--- category: literaturenote citekey: zouevaluatingdoublepoissongeneralized2013 title: Evaluating the double Poisson generalized linear model authors: "Zou, Yaotian; Geedipally, Srinivas Reddy; Lord, Dominique" year: 2013 date: 2013-00-00 2013 doi: 10.1016/j.aap.2013.07.017 publication: "Accident Analysis & Prevention" url: "https://doi.org/10.1016/j.aap.2013.07.017" zotero_key: VCZ27CJK zotero_storage: JKNS82TJ collections: transpordi_planeerimine folder: 001_artiklid firstAuthor: "Zou, Yaotian" status: converted --- Contents lists available at [ScienceDirect](http://www.sciencedirect.com/science/journal/00014575) # Accident Analysis and Prevention journal homepage: [www.elsevier.com/locate/aap](http://www.elsevier.com/locate/aap) ![](_page_0_Picture_5.jpeg) # Evaluating the double Poisson generalized linear model Yaotian Zoua,1, Srinivas Reddy Geedipally b,∗, Dominique Lordc,2 - a School of Civil Engineering, Purdue University, 550 Stadium Mall Drive, West Lafayette, IN 47907-2051, United States - b Texas Transportation Institute, Texas A&M University System, 110N, Davis Dr. 101, Arlington, TX 76013, United States - c Zachry Department of Civil Engineering, Texas A&M University, 3136 TAMU, College Station, TX 77843-3136, United States Article history: Received 29 August 2012 Received in revised form 9 July 2013 Accepted 12 July 2013 a r t i c l e i n f o Keywords: Double Poisson Normalizing constant Conway–Maxwell–Poisson Gamma model Generalized linear model # a b s t r a c t The objectives of this study are to: (1) examine the applicability of the double Poisson (DP) generalized linear model (GLM) for analyzing motor vehicle crash data characterized by over- and under-dispersion and (2) compare the performance of the DP GLM with the Conway–Maxwell–Poisson (COM-Poisson) GLM in terms of goodness-of-fit and theoretical soundness. The DP distribution has seldom been investigated and applied since its first introduction two decades ago. The hurdle for applying the DP is related to its normalizing constant (or multiplicative constant) which is not available in closed form. This study proposed a new method to approximate the normalizing constant of the DP with high accuracy and reliability. The DP GLM and COM-Poisson GLM were developed using two observed over-dispersed datasets and one observed under-dispersed dataset. The modeling results indicate that the DP GLM with its normalizing constant approximated by the new method can handle crash data characterized by over- and under-dispersion. Its performance is comparable to the COM-Poisson GLM in terms of goodness-of-fit (GOF), although COM-Poisson GLM provides a slightly better fit. For the over-dispersed data, the DP GLM performs similar to the NB GLM. Considering the fact that the DP GLM can be easily estimated with inexpensive computation and that it is simpler to interpret coefficients, it offers a flexible and efficient alternative for researchers to model count data. © 2013 Elsevier Ltd. All rights reserved. #### **1. Introduction** Due to the limited access to driving information, it is very difficult to identify factors (e.g., brake reaction time, inattention, etc.) directly influencing the number and severity of motor vehicle crashes in traffic safety analysis. Thus, instead of focusing on individual driver's information,most researchers approachcrashcausal or correlation analyses from a long term statistical view. In this regards, researchers try to associate the factors of interest with the frequency of crashes that occurs in a given space (roadway or intersection) and time period [\(Lord](#page-7-0) [and](#page-7-0) [Mannering,](#page-7-0) [2010\).](#page-7-0) Therefore, statistical models have been the analysis tool of choice for analyzing the relationship between motor vehicle crashes and factors such as road section geometric design, traffic flow, weather, etc. The traditional poisson distribution has been commonly used to model motor vehicle crashes. Despite its simple probabilistic structure, the traditional Poisson distribution has a strict assumption in thatits singleparameterdoesnot allow for theflexibility of variance varying independently of the mean, which is often violated by its application to the over-dispersed (i.e., the sample variance is larger than the sample mean) and under-dispersed (i.e., the sample variance is smaller than the sample mean) crash data. Under-dispersion occurs on rare occasions and this often happens when the sample mean value is low ([Lord](#page-7-0) [and](#page-7-0) [Mannering,](#page-7-0) [2010\).](#page-7-0) Over-dispersed and under-dispersed data can lead to inconsistent standard errors of parameter estimates when the Poisson model is used [\(Cameron](#page-7-0) [and](#page-7-0) [Trivedi,](#page-7-0) [1998;](#page-7-0) [Park](#page-7-0) [and](#page-7-0) [Lord,](#page-7-0) [2007\).](#page-7-0) Over the last three decades, the negative binomial distribution or model (NB or Poisson-gamma) has been quite popular to handle over-dispersed datasets. Although the mean-variance relationship of the NB is simple to manipulate so as to capture over-dispersion [\(Hauer,](#page-7-0) [1997\),](#page-7-0) it has been found to have difficulties in handling the data characterized by under-dispersion [\(Lord](#page-7-0) et [al.,](#page-7-0) [2008a\).](#page-7-0) Though crash data do not exhibit underdispersion very often, it is observed more frequently in other fields of research (for example, see [Guikema](#page-7-0) [and](#page-7-0) [Coffelt,](#page-7-0) [2008;](#page-7-0) [Sellers](#page-7-0) [and](#page-7-0) [Shmueli,](#page-7-0) [2010;](#page-7-0) [Borle](#page-7-0) et [al.,](#page-7-0) [2006\).](#page-7-0) In order to manage data characterized by over-dispersion and under-dispersion, researchers have proposed alternative models such as the weighed Poisson distribution [\(Castillo](#page-7-0) [and](#page-7-0) [Perez-Casany,](#page-7-0) [2005\),](#page-7-0) the generalized Poisson distribution [\(Consul,](#page-7-0) [1989\)](#page-7-0) and the gamma count distribution [\(Winkelmann,](#page-8-0) [2008\).](#page-8-0) However, these models suffered from important theoretical or logical soundness. For instance, the ∗ Corresponding author. Tel.: +1 817 462 0519. E-mail addresses: [zou21@purdue.edu](mailto:zou21@purdue.edu) (Y. Zou), [srinivas-g@ttimail.tamu.edu](mailto:srinivas-g@ttimail.tamu.edu) (S.R. Geedipally), [d-lord@tamu.edu](mailto:d-lord@tamu.edu) (D. Lord). 1 The work was done while the author was at Texas A&M University. 2 Tel.: +1 979 458 3949. weighted Poisson distribution has convergence restrictions and itis necessary to choose an appropriate function, which is sometimes difficult to do. With the generalized Poisson model, the bounded dispersionparameter whenunder-dispersionoccurs greatlydiminishes its applicability to count data ([Famoye,](#page-7-0) [1993\)](#page-7-0) and the model does not include the Poisson family in the interior of the parameter space ([Castillo](#page-7-0) [and](#page-7-0) [Perez-Casany,](#page-7-0) [2005\).](#page-7-0) The gamma count distribution assumes that observations are not independent where the observation for time t − 1 would affect the observation for time t ([Winkelmann,](#page-8-0) [2008;](#page-8-0) [Cameron](#page-8-0) [and](#page-8-0) [Trivedi,](#page-8-0) [1998\).](#page-8-0) This would become unrealistic if the time gap between the two observations is large, which can be problematic for analyzing crash data. Among the distributions or models that have been documented in the literature, two distributions that can handle both under- and over-dispersion are particularly noteworthy. One is the Conway–Maxwell–Poisson (COM-Poisson or CMP) distribution ([Conway](#page-7-0) [and](#page-7-0) [Maxwell,](#page-7-0) [1962;](#page-7-0) [Shmueli](#page-7-0) et [al.,](#page-7-0) [2005;](#page-7-0) [Kadane](#page-7-0) et [al.,](#page-7-0) [2006\)](#page-7-0) and the other is the double Poisson (DP) distribution ([Efron,](#page-7-0) [1986\).](#page-7-0) Albeit first introduced in 1962, the statistical properties of the COM-Poisson have not been extensively investigated until recently. The COM-Poisson distribution and its generalized linear model (GLM) have been found to be very flexible to handle count data ([Guikema](#page-7-0) [and](#page-7-0) [Coffelt,](#page-7-0) [2008;](#page-7-0) [Lord](#page-7-0) et [al.,](#page-7-0) [2008a;](#page-7-0) [Sellers](#page-7-0) et [al.,](#page-7-0) [2012;](#page-7-0) [Francis](#page-7-0) et [al.,](#page-7-0) [2012\).](#page-7-0) As for the DP, its distribution has seldom been investigated and applied since its first introduction about 25 years ago. A handful of research studies have mentioned that the hurdle for applying the DP is its normalizing constant (or multiplicative constant), which is not available in the closed form [\(Winkelmann,](#page-8-0) [2008;](#page-8-0) [Hilbe,](#page-8-0) [2011;](#page-8-0) [Zhu,](#page-8-0) [2012\).](#page-8-0) They found that the results of the DP with its normalizing constant approximated by Efron's original method are not exact. Instead of using Efron's approximation method, this study documents a different method for handling the normalizing constant. The objectives of this study are to: (1) examine the applicability of the DP GLM for analyzing motor vehicle crash data characterized by over- and under-dispersion and (2) compare the performance of the DP GLM with COM-Poisson GLM in terms of goodness-of-fit and theoretical soundness. Two empirical over-dispersed datasets (one for the high mean scenario and one for the low mean scenario) and one empirical under-dispersed dataset were used. #### **2. Background** This section describes the characterization and GLM framework of the DP and COM-Poisson models. # 2.1. Double Poisson model Based on the double exponential family, [Efron](#page-7-0) [\(1986\)](#page-7-0) proposed the double Poisson distribution. The double Poisson model, based on the distribution, has two parameters and . The approximate probability mass function (PMF) is given as: $$P(Y = y) = f_{\mu,\theta}(y) = (\theta^{1/2}e^{-\theta\mu}) \left(\frac{e^{-y}y^y}{y!