3.3 The effect of CHMSL in fatal crashes

NHTSA's previous evaluation did not find any effect for CHMSL on fatalities in CY 1986-87 data [19], pp. 26-27. The report suspended any definitive conclusion, however, because the quantity of data available for that analysis was small. Relatively few people die in rear-impact crashes.

With CY 1986-95 data now available, a new analysis can be performed with about 5 times as many crash records. The Fatality Analysis and Reporting System (FARS), a census of the nation's fatal crashes, is the data source [10]. (Of course, the eight State files used in the other analyses of this report contain only a fraction of the fatality cases in FARS.) The analysis procedure with FARS data is essentially identical to the process for any of the individual State files in Chapter 2.

The first step is to identify all records of passenger cars (body_type = 1-9,12 in 1991-95 FARS and body_type = 1-9,13 in 1986-90 FARS) involved in a multivehicle fatal crash in which this car's "most harmful event" was a collision with a motor vehicle in transport (m_harm = 12). Parked cars (veh_man = 7) are excluded. Cadillacs (make = 19) are also excluded, because they got CHMSL as standard equipment in MY 1985 rather than 1986. Rear impacts need to be distinguished from other impacts. As in Chapter 2, the definition of "rear impact" is quite inclusive: any vehicle whose initial damage (impact1) and/or principal damage (impact2) is on the rear, the rear-corner, or the side towards the rear (impact1 = 4-8 or impact2 = 4-8). Intuitively, impact1 is the more important variable, but since it is less completely reported than impact2 on FARS, the latter variable has been added to minimize missing data. Cases with unknown principal damage location (impact2 = 99) are excluded. The 1986-95 FARS files comprise 17,735 records of cars 0-15 years old involved in fatal rear-impact crashes: substantial data, but not many in comparison to the 3,000,000 nonfatal rear-impact cases available for the analyses of Chapter 2.

The next step is to calibrate the "vehicle age effect." Section 2.3 describes this effect in nonfatal crash data and explains how it is calibrated. A similar effect exists in fatal crashes: older cars have fewer rear-impact crashes relative to frontal and side impacts with other vehicles. The 1986-95 FARS data are tabulated by calendar year and model year (cars 0-15 years old). For example, Table 3-1 shows the percentage of rear impacts for MY 1977-92 cars in CY 1992. Three phenomena may be detected when these numbers are compared to Table 2-1 (nonfatal Florida data): (1) The proportion of fatal involvements that are rear impacts is only about half as large as the proportion of nonfatal crashes that are rear impacts, even after excluding fatal single-vehicle crashes. (2) Because there are fewer cases in each cell, the proportion of fatalities that are rear impacts varies with little pattern from one MY to the next; based on one CY of data, it is hard to discern the vehicle age trend. (3) Needless to say, it is also impossible to detect the effect of CHMSL in MY 1986.

However, when the data are pooled across the ten calendar years and graphed by vehicle age, as in Figure 3-1, a reasonably clear pattern emerges. The vehicle age effect - i.e., the rate of decline in the percentage of crashes that are rear impacts - is strongest when cars are new (unlike Florida nonfatal crashes, where the effect is strongest for 5-8 year old cars; see Figure 2-1). A quadratic regression is used to calibrate D_LOGODDS, the rate of change in the percentage of crashes that are rear impacts, by vehicle age. The regression equation is

It somewhat resembles the equation for the Indiana nonfatal data (see Table 2-3).

As in Sections 2.4 and 2.5, the preceding equation is used to adjust the log odds of a rear impact for j-year-old cars to what it would have been if the cars had been brand new instead of j years old. The adjusted log odds of a rear impact are computed in each of the model years 1982-89 in each of the calendar years. The "preliminary effectiveness estimate" CHMSLAVG, a weighted average of the difference in adjusted log odds for MY 82-85 (without CHMSL) and MY 86-89 (with CHMSL) is computed in each calendar year. So is DELAVG, the average annual reduction in the adjusted log odds of a rear impact from MY 83 to 85 and from MY 86 to 88 - when cars' CHMSL status remained the same. CHMSLAVG measures the effectiveness of CHMSL (positive numbers indicate a benefit, a reduction of rear impact crashes). DELAVG measures whether the model is producing spurious benefits. The values of these statistics in the ten individual calendar years, and their averages over the ten years, are the following:

