|
Abstract
|
Meyer and Finney constructed a comprehensive logistic regression model that included seating position and gender as potential factors, but they reported that neither factor was statistically significant. However, when Table 7A is restricted to front- seat passengers in frontal impacts, the estimated effect of airbags is a 58% increase in fatality risk (116% for male passengers). For drivers, the estimated effect is a 16% decline in fatality risk. The point is that the effectiveness estimates continue to change, sometimes dramatically, as more factors are considered--and there is no good rule for when to stop. Theoretically, one stops adding factors when they no longer have a statistically significant effect on the results (the approach taken by Meyer and Finney). In reality, one must keep in mind that even important factors may turn up nonsignificant if the samples are small. NASS/CDS is a well-designed, but relatively small, sample of crashes.
Missing values (especially for delta- V) make the sample even smaller. The crashes excluded due to missing values seem to have been among the more severe crashes sampled, precisely those severity categories responsible for Meyer and Finney’s negative effects. If it were possible to classify these excluded crashes by delta-V, they could easily change the results. Without considering these excluded crashes, one cannot make reliable conclusions.
As the sample shrinks, the national weighting factors become less reliable, even when adjusted for nonresponse. For example, if raw counts are used instead of the nationally weighted counts in the corrected data of Meyer and Finney, the estimated effect of airbags for front-seat passengers in frontal impacts is a 3% decline in fatal- ity risk (36 deaths compared with 37 expected). The weights change this effect to a 30% increase (1,678 deaths compared with 1,292 expected).
One might suspect that the variability of results is due to the low number of fatalities in NASS. However, the same problems occur when Meyer and Finney’s methods are used to look at the risk of serious injury. Serious injury is defined as any injury with an Abbreviated Injury Scale greater than or equal to 3, according to the Association for the Advancement of Automotive Medicine. If the methodology of Table 5 is used to estimate the effect of airbags on serious injury risk, the raw counts again contradict the weights. Using the raw counts, airbags are associated with an estimated 9% decline in serious injury risk. The weights change this effect to a 12% increase.
There is no logical reason for the sampling weights--which are based on geography, population size, and crash severity (also a factor in the above analyses)--to change the effectiveness estimates so dramatically. Thus, it seems clear that this methodology of computing fatality or serious injury rates per crash cannot be applied reliably to a NASS database severely depleted by missing data. |