}\right) \left(\frac{e\mu}{y}\right)^{\theta y}, y = 0, 1, 2, ...,$$ (1) The exact double Poisson density is given as: $$P(Y = y) = \tilde{f}_{\mu,\theta}(y) = c(\mu,\theta)f_{\mu,\theta}(y)$$ (2) where the factor c(-, ) can be calculated as: $$\frac{1}{c(\mu,\theta)} = \sum_{y=0}^{\infty} f_{\mu,\theta}(y) \approx 1 + \frac{1-\theta}{12\mu\theta} (1 + \frac{1}{\mu\theta})$$ (3) With c(-, ) being the normalizing constant nearly equal to 1. The constant c(-, ) ensures that the density sums to unity. The expected value and the standard deviation (SD) referring to the exact density ˜f-,(y) are estimated as follows: $$E(Y) \approx \mu$$ , (4) $$SD(Y) \approx \left(\frac{\mu}{\theta}\right)^{1/2}$$ (5) Thus, the double Poisson model allows for both over-dispersion ( < 1) and under-dispersion ( > 1). When = 1, the double Poisson distribution collapses to the Poisson distribution. Particular focus should be given to the use of the normalizing constant which includes an infinite series ∞ y=0f-,(y). Although [Efron](#page-7-0) [\(1986\)](#page-7-0) demonstrated the closed form approximation to the infinite series in Eq. (3) and even the approximate density in Eq. (2) are reasonably good for the case - = 10, they are highly unreliable when is small (i.e., the sample mean is small). For instance, when - = 0.1 and = 1.5, the closed form approximation solution turns into a negative value (1 + ((1 − )/12-)(1 + (1/-)) = −1.13) and the sum of the approximate densities is not fairly close to unity ∞ y=0f-,(y) = 1.11 . [Winkelmann](#page-8-0) [\(2008\)](#page-8-0) and [Hilbe](#page-7-0) [\(2011\)](#page-7-0) also indicated that the normalizing constant is the hurdle in applying the DP. They found that the results of the DP with its normalizing constant approximated by Efron's original method are not exact. Some researchers approached the DP model by completely ignoring the normalizing constant. However, this method is associated with both theoretical and practical limitations. It should be noted that the removal of the normalizing constant in the PMF will make the sum of all the likelihoods not equal to unity, which substantially diminishes the DP's mathematical appeal. Corresponding application based on this method also results in unreliable estimates. [Zhu](#page-8-0) [\(2012\)](#page-8-0) recently tested the DP without the normalizing constant, and found thatthe model provides a good fitfor the mean, but does a terrible job for adequately capturing the variance. In light of the inadequacy of the approximate density function and the aforementioned approximation to the normalizing constant in handling low mean data, it is important to approximate the normalizing constant with high accuracy and reliability. Given the factthatthe infinite series∞ y=0f-,(y)is similar to the Poisson sum and that it converges very quickly especially when is small, this study takes the kth partial sum of the infinite series (i.e., the sum of the first k terms) to approximate its sum. This method makes it possible to compute the normalizing constant to some modest accuracy by adjusting the number k according to user's preference. After multiple trials on the selection of k values, we recommend the value of k to be as least twice as large as the sample mean. For the DP GLM, the expected number of crashes per year is linked to the explanatory variables xj by the following link function (similar to the traditional Poisson): $$E(Y) \approx \mu = \exp\left(\beta_0 + \sum_{j=1}^p \beta_j x_j\right)$$ (6) where the vector ˇj is the coefficients to be estimated. #### 2.2. Conway–Maxwell–Poisson model Inorder tomodel queues andservice rates,[Conway](#page-7-0) [andMaxwell](#page-7-0) [\(1962\)](#page-7-0) first introduced the COM-Poisson distribution as a generation of the Poisson distribution. However, this distribution was not widely used until [Shmueli](#page-8-0) et [al.](#page-8-0) [\(2005\)](#page-8-0) further examined its statistical and probabilistic properties. [Kadane](#page-7-0) et [al.](#page-7-0) [\(2006\)](#page-7-0) developed the conjugate distributions for the parameters of the COM-Poisson distribution. The COM-Poisson distribution has two parameters with as the centering parameter and as the dispersion parameter. When is close to one, the centering parameter is approximately equal to the mean. When gets small, differs substantially from the mean. For over-dispersed data, would be expected to be small and thus a COM-Poisson GLM based on the original COM-Poisson formulation would be very difficult to interpret and use for over-dispersed data. To circumvent the problem, [Guikema](#page-7-0) [and](#page-7-0) [Coffelt](#page-7-0) [\(2008\)](#page-7-0) proposed a re-parameterization of the COM-Poisson distribution to provide a clear centering parameter. They substituted - = 1/*v* and then the new formulation of the COM-Poisson distribution is summarized in Eqs. (7) and (8): $$P(Y = y) = \frac{1}{S(\mu, \nu)} \left(\frac{\mu^y}{y!}\right)^{\nu} \tag{7}$$ $$S(\mu, \nu) = \sum_{n=0}^{\infty} \left(\frac{\mu^n}{n!}\right)^{\nu} \tag{8}$$ Correspondingly, the mean and variance of Y are given by Eqs. (9) and (10) in terms of the new information and the asymptotic approximations ofthe mean and variance of Y are given by Eqs.(11) and (12): $$E[Y] = \frac{1}{v} \frac{\partial \log S}{\partial \log \mu} \tag{9}$$ $$V[Y] = \frac{1}{v^2} \frac{\partial^2 \log S}{\partial \log^2 \mu}$$ (10) $$E[Y] \approx \mu + 1/2\nu - 1/2$$ (11) $$Var[Y] \approx \mu/\nu$$ (12) The approximations are especially accurate once - > 10. However, [Francis](#page-7-0) et [al.](#page-7-0) [\(2012\)](#page-7-0) recently found that the asymptotic mean can accurately approximate the true mean for E(Y) > 5. This new parameterization makes the integral part of the mode leaving as a reasonable centering parameter. The substitution allows to keep its role as a shape parameter. That is, < 1 leads to over-dispersion and > 1 to under-dispersion. Based on the new parameterization, [Guikema](#page-7-0) [and](#page-7-0) [Coffelt\(2008\)](#page-7-0) developed a COM-Poisson GLM framework to model discrete count data using Bayesian framework in WinBUGS [\(Spiegelhalter](#page-8-0) et [al.,](#page-8-0) [2003\).](#page-8-0) Their modeling framework is a dual-link GLM in which both the mean and variance depend on the covariates. It should be noted that parameter estimation for the dual-link GLM is complex and difficult. To make all the models in this paper comparable, this paper used the single-link COM-Poisson GLM framework [\(Lord](#page-7-0) et [al.,](#page-7-0) [2008a\).](#page-7-0) The link function is given as: $$\ln(\mu) = \beta_0 + \sum_{j=1}^{p} \beta_j x_j$$ (13) where xj are the explanatory variables (traffic and geometric variables) and ˇj are the coefficients to be estimated. With the derivation of the likelihood function of the COM-Poisson GLM by [Sellers](#page-8-0) [and](#page-8-0) [Shmueli](#page-8-0) [\(2010\),](#page-8-0) the maximum likelihood estimation (MLE) of the parameters of a COM-Poisson GLM was greatly simplified when compared to the Bayesian estimating method. The MLE-based COM-Poisson GLM is available in R package (COMPoissonReg on CRAN). Note that the MLE-based COM-Poisson GLM was developed under the original parameterization and thus it is difficult to interpret the coefficients in terms of their impact on the mean. # **3. Methodology** This section describes the link functions and parameter estimation methods for different models as well as the goodness-of-fit (GOF) statistics for comparing different models. # 3.1. Link function # 3.1.1. Toronto data For the first over-dispersed dataset (Toronto data), the linear predictor for all the models are linked to their respective centering parameter i. Note i here always do not denote the expectation of a variable (i.e. E(Yi)). For the COM-Poisson GLM, it is not directly linked to the expected mean as is the case with the NB GLM and DP GLM. The link function for the DP GLM is given as: $$E(Y_i) \approx \mu_i = \beta_0 F_{Maj\_i}^{\beta_1} F_{Min\_i}^{\beta_2}$$ (14) The link function for the NB GLM is given as: $$E(Y_i) = \beta_0 F_{Maj\_i}^{\beta_1} F_{Min\_i}^{\beta_2}$$ (15) The link function for the COM-Poisson GLM is given as: $$\mu_i = \beta_0 F_{Maj,i}^{\beta_1} F_{Min,i}^{\beta_2} \tag{16}$$ where E(Yi) = the mean number of crashes per year for intersection i; i = the estimated centering parameter for the DP GLM or COM-Poisson GLM for intersection i; Fmaj i = entering flow for the major approach (or AADT) for intersection i; Fmin i = entering flow for the minor approach (or AADT) for intersection i; ˇ0, ˇ1, ˇ2 = coefficients. Although the functional form may be affected by the omitted variable bias ([Lord](#page-7-0) [and](#page-7-0) [Mannering,](#page-7-0) [2010\),](#page-7-0) it is still adequate for the purpose of this exercise. With the parameters -ˆ i and *v*ˆ, the mean response for the COM-Poisson GLM can be obtained using the asymptotic approximations given in Eq. (11) ([Lord](#page-7-0) et [al.,](#page-7-0) [2008a\).](#page-7-0) However, it should be noted that the approximation is accurate only when - > 10. Given the fact thatthe sample mean is below 10, 100,000 samples were simulated for estimating -ˆ i and *v*ˆ for the COM-Poisson GLMs. Then, the mean of those samples was as the mean response. For other methods on obtaining fitted values,the reader is referred to [Sellers](#page-8-0) [and](#page-8-0) [Shmueli](#page-8-0) #### 3.1.2. Texas data Similar to the Toronto data, the link function for the DP GLM whenfitting the second over-dispersed dataset(Texas data)is given as: $$E(Y_i) \approx \mu_i = \beta_0 L_i