NIOSHTIC-2 Publications Search

Re: "Estimating the relative risk in cohort studies and clinical trials of common outcomes."

Deddens JA; Petersen MR

Am J Epidemiol 2004 Jan; 159(2):213-214

https://doi.org/10.1093/aje/kwh022

20031701

McNutt et al. (1) recently pointed out that Poisson regression can be used to estimate adjusted risk ratios from cumulative incidence cohort data. Although other investigators have made the same point (2-5), this use of Poisson regression is not widely known. In addition, McNutt et al. provided a useful comparison of several methods and gave valuable information about confidence intervals (1). In cohort studies, researchers sometimes match exposed and unexposed subjects on certain characteristics that may be potential confounders-studies of twins (6), for example. Data from such studies are sometimes analyzed using conditional logistic regression. With this method, investigators can estimate odds ratios using information only from matched sets with a subject who had the study outcome. However, if the outcome is common, the resulting odds ratios may not closely approximate risk ratios (7). Adjusted risk ratios, rather than odds ratios, can be estimated in matched cohort data using commercially available software. Information is needed only from the matched sets in which one or more members had the study outcome. One method is conditional Poisson regression (5, 8-10), which is implemented in Stata software (11). For matched-pair data, the Cox proportional hazards model can be used if follow-up time is the same for all subjects, the regression analysis is stratified on the pairs, and the Breslow or Efron method is used to account for the tied survival times; the likelihood is the same as the conditional Poisson likelihood (12, 13). (Breaking the tied survival times with the exact marginal or exact partial likelihoods will permit estimation of the odds ratio; the likelihood is the same as the conditional logistic likelihood.) My colleagues and I have reported that the 95 percent confidence intervals estimated with these methods cover approximately 98 percent of simulated risk ratio estimates (12), which is similar to the findings of McNutt et al. for unmatched data. Bootstrap methods can be used if more accurate confidence intervals are desired. These methods have been used in studies of traffic crashes, since persons involved in such crashes can be thought of as matched when they are in the same vehicle (14-16).

Mathematical-models; Risk-analysis; Risk-factors; Hazards-Confirmed; Drivers; Accident-analysis; Accident-statistics

AJEPAS

20040115

Other

2004

2

0002-9262

HELD

American Journal of Epidemiology

WV; NY; IL; MD