Safety Evaluation of Access Management Policies and Techniques

CHAPTER 10. CALIBRATION

Calibration of the models involves the estimation and application of adjustment factors for applying the models in a jurisdiction and time period that are different from those used to develop the original models. The factors reflect differences in crash experience due to differences in terrain, climate, crash definition, and reporting. The following procedure, based on AASHTO’s Highway Safety Manual, is suggested for approximating calibration factors, recognizing that there will be limitations in availability of data required for more thorough procedures:

1. For a given crash type and land use, consider which alternative model requires data for which a suitably large dataset can be assembled. Note that models with the most variables are preferred. The calibration sample should average at least 100 crashes/yr for the most recent 3-yr period and contain at least 10 corridors. Note that if the data cannot be assembled for all crash and land use types, the procedure is applied to datasets that are available, and the calibration factor estimated in step 3 is assumed for the crash type and land use types for which calibration data are unavailable.⁽²⁰⁾
   
   
2. Use the model for the region deemed most similar to the jurisdiction of interest to estimate the sum of predicted crashes over all corridors in the dataset for the 3-yr period. Note that the assessment of which region is most similar is not critical; the region is mainly used to define a base condition, but experience-based judgment may be used in assessing the reasonableness of the calibration factor estimated in step 3.⁽²⁰⁾
   
   
3. Estimate the calibration factor as the ratio of the sum of the observed crashes in the calibration dataset to the sum of the predicted crashes from the model output.⁽²⁰⁾
   
   
4. Apply the calibration factor to the base region multiplier to obtain the multiplier for the jurisdiction of interest.⁽²⁰⁾

Example: Suppose it is desired to calibrate the model for predicting total crashes on commercial corridors in a jurisdiction. A total of 10 commercial corridors are identified for use in the calibration process. The dataset includes 328 reported crashes for the 3-yr period. Note that there are more than 100 crashes/yr for the 10 corridors combined, so this satisfies the requirements in step 1 of the calibration process.

Step 1. Based on a review of table 25, there are two alternative models for predicting total crashes on commercial corridors. Model 1 includes ACCDENS and SIGDENS, and model 2 includes PROPNODEV. If data are available for ACCDENS and SIGDENS for each of the 10corridors, then model 1 would be selected because it includes more variables than model 2. In this example, assume that data are not available for ACCDENS, but data are available for PROPNODEV. Thus, model 2 is selected for use in calibration based on availability of data. The detailed information for model 2 is provided in table 48, and the model form is given by the equation in figure 18, where exp^{(intercept + region)} is the regional multiplier.

Step 2. Based on a comparison of local roadway characteristics and crash statistics, it was determined that Minnesota is most similar to the jurisdiction of interest. The model is applied to predict crashes for each of the 10 corridors in each year, and the results are summed. A total of 360.25 crashes are predicted for the 10 corridors over the 3-yr period.

Step 3. The estimated calibration factor is calculated using the equation in figure 50:

Figure 50. Equation. Estimated calibration factor.

Step 4. The original multiplier from table 48 is shown in figure 51:

Figure 51. Equation. Original multiplier.

The calibrated multiplier for use in the jurisdiction of interest is shown in figure 52:

Figure 52. Equation. Estimation of calibrated multiplier.

The calibrated model for the jurisdiction is shown in figure 53:

Figure 53. Equations. Minnesota crash prediction model.

In the same way, multipliers can be obtained for all other crash types and land use types of interest. If data are unavailable or insufficient to estimate a calibration factor for specific crash types for the same land use type (commercial in this example), the calibration factor for total crashes (0.910) can be applied to the models for other crash types of interest for commercial corridors. Similarly, if data are unavailable or insufficient to estimate a calibration factor for other land use categories for the same crash type (total crashes in this example), the calibration factor for total crashes (0.910) can be applied to the models for total crashes for other land use categories of interest.