F_i^{\beta_1} e^{\{\beta_2 LW_i + \beta_3 SW_i + \beta_4 CURVE\_DEN_i\}}$$ $$\tag{17}$$ The link function for the NB GLM is given as: $$E(Y_i) = \beta_0 L_i F_i^{\beta_1} e^{\{\beta_2 LW_i + \beta_3 SW_i + \beta_4 CURVE\_DEN_i\}}$$ (18) The link function for the COM-Poisson GLM is given as: $$\mu_i = \beta_0 L_i F_i^{\beta_1} e^{\{\beta_2 L W_i + \beta_3 S W_i + \beta_4 CURVE\_DEN_i\}}$$ $$\tag{19}$$ where E(Yi) = the mean number of KAB (Killed, Injury A, Injury B) crashes per year for segment i; i = the estimated centering parameter for DP GLM or COM-Poisson GLM for segment i; Li = length of the segment in miles for segment i; Fi = the traffic flow volume (or AADT) in veh/day for segment i; LWi = lane width in ft for segment i; SWi = total shoulder width (both sides) in ft for segment i; CURVE DENi = number of horizontal curves per mile located on the segment i; ˇ0, ˇ1, ˇ2, ˇ3, ˇ4 = coefficients. # 3.1.3. Under-dispersed data For the under-dispersed data, the link function for fitting all the models are the same with previous studies using this dataset ([Oh](#page-8-0) et [al.,](#page-8-0) [2006;](#page-8-0) [Lord](#page-8-0) et [al.,](#page-8-0) [2010\):](#page-8-0) $$\mu_{i} = e^{\left\{\beta_{0} + \beta_{1} \ln(F_{i}) + \sum_{i=1}^{n} \beta_{i} x_{i}\right\}}$$ (20) where i = the estimated centering parameter; Fi = the average daily traffic (ADT) in veh/day for segment i; xi = the estimated covariates such as average daily railway traffic, presence of commercial area, etc.; ˇi = coefficients. # 3.2. Parameter estimation The parameters of the DP GLMs were estimated using the NLMIXED procedure in SAS [\(SAS,](#page-8-0) [2002\).3](#page-8-0) The default quasi-Newton algorithm was applied to obtain the estimated parameters along with their approximate standard errors. For the COM-Poisson GLM, the coefficients were estimated using the Bayesian framework in WinBUGS ([Spiegelhalter](#page-8-0) et [al.,](#page-8-0) [2003\).](#page-8-0) More information on estimating the coefficients ofthe COM-Poisson GLMs using WinBUGS can be found in [Lord](#page-7-0) et [al.](#page-7-0) [\(2008a\).](#page-7-0) ## 3.3. Goodness-of-fit statistics The following GOF statistics were used to compare different models for both over- and under-dispersed data: • Akaike Information Criterion (AIC): AIC is a measure that is used to describe the tradeoff between accuracy and complexity of the model. The model with smaller AIC is preferred to others. AIC is defined as: $$AIC = -2 \log L + 2p \tag{21}$$ where L is the maximized value of the likelihood function for the estimated model, and p is the number of parameters in the statistical model. By penalizing models with a large number of parameters, the AIC attempts to select the model that best explains the data with a minimum of parameters. Lower the AIC, better the model. • Deviance Information Criterion (DIC): The DIC is a hierarchical modeling generalization of the AIC. The model with smaller DIC is preferred to others. DIC is defined as: $$DIC = \bar{D} + p_D \tag{22}$$ where D¯ = −2Log L is the posterior mean of the deviance of the unstandardized model and L is the mean of the model log likelihood. pD = D¯ − D¯ (y|¯) is the effective model parameters and ¯ is the expectation of . • MeanPredictionBias (MPB): TheMPBisusedtomeasure themagnitude and direction of the average model bias. MPB is calculated using the following equation ([Oh](#page-8-0) et [al.,](#page-8-0) [2003\):](#page-8-0) $$MPB = \frac{1}{n} \sum_{i=1}^{n} (\hat{y}_i - y_i)$$ (23) where n is the sample size, yˆi and yi are the predicted and observed crashes at site i respectively. When the model - over-predicts crashes, MPB is positive and when the model under-predicts crashes, MPB is negative. - Mean Absolute Deviance (MAD): The MAD is the average of the absolute deviations and it measures the average mis-prediction of the model. The model closest to 0 is considered to be the best. It is computed by the following equation ([Oh](#page-8-0) et [al.,](#page-8-0) [2003\):](#page-8-0) $$MAD = \frac{1}{n} \sum_{i=1}^{n} |\hat{y}_i - y_i|$$ (24) • Mean Squared Predictive Error (MSPE): The MSPE is often used to access the error associated with a validation or external data set. The model closest to 0 is considered to be the best. It can be calculated by the following equation ([Oh](#page-8-0) et [al.,](#page-8-0) [2003\):](#page-8-0) MSPE = $$\frac{1}{n} \sum_{i=1}^{n} (\hat{y}_i - y_i)^2$$ (25) #### **4. Data description** This section describes the summary characteristics of the two observed over-dispersed datasets and one observed underdispersed dataset. # 4.1. Over-dispersed data The first over-dispersed dataset, Toronto data, contained crashes collected at urban 4-legged signalized intersections in Ontario, Toronto, from the year 1990 to 1995. A total number of 54,869 crashes that occurred at 868 intersections were reported. The data were found to be of good quality and then applied to many studies [\(Lord,](#page-7-0) [2000;](#page-7-0) [Miaou](#page-7-0) [and](#page-7-0) [Lord,](#page-7-0) [2003\).](#page-7-0) The crash data for the year 1995 with a total number of 10,030 crashes were used in this study. Traffic flow data from both major approaches (Major AADT) and minor approaches (Minor AADT) were also collected. [Table](#page-4-0) 1 shows the summary statistics for the key variables in this dataset.The second dataset, Texas data, contained crashes that occurred on 4-lane rural undivided and divided roadway segments for five years (from 1997 to 2001) in Texas. The dataset was used to develop the statistical models and accident modification factors in the project NCHRP 17-29 [\(Lord](#page-8-0) et [al.,](#page-8-0) [2008b\).](#page-8-0) The data were provided by the Texas Department of Public Safety (DPS) and the Texas Department of Transportation (TxDOT). Only undivided segments crash data were used in this study. Other than the crash frequency classified by the year and severity level, information about variables such as annual average daily traffic (AADT), shoulder width, number of intersections was also collected for those segments. [Table](#page-4-0) 2 shows the summary statistics of the key variables in this dataset. #### 4.2. Under-dispersed data The dataset used for modeling under-dispersion were collected at railway-highway crossings in Korea. This dataset were found to be under-dispersed and were used for establishing the gamma count models to examine factors associated with railroad crossing crashes [\(Oh](#page-8-0) et [al.,](#page-8-0) [2006\).](#page-8-0) Traffic accident records were recorded at a total of 162 railway-highway crossings in the 5-year period from 1998 to 2002. A total of 56 continuous and categorical explanatory variables including average daily traffic (ADT) were collected through site visits and investigations. [Tables](#page-4-0) 3 and 4 summarize the key continuous and categorical variables of that dataset respectively. 3 The NLMIXED procedure was designed to fit nonlinear mixed models, that is, models with nonlinear random and fixed effects. The statement PROC NLMIXED offers an interface for specifying and coding a user-defined conditional distribution. Besides, PROC NLMIXED provides a variety of optimization techniques to maximize an approximation to the likelihood. Although NLMIXED procedure was often intended for mixed effects model, it was used to fit models with only fixed effects in this study. **Table 1** Summary statistics of variables for the Toronto data. | Variables | | Min. | Max. | Mean | Std. err. | |-------------|-----------------------------------|------|-------|---------|-----------| | Response | Crashes | 0 | 54 | 11.56 | 10.02 | | Explanatory | AADT for Major Approach (veh/day) | 5469 | 72178 | 28044.8 | 10660.4 | | | AADT for Minor Approach (veh/day) | 53 | 42644 | 11010.2 | 8599.4 | **Table 2** Summary statistics of variables for the Texas data. | Variables | | Min. | Max. | Mean | Std. err. | |-------------|------------------------------------|------|-------|---------|-----------| | Response | KABa crashes for five years | 0 | 19 | 0.63 | 1.60 | | Offset | Segment length (miles) | 0.1 | 6.275 | 0.55 | 0.67 | | Explanatory | AADT (veh/day) | 402 | 24800 | 6613.61 | 4010.01 | | | Lane width (ft) | 9.75 | 16.5 | 12.57 | 1.59 | | | Shoulder width (right + left) (ft) | 0 | 40 | 9.96 | 8.02 | | | Right shoulder width (ft) | 0 | 24 | 13.65 | 3.65 | | | Median width (ft) | 1 | 240 | 47.71 | 28.87 | | | Number of intersections | 0 | 47 | 2.33 | 2.62 | | | Number of horizontal curves | 0 | 16 | 0.70 | 1.32 | a KAB crashes: crashes with severity level of fatality, injury type A, and injury type B. **Table 3** Summary statistics of continuous variables for the Korea data. | Variables | | Min. | Max. | Mean | Std. err. | |-------------|-----------------------------------------------------------------------|------|-------|---------|-----------| | Response | Crashes | 0 | 3 | 0.33 | 0.60 | | Explanatory | ADT (vehicles per day) | 10 | 61199 | 4617.00 | 10391.57 | | | Average daily railway traffic (trains per day) | 32 | 203 | 70.29 | −37.34 | | | Gradient of Road | −20 | 10.5 | −1.30 | 3.73 | | | Train detector distance (m) | 0 | 1329 | 824.50 | 328.38 | | | Time duration between the activation of warning signals and gates (s) | 0 | 232 | 25.46 | 25.71 | #### **5. Modeling results** This section presents the modeling results of the two observed over-disperseddatasets andone observedunder-disperseddataset. The results of the DP without the normalizing constant were not includedinthis studydue to its theoretical andpracticallimitations, as mentioned in Section [2.1.](#page-1-0) The level of significance is 10%. # 5.1. Toronto data [Table](#page-5-0) 5 summarizes the results of the DP GLMs, NB GLM and COM-Poisson GLM for the Toronto data. Two types of DP GLMs were investigated and each type represents a method to approximate