Zero Truncated Bivariate Poisson Model : Marginal-Conditional Modeling Approach with an Application to Traffic Accident Data

A new covariate dependent zero-truncated bivariate Poisson model is proposed in this paper employing generalized linear model. A marginal-conditional approach is used to show the bivariate model. The proposed model with estimation procedure and tests for goodness-of-fit and under (or over) dispersion are shown and applied to road safety data. Two correlated outcome variables considered in this study are number of cars involved in an accident and number of casualties for given number of cars.


Introduction
The count data analysis occupies an important role in applied statistics in various fields.When the observed outcomes are count and the desire is to estimate the covariate effects on outcomes, covariate dependent Bivariate Poisson (BVP) model is a tool of natural choice.It is expected that the observed outcomes on the same subject are be correlated.This type of data arises in many fields, for example, traffic accidents, health sciences, economics, social sciences, environmental studies among others.A typical example of such dependence arises in the number of traffic accidents and the number of injuries or fatalities during a specified period.However, in some situations outcomes may be truncated as zero values of counts may not be observed or may be missing for one or both of the outcomes.For example, in a sample drawn from hospital admission records, frequencies of zero accidents and length of stay are not available.Another example is the case where the data on number of traffic accidents and related injuries or fatalities and related risk factors are collected from records and, naturally, zero counts are not available.As an example, road safety data from data.gov.ukwebsite provides detailed information about the conditions of personal injury road accidents in Great Britain including the types of vehicles involved and the consequential casualties on public roads along with other background information.Only those accidents that involve personal injury reported to the police using the accident reporting form are recorded.Damage-only accidents, with no human casualties or accidents on private roads or car parks, are not included generating zero-truncated count data.To investigate the effect of risk factors on this type of outcomes, zerotruncated BVP regression is the appropriate model.
Campbell [1] introduced BVP distribution.Various assumptions have been used to develop BVP distribution.The most comprehensive one has been proposed by Kocherlakota and Kocherlakota [2].Leiter and Hamdan [3] suggested bivariate probability models applicable to traffic accidents and fatalities.A similar problem was addressed by Cacoullos and Papageorgiou [4].Several other attempts were made to define and study the BVP distribution [5]- [9].Jung and Winkelmann [10] showed bivariate Poisson form using a trivariate reduction method allowing for correlation between the variables, which is considered as a nuisance parameter.This bivariate Poisson regression is used by others [11] [12].Islam and Chowdhury [13] suggested covariate dependent BVP model using generalized linear modeling approach based on Leiter and Hamdan [3] bivariate probability models.They used marginal and conditional models to obtain BVP model.
Studies on the covariate dependent zero-truncated BVP model are scarce.Different techniques of the parameter estimation of BVP distribution are presented in [14]- [16].A unified treatment of three types of zero-truncated BVP discrete distribution based on probability generating function is shown elsewhere [17].Properties of BVP distribution truncated from below at an arbitrary point were studied by others [18] [19].At this backdrop, we proposed a zero-truncated covariate dependent BVP model based on the work of Islam and Chowdhury [13].The exposition of the following sections of the paper is as follows.Firstly in Section 2, we present briefly the marginal, conditional and BVP distribution for two outcomes without zero truncation as shown in [13].In Section 3, we have shown the zero-truncated marginal and conditional Poisson distribution and obtained the joint model for both outcomes zero-truncated.The estimation and the related procedures are also shown.In Section 4, applications of the proposed models are illustrated using road safety data for both outcomes zero-truncated published by the Department for Transport, United Kingdom.Finally, concluding remarks can be found in Section 5.

