Clustered data, multiple observations collected on the same experimental unit, is common in epidemiological studies. Bivariate outcome data is often the result of interest in two correlated response variables. An efficient method is presented for dealing with bivariate outcomes when one outcome is continuous and the other is a count using a simple transformation to handle over-dispersed Poisson data. A multilevel analysis was performed on data from the National Health Interview Survey (NHIS) with body mass index (BMI) and the number of cigarettes smoked per day (NCS) as responses. Results show that these random effects models yield misleading results in cases where the data is not transformed.
Atem, Folefac; Ngwa, Julius S.; and Adeniji, Abidemi
"Regression Models for Mixed Over-Dispersed Poisson and Continuous Clustered Data: Modeling BMI and Number of Cigarettes Smoked Per Day,"
Journal of Modern Applied Statistical Methods:
1, Article 19.
Available at: http://digitalcommons.wayne.edu/jmasm/vol11/iss1/19