*Geek Box: Adjustment Models
Reading nutritional epidemiology, you will continually encounter a long list of variables [generally referred to as ‘covariates’] that the investigators adjusted for. In statistics, the term ‘adjustment’ means to control, i.e., remove from the equation, a variable that might influence the association between the exposure of interest and the outcome of interest. By adjusting for these factors, it provides a more direct estimate of the effects of the exposure on the outcome.
The term ‘model’ refers to the particular set of variables which have been added to the analysis. There may be different combinations of variables included. Every specific group of variables that is adjusted for is referred to as a model. Basic models tend to adjust for factors like age, gender, ethnicity, BMI, or whichever combinations of these basic factors the authors deem relevant. When multiple variables are included in a model, this is known as ‘multivariate analysis’ or a ‘multivariate model’.
In nutritional epidemiology, there are two levels of variables we want to account for in a multivariate analysis: potential confounding by lifestyle factors [smoking, alcohol, BMI], and potential confounding by correlated dietary factors [other nutrients]. Dietary factors can be analysed by using a substitution analysis: this is where the effects of replacing one nutrient with another are modelled, for example, the effect of replacing saturated fat with polyunsaturated fat. A single study might have three models; an ‘Unadjusted Model’, where the results are displayed before adjusting for any variables; a ‘Basic Model’ generally adjusts for typical lifestyle factors, like smoking, alcohol, and/or BMI; and a ‘Fully Adjusted Model’, where all variables that the investigators have deemed important to control for are added.