The National Academies Press

Currently Skimming:

Pages 127-135

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 127... ... 111 7 FINDINGS AND CONCLUSIONS 7.1 PROPOSED MODELS AND PROCEDURES FOR MANUAL This report presents SPFs that were estimated to predict crashes by type and severity for the facility types covered by the HSM. To optimize the accuracy of the crash predictions, it would have been ideal to estimate SPFs for all of the crash types and severities need as base conditions for applying CMFs. Unfortunately, because the numbers of crashes of various types and severities were limited in the databases available for the project, we could not estimate models for all the specific types -- for example, for head‐on and rear‐end crashes within the opposite‐ and same‐direction crash type categories, respectively. Some models also have overdispersion parameters high enough to cast doubt on their accuracy for prediction or coefficients on the volume predictors (AADT) that are not statistically significant at 90 percent confidence. We address these cases by reporting average proportions of specific crash types within the broader crash type category for which models were estimated. These proportions can be used where the predicted models are not available, or where the analyst chooses not to use them. This approach will still be more accurate than that provided in the current HSM. While the initial scope of work had proposed to estimate probabilistic models for crash severity, based on theoretical and practical considerations about the application of such models in the HSM procedures, we chose not to use this approach. Predictions of crash severity may be calculated using the count models for severity that have been estimated and presented here. The data sources for estimation and validation for each model by facility are listed in Table 7‐1. The rest of this section identifies the models that were estimated and will be proposed for inclusion in the HSM, by facility type. It also lists the crash types and severity for which models were NOT estimated, for each facility type that might require proportions to be estimated. For these situations, we provide default proportions from the data for each facility type, although individual jurisdictions may calculate proportions from their own data for more accurate local predictions. Finally, the section summarizes findings from revisiting the calibration procedures in the HSM in light of the newly estimated models. Table 7‐1: Data Sources (States) Read the entire page →
From page 128... ... 112 direction, and single‐vehicle crashes. Models for intersecting‐direction crashes were not estimated, as they were all assigned to intersections. Figure 7‐2 lists the specific crash types included in each of the broader crash type categories that were estimated. Total crashes include all of these crash types. If predictions of crashes of any of these specific types are needed for applying CMFs, proportions of them within the aggregate crash types (those in the left column of the figure) will be required and provided in the proposed HSM content. Figure 7‐3 illustrates the crash models estimated for intersections (all types) Read the entire page →
From page 129... ... 113 Figure 7‐2: Specific Crash Types Included in the Estimated Crash Types (Rural 2U) . Figure 7‐3: Crash Types Estimated for Intersections on Two‐Lane Rural Highways. Same Direction Crashes • Rear End • Sideswipe Same Direction • Turning Same Direction Opposite Direction Crashes • Head On • Sideswipe Opposite Direction • Turning Opposite Direction Single Vehicle Crashes • Fixed Object • Roll Over • Moving Object Total Crashes KABC0 KABC KAB KA Same Direction Crashes KABC0 KABC KAB KA Intersecting Direction Crashes KABC0 KABC KAB KA Opposite Direction Crashes KABC0 KABC KAB KA Single Vehicle Crashes KABC0 KABC KAB KA Read the entire page →
From page 130... ... 114 Figure 7‐4: Specific Crash Types included in the Estimated Crash Types (Rural 3ST, 4ST and 4SG) Multilane Rural Highway Models Figure 7‐5 identifies the crash models estimated for divided and undivided segments on multilane rural highways. Models were estimated for the same combinations of type and severity as for two‐lane rural highways, with two exceptions. First, due to the small number of same‐direction KA crashes, we could not estimate a model for that combination. Second, we attempted models for intersecting‐direction crashes for all severity levels, but due to the small number of crashes, only the model for all severity levels was successfully estimated. Figure 7‐4 lists the specific crash types included in each of these aggregated types (same as for two‐lane rural highway intersection models) Read the entire page →
