One important property of Poisson distribution is that its mean and variance are equal, Var(yi xi j Þ¼E(yi xi j Þ¼ki. In fact, the Poisson distribution is parameterizedby a single scalar parameter ( ki) so that all moments of yi are a function of ki. In practice, the assumption of equidispersion may be violated for two main reasons. First, the frequency of zero counts is greater than the number of expected zeroes generated by the Poisson distribution. Second, the variance of observed counts data may exceed the mean due to unobserved heterogeneity. It is important to control overdispersion because, if it is large, it leads to grossly deflated standard errors and grossly inflated t-statistics in maximum likelihood estimation. For these reasons, many statistical tests are developed in order to detect overdispersion in data. Cameron and Trivedi [4] set out a test for overdispersion based on a linear regression without the intercept. The test is designed so as to choose one of the following null and alternative hypotheses: H0 : VarðyiÞ¼ki H1 : VarðyiÞ¼ki þagðkiÞ where a is an unknown parameter and g() is a definite function, most commonly gðkiÞ¼k2 i or gðkiÞ¼ki. This test can be computed by estimating the Poisson model, constructing the fitted value of ^ ki ¼ expðx0 i ^ bÞ and running the OLS regression without the intercept: ðyi ^ kiÞ2 yi ^ ki ¼a gð^ kiÞ ^ ki þei
where ei is an error term. The significance of the coefficient a in the OLS regression model implies the existence of overdispersion in data. Note that this test can also be used for detecting under-dispersion which is the situation where conditional variance is less than conditional mean. Another test for overdispersion, introduced by Greene [8], is based on the Lagrange multiplier (LM) statistics. The LM statistic is defined by:
LM ¼
P n i¼1
^ wi½ð^ yi ^ kiÞ2 yi ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2P n i¼1 ^ wi^ k2 i s
2 6 6 6 6 4
3 7 7 7 7 5 2
where ^ wi define the weight of the alternative distribution. When the alternative is a Negative binomial negative, the weights equals to 1. In this case, the LM is given by:LM ¼ðe0enyÞ2 2^ k0^ k Note that the limiting distribution of the LM statistic is Chi-Squared with one degree of freedom [8].
One important property of Poisson distribution is that its mean and variance are equal, Var(yi xi j Þ¼E(yi xi j Þ¼ki. In fact, the Poisson distribution is parameterizedby a single scalar parameter ( ki) so that all moments of yi are a function of ki. In practice, the assumption of equidispersion may be violated for two main reasons. First, the frequency of zero counts is greater than the number of expected zeroes generated by the Poisson distribution. Second, the variance of observed counts data may exceed the mean due to unobserved heterogeneity. It is important to control overdispersion because, if it is large, it leads to grossly deflated standard errors and grossly inflated t-statistics in maximum likelihood estimation. For these reasons, many statistical tests are developed in order to detect overdispersion in data. Cameron and Trivedi [4] set out a test for overdispersion based on a linear regression without the intercept. The test is designed so as to choose one of the following null and alternative hypotheses: H0 : VarðyiÞ¼ki H1 : VarðyiÞ¼ki þagðkiÞ where a is an unknown parameter and g() is a definite function, most commonly gðkiÞ¼k2 i or gðkiÞ¼ki. This test can be computed by estimating the Poisson model, constructing the fitted value of ^ ki ¼ expðx0 i ^ bÞ and running the OLS regression without the intercept: ðyi ^ kiÞ2 yi ^ ki ¼a gð^ kiÞ ^ ki þeiwhere ei is an error term. The significance of the coefficient a in the OLS regression model implies the existence of overdispersion in data. Note that this test can also be used for detecting under-dispersion which is the situation where conditional variance is less than conditional mean. Another test for overdispersion, introduced by Greene [8], is based on the Lagrange multiplier (LM) statistics. The LM statistic is defined by:LM ¼P n i¼1^ wi½ð^ yi ^ kiÞ2 yi ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2P n i¼1 ^ wi^ k2 i s2 6 6 6 6 43 7 7 7 7 5 2where ^ wi define the weight of the alternative distribution. When the alternative is a Negative binomial negative, the weights equals to 1. In this case, the LM is given by:LM ¼ðe0enyÞ2 2^ k0^ k Note that the limiting distribution of the LM statistic is Chi-Squared with one degree of freedom [8].
การแปล กรุณารอสักครู่..