the infinite series of the normalizing constant in the PMF. The "DP Exact" denotes the DP GLM in which the PMF includes a normalizing constant which is approximated by its partial sum and ensure the coefficients accurate to four decimal places. The "DP Approximate" denotes the DP GLM in which the PMF includes a normalizing constant which was originally approximated using a closed form formula given in Eq. [\(3\)](#page-1-0) [\(Efron,](#page-7-0) [1986\).](#page-7-0) As shown in [Table](#page-5-0) 5, both the DP Exact and DP Approximate provide similar coefficients with those of the NB GLM whereas the coefficients estimated by the COM-Poisson GLM are different from the other three models. The difference can be explained by the fact **Table 4** Summary statistics of categorical variables for the Korea data. | Variables | Coding | Frequency | Percentage | |-----------------------------|---------|-----------|------------| | Presence of commercial area | 1 (no) | 149 | 91.98% | | | 0 (yes) | 13 | 8.02% | | Presence of speed hump | 1 (no) | 134 | 82.72% | | | 0 (yes) | 28 | 17.28% | | Presence of track circuit | 1 (no) | 113 | 69.75% | | | 0 (yes) | 49 | 30.25% | | Presence of guide | 1 (no) | 126 | 77.78% | | | 0 (yes) | 36 | 22.22% | thatthe coefficients oftheCOM-Poisson GLMare linkedto themode rather than the mean, as is the case with the other three models. The standard errors of those coefficients for the DP Approximate have shown to be slightly smaller than those for the DP Exact. In terms of GOF, the COM-Poisson GLM fits the data slightly better than the other three with the smallest DIC, MAD and MPB value. The DP Exact fits the data almost the same as the NB GLM but better than the DP Approximate. # 5.2. Texas data [Table](#page-5-0) 6 summarizes the results of the DP GLM, NB GLM and COM-Poisson GLM for the Texas data. The results of the DP Approximate were not included because it had a convergence issue related to the low mean of this dataset. The absolute values of all the coefficients and their standard errors for the DP Exact and COM-Poisson GLM are larger than those for the NB GLM. Same as found in the Toronto data, the DP Exact provides estimated coefficients similar to the NB GLM. The GOF statistics in [Table](#page-5-0) 6 indicate that the three models perform equally well. #### 5.3. Under-dispersed data [Table](#page-5-0) 7 summarizes the results of the DP GLM (i.e. DP Exact), COM-Poisson GLM and Gamma count model for the underdispersed data. Again, the results for the DP Approximate are not presented because of its failure on convergence for the low mean data. As shown in [Table](#page-5-0) 7, variables significant in one model are found to be not necessarily significant in another model. The estimated shape parameter of the DP GLM is larger than 1, indicating some evidence of under-dispersion but not very strong; note that for the DP, the shape parameter being 1 is the cutoff point between over- and under-dispersion. It should be noted, however, that all three models are consistent in showing that under-dispersion is present (when conditioned on the mean). **Table 5** Modeling results for the Toronto data. | Variables | DP Exact | DP Approximate | NBa | COM-Poisson | |---------------------|--------------------|-------------------|------------------|------------------| | Constant | −10.2178 (0.4435)b | −10.0000 (0.4248) | −10.2458 (0.465) | −11.53 (0.4159) | | Major AADT | 0.6014 (0.0449) | 0.5988 (0.0431) | 0.6207 (0.046) | 0.635 (0.04742) | | Minor AADT | 0.7040 (0.022) | 0.6843 (0.0207) | 0.6853 (0.0211) | 0.795 (0.03101) | | Shape parameter | 0.4121 (0.0197) | 0.4334 (0.0189) | 0.1398 (0.0122) | 0.3408 (0.02083) | | AIC | 5100.4 | 5115.8 | 5077.3 | − | | DIC | – | – | – | 4953.7 | | MAD | 4.140 | 4.152 | 4.142 | 4.129 | | MSPE | 32.597 | 32.685 | 32.699 | 33.664 | | MPB | 0.028 | 0.074 | −0.045 | −0.025 | | No. of observations | 868 | 868 | 868 | 868 | a The results for the NB GLM and COM-Poisson GLM are directly taken from [Lord](#page-7-0) et [al.](#page-7-0) [\(2008a\).](#page-7-0) **Table 6** Modeling results for the Texas data. | Variables | DP exact | NBa | COM-Poisson | |--------------------------------------|------------------|-----------------|------------------| | Constant | −7.7514 (0.6015) | −6.8242 (0.547) | −8.057 (0.6601) | | AADT | 0.9193 (0.0644) | 0.7768 (0.058) | 0.955 (0.071) | | Lane width | −0.1137 (0.0245) | −0.0844 (0.023) | −0.1222 (0.0233) | | Total shoulder width | −0.0161 (0.0047) | −0.0114 (0.005) | −0.0171 (0.0045) | | Number of horizontal Curves per mile | 0.0888 (0.0169) | 0.0635 (0.016) | 0.0866 (0.0161) | | Shape parameter | 0.6209 (0.0359) | 0.3793 (0.057) | 0.7235 (0.0447) | | AIC | 3268.2 | 3199.2 | – | | DIC | – | – | 3200.4 | | MAD | 0.82 | 0.826 | 0.862 | | MSPE | 2.61 | 2.727 | 2.73 | | MPB | −0.074 | −0.067 | −0.001 | | No. of observations | 1499 | 1499 | 1499 | a The results for the NB GLM are directly taken from the project NCHRP 17-29 [\(Lord](#page-8-0) et [al.,](#page-8-0) [2008b\).](#page-8-0) For those variables significant in two or three models, their estimated coefficients and standard errors are shown to differ much across models. Those differences between the DP GLM and COM-Poisson GLM could be explained by the fact that the coefficients estimated by the COM-Poisson GLM are for the centering parameter , rather than the mean as in the case ofthe DP GLM and Gamma count model ([Lord](#page-8-0) et [al.,](#page-8-0) [2010\).](#page-8-0) In spite of the difference from the coefficients, the DP GLM and COM-Poisson GLM provide very close GOF statistics and the COM-Poisson GLM is slightly better with its AIC, MAD and MSPE being the smallest. For the differences on coefficients between the DP GLM and Gamma count model, however, the inferior performance of the Gamma count model might be a major contribution, which is illustrated by its overall GOFs being the worst among three models and not quite close to the other two (at least for three GOF measures). So we could further conclude that the Gamma count model could not provide the precise estimated coefficients as the DP GLM and COM-Poisson GLM do. #### 5.4. Marginal effects With the aim to better interpret the effects of corresponding variables on crash frequency, this study investigated the relative marginal effects of significant variables produced by different models for the three datasets. (Note: the discussion results about the various relationships between the number of crashes and the covariates can be found in the original documents from which the data were used.) An example using the traffic flow as the interested variable has been provided, as shown in [Fig.](#page-6-0) 1. The vertical axis is based on the logarithmic scale. The relative marginal effects of other variables can be generated and interpreted similarly. The **Table 7** Comparison results for the Korea data. | Variables | DP exact | COM-Poissona | Gammab | |-------------------------------------------------------------------|-----------------|-----------------|----------------| | Constant | −5.234 (0.911) | −6.657 (1.206)c | −3.438 (1.008) | | Ln(ADT) | 0.383 (0.075) | 0.648 (0.139) | 0.230 (0.076) | | Average daily railway traffic | – | – | 0.004 (0.0024) | | Presence of commercial area | 1.094 (0.348) | 1.474 (0.513) | 0.651 (0.287) | | Train detector distance | 0.0019 (0.0006) | 0.0021 (0.0007) | 0.001 (0.0004) | | Time duration between the activation of warning signals and gates | – | – | 0.004 (0.002) | | Presence of track circuit controller | −0.800 (0.317) | −1.305 (0.431) | – | | Presence of guide | – | −0.998 (0.512) | – | | Presence of speed hump | −1.009 (0.399) | −1.495 (0.531) | −1.58 (0.859) | | Shape parameter | 1.142 (0.226) | 2.349 (0.634) | 2.062 (0.758) | | AIC | 212.1 | 210.7 | 211.38 | | MPB | 0.002 | −0.007 | 0.179 | | MAD | 0.363 | 0.348 | 0.459 | | MSPE | 0.252 | 0.236 | 0.308 | | No. of observations | 162 | 162 | 162 | a The results for the COM-Poisson GLM and Gamma count model are directly taken from [Lord](#page-8-0) et [al.](#page-8-0) [\(2010\).](#page-8-0) b Values in parentheses are the standard errors. Same format is used for Tables 6 and 7. b For model details about the Gamma count model and the results, the reader is referred to [Oh](#page-8-0) et [al.](#page-8-0) [\(2006\)](#page-8-0) and [Lord](#page-8-0) et [al.](#page-8-0) [\(2010\).](#page-8-0) ![](_page_6_Figure_2.jpeg) **Fig. 1.