7 Concluding remarks
In this paper, we have introduced two models usually fitted to counts data: Poisson regression and Zero-inflated Poisson regression. Maximum likelihood techniques are used to estimate the parameters of both models. Since the Hessian matrix associated to the Poisson regression is negative, the Newton–Raphson iterative algorithm converges rapidly and provides unique parameters estimates. The EM algorithm, used to maximize the likelihood of Zero-inflated Poisson regression, proceeds iteratively via three steps: E step, M step for truncated Poisson model and M step for binomial model. The EM algorithm estimates the expectations of missing data and iterates until convergence. Given a significance level (e.g. 5 %) Wald test, Lagrange Multiplier test or Likelihood Ratio test are used to measure the significance of factors incorporated in each model. The equidispersion assumption of Poisson regression was tested using LM test introduced by [8] and regression test set by [4]. These tests have proved the overdispersion of the number of claims in a Moroccan private health insurance scheme. According to histogram, which is highly peaked at zero, we state that this overdispersion is due to the preponderance of zeroes in the population. In such case, we have shown that standard Poisson model is unable to reproduce the number of zeroes in the data and therefore, underestimates the dispersion of the population. Alternatively, we have fitted Zero-Inflated Poisson model. We have shown that this model simulates well the data and the number of zeroes reproduced by it is very close to the number of zeroes in the population. Finally, we have computed Vuong’s test and the probability integral transforms for selecting the best model in the case of excess of zeroes. We can conclude that Zero-inflated Poisson regression fits excess of zeroes counts data better than standard Poisson regression.
หมายเหตุ Concluding 7In this paper, we have introduced two models usually fitted to counts data: Poisson regression and Zero-inflated Poisson regression. Maximum likelihood techniques are used to estimate the parameters of both models. Since the Hessian matrix associated to the Poisson regression is negative, the Newton–Raphson iterative algorithm converges rapidly and provides unique parameters estimates. The EM algorithm, used to maximize the likelihood of Zero-inflated Poisson regression, proceeds iteratively via three steps: E step, M step for truncated Poisson model and M step for binomial model. The EM algorithm estimates the expectations of missing data and iterates until convergence. Given a significance level (e.g. 5 %) Wald test, Lagrange Multiplier test or Likelihood Ratio test are used to measure the significance of factors incorporated in each model. The equidispersion assumption of Poisson regression was tested using LM test introduced by [8] and regression test set by [4]. These tests have proved the overdispersion of the number of claims in a Moroccan private health insurance scheme. According to histogram, which is highly peaked at zero, we state that this overdispersion is due to the preponderance of zeroes in the population. In such case, we have shown that standard Poisson model is unable to reproduce the number of zeroes in the data and therefore, underestimates the dispersion of the population. Alternatively, we have fitted Zero-Inflated Poisson model. We have shown that this model simulates well the data and the number of zeroes reproduced by it is very close to the number of zeroes in the population. Finally, we have computed Vuong’s test and the probability integral transforms for selecting the best model in the case of excess of zeroes. We can conclude that Zero-inflated Poisson regression fits excess of zeroes counts data better than standard Poisson regression.
การแปล กรุณารอสักครู่..