The count regression model for panel data can be derived from the Poisson regression model. Consider the multiplicative one-way panel data model,
|
where
|
Here, are the individual effects.
In the fixed effects model, the are unknown parameters. The fixed effects model can be estimated by eliminating by conditioning on .
In the random effects model, the are independent and identically distributed (iid) random variables, in contrast to the fixed effects model. The random effects model can then be estimated by assuming a distribution for .
In the Poisson fixed effects model, conditional on and parameter , is iid Poisson distributed with parameter , and does not include an intercept. Then, the conditional joint density for the outcomes within the th panel is
|
|
|
|
|
|
Since is iid Poisson(), is the product of Poisson densities. Also, is Poisson(). Then,
|
|
|
|
|
|
|
|
|
|
|
|
Thus, the conditional log-likelihood function of the fixed effects Poisson model is given by
|
The gradient is
|
|
|
|
|
|
where
|
In the Poisson random effects model, conditional on and parameter , is iid Poisson distributed with parameter , and the individual effects, , are assumed to be iid random variables. The joint density for observations in all time periods for the th individual, , can be obtained after the density of is specified.
Let
|
so that and :
|
Let . Since is conditional on and parameter is iid Poisson(), the conditional joint probability for observations in all time periods for the th individual, , is the product of Poisson densities:
|
|
|
|
|
|
|
|
|
|
|
|
Then, the joint density for the th panel conditional on just the can be obtained by integrating out :
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
where is the overdispersion parameter. This is the density of the Poisson random effects model with gamma-distributed random effects. For this distribution, and ; that is, there is overdispersion.
Then the log-likelihood function is written as
|
|
|
|
|
|
|
|
|
The gradient is
|
|
|
|
|
|
|
|
|
and
|
|
|
|
|
|
where , and is the digamma function.
This section shows the derivation of a negative binomial model with fixed effects. Keep the assumptions of the Poisson-distributed dependent variable
|
But now let the Poisson parameter be random with gamma distribution and parameters ,
|
where one of the parameters is the exponentially affine function of independent variables . Use integration by parts to obtain the distribution of ,
|
|
|
|
|
|
which is a negative binomial distribution with parameters . Conditional joint distribution is given as
|
|
|
|
|
|
Hence, the conditional fixed-effects negative binomial log-likelihood is
|
|
|
|
|
|
The gradient is
|
|
|
|
|
|
|
|
|
This section describes the derivation of negative binomial model with random effects. Suppose
|
with the Poisson parameter distributed as gamma,
|
where its parameters are also random:
|
Assume that the distribution of a function of is beta with parameters :
|
Explicitly, the beta density with domain is
|
where is the beta function. Then, conditional joint distribution of dependent variables is
|
Integrating out the variable yields the following conditional distribution function:
|
|
|
|
|
|
|
|
|
Consequently, the conditional log-likelihood function for a negative binomial model with random effects is
|
|
|
|
|
|
|
|
|
The gradient is
|
|
|
|
|
|
|
|
|
and
|
|
|
|
|
|
and
|
|
|
|
|
|