** Relative marginal effects of the traffic flows for three datasets: (a) marginal effect for the Toronto data, (b) marginal effect for the Texas data and (c) marginal effect for the Korea data. relative marginal effect of a particular variable Xi is estimated as the following ([Cameron](#page-7-0) [and](#page-7-0) [Trivedi,](#page-7-0) [1998\):](#page-7-0) $$RME_{X_i} = \frac{1}{E[Y|X_i]} \frac{\partial E[Y|X_i]}{\partial X_i}$$ (26) Given that the mean for the COM-Poisson is not directly available from its estimated parameters as the case with the DP GLM, NB GLM and Gamma count model, Eq. [\(11\)](#page-2-0) was used to approximate the mean and thus the relative marginal effectfor the COM-Poisson GLM. From the relative marginal plot for the Toronto data (Fig. 1a), all three models show very similar results and they tend to be closer to each other with the increase on the flow variable for the major approach. For the relative marginal effects for the Texas data shown in Fig. 1b,the DP GLM and NB GLM provide similar results while the COM-Poisson GLM tends to give smaller estimated marginal effect for one unit change on the flow variable. The difference between the COM-Poisson GLM and the other two models is obvious when the flow variable is less than 10,000 veh/day and the difference becomes smaller with the increase of the flow variable. This difference could be explained by the lack of accuracy in applying the mean approximation of the COM-Poisson model to the low mean data (see Eq. [\(11\)](#page-2-0) and the discussion of its use documented in Section [2.2\).](#page-1-0) Fig. 1c shows themarginal effects for theKorea data. The Gamma count model is shown to always provide the smallest marginal effect for a particular flow value, which is consistent with the fact thatits coefficientfor the flow variable is considerably smaller than those for the other two. The difference on the results between the DP GLM and COM-Poisson GLM decreases with the increase of the flow variable, which is consistent with our findings for the other two datasets. The difference is hardly detectable when the flow variable is larger than 20,000 veh/day. Overall, the marginal effects for the three datasets in Fig. 1 indicate that the DP GLM and COM-Poisson GLM provide similar results unless when the flow variable is so small that the mean approximation of the COM-Poisson GLM is not able to work properly. #### **6. Discussion** First, this study has demonstrated that the DP GLM with its normalizing constant approximated by the new method can handle both over-dispersed and under-dispersed datasets. Its performance is comparable to the COM-Poisson GLM in terms of GOF, although the COM-Poisson GLM provides a slightly better fit in all types of datasets. Second, the DP distribution has seldom been investigated and applied since its first introduction more than two decades ago. The hurdle of applying the DP is related to the approximation of its normalizing constant by the closed form equation originally proposed by [Efron](#page-7-0) [\(1986\).](#page-7-0) This study has found that this approximation method is inaccurate and not reliable enough for the low mean count data. Imposing this approximation leads to convergence issues as evidentfrom the two low mean datasets used in this study. More importantly, this study has documented the use of the partial sum to approximate the infinite series for the normalizing constant ofthe DP distribution. The newly proposed approximation method not only improved the GOF performance of the DP GLM but also solved the convergence issue when applying to datasets with a relatively low sample mean. Besides, this method makes it possible to compute the normalizing constant to some modest accuracy according to user's preference. With more terms included in the partial sum, the values for the coefficients and deviance of the DP GLM approach their limits more closely. Thus, it is possible to approximate those values to the desired decimal places. Although more accuracy is achieved by increasing the nonlinearity of the PMF, this method does not significantly increase computational complexity. For example,the DP GLM for the Texas data using the first 100 terms partial sum converged in less than 10 s in SAS 9.3. In comparison, the COM-Poisson GLM also converged in less than 10 s in the software R ([R](#page-8-0) [Development](#page-8-0) [Core](#page-8-0) [Team,](#page-8-0) [2006\).](#page-8-0) Third, the DP GLM provides very similar estimates to those coefficients provided by the NB GLM. The mathematics to manipulate the DP models is very simple and very similar to the NB models, since both of the two models use the mean as the centering parameter and have a shape parameter to capture dispersion. Moreover, both the DP GLM and NB GLM in this study were developed based on the MLE in SAS. However, the application of the NB GLM is only limited to modeling the over-dispersed data and provides unreliable parameter estimates when fitting the under-dispersed data ([Lord](#page-8-0) et [al.,](#page-8-0) [2010\).](#page-8-0) On the other hand, the shape parameter of the DP GLM has been found to be very flexible and it can handle under-, equi-, and over-dispersion (Var(Y) ≈ E(Y)/, > 1 for underdispersion; = 1 for equi-dispersion; < 1 for over-dispersion). The simplicity and flexibility of the DP mathematics also allows for the further development of multilevel models such as the random parameter models (Anastasopoulos and Mannering, 2009), finite mixture models ([Park](#page-8-0) [and](#page-8-0) [Lord,](#page-8-0) [2009\),](#page-8-0) Markov switching models ([Malyshkina](#page-8-0) et [al.,](#page-8-0) [2009\),](#page-8-0) etc., which will better account for the unobserved heterogeneity in crash datasets. This will encourage the potential use of the DP models to handle more complex crash dataset. Fourth, the DP GLM can be more easily estimated and computationally inexpensive than the COM-Poisson GLM although the former fits the data slightly worse than the latter (at least for these datasets). It is simple to use the DP GLM for interpreting the coefficients in terms of their impact on the mean since the mean parameter is linked to the linear predictor through the logarithmic link. On the other hand, certain mean approximation methods have to be made for applying the COM-Poisson GLM to obtain its mean and thus the marginal effects. However,those approximation methods may not work well for low mean datasets. Previous studieshave investigatedthe accuracy ofthose approximationmethods. For the COM-Poisson GLM based on its original parameterization, the estimated means can be approximated using the asymptotic expression but is only accurate for *v* ≤ 1 or > 10*v* ([Sellers](#page-8-0) [and](#page-8-0) [Shmueli,](#page-8-0) [2010\).](#page-8-0) Although the re-parameterization of the COM-Poisson distribution proposed by Guikema and Coffelt (2008) was based on a more clear centering parameter (i.e., the mode), it is still not directly comparable to the mean in the DP GLM. Despite of the aforementioned advantages of the DP GLM, the application of the DP GLM is associated with several limitations. Due to the scarcity of the observed under-dispersed crash data, this study investigated the performance of the DP GLM using only one observed dataset. Further investigation should be conducted on how the DP GLM will handle under-dispersed count data in a broader context. Besides, to make a fair comparison, the model specifications of the DP GLMs are consistent with those of the COM-Poisson GLMs, which have been documented in past studies. However, doing this does not guarantee the models are not affected by the issue of omitting variable bias. Future work should investigate how including more variables in the model will affect the comparison in terms of performance between these two models, especially for variables that are known to significantly influence the frequency of crashes. Further work is also recommended to better understand under which circumstances the DP GLM is preferred over the COM-Poisson GLM. ### **7. Summary and conclusions** This study has examined the applicability of the DP GLM for analyzing motor vehicle crash data characterized by over- and under-dispersion and compared its performance to that of the COM-Poisson GLM. The DP GLM was developed based on a new method to approximate its normalizing constant with high accuracy and reliability. The normalizing constant ofthe DP distribution used to be left out or approximated using a closed formformula, and neither of them can provide accurate enough results especially for the low mean data. The DP GLM and COM-Poisson GLM were developed using two observed over-dispersed datasets and one observed under-dispersed dataset. The results of the NB GLM and Gamma count model were also provided as reference. The modeling results for both over- and under-dispersed datasets indicate that the DP GLM with its normalizing constant approximated by the new method performs slightly worse than the COM-Poisson GLM in terms of the GOF statistics. The differences, however, are not large enough to disregard its use for analyzing crash data. Indeed, the former is much easier to use and estimate than the latter. The coefficients in terms of their impact on the means are simpler to interpret with the DP GLM than for the COM-Poisson GLM. The computational time needed for the DP GLM to converge is so small that it only takes several seconds to generate the modeling results. In sum, the DP GLM with its properly approximated normalizing constant offers a very flexible and efficient alternative for researchers for modeling count data. #### **Acknowledgement** The authors would like to thank Dr. Thomas Wehrly from Texas A&M University for his insightful statistics-related comments. # **References** Anastasopoulos, P., Mannering, F., 2009. [A](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [note](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [on](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [modeling](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [vehicle-accident](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [frequencies](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [with](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [random-parameters](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [count](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [models.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [Accident](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [Analysis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [Pre](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005)[vention](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [41](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [\(1\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) [153](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005)–[159.