Examining the Bias of the Difference in Coefficients Method for Evaluating Intermediate Endpoints Using Logistic Regression

• Activity Number: 71
• Type: Contributed
• Date/Time: Sunday, August 1, 2010 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Miranda Elaine Kroehl*+ and Lorraine Ogden
• Companies: University of Colorado, Denver and University of Colorado, Denver
• Keywords: Intermediate Endpoint ; Mediator ; Mediation ; Difference in Coefficients ; Product of Coefficients

Abstract

The evaluation of intermediate endpoints (IE) is becoming increasingly popular in both clinical and epidemiological research. Two common methods are often used for estimating the IE effect: the product of coefficients method and the difference in coefficients method. It has been noted that the point estimates from these two methods are not usually equal for binary outcome data, however, the magnitude of this discrepancy has not been well described. This paper states the conditions under which the difference in coefficient method is biased when using logistic regression. For instance, (1) the estimate is usually biased towards no effect when a true IE effect exists; (2) the magnitude of the bias increases with increasing strength of association between the IE and the outcome; and (3) the bias is negligible when the outcome is rare.