BIAS IN DIET ASSESSMENT METHODS - CONSEQUENCES OF COLLINEARITY AND MEASUREMENT ERRORS ON POWER AND OBSERVED RELATIVE RISKS

Background. if several risk factors for disease are considered in a re gression model and these factors are affected by measurement errors, t he observed relative risk will be attenuated. In nutritional epidemiol ogy, several nutrient variables show strong correlation, described as collinearity. The observed relative risk will then depend not only on the validity of the chosen diet assessment method but also on collinea rity between variables in the model. Methods. The validity of differen t diet assessment methods are compared. The correlation coefficients b etween common nutrients and foods are given using data from the Malmo Food Study. Intake of nutrients and foods were assessed with a modifie d diet history method, combining a 2-week food record for beverages an d lunch/dinner meals and a food frequency questionnaire for other food s. The study population comprised 165 men and women aged 50-65 years. A multivariate logistic regression model is used to illustrate the eff ect of collinearity on observed relative risk (RRo). Results. A modera te to high correlation between risk factors will substantially influen ce RRo even when using diet assessment methods with high validity. Met hods with low validity might even give inverse RRo. Conclusion. It is stressed that caution must be exercised and only a selected number of variables should be included in the model, especially when they are hi ghly intercorrelated, since RRo might be severely biased.