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0005) Borle, S., Boatwright, P., Kadane, J.B., 2006. [The](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [timing](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [bid](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [placement](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [extent](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [multiple](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [bidding:](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [an](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [empirical](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [investigation](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [using](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [eBay](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [online](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [auctions.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [Sta](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010)[tistical](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [Science](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [21](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [\(2\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) [194–205.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0010) Cameron, A.C., Trivedi, P.K., 1998. [Regression](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0015) [Analysis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0015) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0015) [Count](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0015) [Data.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0015) [Cambridge](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0015) [University](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0015) [Press,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0015) [Cambridge,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0015) [UK.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0015) Castillo, J., Perez-Casany, M., 2005. [Over-dispersed](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) [under-dispersed](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) [Poisson](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) [generalizations.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) [Journal](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) [Statistical](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) [Planning](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) [Inference](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) [134](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) [\(2\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) [486](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020)–[500.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0020) Consul, P.C., 1989. [Generalized](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0025) [Poisson](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0025) [Distributions:](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0025) [Properties](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0025) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0025) [Applications.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0025) [Marcel](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0025) [Dekker,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0025) [New](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0025) [York.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0025) Conway, R.W., Maxwell, W.L., 1962. [A](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) [queuing](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) [model](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) [with](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) [state](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) [dependent](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) [service](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) [rates.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) [Journal](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) [Industrial](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) [Engineering](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) [12,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) [132–136.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0030) Efron, B., 1986. [Double](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [exponential](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [families](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [their](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [use](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [in](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [generalized](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [linear](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [Regression.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [Journal](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [the](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [American](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [Statistical](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [Association](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [81](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [\(395\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) [709–721.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0035) Famoye, F., 1993. [Restricted](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [generalized](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [Poisson](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [regression](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [model.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [Communications](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [in](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [Statistics](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [–](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [Theory](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [Methods](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [22](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [\(5\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) [1335–1354.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0040) Francis, R.A., Geedipally, S.R., Guikema, S.D., Dhavala, S.S., Lord, D., Larocca, S., 2012. [Characterizing](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) [the](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) [performance](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) [the](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) [Conway–Maxwell–Poisson](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) [generalized](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) [linear](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) [model.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) [Risk](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) [Analysis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) [32](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) [\(1\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) [167–183.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0045) Guikema, S.D., Coffelt, J.P., 2008. [A](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) [flexible](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) [count](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) [data](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) [regression](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) [model](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) [for](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) [risk](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) [analysis.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) [Risk](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) [Analysis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) [28](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) [\(1\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) [213–223.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0050) Hauer, E., 1997. [Observation](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [Before-After](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [Studies](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [in](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [Road](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [Safety:](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [Estimating](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [the](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [Effect](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [Highway](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [Traffic](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [Engineering](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [Measures](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [on](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [Road](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [Safety.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) [Elsevier](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0055) Hilbe, J.M., 2011. [Negative](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0060) [Binomial](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0060) [Regression,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0060) [second](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0060) [ed.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0060) [Cambridge](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0060) [University](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0060) [Press,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0060) [Cambridge,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0060) [UK.