Safety Evaluation of Corner Clearance at Signalized Intersections

CHAPTER 4. METHODOLOGY

The research team used an observational cross-sectional study design for the evaluation. At the most basic level, the safety effect was estimated by taking the ratio of the average crash frequency for two groups, one with the treatment and the other without the treatment. The two groups of sites should be similar in all regards except for the presence of the treatment. This is difficult to accomplish in practice, and the research team adopted the propensity score matching technique to match treatment and reference sites while using multivariable regression modeling to control for other characteristics that vary among sites.

The research team employed multivariable regression to develop the statistical relationships between the dependent variables and a set of predictor variables. In this case, crash frequency was the dependent variable; the research team considered several predictor variables, including treatment presence, traffic volume, and other roadway characteristics. The regression coefficients for each predictor variable represented the expected change in crash frequency due to a unit change in the predictor variable with all else being equal.

The research team applied generalized linear modeling techniques to develop the crash prediction models and specified a log-linear relationship using a negative binomial error structure. The negative binomial error structure has advantages over the Poisson distribution in that it allows for overdispersion of the variance that is often present in crash data.

After developing a propensity score-matched dataset, the research team employed the following protocol to develop the multivariable models:

• Step 1—Develop base models with traffic volume only.
  
  
• Step 2—Explore the value of including other predictor variables.
  
  
• Step 3—Select the final model with the variable of interest (corner clearance), traffic volume, and other predictor variables as appropriate.

The research team determined the appropriate form for the base models (Step 1) according to the procedure outlined in Hauer.⁽⁸⁾ The research team added predictor variables to the base models and assessed them one at a time to determine the appropriate functional form and value added. The team then used various functional forms to assess potential relationships between crash frequency and continuous variables (e.g., speed limit) and to determine if the continuous variables could be best represented as continuous or indicator variables (e.g., use indicator variables for different speed limits). In this process, the research team also used a correlation matrix to consider correlations among predictor variables and prioritize the inclusion of correlated variables in the final models. Once the research team had included a variable in the model, they examined estimated parameters and associated standard errors (SEs) to determine the following:

• Is the direction of effect (i.e., expected decrease or increase in crashes) in general agreement with expectations?
  
  
• Does the magnitude of the effect seem reasonable?
  
  
• Are the parameters of the model estimated with statistical significance?
  
  
• Does the estimated overdispersion parameter improve significantly?