ims.22

IMS

Session Slot: August 12 8:30- 10:20 Tuesday

Estimated Audience Size: 125-175

AudioVisual Request: Two Overheads

Session Title: Conditional Independence Properties of Structural Equation Models

Theme Session: No

Applied Session: No

Session Organizer: Tritchler, David Ontario Cancer Institute and University of Toronto

Address: Ontario Cancer Institute 610 University Avenue Toronto, Ontario M5G 2M9 Canada

Phone: (416) 946 2064

Fax: (416) 946 2024

Email: tritchle@oci.utoronto.ca

Session Timing: 110 minutes total (Sorry about format):

Opening Remarks by Chair - 0 minutes First Speaker - 30 minutes Second Speaker - 30 minutes Third Speaker - 30 minutes Discussant - 15 minutes Floor Discusion - 5 minutes

Session Chair: Tritchler, David Ontario Cancer Institute and University of Toronto

Address: Ontario Cancer Institute 610 University Avenue Toronto, Ontario M5G 2M9 Canada

Phone: (416) 946 2064

Fax: (416) 946 2024

Email: tritchle@oci.utoronto.ca

1. Marginalization and Conditioning in Graphical Models

Koster, Jan,   Erasmus University Rotterdam

Address: Erasmus University Rotterdam Department of Sociology P.O. Box 1738 Rotterdam NL-3000 DR The Netherlands

Phone: +31.10.408.20.77

Fax: +31.10.452.58.70

Email: koster@soc.fsw.eur.nl

Abstract: A class of graphs is introduced which is closed under marginalization and conditioning. It contains all graphs associated with linear structural equation models. Using earlier consistency results, conditional independence properties of subsystems of linear structural equation models are discussed.

2. Analyzing Structural Underdetermination and Model Equivalence Via Graphical Models

Richardson, Thomas,   University of Washington

Address: University of Washington Department of Statistics P.O. Box 35-4322 Seattle, WA 98195-7232

Phone: +1 206 685 8488

Fax: +1 206 685 7419

Email: tsr@stat.washington.edu

Spirtes, Peter, Carnegie-Mellon University

Abstract: Analyses based on structural equation models often claim to provide insight into the nature of the mechanism which generated a given set of data. A fundamental criticism of structural equation modelling is that, without background knowledge, there may be many plausible data generating structures which give rise to statistically equivalent models. For example, there may be two SEM models which fit the data well, in one of which $A\rightarrow B$, while in the other $B\rightarrow A$. These problems are particularly severe in cases where there may be unmeasured confounding variables. One approach to addressing this problem is to consider equivalence classes of structural equation models. This talk will describe recent work which uses graphical methods to characterize equivalence classes of structural equation models with latent variables and/or correlated errors.

3. Testing the Testable and Identifying the Identifiable in Structural Equation Models

Pearl, Judea,   UCLA

Address: UCLA Department of Computer Science Room 4515 BH Los Angeles CA 90024

Phone: (310) 825 3243

Fax: (818) 789 6311

Email: judea@cs.ucla.edu

Abstract: Recent progress in graphical models has brought clarity, precision and legitimacy to structural equation modeling (SEM). Fundamental concepts that were left ambiguous by decades of inadequate formulations can now be given formal and operational meaning . What makes a set of equations ``structural,'' what assumptions should be ascertained by the authors of such equations, what policy claims are advertised by a given set of structural equations, and what portions of these claims are testable are some of the questions that now receive simple and mathematically precise answers. The talk will survey graphical techniques for identifying the testable claims of SEM, for recognizing the identifiable parameters in a given SEM, and for determining when a variable is exogenous, when a variable is an ``instrument,'' and what those ``so-called disturbance terms'' are.

Discussant: TBN  

Address:

Phone:

Fax:

Email:

List of speakers who are nonmembers: None