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0060) Kadane, J.B., Shmueli, G., Minka, T.P., Borle, S., Boatwright, P., 2006. [Conjugate](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0065) [analy](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0065)[sis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0065) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0065) [the](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0065) [Conway](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0065)–[Maxwell–Poisson](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0065) [distribution.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0065) [Bayesian](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0065) [Analysis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0065) [1,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0065) [363](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0065)–[374.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0065) Lord, D., 2000. [The](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [prediction](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [accidents](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [on](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [digital](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [networks:](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [characteristics](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [issues](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [related](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [to](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [the](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [application](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [accident](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [prediction](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [models.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [Department](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [Civil](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [Engineering,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [University](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [Toronto,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [Toronto,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [ON](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [\(Ph.D.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) [Dissertation\).](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0070) Lord, D., Mannering, F.L., 2010. [The](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [statistical](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [analysis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [crash-frequency](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [data:](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [a](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [review](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [assessment](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [methodological](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [alternatives.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [Transportation](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [Research](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [Part](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [A:](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [Policy](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [Practice](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [44](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [\(5\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) [291–305.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0075) Lord, D., Guikema, S.D., Geedipally, S.R., 2008a. [Application](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [the](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [Conway–Maxwell–Poisson](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [generalized](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [linear](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [model](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [for](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [analyzing](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [motor](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [vehicle](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [crashes.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [Accident](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [Analysis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [Prevention](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [40](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [\(3\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) [1123](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080)–[1134.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0080) - Lord, D., Geedipally, S.R., Persaud, B.N., Washington, S.P., van Schalkwyk, I., Ivan, J.N., Lyon, C., Jonsson, T., 2008b. Methodology for estimating the safety performance of multilane rural highways. NCHRP Web-Only Document 126. National Cooperation Highway Research Program, Washington, D.C. [http://onlinepubs.](http://onlinepubs.trb.org/onlinepubs/nchrp/nchrp w126.pdf) [trb.org/onlinepubs/nchrp/nchrp](http://onlinepubs.trb.org/onlinepubs/nchrp/nchrp w126.pdf) w126.pdf (retrieved February 2012). - Lord, D., Geedipally, S.R., Guikema, S., 2010. [Extension](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [the](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [application](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [Conway–Maxwell](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090)–[Poisson](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [models:](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [analyzing](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [traffic](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [crash](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [data](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [exhibiting](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [under-dispersion.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [Risk](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [Analysis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [30](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [\(8\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) [1268](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090)–[1276.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0090) - Malyshkina, N., Mannering, F., Tarko, A., 2009. [Markov](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [switching](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [negative](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [binomial](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [models:](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [an](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [application](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [to](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [vehicle](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [accident](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [frequencies.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [Accident](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [Analysis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [Prevention](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [41](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [\(2\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) [217–226.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0095) - Miaou, S.-P., Lord, D., 2003. [Modeling](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [traffic](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [crash-flow](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [relationships](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [for](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [intersec](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100)[tions:](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [dispersion](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [parameter,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [functional](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [form,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [Bayes](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [versus](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [empirical](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [Bayes.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [Transportation](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [Research](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [Record](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [1840,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) [31–40.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0100) - Oh, J., Lyon, C.,Washington, S.P., Persaud, B.N., Bared, J., 2003.[Validation](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [ofthe](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [FHWA](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [crash](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [models](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [for](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [rural](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [intersections:](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [lessons](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [learned.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [Transportation](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [Research](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [Record](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [1840,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) [41–49.