The National Academies Press

Currently Skimming:

Pages 118-126

The Chapter Skim interface presents what we've algorithmically identified as the most significant single chunk of text within every page in the chapter.
Select key terms on the right to highlight them within pages of the chapter.

From page 118... ... 102 6 REVISITING THE HSM CALIBRATION APPROACH 6.1 APPROACHES CONSIDERED Background on HSM Approach The development of new models for the HSM, taken together with research conducted since its release in 2010 on key issues pertaining to the calibration procedure, provided the need and the opportunity to revisit that procedure in this research project with a view to updating it. The key issues, which are interrelated with others, pertain to the sample size for calibration data and to whether and how to capture the variation of the calibration factor with site characteristics. To address the latter issue, we investigated a procedure based on calibration functions. A review of the research on establishing minimum sample sizes and estimating calibration functions, along with the results of an empirical investigation in this project, led to the proposed calibration procedure update documented here. The research review suggested that required samples will, indeed, vary across site types, jurisdictions, and crash types and severities. In particular, a consensus seemed apparent that the desirable minimum suggested in the HSM of 30–50 sites with at least 100 crashes a year might not be universally applicable. The research carried out since 2010 has not, however, provided any consistent guidance on what does constitute an appropriate sample. In some cases, recommended sample sizes are so large that a jurisdiction may be better off acquiring (or hiring) personnel with the skill sets required to estimate their own models directly rather than calibrate an external one. The sample size guidance in the procedure recommended here is based on a report by Bahar et al. (2014) Read the entire page →
From page 119... ... 103 Approach 1 We used three representative site types for this investigation: urban four‐lane divided segments; urban two‐lane divided segments; and rural two‐lane, three‐leg stop‐controlled intersections. The final models estimated and presented in earlier chapters were calibrated to randomly selected sites from another jurisdiction to increase sample sizes. We also directly estimated models with model forms identical to those being calibrated, with the exception that we used a constant overdispersion parameter. The logic behind this "iterative" approach was that, at small sample sizes, applying either a calibration factor or function to an original model would prove superior to using a directly estimated model. As sample sizes increased, there would be a point at which a directly calibrated model would perform better. At the other end of the spectrum, there would also be a point at which the sample size would be too small even to estimate a reliable calibration factor. We evaluated the performance of a calibrated model using several criteria provided by the FHWA Calibrator spreadsheet tool (Lyon et al. 2016) . The guidance this tool provided indicated a calibrated model is reasonable if either the coefficient of variation (CV) Read the entire page →
From page 120... ... 104 3. The calibration function does perform better in general than a calibration factor, although the differences are not very large for these data. 4. Read the entire page →
From page 121... ... 105 indicate that at small sample sizes the percentage outside these limits may be small simply due to the small sample. 6.3 APPROACH 2 RESULTS The investigation for this approach and site type involved an assessment of the temporal and spatial transferability and calibration of the models based on the CV of the calibration factor. In this case, we used all of the data available for the calibration rather than samples of various sizes. First, we applied Texas 2012 data for calibration of the SPFs, using Texas 2009–11 data for undivided highway segments. Table 6‐4 shows the results. Then, we used Ohio 2009–11, Washington 2009–11, and Illinois 2009–10 data for calibration of the California SPFs for divided highway segments. The results are shown in Table 6.5. The results in Table 6‐5 indicate that, for Ohio and Illinois, the calibration function would provide predictions similar to those provided by a single calibration factor, since parameter b (Equation 6.1) was close to 1.0. No insights could be obtained on sample sizes of sites and crashes, as the results were not only inconsistent but very jurisdiction‐specific. The lowest MAD value, for example, was for the data with the largest number of sites but the fewest crashes and the highest value of the CV of the calibration factor. The temporal calibration results in Table 6‐4 show parameter b of the calibration function was also close to 1.0, but even with a relatively large sample of sites and crashes for the same state, the CV of the calibration factor was beyond the threshold of 0.15 recommended for a successful calibration. Read the entire page →
From page 122... ... 106 Table 6‐1: Results for Urban Four‐Lane Divided Segments No. Sites Observed Crashes C (CV) Read the entire page →
From page 123... ... 107 Table 6‐3: Results for Rural Two‐Lane Three‐Leg Stop‐controlled Intersections No. Sites Observed Crashes C (CV) Read the entire page →
From page 124... ... 108 6.4 CONCLUSIONS ON CALIBRATION EXERCISE Summary of Findings The results of the analyses indicate no consistency with regard to which option (calibration factor, calibration function, or directly estimated model) will perform best for a given sample size. For some cases, a small sample that is estimated using some criterion (for example, maximum CV of the calibration factor) Read the entire page →
From page 125... ... 109 o If a successful calibration cannot be achieved with the entire sample available for total crashes, then the calibration results for a similar site type (from which a successful calibration was achieved) may be assumed to apply. o If a successful calibration cannot be achieved with the entire sample available for a specific crash type or severity, then the calibration results for total crashes, however obtained, may be assumed to apply. 6. Read the entire page →
From page 126... ... 110 Figure 6‐1: Suggested Calibration Process Read the entire page →

From page 118...

