كاربرد مدل هاي رگرسيون پرصفر در مدل سازي تصادفات بزرگراههاي درون شهري

Application of Zero-Inflated Regression Models in Modeling Accidents on Urban Highways

چكيده- در اين نوشتار به پژوهشي براي شناسايي عوامل موثر بر فراواني و شدت تصادفات بزرگراههاي درون شهري انجام شده پرداخته مي شود و به عنوان مطالعه موردي از داده هاي تصادفات بزرگراههاي شهري مشهد استفاده مي شود. مدل هاي آماري كه در اين پژوهش استفاده شده، مدل هاي رگرسيون پواسن، دوجمله اي منفي، پواسن پرصفر و دوجمله اي منفي پرصفر است، از متغيرهاي مربوط به جريان ترافيك و طرح هندسي راه به عنوان متغيرهاي ناوابسته مدل ها استفاده شده است. كار ويژه اي كه در اين پژوهش انجام شده، جداسازي حجم كل ترافيك به حجم خودروهاي سواري، خودروهاي سنگين و خودروهاي غير سواري سبك است تانقش حجم ترافيك در رخداد تصادفات به گونه اي ويژه و موشكافانه بررسي و روشن شود دقيقاً كدام بخش از ترافيك نقش موثر ياموثرتري در رخداد تصادفات دارد.در اين پژوهش كارايي مدل هاي رگرسيون پرصفر در برابر مدل هاي رگرسيون ساده پواسن و دوجمله اي منفي در مدل سازي تصادفات بزرگراه هاي درون شهري بررسي مي شود. نتايج پژوهش نشان مي دهد شمار راه هاي دسترسي و قوس هاي افقي دربخش هاي بزرگراه و حجم خوروهاي غيرسواري سبك، نقش افزاينده اي در احتمال رخداد تصادفات دارد، همچنين مدل رگرسيوندوجمله اي منفي پرصفر بهترين و برازنده ترين مدل، هم براي مدل سازي تصادفات با زيان مالي و هم براي تصادفات با زيان جاني است و بر اين اساس مي توان كفايت مدل هاي رگرسيون پرصفر را در مدل سازي تصادفات بزرگراه هاي درون شهري مشهد تاييد كرد.

Abstract: We handle in this paper, the research that have been performed to recognize the factors that affect crash frequency and severity in urban highways and use crash data of Mashhad urban highways as a case study. Statistical models that have been used in this research include Poisson, Negative binomial, Zero-inflated poisson and Zero-inflated negative binomial regression models. Traffic flow related variables and road geometric related variables have been used as independent variables of models. We are interested in this study, to inspect the efficiency of Zero-inflated models against simple Poisson and Negative binomial regression models in modeling accidents on urban highways. Special task that have been done in this research, is separation of total traffic volume into passenger cars, heavy vehicles and light non-passenger vehicles volume. Through this special, Researcher intend to have an especial look at the role of traffic volume in accident occurrence to see precisely, which part of traffic have effective role or more effective role in crash occurrence. Accident data are two-level data, the first level is road segments i.e. highway is divided into several segments. The segmentation is based on total traffic volume i.e. each segment has a constant volume. The second level is daily hours; peak hour traffic considered as the first sublevel, day non-peak hour traffic the second and night non-peak hour traffic as the third sublevel. SAS 9.1 software has been used to fulfill statistical computations. It turns up, after statistical analyses, which factors affect crash occurrence and which do not have much effect. Comparisons between obtained results and other researchers’ results have been made then. The main object of researcher is to assess the efficiency of Zero-inflated models against Poisson and Negative binomial regression models in modeling urban highways crashes. This aim is followed by, with evaluating goodness of fit and making comparison between models. The Results of study show that the presence and number of access roads and horizontal curves on highway segments increase the likelihood of accidents, both no injury and more severe. Also increment of speed and number of lanes increase the likelihood of no injury accidents, but not more severity ones. The conclusions also demonstrate that the volume of passenger cars and light nonpassenger car vehicles increase the likelihood of no injury accidents, but heavy vehicles volume does not have much effect on occurrence of no injury accidents, also light vehicles increase the likelihood of more severe accidents, but passenger cars and heavy vehicles volume does not have much effect on occurrence of severe accidents. Finally, the results of research indicate that Zero-inflated negative binomial regression model is best fitting the modeling of accidents, whether no injury or more severe and consequently, the efficiency of zero-inflated models in modeling accidents on urban highways is approved.