Driver injury probabilities in real-world frontal crashes were statistically modeled to estimate the relative roles of five variables of topical interest. One variable pertained to behavior (belt-wearing rate), one pertained to crash circumstances (speed change), and three pertained to occupant demographics (sex, age, and body mass index). The attendant analysis was composed of two parts: (1) baseline statistical modeling to help recover the past, and (2) sensitivity analyses to help consider the future.
In Part 1, risk functions were generated from statistical analysis of real-world data pertaining to 1998-2014 model-year light passenger cars/trucks in 11-1 o’clock, full-engagement frontal crashes documented in the National Automotive Sampling System (NASS, 1997-2014). The selected data yielded a weighted estimate of 1,269,178 crash-involved drivers. Those data were parsed for four subpopulations: two levels of belt use (properly-belted vs. unbelted) and two levels of driver injury (moderate-to-maximum, MAIS2+ vs. serious-to-maximum, MAIS3+). For each subpopulation, a baseline statistical model was generated via logistic regression, cast as a function of the studied variables. Each risk function was assessed for statistical significance (p-value for each term) and statistical associativity (Goodman-Kruskal Gamma). The four resulting risk functions had some statistical insignificance and fair fidelity, with Gammas ranging from 0.54 to 0.73. However, the risk functions demonstrated excellent fidelity for estimating aggregate injury rates (function-estimated vs. directly-estimated). They were accordingly applied in Part 2.
In Part 2, sensitivity studies were conducted by (a) perturbing the studied variables in the NASS dataset to generate thousands of hypothetical NASS files, (b) applying the risk functions to estimate attendant net injury rates, and (c) relating the net injury rates to the variations. Specifically, net injury rates and mean statistics were generated for 15,552 hypothetical NASS datasets involving both belted and unbelted drivers. Those data were then normalized by the means of the baseline NASS file. Finally, power functions were developed to relate the resulting dimensionless net injury-rate data to the five dimensionless predictor variables. Those functions demonstrated excellent fidelity (R2≥0.95), and their exponents helped quantify the relative role of the five studied variables. Belt-wearing rate and speed change were determined to be the most influential, followed by age, body mass index, and sex. These findings might help guide engineers and regulators.