It is Time to Say “Good bye” to Poisson Regression: Application of Individual Level Data to Nuclear Worker Data

Thursday, 21 August 2014: 9:00 AM
Tubughnenq 4 (Dena'ina Center)
Yutaka Hamaoka , Keio University, Tokyo, Japan
INTRODUCTION:  Although individual level data are recorded, most of the radiation-epidemiological studies apply the Mantel-Haenszel score test or the Poisson regression model to tabulated data by age, sex, dose, and other covariates. This aggregation can lead to a loss of information, inefficient estimation, and weaker statistical power when detecting the risk of a low dose. 

METHODS:  US DOE nuclear worker data provided by CEDR project was re-analyzed. Multinomial logit model that explain death probabilities among some causes, such as solid cancer, leukemia, other cancer, non-cancer, external cause, and other cause was applied to the data with explanatory variables:age, sex, race, calendar year of first employment, age at first employment, site dummy, length of employment, latency dummy, and cumulative dose.

RESULTS:   Radiation cumulative dose is positive and significant for solid cancer (beta=1.70, p<0.05), other cancer (beta=2.22, p<0.05), and non-cancer (beta=2.50, p<0.05) and insignificant for leukemia (beta=-0.38, p>0.1). 

CONCLUSIONS:  For the same data, the Mantel-Haenszel score test failed to detect this relationship (Gilbert et al. 1993). Using the individual level model, a statistically significant effect of a radiation dose was detected. To detect low does effects, models that utilize individual data are more effective.