6. FINDINGS AND CONCLUSIONS

The following findings and conclusions were reached as a result of the statistical analysis of relationships between traffic accidents and geometrics of interchange ramps and speed-change lanes conducted in this research:

  1. Traditional multiple linear regression is generally not an appropriate statistical approach to modeling accident relationships because accidents are discrete, non- negative events that often do not follow a normal distribution.

  2. The Poisson and negative binomial distributions appear to be better suited to the modeling of accident relationships than the normal distribution. In all cases, the form of the statistical distribution selected for any particular modeling effort should be chosen based on a review of the data to be modeled.

  3. The choice between the Poisson and the negative binomial distributions should be based on the overdispersion observed in the accident data. Overdispersion results when the variance of the accident data exceeds the mean of the Poisson distribution. Extra variation or overdispersion in a Poisson model causes underestimation of the variance of the model coefficients. This, in turn, results in overstating the significance of the coefficients; in other words, some coefficients may be found to be statistically significant when, in fact, they are not. In the modeling of accidents for interchange ramps and speed-change lanes with Poisson regression, overdispersion was commonly observed and, therefore, the negative binomial distribution was preferred.

  4. Regression models to determine relationships between accidents and the geometric design and traffic volume characteristics of ramps, based on the negative binomial distribution, explained between 10 percent and 42 percent of the variability in the accident data.

  5. Accident frequencies on interchange ramps and speed-change lanes are so low at most locations that they are very difficult to model. Between 50 percent and 80 percent of the ramps studied experienced no accidents or only one accident in the 3-year study period. Only a very few ramps experienced a substantial number of accidents during the 3-year period.

  6. Negative binomial regression models developed to predict total accidents generally performed slightly better than did models to predict fatal and injury accidents.

  7. Negative binomial regression models developed to predict the combined accident frequency for an entire ramp, together with its adjacent speed-change lane, generally fit the available accident data better than separate models for ramps and speed-change lanes.

  8. The independent variables, whose effects on accident frequency were most often found to be statistically significant, were:

    The ramp AADT was the strongest predictor of accident frequency; the other variables, while they were generally statistically significant, had much less predictive ability.

  9. A number of other geometric design variables for ramps and speed-change lanes were considered in modeling. These included:

    However, none of these geometric design variables was found to have a statistically significant relationship to accident frequency, except in limited situations in models that were not ultimately recommended for use.

  10. The best models obtained for predicting accident frequencies for ramps and speed- change lanes are the model presented in table 38 for total accidents and the model presented in table 39 for fatal and injury accidents. These models are also presented in equations (15) and (16), respectively.

  11. A review of hard-copy police accident reports found that rear-end accidents on urban off-ramps of four configurations (diamond, parclo loop, free-flow loop, and outer connection ramps) were generally related to the operation of the cross-road ramp terminal, rather than to the geometric design of the ramp itself. Only 5 percent of the rear-end accidents reviewed were not related to the operation of the cross-road ramp terminals.