Back to Annual Meeting

Back to Annual Meeting

APHA Scientific Session and Event Listing

Recorded presentation

Thomas M. Rice, MPH, PhD¹, Craig Anderson, DHSc, PhD², and David R. Ragland, PhD, MPH¹. (1) Traffic Safety Center, University of California Berkeley, 140 Warren Hall, #7360, Berkeley, CA 94720-7360, 510-643-7625, tomrice@berkeley.edu, (2) Department of Emergency Medicine, University of California Irvine, 101 The City Drive, Building 200, Suite 715, Orange, CA 92868

Background / Objectives: A severe selection bias can result from the application of unconditional analysis methods (e.g., unconditional logistic regression) to data from the Fatality Analysis Reporting System (FARS), a data system that includes information on virtually all fatal traffic collisions in the United States. We explore this problem and discuss correct analysis options.

Method: Using literature search methods, we identified and discussed the various methods that have been used to analyze data from matched cohort studies. We then applied the methods to data from a study of seat belt effectiveness among occupants of passenger vehicles involved in fatal crashes. We also simulated data with known parameters (e.g., we set the seat belt risk ratio to 0.5) using Monte Carlo methods, applied the various methods, and compared the model coefficients and variances.

Results: We propose that a widely overlooked bias can be avoided by applying conditional regression methods to matched cohort FARS data. All unconditional methods that we tested gave incorrect coefficient and variance estimates, while most conditional methods gave correct coefficient estimates but did not estimate correct variances. Pseudoscore and maximum pseudolikelihood estimating equation regression methods supplied correct estimates for coefficients and variances.

Discussion: Many traffic safety researchers are unaware of the bias that can result from the application of unconditional analysis methods to FARS data. A few recent studies have employed conditional Poisson regression, which produces correct risk ratios estimates but incorrect variance estimates. We recommend two methods and offer Stata procedures to implement the methods.

Learning Objectives:

• Differentiate between “conditional” and “unconditional” methods for analyzing crash data
• Describe the potential bias due to the application of unconditional methods to FARS data
• Recognize the inherent “matching” in most FARS data samples
• Identify conditional methods that can be applied to these data to avoid such bias

Keywords: Motor Vehicles, Methodology

Presenting author's disclosure statement:

Any relevant financial relationships? No

Recorded presentation