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0105) - Oh, J., Washington, S.P., Nam, D., 2006. [Accident](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110) [prediction](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110) [model](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110) [for](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110) [railway](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110)[highway](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110) [interfaces.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110) [Accident](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110) [Analysis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110) [&](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110) [Prevention](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110) [38](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110) [\(2\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110) [346–356.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0110) - Park, E.S., Lord, D., 2007.[Multivariate](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0115) [Poisson-lognormalmodels](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0115) [for](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0115) [jointlymodeling](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0115) [crash](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0115) [frequency](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0115) [by](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0115) [severity.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0115) [Transportation](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0115) [Research](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0115) [Record](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0115) [2019,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0115) [1–6.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0115) - Park, B.-J., Lord, D., 2009. [Application](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [finite](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [mixture](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [models](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [for](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [vehicle](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [crash](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [data](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [analysis.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [Accident](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [Analysis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [Prevention](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [41](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [\(4\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) [683](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120)–[691.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0120) - R Development Core Team, 2006. [R:](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [A](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [Language](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [And](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [Environment](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [for](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [Statistical](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [Computing.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [R](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [Foundation](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [for](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [Statistical](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [Computing,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [Vienna,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [Austria,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [ISBN](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [3-](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) [900051-07-0.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0125) - SAS Institute Inc, 2002. [Version](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0130) [9](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0130) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0130) [the](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0130) [SAS](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0130) [System](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0130) [for](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0130) [Windows,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0130) [Cary,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0130) [NC.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0130) - Sellers, K.F., Shmueli, G., 2010. [A](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) [flexible](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) [regression](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) [model](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) [for](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) [count](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) [data.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) [Annals](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) [Applied](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) [Statistics](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) [4](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) [\(2\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) [943](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135)–[961.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0135) - Sellers, K.F., Borle, S., Shmueli, G., 2012. [The](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [COM-Poisson](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [model](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [for](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [count](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [data:](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [a](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [survey](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [methods](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [applications.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [Applied](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [Stochastic](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [Models](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [in](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [Business](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [Industry](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [28](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) [\(2\).](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0140) - Shmueli, G., Minka, T.P., Kadane, J.B., Borle, S., Boatwright, P., 2005. [A](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [useful](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [dis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145)[tribution](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [for](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [fitting](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [discrete](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [data:](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [revival](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [the](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [Conway–Maxwell–Poisson](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [distribution.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [Journal](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [the](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [Royal](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [Statistical](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [Society](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [Part](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [C](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [54,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) [127–142.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0145) - Spiegelhalter, D.J., Thomas, A., Best, N.G., Lun, D., 2003. WinBUGS Version 1.4 1 User Manual. MRC Biostatistics Unit, Cambridge, Available from: - Winkelmann, R., 2008. [EconometricAnalysis](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0155) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0155) [Count](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0155) [Data,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0155) [fifth](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0155) [ed.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0155) [Springer-Verlag,](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0155) [Berlin.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0155) - Zhu, F., 2012. [Modeling](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [Time](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [Series](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [of](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [Counts](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [with](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [COM-Poisson](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [INGARCH](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [Models.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [Mathematical](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [and](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [Computer](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [Modelling](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [56](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [\(9\),](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160) [191–203.](http://refhub.elsevier.com/S0001-4575(13)00285-6/sbref0160)