• Federal Highway Administration >
• Publications >
• Research Publications >
• 05042 >
• Appendh.Cfm
• Safety Effects of Differential Speed Limits

Safety Effects of Differential Speed Limits

PDF files can be viewed with the Acrobat® Reader®

APPENDIX H. EXAMPLE APPLICATION OF THE CRASH ESTIMATION MODEL TO THE AFTER DATA

Once the crash estimation model of the form shown in figure 8 has been developed using the before data, the next step is to apply the equations in figures 17-25 to determine whether the policy change has had a net reduction in the expected number of crashes. Appendix H illustrates this procedure, using a single site as an example.

In 1994, Virginia changed speed limits on rural interstate highways from a differential to a uniform limit. For this illustration, suppose that the investigator has only one site in the group, such that the before data are represented by the years 1991, 1992, and 1993. Because of data limitations, the only after data that are available are 1995, 1996, and 1997.

Thus, the investigator begins with the crash estimation model, shown as figure 54, which from just one site has been calibrated as shown below, using the Virginia data from 1991-1993. Table 37 shows the before and after crash and volume data for this site, which measures 7.16 mi in length.

In the derivation that follows, the ratio using the first year as base in the denominator as C¹_i,y, while the ratio with the third year as base in the denominator is C³_i,y. Thus, the equation in figure 12 may be rewritten for each case as:

Estimation of Expected Crash Frequency M₁, M₂,... M_y for the Before Period

The crash estimation model shown above is used with the data in table 37 to calculate the E(m_1,y), that is, the mean of the estimated crash frequency of site i for each year, as shown in figures 55-57. (Normally, i will range from 1 to the number of sites (e.g., for Virginia, with 266 sites, there would be equations with i=1, i=2, ... i=266. In this example with only one site, however, i will always be 1.)

Next, the ratios C_1,y which are the ratios of E(m_1,y) to E(m_1,1) for each before year y, are calculated using the form of figure 58 and applied in figures 59-61.

The next step is to calculate the expected crash counts m_i,y on this site for each before year with their variance VAR(m_i,y) using the equations in the figures below.

The application of these four expressions is shown in the equations in figures 66-71.

These results are summarized in table 38.

Prediction of M_y+1, M_y+2,... M_y+Z for the After Period

The next step is to use the crash estimation model from figure 56 to compute the mean of the expected would-have-been crash frequency for the after period years. That is, even though Virginia changed from a differential to a uniform limit in 1994, there question remains of "what would the mean of the expected crash frequency have been had Virginia not changed its speed limit policy." Computation of these E(m_1,y) values for the after period, as shown in figure 72 and 73, answers this question.

This process is repeated for the 1997 and 1999 years.

One then computes the ratio C_i,y using figure 60, as illustrated in figures 74 and 75.

Finally, figures 76 and 77 allow one to predict the would-have-been expected crash frequencies for the after years. Application of these methods is shown in figures 78-81.

These results are summarized in table 39.

^aK_i,y: The actual after crashes for year y.
^bE(m_1,y): The mean of the expected would-have-been after crashes for year y.
^cC_{1, y}: The changing ratio for the would-have-been after crashes.
^dm_1,y: The expected would-have-been after crashes for year y.
^eVAR(m_1,y): The variance of the expected would-have-been after crashes for year y.

Evaluation of Safety Effects of Changing the Speed Limit for This Particular Site

Evaluation of Safety Effects of Changing the Speed Limit for This Particular Site The effect of the treatment (that is, changing from a differential speed limit to a uniform speed limit) is determined by comparing the actual after crashes K_i,y with the predicted after crashes m_i,y for each year of the after period. The cumulative differences, shaded in table 40, are computed by applying the equations in figures 82-84.

An interpretation of table 40 is that, over the 4-year after period, the actual number of crashes at this site was 5.76 less than the predicted number of crashes that would have resulted had there been no change in the speed limit. Alternatively, the investigator could use the equation in figure 85 to indicate that the speed limit change decreased crashes by approximately 16 percent at the site, since the ratio of the actual crashes to "would-have-been crashes had no change occurred" is 0.84.

The investigator can graphically portray these cumulative differences as shown in figure 86.

Evaluation of Safety Effects of Changing the Speed Limit for All Sites

The example shown above has been applied to only one particular site, but realistically a researcher would apply these concepts to all sites (e.g., in Virginia, 266 sites). Thus, using similar computations from all 266 sites, the investigator compares (a) the actual after crashes for the after period 1995 through 1999 that resulted under the uniform speed limit imposed in 1994, and (b) the would-have-been crashes for the after period 1995 through 1999 that would have resulted had the speed limit not been changed in 1994. Two basic performance measures, the reduction in the expected number of crashes (δ) and the index of effectiveness (θ), are computed and tested for statistical significance.

The equations in figures 87-88 compute, respectively, the would-have-been crashes (those that would have occurred had no changes taken place) and the actual crashes that did occur. Since the equation in figure 89 shows that the actual number of crashes (λ) is larger than these wouldhave- been crashes (Π), the negative value of δ suggests that the speed limit change had an adverse impact on safety, and it would have been better not to make the change. This statement, however, needs to be tested for statistical significance.

To determine statistical significance, the starting point is the variance as calculated in figure 92. The term ΣVar(Π_i) is obtained by summing the values of VAR(m_1,y) for each site and each year. Thus, the italicized values shown in the sixth column of table 40 (e.g., 2.39, 2.48, 2.57, and 2.58) would be computed for each site and for each year, and with 266 sites and 4 years of data, the 1,064 values of VAR(m_1,y) are summed to equal 4,246.913. Because of statistical properties appropriate to the Poisson distribution, the summation of the λ_i values are equivalent to the variance of these values, which is 15,377 as tabulated in figure 89 above. These two items are used in figure 90 below to obtain Var(δ), which intuitively may be described as the variation associated with the difference between would-have-been crashes and actual crashes.

The standard deviation is thus the square root of this variance in figure 91, such that:

Empirical confidence bounds are thus δ ± 2σ(δ) or -2011 ± 2(140).

Computation of the index of effectiveness is accomplished via the equation below as:

The variance of θ is given below as:

The empirical confidence bounds are shown below:

Thus, the change from DSL to USL in Virginia increased the number of crashes by about 15 percent, with the full response being that this 15 percent is not a perfect estimator but instead should be given as a range, such that the increase was between approximately 12.9 percent and 17.2 percent according to this application of the empirical Bayes method. Assuming all the other factors remained constant except for the change of speed limit from differential to uniform, therefore, the value of θ being greater than 1.0 means that the change had an adverse impact on safety, as reflected in the number of crashes on rural interstates in Virginia.

Previous | Table of Contents | Next