Safety Evaluation of Access Management Policies and Techniques

CHAPTER 3. METHODOLOGY

Study designs fall into one of two general study types: experimental and observational. Experimental studies are planned; that is, entities identified for some treatment are then randomly assigned to either a treatment or to a control group that is left untreated. Observational studies are not planned; that is, data are collected by observing the performance of entities, where the treatment is implemented at some sites, not on the basis of a planned experiment. While experimental studies are useful to control for factors other than the treatment of interest, they are often excluded in highway safety research because of ethical concerns regarding experimentation in road safety. Thus, observational studies are more common in road safety research and are the basis for this study.

Several observational study designs are available to assess the safety impacts of AM strategies. Well-designed before–after studies are often preferred to estimate crash modification factors (CMFs), while cross-sectional models are often necessary to develop crash prediction models. In this case, the objective was to develop crash prediction models, so a cross-sectional approach was selected.

The safety impact of a given feature can be derived from a cross-sectional study by comparing the safety of a group of sites with that feature with the safety of a group of sites without that feature. This type of comparison directly relates to the investigation of AM strategies (e.g., TWLTL versus undivided road). The safety effect is estimated by taking the ratio of the average crash frequency for the two groups. For this method to work, the two groups should be similar in all ways except the feature of interest. In practice, this is difficult to accomplish, and multiple variable regression models are used to estimate the effects of one feature while controlling for other characteristics that vary among the sites. These cross-sectional models are also called “crash prediction models,” which are mathematical equations that relate crash frequency with site characteristics. While cross-sectional models provide a means to estimate the safety impacts of AM strategies, there are potential issues that need to be addressed. Table 4 identifies potential issues and biases associated with cross-sectional models and opportunities to overcome these limitations.

Table 4. Potential issues and opportunities related to cross-sectional studies.

The following potential biases were identified in this study with an explanation of how they were addressed or dismissed:

• Selection of appropriate functional form. Generalized linear modeling (GLM) techniques were applied to develop corridor-level crash prediction models. A log-linear relationship was specified using a negative binomial error structure following the state of the art in modeling crash data. The negative binomial error structure is now recognized as more appropriate for crash counts than the normal distribution assumed in conventional regression modeling. The negative binomial error structure also has advantages over the Poisson distribution in that it allows for the overdispersion that is often present in crash data. The appropriate model form for each variable was determined following a review of the data.
  
  
• Accounting for State-to-State differences. Data from four regions were used to develop the final crash prediction models, and indicator variables were included in the models to identify the respective region for each corridor.
  
  
• Correlation among independent variables. The correlation matrix of the estimated parameters was examined to determine the extent of correlation among independent variables. Several AM strategies are highly correlated, and it was necessary to develop multiple models with subsets of the independent variables rather than one single model with all variables.
  
  
• Overfitting of prediction models. Relatively few parameters were included in the final models because of the correlation issue. Consequently, overfitting was dismissed as a potential issue.
  
  
• Low sample mean and sample size. Corridor-level models help to overcome issues related to low sample mean because each site (i.e., corridor) typically experiences multiple crashes per year. Sample size was addressed during the early planning stages of the study, and more than 600 mi of data were obtained to provide a large database for analysis. When possible, 3 yr of crash data were obtained to further increase the sample size.
  
  
• Aggregation, averaging, or incompleteness in data. Data were obtained from various sources and supplemented with field measurements to ensure a relatively complete and accurate dataset. While multiple years of crash data were used in the analysis, the number of years was also included as an independent variable to account for the multiple years of data (i.e., the model prediction results in crashes per year). Only a maximum of 3 yr of data were used for any site.
  
  
• Temporal and spatial correlation. Temporal correlation may arise if multiple observations are used for the same entity. In this case, multiple years of data were used for each corridor, but these data were aggregated into a single observation because the maximum was limited to 3 yr, and the number of years was included as an independent variable. Spatial correlation is a potential issue, but the corridors were selected from four regions and were relatively dispersed within each of the regions. Indicator variables were also included to account for similarities within regions.
  
  
• Endogenous independent variables. Endogeneity arises when one or more of the independent variables depend on the dependent variable. For example, left-turn lanes may be installed because of the frequency of left-turn crashes at an intersection. A cross-sectional model that predicts crash frequency based on the presence of left-turn lanes and other factors may conclude that left-turn lanes increase crashes. Similar examples could be drawn for other AM strategies. In this study, endogeneity was not considered to be a substantial threat because data were aggregated at the corridor level.
  
  
• Omitted variable bias. It is difficult, if not impossible, to completely account for the potential effects of omitted variable bias in an observational cross-sectional study. In this case, omitted variable bias was addressed to the extent possible by carefully considering the roadway and traffic characteristics that should be included in the models. Detailed data were collected for each corridor, and numerous variables were tested for suitability in the models. There is the potential for omitted variable bias due to other factors such as weather, driver population, and vehicle fleet, but a regional indicator variable was included in the models to help to account for these differences among the regions.