CHMSLAVG DELAVG

Figure 3-2 graphs CHMSLAVG () and DELAVG (O) by calendar year. It makes quite a contrast with Figure 2-7, the corresponding results for (almost exclusively) nonfatal crashes from the eight State files. In Figure 3-2, the 's and the O's are scattered all over the diagram, without any recognizable pattern (except, perhaps, just a hint of positive CHMSLAVG values in 1987-88). In Figure 2-7, the 's were clearly positive, with a peak in 1987 and flat after 1989, while the O's were always close to zero.

In the fatal crashes, the observed effectiveness of CHMSL is positive in four years and negative in six. In the nonfatal crashes, effectiveness was positive in all ten years.

In the fatal crashes, CHMSLAVG is more positive than DELAVG in four years and less positive in six. In other words, there is no evidence that the effect from MY 85 to 86 (when CHMSL became standard equipment) is any different than the effect from MY 84 to 85, or 86 to 87, etc. In the nonfatal crashes, CHMSLAVG was always more positive than DELAVG; in fact, DELAVG was always much closer to zero than to CHMSLAVG.

The facts that CHMSLAVG is positive in four of ten years, and that CHMSLAVG is more positive than DELAVG in four of ten years are nonparametric tests suggesting that the effect of CHMSL is nonsignificant. Additional insight may be gained by statistical testing and confidence intervals based on the variance of the ten individual CHMSLAVG values. The first question is whether the slightly more positive results for CY 1987 and 1988, as seen in Figure 3-2, are significantly more positive than the CHMSLAVG values in the other calendar years. The GLM procedure of SAS [32] is used to run an analysis of variance on CHMSLAVG across two CY groups: "1987-88" and "other." The difference between the two groups is not statistically

significant (F = 1.49; df = 1,8; p > .05), and we may assume that CHMSLAVG is essentially invariant throughout CY 1986-95, except for "noise."

Thus, the ten individual values of CHMSLAVG may be treated as repeated measurements and used to compute a point estimate and confidence bounds. The point estimate of CHMSL effectiveness is the arithmetic average of the ten values: a 0.29 percent reduction of rear-impact crashes. This point estimate is, of course, not significantly different from zero: the t value for the ten individual estimates is 0.12 (p > .05). One standard deviation of the point estimate is 2.63 percent. With 9 degrees of freedom, the 95 percent confidence bounds for effectiveness are 0.29 ±2.262 standard deviations:

The most appropriate conclusion continues to be that CHMSL had little or no effect on fatal crashes.

Given that CHMSL are about equally effective in property-damage-only and nonfatal-injury crashes, it might seem surprising, at first glance, that effectiveness is nonsignificant in fatal crashes. Several characteristics of fatal rear impacts, however, substantially reduce the potential effect of CHMSL, as evidenced by a comparison of FARS and Florida (nonfatal crash) data:

The struck car was stopped, slowing or turning (and thereby presumably activating the CHMSL) in only 37 percent of the fatal crashes, as opposed to 73 percent of the nonfatal crashes.

The striking vehicle was traveling 55 mph or faster in 51 percent of the fatal crashes, as opposed to about 10 percent of the nonfatal crashes. A modest reduction in the collision severity with CHMSL might not prevent the fatality.

The striking vehicle was a heavy truck or a motorcycle in 25 percent of the fatal crashes, but only 5 percent of the nonfatal crashes. When there is a severe weight mismatch (the struck vehicle being a passenger car), there is a high probability of a fatality in the lighter vehicle, and it is unlikely to be prevented by a modest reduction in the collision severity.

The driver of the striking vehicle had been drinking, and might have become inattentive to the CHMSL, in 23 percent of the fatal crashes but only 10 percent of the nonfatal crashes.

44 percent of the fatal crashes happened at night, when CHMSL is least effective, but only 25 percent of the nonfatal crashes.

While these factors, by themselves, might not necessarily reduce CHMSL effectiveness all the way to zero, they could easily lower it into the "noise" range, undetectable given the quantity of data available for the FARS analyses.