Bias in diet assessment methods - Consequences of collinearity and measurement errors on power and observed relative risks

Abstract

Background. If several risk factors for disease are considered in a regression model and these factors are affected by measurement errors, the observed relative risk will be attenuated. In nutritional epidemiology, 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 collinearity between variables in the model. Methods. The validity of different diet assessment methods are compared. The correlation coefficients between common nutrients and foods are given using data from the Malmo Food Study. Intake of nutrients and foods were assessed with a modified diet history method, combining a 2-week food record for beverages and lunch/dinner meals and a food frequency questionnaire for other foods. The study population comprised 165 men and women aged 50-65 years. A multivariate logistic regression model is used to illustrate the effect of collinearity on observed relative risk (RRo). Results. A moderate to high correlation between risk factors will substantially influence RRo even when using diet assessment methods with high validity. Methods 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 highly intercorrelated, since RRo might be severely biased.