From page 131... ... 115 Figure 7‐5: Crash Types Estimated for Divided and Undivided Segments on Multilane Rural Highways Intersection models cover the same crash types as for rural two‐lane highway intersections, as depicted in Figure 7‐3 and Figure 7‐4. Urban/Suburban Arterial Models Final base condition models were estimated for urban/suburban arterial segments for the following crash types:  Total  KABC  KAB  KA  Multiple‐vehicle non‐driveway related  Rear end  Sideswipe same direction  Head‐on + sideswipe opposite direction  Multiple‐vehicle non‐driveway other  Single vehicle  Nighttime  Multiple‐vehicle driveway Total Crashes KABC0 KABC KAB KA Same Direction Crashes KABC0 KABC KAB Intersecting Direction Crashes KABC0 Opposite Direction Crashes KABC0 KABC KAB KA Single Vehicle Crashes KABC0 KABC KAB KA Read the entire page →
From page 132... ... 116 Note that crashes were estimated by type or severity level, not in combination, as was done for the rural facility types. The reason for this is that, for many combinations of crash type and severity, there simply were not enough crashes to estimate viable models. The combination of crash type and severity is not frequently needed to apply HSM methods, so it is recommended that when predictions of such combinations are needed, proportions may be calculated to allocate crash type predictions among the various severity levels. For urban/suburban intersections, we estimated models for the same combinations of crash type and severity as for the rural intersection facilities (see Figure 7‐3) . As noted, however, models for many of the combinations could not be estimated due to small sample sizes or odd estimation results. Figure 7‐6 depicts the combinations that could not be estimated for each type of intersection. Read the entire page →
From page 133... ... 117 Figure 7‐6: Crash Type and Severity SPFs that Were Not Estimated for Urban/Suburban Intersections 3ST OD KA 4ST Total KAB KA SD KA OD KA ID KABC0 KABC KAB KA 3SG Total KA SV KABC KAB KA SD KA OD KA 4SG SV KA ID KA Read the entire page →
From page 134... ... 118 Revisit of the Calibration Procedure The project team revisited the current HSM calibration procedure and evaluated its performance relative to sample size and to using a constant or variable calibration factor (or calibration function) . In addition, we considered the issues of calibrating models based on crash predictions with and without CMFs and calibration of the overdispersion parameter. The recommendation is to continue the current procedure in the HSM of calibrating with the CMFs, assuming most of the CMFs for doing so are available. Otherwise, as may be the case for many crash type and severity models at the moment, the calibration may be done without applying CMFs. The issue of calibrating the overdispersion parameter requires further research. The findings show that the calibration results are definitely sensitive to sample size, but not always in ways that might be expected. The calibration function did not work well with small sample sizes because the optimization procedure to estimate it failed to converge. In general, the more complex the calibration approach, the more data are required to apply it successfully. Unsurprisingly, the sample sizes that resulted in the best calibration results would also be large enough to estimate jurisdiction‐specific SPFs. This latter option would be preferred, when possible, to get the most accurate predictions for application in a given jurisdiction. But the findings here show that, in many cases, reasonable predictions are also possible following the HSM procedures, with even a constant calibration factor and modest calibration sample sizes. 7.2 CONCLUSIONS In conclusion, this project has estimated new prediction models for crash types and severity that promise better predictive results than the current HSM‐recommended combination of base models for total crashes with proportional factors for allocating among crash type and severity. When sample size permitted, extensive SPFs developed by severity and type are provided for detailed analytics of safety rather than fixed proportions as the current HSM provides. They are estimated with much newer crash data than the models in the HSM, which were estimated using data 10 to 15 years old. These updated models, including ones for total crashes, reflect more current relationships between traffic exposure and crash occurrence, as well as differences in the shape of the SPF viz. the traffic exposure from one crash type or severity level to another. The project has also revisited the predictive method calibration procedure in the HSM and offers refinements to the recommended calibration procedure. It is noted that estimation and application of crash prediction models is dependent upon having datasets of sufficient size and quality. It was not possible to estimate models for K only crashes for any crash types or in total for any facility type due to the small number of these crashes in any of the data sets. For some crash types, such as same direction crashes, KA crash models also could not be estimated. Some of these crash type and severity combinations are extremely rare due to their nature (e.g., same direction KA crashes) Read the entire page →
From page 135... ... 119 This project also provides, in Appendix C, content for incorporating the new estimated models and calibration recommendations into a second edition of the HSM. Read the entire page →

From page 127...

