A Zero-expansion Counting Model For Research On Factors Affecting Road Traffic Injuries

Objective With the growth in the living standard of residents, mechanization is increasing rapid, which results in amounts of questions about the road traffic injury (RTI) and the risk assessment of going out.becomes increasingly serious. Dead number involve RTI in China is in the lead since ten years ago around the world, although the household cars number is not the most. Based on a incomplete statistical data, the cars number in China takes up3%of the whole world, while the dead number takes up16%. This paper collects the RTI data occurred from April1st to October4th in2010, with the zero-inflated count model being used which can depict the regular of RTI as well as its distribution feature and economic losses. It can give forward feasible opinions for the prevention and control of localized RTI and methodology basis on the application of zero-inflated model in injury epidemiology.Method An epidata3.1database is designed with the data comes from the traffic accident cases occurred in Hejin from April1st to October4th. What's more, the SAS9.13software is used to construct the Poisson regression, Negative binomial regression, Zero-inflated Poisson regression and Zero-inflated Negative binomial regression model with dead number and injury number in accidents.Results and discussion1Most traffic accidents in county and city have none injury or dead, while little accidents occurred.In the analysis of509RTIS, the number of the RTI reaches a peak in May and the incidence of RTI at this place is related to the season. The median money lost in one RTI is1500yuan, while89.78%RTIs less than5000yuan and3.34%more than10000yuan. Although there are a large number of RTIs during April and September, dead number is very small, which means there are a lot of zero cases of RTIs that account for49.31%, and the drivers in accidence are most man about30years old.2Count model that affect the injured number of RTIsThe injured number in one RTI is integer, y=0,1,2,3,..., which can't meat the "LINE", so a count model should be used. The Poisson regression and negative binomial regression get similar output in fitting the injured number with the likelihood ratio estimate being0.441which means the negative binomial regression fits better than the Poisson regression. However, the predicted zero percentage in two models are44.81%and45.15respectively, which indicate that the real zero percentage have exceeded the predictive ability of the two models. The score test indicate that the injured number of RTIs exists exceed zeros. When fitting the zero inflated model, the Vuong test indicate that ZIP model is better than the Poisson regression model and the ZINB model is better than the negative binomial model. The comparison output of the four models indicate that ZINB can solve the real question best, while ZIP is in the second place and the negative binomial regression and Poisson regression are worst.With the injured number as the response variable in the ZINB model, the results of the negative binomial regression part shows that the automotive vehicle type and the damage situation are statistical significance. The injured number of non-automotive vehicle is1.47times of that of automotive vehicle. The injured number of serious damage is more than that of slight damage, which is1.36times of the later one. The drivers'age of aspect A is the only variable that included in the logit part, which shows that when the drivers'age is more than age40, the probability of injured number occurred becomes larger, which is1.81times of those under age40.3Count model that affect the dead number of RTIsBecause of the abundant zero of dead number in RTIs, predicted results of traditional count regression have large difference with the real zero percentage. ZIP and ZINB models are used to this data.The ZIP output shows that the responsible party,the type of aspect A and the damage situation are statistical significance. Dead number of non-automotive vehicle is more than that of automotive vehicle and dead number of damage serious is more than that of slight damage. Only the intercept of logit part is statistical significant. The predicted zero percentage of ZIP is91.51%which is better than the Poisson regression. The Vuong test shows that the ZIP is better than the Poisson regression.With the died number as the response variable in the ZINB model, the result shows that the responsible party,the motorcycle type of party A, the drivers'age of party A and the damage situation have statistical significance from the negative binomial regression part. The died number of which the accident involved party A and party B is1.23more times than the accident involved only one party. The died number of nonmotor vehicle is more than the motor vehicle,which is2.66times of motor vehicle'The drivers'age is also the influence factor of died number, the died number of the accident in which the drivers'age is more than age40is8.79times of which the drivers'age is under age40. The died number of serious damage is more than the slight damage, which is1.50times of slight damage. he drivers'age of party A is only included in the logit part, which shows that when the drivers'age is more than age40, the probability of died number occurred becomes bigger, which is4.27times of those under age40.Overall, ZIP and ZINB are a kind of meaningful methods for traffic hurt study. They can make the best of information in the count data as well as cope with the exceed zero of the injured and dead number in RTIs.