... 102 6 REVISITING THE HSM CALIBRATION APPROACH 6.1 APPROACHES CONSIDERED Background on HSM Approach The development of new models for the HSM, taken together with research conducted since its release in 2010 on key issues pertaining to the calibration procedure, provided the need and the opportunity to revisit that procedure in this research project with a view to updating it. The key issues, which are interrelated with others, pertain to the sample size for calibration data and to whether and how to capture the variation of the calibration factor with site characteristics. To address the latter issue, we investigated a procedure based on calibration functions. A review of the research on establishing minimum sample sizes and estimating calibration functions, along with the results of an empirical investigation in this project, led to the proposed calibration procedure update documented here. The research review suggested that required samples will, indeed, vary across site types, jurisdictions, and crash types and severities. In particular, a consensus seemed apparent that the desirable minimum suggested in the HSM of 30–50 sites with at least 100 crashes a year might not be universally applicable. The research carried out since 2010 has not, however, provided any consistent guidance on what does constitute an appropriate sample. In some cases, recommended sample sizes are so large that a jurisdiction may be better off acquiring (or hiring) personnel with the skill sets required to estimate their own models directly rather than calibrate an external one. The sample size guidance in the procedure recommended here is based on a report by Bahar et al. (2014)

Read the entire page →

From page 119...

... 103 Approach 1 We used three representative site types for this investigation: urban four‐lane divided segments; urban two‐lane divided segments; and rural two‐lane, three‐leg stop‐controlled intersections. The final models estimated and presented in earlier chapters were calibrated to randomly selected sites from another jurisdiction to increase sample sizes. We also directly estimated models with model forms identical to those being calibrated, with the exception that we used a constant overdispersion parameter. The logic behind this "iterative" approach was that, at small sample sizes, applying either a calibration factor or function to an original model would prove superior to using a directly estimated model. As sample sizes increased, there would be a point at which a directly calibrated model would perform better. At the other end of the spectrum, there would also be a point at which the sample size would be too small even to estimate a reliable calibration factor. We evaluated the performance of a calibrated model using several criteria provided by the FHWA Calibrator spreadsheet tool (Lyon et al. 2016) . The guidance this tool provided indicated a calibrated model is reasonable if either the coefficient of variation (CV)

Read the entire page →

From page 120...

... 104 3. The calibration function does perform better in general than a calibration factor, although the differences are not very large for these data. 4.

Read the entire page →

From page 121...

... 105 indicate that at small sample sizes the percentage outside these limits may be small simply due to the small sample. 6.3 APPROACH 2 RESULTS The investigation for this approach and site type involved an assessment of the temporal and spatial transferability and calibration of the models based on the CV of the calibration factor. In this case, we used all of the data available for the calibration rather than samples of various sizes. First, we applied Texas 2012 data for calibration of the SPFs, using Texas 2009–11 data for undivided highway segments. Table 6‐4 shows the results. Then, we used Ohio 2009–11, Washington 2009–11, and Illinois 2009–10 data for calibration of the California SPFs for divided highway segments. The results are shown in Table 6.5. The results in Table 6‐5 indicate that, for Ohio and Illinois, the calibration function would provide predictions similar to those provided by a single calibration factor, since parameter b (Equation 6.1) was close to 1.0. No insights could be obtained on sample sizes of sites and crashes, as the results were not only inconsistent but very jurisdiction‐specific. The lowest MAD value, for example, was for the data with the largest number of sites but the fewest crashes and the highest value of the CV of the calibration factor. The temporal calibration results in Table 6‐4 show parameter b of the calibration function was also close to 1.0, but even with a relatively large sample of sites and crashes for the same state, the CV of the calibration factor was beyond the threshold of 0.15 recommended for a successful calibration.

Read the entire page →

From page 122...

... 106 Table 6‐1: Results for Urban Four‐Lane Divided Segments No. Sites Observed Crashes C (CV)

Read the entire page →

From page 123...

... 107 Table 6‐3: Results for Rural Two‐Lane Three‐Leg Stop‐controlled Intersections No. Sites Observed Crashes C (CV)

Read the entire page →

From page 124...

... 108 6.4 CONCLUSIONS ON CALIBRATION EXERCISE Summary of Findings The results of the analyses indicate no consistency with regard to which option (calibration factor, calibration function, or directly estimated model) will perform best for a given sample size. For some cases, a small sample that is estimated using some criterion (for example, maximum CV of the calibration factor)

Read the entire page →

From page 125...

... 109 o If a successful calibration cannot be achieved with the entire sample available for total crashes, then the calibration results for a similar site type (from which a successful calibration was achieved) may be assumed to apply. o If a successful calibration cannot be achieved with the entire sample available for a specific crash type or severity, then the calibration results for total crashes, however obtained, may be assumed to apply. 6.

Read the entire page →

From page 126...

... 110 Figure 6‐1: Suggested Calibration Process

Read the entire page →

Key Terms

← Previous Chapter Skim

Next Chapter Skim →

This material may be derived from roughly machine-read images, and so is provided only to facilitate research.
More information on Chapter Skim is available.