... 111 7 FINDINGS AND CONCLUSIONS 7.1 PROPOSED MODELS AND PROCEDURES FOR MANUAL This report presents SPFs that were estimated to predict crashes by type and severity for the facility types covered by the HSM. To optimize the accuracy of the crash predictions, it would have been ideal to estimate SPFs for all of the crash types and severities need as base conditions for applying CMFs. Unfortunately, because the numbers of crashes of various types and severities were limited in the databases available for the project, we could not estimate models for all the specific types -- for example, for head‐on and rear‐end crashes within the opposite‐ and same‐direction crash type categories, respectively. Some models also have overdispersion parameters high enough to cast doubt on their accuracy for prediction or coefficients on the volume predictors (AADT) that are not statistically significant at 90 percent confidence. We address these cases by reporting average proportions of specific crash types within the broader crash type category for which models were estimated. These proportions can be used where the predicted models are not available, or where the analyst chooses not to use them. This approach will still be more accurate than that provided in the current HSM. While the initial scope of work had proposed to estimate probabilistic models for crash severity, based on theoretical and practical considerations about the application of such models in the HSM procedures, we chose not to use this approach. Predictions of crash severity may be calculated using the count models for severity that have been estimated and presented here. The data sources for estimation and validation for each model by facility are listed in Table 7‐1. The rest of this section identifies the models that were estimated and will be proposed for inclusion in the HSM, by facility type. It also lists the crash types and severity for which models were NOT estimated, for each facility type that might require proportions to be estimated. For these situations, we provide default proportions from the data for each facility type, although individual jurisdictions may calculate proportions from their own data for more accurate local predictions. Finally, the section summarizes findings from revisiting the calibration procedures in the HSM in light of the newly estimated models. Table 7‐1: Data Sources (States)

Read the entire page →

From page 128...

... 112 direction, and single‐vehicle crashes. Models for intersecting‐direction crashes were not estimated, as they were all assigned to intersections. Figure 7‐2 lists the specific crash types included in each of the broader crash type categories that were estimated. Total crashes include all of these crash types. If predictions of crashes of any of these specific types are needed for applying CMFs, proportions of them within the aggregate crash types (those in the left column of the figure) will be required and provided in the proposed HSM content. Figure 7‐3 illustrates the crash models estimated for intersections (all types)

Read the entire page →

From page 129...

... 113 Figure 7‐2: Specific Crash Types Included in the Estimated Crash Types (Rural 2U) . Figure 7‐3: Crash Types Estimated for Intersections on Two‐Lane Rural Highways. Same Direction Crashes • Rear End • Sideswipe Same Direction • Turning Same Direction Opposite Direction Crashes • Head On • Sideswipe Opposite Direction • Turning Opposite Direction Single Vehicle Crashes • Fixed Object • Roll Over • Moving Object Total Crashes KABC0 KABC KAB KA Same Direction Crashes KABC0 KABC KAB KA Intersecting Direction Crashes KABC0 KABC KAB KA Opposite Direction Crashes KABC0 KABC KAB KA Single Vehicle Crashes KABC0 KABC KAB KA

Read the entire page →

From page 130...

... 114 Figure 7‐4: Specific Crash Types included in the Estimated Crash Types (Rural 3ST, 4ST and 4SG) Multilane Rural Highway Models Figure 7‐5 identifies the crash models estimated for divided and undivided segments on multilane rural highways. Models were estimated for the same combinations of type and severity as for two‐lane rural highways, with two exceptions. First, due to the small number of same‐direction KA crashes, we could not estimate a model for that combination. Second, we attempted models for intersecting‐direction crashes for all severity levels, but due to the small number of crashes, only the model for all severity levels was successfully estimated. Figure 7‐4 lists the specific crash types included in each of these aggregated types (same as for two‐lane rural highway intersection models)

Read the entire page →

From page 131...