Poisson Distribution without Zero Truncation
In this section bivariate Poisson model without zero truncation is shown.For simplicity, we shall follow the notations used in [13].Let Y 1 be the number of accidents at a specific location in a given interval that has a Poisson distribution with density ( ) ( ) and the corresponding link function is , where 1, , , , , , , .
Y 's are assumed to be mutually independent, then the conditional distribution of
=   Then following [13] the joint distribution of number of accidents and number of fatalities can be shown as

Zero-Truncated Poisson Distribution
The probability of 1 0 Y = is 1 e λ − , using Equation (1).Hence Y 1 is observed conditional on Y 1 > 0. Thus, we have the conditional probability mass function ) Now, using Equation (1) the zero-truncated Poisson probability mass function for Then the exponential form of the mass function is The mean and variance can be shown as Similarly, the zero-

P Y y y P Y y y P Y y y Y P Y y P Y y
Then the zero-truncated conditional Poisson distribution is ( The exponential form of Equation ( 9) can be shown as Then the mean and variance are

Zero Truncated Bivariate Poisson (ZTBVP) Model
Now using the marginal and conditional distribution for zero truncation derived above the joint distribution of ZTBVP can be obtained as follows y y y y g y y g y y g y y y The ZTBVP expression in Equation ( 12) can be expressed in bivariate exponential form as where the link functions are ! ln e 1 ln ln !ln e 1 .
x ij The estimating equations are ln e e 0, 0,1, , ; e 1 and e e ln 0, 0,1, , .e 1 Then the score vector is ( ) The second derivatives are: ( 2 e e e 1 1 1 e e e e ln e , , x ij The observed information matrix is ( ) and the approximate variance-covariance matrix for ( ) The estimates of the regression parameters vectors 1 β and 2 β can be obtained iteratively by using Newton-Raphson method as follows ( ) ( ) where ˆt β denotes the estimate at t-th iteration.

Test for Significance of Parameters
We can use the likelihood ratio tests for testing 0 1 : 0 H β = and 2 0 β = model fit using full model and re- duced model.The test statistic is asymptotically chi-square as follows 2 log likelihood reduced model log likelihood full model .
For independence, we can test the equality of zero-truncated bivariate models under independence.The inde-pendence model can be shown as , g y y g y g y = × .

Deviance and Goodness of Fit
The deviance measures the difference in log-likelihood based on observed and fitted values.Let ; , where ( ) .,. l represents log-likelihood functions, as follows: and ; ln ln e 1 ln !ln ln ln e 1 ln ! .
After some algebra we get the deviance as We can use following test for goodness-of-fit proposed by Islam and Chowdhury (2015). where,

Test for Over or Underdispersion
The presence of overdispersion or underdispersion may influence the standard error of parameter estimates, hence, the significance level of the estimates.Test for the goodness of fit as shown in Equation ( 26) is modified to test the overdispersion or underdispersion.The method of moments estimator suggested by [20] is used to estimate the dispersion parameter, r φ , as shown below Using the mean, variance and correction factor as shown in [21] for truncated marginal and conditional Poisson models for and then using these values we can estimate r φ .
Then the test for dispersion 2 T for zero-truncated bivariate Poisson regression model is: where,

