
• RQ4 What happens to the model structure when the order of variable generation is changed?
• RQ5 How can regression graph models be used to predict and compare alternative data-generating processes?

Key findings

• Regression graph models provide a unified framework for modeling causal pathways in longitudinal and intervention studies by combining directed and undirected edges.
• The models allow for the derivation of structural effects resulting from the omission of variables, enabling analysis of missing data impacts.
• Subpopulation selection via fixed levels of variables leads to identifiable changes in conditional independence structures, which can be traced through the graph.
• Changing the order of variable generation alters the dependency structure, and these changes can be systematically analyzed using the graphical model.
• The models preserve key properties of joint Gaussian distributions, ensuring reliable prediction and inference under various data-generating assumptions.
• These models enable the integration of knowledge from related studies by modeling developmental pathways and structural changes across different empirical contexts.