Statistical analyses of calendar year 1989-95 crash records from five States indicate that the long-term effectiveness of CHMSL is higher in 2-vehicle crashes than in 3-vehicle crashes; during the daytime than at night; and at non-signalized places than at locations equipped with traffic signals. CHMSL effectiveness in the struck vehicle in a front-to-rear collision is about the same whether the driver of the striking vehicle is young or old, male or female. Effectiveness is about equal in injury and noninjury crashes.
Analyses of 1986-95 FARS data suggest that CHMSL have had little or no life-saving benefit.
3.1 Analyses of five State files: method
The overall effectiveness of passenger car CHMSL had declined to 4 percent by 1989, but it stayed close to that "long-term" level throughout 1989-95, as was shown in Chapter 2. In other words, when all types of roadways, drivers, weather conditions, etc. are taken into consideration, a car equipped with CHMSL is 4 percent less likely to have a rear-impact crash than a car without CHMSL. We will now pool the CY 1989-95 data and compute the long-term CHMSL effectiveness separately for daytime vs. nighttime crashes, wet vs. dry roads, etc.
In Chapter 2, where effectiveness had to be separately calculated and meaningfully compared from one calendar year to another, a premium was placed on obtaining as many crash records from as many States as possible (eight States). Here, with the data pooled across seven calendar years, sample size is less of an issue, and it is worthwhile to take several steps that improve data quality or reduce the potential for biases in the analysis, even if those steps reduce available sample size. The analyses in this chapter will be:
Limited to the subset of five States that encode VINs on their crash files (Florida, Maryland, Missouri, Pennsylvania, Utah), and limited to records of passenger cars with good VINs, and whose VIN-defined model year agrees with the police-reported model year.
Limited to cars of model year 1985 (the last year before all cars were CHMSL-equipped) and 1986 (the first year that CHMSL was standard equipment)
Limited to crashes involving two or more vehicles, and where the first harmful event is a collision between vehicles.
The advantage of using only the records of passenger cars with good VINs is that the findings are not obscured by the inclusion of light trucks inadvertently coded as "cars," or vehicles coded as MY 1986 or later (i.e., with CHMSL) when they were actually pre-1986, or vice-versa. (It was shown in Section 2.1, however, that only a small percentage of vehicles are thus misclassified.) Comparing only MY 1985 and 1986 cars simplifies the analysis and avoids biases that may enter when longer series of model years are included. In Chapter 2, the "control group" in the analysis (vehicles with other than rear impact damage) included cars involved in single-vehicle crashes, or in multivehicle crashes whose first harmful event was not necessarily a collision between vehicles. Excluding those cars perhaps reduces bias because it makes the control group (collision-involved vehicles with side or front damage) more directly comparable to the CHMSL-affected group (collision-involved vehicles with rear damage). The codes used to limit the data to crashes in which the first harmful event was a collision between vehicles are:
|Florida||exclude event1 = 4,11,12,14,15,34|
|Maryland||include event1 = 1|
|Missouri||include event1 = 7|
|Pennsylvania||include event1 = 13-19|
|Utah||include event1 = 2|
The method for comparing CHMSL effectiveness in specific situations will be illustrated by an example: daylight vs. nighttime crashes in Pennsylvania. The starting point for the analysis is the basic contingency table for MY 1985 vs. 1986, pooling all the data from CY 1989-95:
|MODEL YEAR||REAR IMPACT?|
The log odds of a rear impact is log(15607/45775) = -1.0760 in MY 85 and -1.0928 in MY 86. The observed reduction in the log odds of a rear impact is
However, this observed D_LOGODDS somewhat underestimates the overall effectiveness of CHMSL in 1989-95 because it has not been adjusted for the "vehicle age effect" described in Section 2.3. Our best effectiveness estimate, as shown in the last rows of Table 2-6, is that CHMSL reduced rear impacts by 3.45 percent in Pennsylvania during 1989-95. That percentage reduction corresponds to a log-odds reduction:
The "correction factor"
will be added to the D_LOGODDS OBS in any contingency table based on subsets of the Pennsylvania data to "bring the effectiveness up" to the level estimated in Table 2-6. For example, in Pennsylvania crashes during daylight:
|MODEL YEAR||REAR IMPACT?|
D_LOGODDS OBS = .0206 is somewhat higher than it was in all crashes (.0168). After adding the correction factor .0183 and converting the reduction in log odds to a percentage reduction in crashes the effectiveness estimate for CHMSL in daylight becomes
During times of reduced light (nighttime, dawn or dusk):
|MODEL YEAR||REAR IMPACT?|
D_LOGODDS OBS = .0053 is lower than it was in daytime crashes. Even after adding the correction factor, the effectiveness estimate for CHMSL in at night is only
Correction factors are similarly computed in the other State files, and they are:
|Effectiveness estimate % (Table 2-6)||4.96||3.35||4.12||3.45||6.65|
|Equivalent log odds||.0508||.0341||.0421||.0351||.0688|
|Observed 85-86 log odds||.0088||.0301||.0143||.0168||.0641|
|Correction factor: ADD||.0420||.0040||.0278||.0183||.0047|
One caveat for this method is that it assumes the same correction factor (adjustment for the vehicle age effect) in all the various subsets of crashes within a State. It is conceivable that the vehicle age effect could vary among subsets. For example, the effect in Pennsylvania nighttime crashes could be larger (or smaller) than in daytime crashes. That would be a "second-order" difference. We are not asking here if the proportion of daytime to nighttime crashes changes as cars get older - nor if the proportion of crash involvements that are rear impacts is different in the daytime and the nighttime. We are only asking if the trend towards proportionately fewer rear impacts as cars get older is stronger (or weaker) in daytime crashes than in nighttime crashes.
In general, the vehicle age effect does not vary between subsets. The effect was examined by averaging D_LOGODDS for MY 1985 vs. 1984 and for MY 1987 vs. 1986 (i.e., in both cases, two adjacent model years where the CHMSL status did not change). For all types of crashes, the population-weighted average of this quantity for the five States was -.040: approximately a 4 percent "age effect" per year in the pooled 1989-95 data. Within each of the subsets considered in Section 3.2 - daytime, nighttime, injury crashes, noninjury crashes, etc. - this quantity was always between -.049 and -.031, except in rural crashes, where it was -.020. These small variations in the observed "vehicle age effect" are well within the "noise" range of the data.
The strategy in the next section is to compute CHMSL effectiveness in the various subsets (e.g., daytime vs. nighttime crashes): separately in each State, and population-weighted averages for the five States. In general, it will only be concluded that effectiveness is higher in one subset than another if it is higher in all five States, or in every State except one of the smaller ones (Utah or Maryland) and if the 5-State average is substantially higher in one subset than the other(s). Otherwise, it is appropriate to conclude that effectiveness is about the same in the two (or more) subsets.