Statistical Power of Relative Risk in Poisson Regression: How Important Is Exposure Data?

• Activity Number: 88
• Type: Contributed
• Date/Time: Sunday, July 31, 2011 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Jake Olivier*+ and Joseph Descallar
• Companies: Prince of Wales Clinical School and University of New South Wales
• Keywords: Relative Risk ; Poisson Regression ; Monte Carlo simulation ; Exposure

Abstract

Estimates of risk rely on the population-at-risk (PAR). However, this information is often not used due to lack of data. Instead, census data of the population is used as a proxy for exposure. This is inappropriate as it assumes equal exposure across the population and results in biased risk estimates. What is less understood is its effect on relevant statistics. That is, how does the use of inappropriate population based denominators (PBD) affect our ability to identify true differences in risk between groups? To this end, Monte Carlo simulation is used to estimate the power of detecting a significant relative risk (RR) from a Poisson regression using either PBD or PAR based offsets. The results show that when using PBDs, the power of detecting differences between two groups, where the lower risk group has higher exposure than the other, is considerably reduced. Conversely, when the group with lower exposure also has lower risk, then the power of detecting these risk differences is higher when using PBDs; however, the estimates of RR are largely inflated. When RR=1, the Type I Error increases rapidly as the difference in exposure increases when using PBDs.