*Geek Box: Statistical Adjustment
Reading research, you’ll come across the term ’adjusted’ or ