Safety Evaluation of Access Management Policies and Techniques

CHAPTER 11. VALIDATION

Traditional validation involves calculating and assessing goodness-of-fit measures for applying the model to an independent dataset not used in the model development process. Often, as is the case here, it is necessary to use all data for the model development, so validity needs to be assessed using alternative methods. The alternative method applied in this report uses previous studies to assess the reasonableness of the effects indicated by the independent variables in the models. The assessment is done for two sets of variables—those for which effects can be implied from AASHTO’s Highway Safety Manual predictive models and those for which CMFs can be obtained from FHWA’s CMF Clearinghouse.^(20,21)

COMPARISON WITH HIGHWAY SAFETY MANUAL PREDICTIVE MODELS

For this exercise, the implied effects of two variables—PROPDIV and ACCDENS—were compared with the implied effects of models for urban and suburban arterials in AASHTO’s Highway Safety Manual.⁽²⁰⁾

For PROPDIV, there was only one model for which this variable was significant and had the intuitively correct sign—right-angle crashes for mixed land use (table 45). The implied CMF for right-angle crashes for installing a raised median (PROPDIV = 1) in a corridor where there was no median (PROPDIV = 0) is exp^(–0.4710) = 0.624. The CMFs inferred for four-lane arterials from the Highway Safety Manual, Part C Predictive Method, depend on AADT, driveway spacing, and driveway type. For total crashes, the inferred CMFs for major commercial land use from the Highway Safety Manual models range from 0.456 for 40 driveways/mi and 5,000 AADT to 0.676 for an arterial with no driveways and AADT of 40,000. Thus, even considering that the comparison is for two different crash types, it can be concluded that the implied CMF from the model in table 45 is reasonably consistent with the CMF implied from the Highway Safety Manual models. ⁽²⁰⁾

For ACCDENS (access density), the CMFs inferred from the Highway Safety Manual, Part C Predictive Method, are shown in table 29, taken from Persaud et al.⁽²²⁾ The Highway Safety Manual provides separate models for multivehicle driveway and nondriveway crashes per mile, considering the AADT, the number and type of driveways, and whether or not the arterial is divided.⁽²⁰⁾ Thus, the inferred CMFs for changing driveway spacing depend on these factors as shown in table 29. Based on table 34, the CMF for total crashes for reducing driveway density by 10/mi for mixed-use corridors is exp^{(0.0053*(–10))} = 0.948. Similarly, the CMFs are 0.933 and 0.969 for commercial and residential land uses based on the coefficients for ACCDENS in table 47 and table 61, respectively. These CMFs are reasonably consistent with those in table 29, considering that the data for this project are a mixture of undivided and divided roads (mean PROPDIV ranges from 0.31 to 0.33 for residential and from 0.44 to 0.73 for other land uses) and that mean AADTs range from 15,000 to 32,000 vpd for residential and from 20,000 to 46,000 vpd for other land uses. In addition, both sets of numbers indicate a larger CMF (i.e., less reduction) for residential than for commercial land use, which is further evidence of consistency.

Table 29. Driveway density CMFs inferred from the Highway Safety Manual predictive models for multivehicle crashes on urban four-lane undivided and divided arterials.⁽²²⁾

COMPARISON WITH EFFECTS FOR OTHER VARIABLES IN THE CMF CLEARINGHOUSE

For this exercise, the implied effects of three variables—MEDOPDENS, SIGDENS, and UNSIGDENS—were compared with the information on effects derived from the CMF Clearinghouse, which were all from a single publication by Mauga and Kaseko.⁽²³⁾ Those effects were also from cross-sectional regression models and pertained to corridors classified as “urban.” It was decided to use the mixed-use models from this project for the comparison.

For MEDOPDENS, there was only one mixed-use model for which this variable was significant and had the intuitively correct sign—right-angle crashes for mixed land use (table 45). The implied CMF for increasing the median openings by 1 per mi is exp^0.1901 = 1.21. Increasing by 2 and 3 per mi gives CMFs of 1.46 and 1.77, respectively. The 3-star CMF from the CMF Clearinghouse is exp^0.0985 = 1.10 for increasing median openings by 1 per mi and 1.22 and 1.34 for increases of 2 and 3 median openings per mi. The standard error of the MEDOPDENS coefficient is 0.0844, which would indicate that the CMF from the Clearinghouse is within the range of approximately one standard error. On this basis, it can be concluded that the implied CMF from the model in table 45 is reasonably consistent with the CMF implied from the CMF Clearinghouse.

For SIGDENS, the CMFs implied from the mixed-use models with the lowest k are shown in table 30. Also shown are the 3-star CMFs from the CMF Clearinghouse for urban environments, which are assumed to be comparable to corridors in mixed land use. The discrepancy between the two sets of implied CMFs is not surprising in that the CMF would be highly variable, depending on the nonintersection crash frequency in the corridor and on the traffic volumes at the signalized intersections. This suggests that caution should be exercised in using models with this variable for corridors that differ greatly from those used to develop the models.

Table 30. Comparison of implied CMFs for SIGDENS.

For UNSIGDENS, there was only one comparable mixed-use model for which this variable was significant and had the intuitively correct sign—total crashes (table 35). The implied CMF for increasing the unsignalized access density by 1 per mi is exp^0.0471 = 1.048. Increasing by 2 and 3unsignalized intersections per mi gives CMFs of 1.099 and 1.151, respectively. The 3-star CMF from the CMF Clearinghouse is exp^0.0126 = 1.013 for increasing unsignalized intersections by 1 per mi. The CMFs from the Clearinghouse for increases of 2 and 3 unsignalized intersections per mi are 1.026 and 1.039, respectively. As was the case for SIGDENS, the discrepancy is not surprising in that the CMF would be highly variable, depending on the nonintersection crash frequency in the corridor and on the traffic volumes at the unsignalized intersections. This suggests that caution should be exercised in using models with this variable for corridors with characteristics that differ greatly from those corridors used to develop the models.