St. James Ballroom
Don't Count on Poisson: COM-Poisson Models for Count Data (303821)
*Darcy Steeg Morris, US Census Bureau
Keywords: count data, Conway-Maxwell-Poisson distribution, data dispersion
Count data is prevalent in many fields of application including economics, demography and epidemiology. The Poisson distribution is the standard go-to distribution, however in practice count data very often exhibit variability inconsistent with the Poisson equi-dispersion assumption. The Conway-Maxwell- (COM-) Poisson distribution allows flexible modeling of count data with over- or under-dispersion. The COM-Poisson distribution is the basis of many extensions for flexible modeling of count data including COM-Poisson regression, a COM-Poisson mixed model, a bivariate COM-Poisson distribution and a COM-multinomial distribution. This work illustrates the utility and flexibility of this COM-Poisson class of modeling techniques through an example of count imputation in decennial census data.