Detecting Confounding and Evaluating the 10% Rule

• 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): Robin Bliss*+ and Janice Weinberg and Verónica Vieira and Thomas Webster
• Companies: Brigham and Women's Hospital and Boston University School of Public Health and Boston University School of Public Health and Boston University School of Public Health
• Keywords: Confounding ; Logistic regression ; 10% Rule

Abstract

In epidemiologic studies researchers are often interested in detecting confounding, i.e. when a third variable is associated with both the outcome and predictor of interest and affects the association between the predictor and outcome. The "10% Rule" is a rule of thumb where confounding is detected if the observed effect sizes in regression models differ by a relative magnitude of at least 10%. In this study we evaluated the performance of the 10% Rule based on odds ratios from a logistic model. We simulated data with a dichotomous outcome and confounder. The predictor of interest was either dichotomous or continuous. We performed logistic regression models including only the predictor and models also adjusting for the confounder. We applied the 10% Rule and found that when confounding existed, over 80 and 60% of models had >10% changes in odds ratios for dichotomous and continuous predictors, respectively. When the confounder was not associated with the outcome, the false positive rate increased as the strength of the association between the predictor and confounder increased. When the confounder and predictor were independent of one another, false positives were rare (most < 10%).