Online Program

283885
Integrating prior knowledge into statistical models using directed acyclic graphs with application to the longitudinal association between life satisfaction and violent and aggressive behavior


Tuesday, November 5, 2013

Michael Regier, PhD, Department of Biostatistics, West Virginia University, Morgantown, WV
Matthew Gurka, PhD, Department of Biostatistics, West Virginia University, Morgantown, WV
Keith Zullig, MSPH, PhD, Social and Behavioral Sciences, West Virginia University, Morgantown, WV
Purpose: The process of designing, analyzing, and reporting a statistical investigation is dynamic, involving a trans-disciplinary team of experts. Frequently, expert knowledge is implicitly incorporated in the statistical design by a biostatistician and is unnoticed by other team members. This study articulates one tool that can move this integration process from an implicit to an explicit exercise in which all researchers can participate.

Methods: We use a study of the effect of life satisfaction (LS) on violent and aggressive behavior among children to illustrate how directed acyclic graphs (DAGs) can be used to explicitly incorporate expert knowledge into the design and analysis of complex data sets. DAGs are a mathematical tool used to identify structural relationships between independent and dependent variables. They can articulate confounding relationships, mediators and potential effect modifiers. The graph itself becomes a conceptual framework that guides the process of statistical model construction.

Results: We constructed a marginal structural model (MSM) predicated on the collaborative design of a DAG to study the effect of LS on violent and aggressive behavior. We were able to identify a subset of time-varying confounders that were explicitly incorporated into the MSM. We compared our results with an unadjusted regression and a confounder-adjusted regression model to illustrate the potential bias that can be introduced by using these traditional methods of identifying relationships between independent and dependent variables.

Conclusions: DAGs are a tool that can facilitate trans-disciplinary conversations related to statistical model construction.

Learning Areas:

Biostatistics, economics

Learning Objectives:
Describe how to construct a basic directed acyclic graph. Demonstrate the structural definition of confounding, time-varying confounders, and baseline variables. Formulate a structural graph that integrates expert knowledge for translation to statistical models.

Keyword(s): Statistics, Collaboration

Presenting author's disclosure statement:

Qualified on the content I am responsible for because: I am funded by the department of Pediatrics and the WVCTSI where I engage in collaborative research: statistical design, analysis and reporting. Prior to WVU, I held a postdoctoral fellowship in biostatistics at McGill University, under the supervision of Dr. Moodie. I researched advanced causal methods focusing on structural models, directed acyclic graphs and time-varying data structures.
Any relevant financial relationships? No

I agree to comply with the American Public Health Association Conflict of Interest and Commercial Support Guidelines, and to disclose to the participants any off-label or experimental uses of a commercial product or service discussed in my presentation.