... 115 Figure 7‐5: Crash Types Estimated for Divided and Undivided Segments on Multilane Rural Highways Intersection models cover the same crash types as for rural two‐lane highway intersections, as depicted in Figure 7‐3 and Figure 7‐4. Urban/Suburban Arterial Models Final base condition models were estimated for urban/suburban arterial segments for the following crash types:  Total  KABC  KAB  KA  Multiple‐vehicle non‐driveway related  Rear end  Sideswipe same direction  Head‐on + sideswipe opposite direction  Multiple‐vehicle non‐driveway other  Single vehicle  Nighttime  Multiple‐vehicle driveway Total Crashes KABC0 KABC KAB KA Same Direction Crashes KABC0 KABC KAB Intersecting Direction Crashes KABC0 Opposite Direction Crashes KABC0 KABC KAB KA Single Vehicle Crashes KABC0 KABC KAB KA

Read the entire page →

From page 132...

... 116 Note that crashes were estimated by type or severity level, not in combination, as was done for the rural facility types. The reason for this is that, for many combinations of crash type and severity, there simply were not enough crashes to estimate viable models. The combination of crash type and severity is not frequently needed to apply HSM methods, so it is recommended that when predictions of such combinations are needed, proportions may be calculated to allocate crash type predictions among the various severity levels. For urban/suburban intersections, we estimated models for the same combinations of crash type and severity as for the rural intersection facilities (see Figure 7‐3) . As noted, however, models for many of the combinations could not be estimated due to small sample sizes or odd estimation results. Figure 7‐6 depicts the combinations that could not be estimated for each type of intersection.

Read the entire page →

From page 133...

... 117 Figure 7‐6: Crash Type and Severity SPFs that Were Not Estimated for Urban/Suburban Intersections 3ST OD KA 4ST Total KAB KA SD KA OD KA ID KABC0 KABC KAB KA 3SG Total KA SV KABC KAB KA SD KA OD KA 4SG SV KA ID KA

Read the entire page →

From page 134...

... 118 Revisit of the Calibration Procedure The project team revisited the current HSM calibration procedure and evaluated its performance relative to sample size and to using a constant or variable calibration factor (or calibration function) . In addition, we considered the issues of calibrating models based on crash predictions with and without CMFs and calibration of the overdispersion parameter. The recommendation is to continue the current procedure in the HSM of calibrating with the CMFs, assuming most of the CMFs for doing so are available. Otherwise, as may be the case for many crash type and severity models at the moment, the calibration may be done without applying CMFs. The issue of calibrating the overdispersion parameter requires further research. The findings show that the calibration results are definitely sensitive to sample size, but not always in ways that might be expected. The calibration function did not work well with small sample sizes because the optimization procedure to estimate it failed to converge. In general, the more complex the calibration approach, the more data are required to apply it successfully. Unsurprisingly, the sample sizes that resulted in the best calibration results would also be large enough to estimate jurisdiction‐specific SPFs. This latter option would be preferred, when possible, to get the most accurate predictions for application in a given jurisdiction. But the findings here show that, in many cases, reasonable predictions are also possible following the HSM procedures, with even a constant calibration factor and modest calibration sample sizes. 7.2 CONCLUSIONS In conclusion, this project has estimated new prediction models for crash types and severity that promise better predictive results than the current HSM‐recommended combination of base models for total crashes with proportional factors for allocating among crash type and severity. When sample size permitted, extensive SPFs developed by severity and type are provided for detailed analytics of safety rather than fixed proportions as the current HSM provides. They are estimated with much newer crash data than the models in the HSM, which were estimated using data 10 to 15 years old. These updated models, including ones for total crashes, reflect more current relationships between traffic exposure and crash occurrence, as well as differences in the shape of the SPF viz. the traffic exposure from one crash type or severity level to another. The project has also revisited the predictive method calibration procedure in the HSM and offers refinements to the recommended calibration procedure. It is noted that estimation and application of crash prediction models is dependent upon having datasets of sufficient size and quality. It was not possible to estimate models for K only crashes for any crash types or in total for any facility type due to the small number of these crashes in any of the data sets. For some crash types, such as same direction crashes, KA crash models also could not be estimated. Some of these crash type and severity combinations are extremely rare due to their nature (e.g., same direction KA crashes)

Read the entire page →

From page 135...

... 119 This project also provides, in Appendix C, content for incorporating the new estimated models and calibration recommendations into a second edition of the HSM.

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.