Application
The models proposed in the paper are illustrated using the road safety data published by Department for Transport, United Kingdom.This data set is publicly available for download from UK givernment website (http://data.gov.uk/dataset/road-accidents-safety-data).The data set includes information about the conditions of personal injury road accidents in Great Britain and the consequential casualties on public roads.Background information about vehicle types, location, road conditions, drivers demographics are also available among others.A total of 1,494,275 accident records were in the data set spanning from 2005 to 2013.We have selected a random sample 14005 accident records approximately 1 percent of all accident records.The outcome variables considered are total number of vehicles involved in the accident (Y 1 ) and the number of casualties (Y 2 ).Due to small frequencies, values five or more were coded as five for both outcomes.Risk factors are sex of the driver (0 = female; 1 = male), area (0 = urban; 1 = rural), two dummy variables for accident severity (fatal severity = 1, else 0; serious severity = 1, else = 0; slight severity is the reference category), light condition (daylight = 1; others = 0) and eight dummy variables for year 2006 to year 2013, where year 2005 is considered as reference category.
The average number of vehicles involved in accident and casualties are 1.83 and 1.37, with standard deviations 0.75 and 0.92, respectively.Table 1 displays the bivariate distribution of the number of vehicles and number of casualties.It is evident that 59 percent of the accidents involved two cars, 30 percent single car, and eight percent three cars.The number of casualties was one in three-fourth of the cases and two in one out of six cases.Descriptive statistics of the number of vehicles involved in accidents and number of casualties by risk factors are presented in Table 2.The mean number of vehicles with fatal injuries was 1.94 compared to 1.70 and 1.85 with serious and slight injuries.The mean number of casualties was 2.15 for fatal cases which appears to be much higher than that of serious and slight injuries.There is not much variation in mean number of vehicles and casualties by sex of driver and area.Although the number of vehicles involved in the accident is higher during daylight, number of casualties appear to be higher during other times.The number of vehicles involved in accidents decreased steadily during the study period, but mean number of cars involved in accidents and casualties remained almost similar.We observe that both numbers of vehicles involved in accidents and number of casualties are heavily underdispersed as displayed in Table 4.In Table 3, the estimates of the parameters are displayed along with standard errors and p-values for both original models as well as for adjustments made for underdispersion.Summary measures of goodness of fit for all the models are summarized in Table 4.The proposed full model of ZTBVP (Table 3) shows a negative association between fatal and serious severity and number of cars involved in accidents, while there is a positive association (p-value < 0.01) between the number of cars involved in an accident and light condition (daytime driving).The number of cars involved in accidents appears to be negatively associated in years 2008-2010 and 2012 as compared to that of 2005.However, the conditional model for the number of casualties given the number of cars involved in an accidents reveals that male drivers compared to females, rural areas compared to urban and daytime compared to night have lower risks.On the other hand, fatal severity and serious severity are positively associated with the number of casualties for given number of accidents compared to light severity.It is also evident that compared to the reference year, 2005, the number of casualties is negatively associated with the years 2012 and 2013.This indicates a significant reduction in the number of casualties for given number of accidents in recent years as compared to that of 2005.The summary results of estimation and tests of different models (proposed model based on marginal-conditional approach and both marginal models) are presented in Table 4.Both the full model and the reduced model under null hypothesis are considered.Both the models indicate that the full models are statistically significant.It is noteworthy that both the outcome variables number of vehicles involved in accidents and number of casualties are substantially underdispersed and adjustments were made accordingly for underdispersion in Table 3.Based on AIC, BIC and deviance we observe that the proposed full model using marginal-conditional approach provides the best fit.The goodness of fit test using the test statistic, T 1 , indicates good fit marginally (p-value = 0.064) for the proposed model.The test for under dispersion reveals the presence of significant deviation from equidispersion in both the variables as observed from T 2 (p-value < 0.001).Adjustments are made for underdispersion and the results are shown in Table 3 (last two columns).
fatalities recorded among the Y 1 accidents occurring in the jt-h time interval is Poisson with pa- of interest as shown before (Section 3.1) and 1i y and 2i y are the observed values under the saturated model.The deviance for zero-truncated bivariate Poisson, (

Table 1 .
Number of vehicles involved in the accident (Y 1 ) and number of casualties (Y 2 ).

Table 2 .
Descriptive statistics of the number of vehicles involved in the accident and the number of casualties by risk factors.

Table 3 .
Parameter estimates of zero truncated BVP model.

Table 4 .
Islam and Chowdhury (2015)r reduced and full models of ZTBVP.zero-truncatedbivariategeneralized linear model for count data is proposed in this paper.This model is based on the bivariate model using marginal-conditional models proposed byIslam and Chowdhury (2015)for count data.Covariate dependent bivariate generalized linear model is shown, and canonical link functions are used to estimate the parameters of the Poisson distribution.The usefulness of the proposed model is demonstrated using road safety data published by Department for Transport, United Kingdom.The proposed ZTBVP model can easily accommodate a varying number of covariates for two outcomes.The joint distribution degenerates into a marginal and conditional distribution that makes estimation problem easier. A