Introduction
Poisson regression is similar to multinomial logistic regression in that the dependent variable can take only non-negative integer values. With this regression method, the dependent variable takes values 0, 1, …, r for some known value of r, while with Poisson regression there is no predetermined r value, i.e. any count value is possible.
In Poisson regression, we assume that we have k independent variables x1, x2, …, xk. We further assume that we have a random sample of size n, whose ith element is xi1, xi2, …, xik. Corresponding to each k-tuple in the sample is an element yi which is taken from a population that has a Poisson distribution with mean μi. Thus, for each i, the probability of yi events occurring is
Since the mean and variance of a Poisson distribution are equal, for each i we also have
Regression Model
We could now study the linear regression equations
and use an ordinary least squares approach, but this would not assure that the values of μi are non-negative. Instead, we use
where exp(x) is the link function. This is equivalent to
We can assume that x1 = 1 (thus, xi1 = 1 for all i) so that we can have a constant term as for other regression models.
For any values of x2, x3, …, xk, we can use the model to predict the risk μ of a rare event occurring in a specified unit of time or space as follows
where b1, b2, …, bk are estimates of the regression coefficients and x1 = 1.
Log-Likelihood Function
The log-likelihood function is
where LL is considered to be a function of the bi (which are used to define the μi) and the xij and yi are considered to be fixed.
Predictions
For each observation, we can also explicitly specify the exposure time or space ti, in which case
Thus, for any values of x2, x3, …, xk, we can use the model to predict the number of occurrences μ of a rare event in time (or space) t by
References
Hintze, J. L. (2007) Poisson regression. NCSS
https://www.ncss.com/wp-content/themes/ncss/pdf/Procedures/NCSS/Poisson_Regression.pdf
Nussbaum, E. M., Elsadat, S., Khago, A. H. (2007) Best practices in evaluating count data, Chapter 21: Poisson regression.
http://www.academia.edu/438746
Penn State (2017) Poisson regression. STAT 504: Analysis of discrete data.
https://online.stat.psu.edu/stat504/lesson/9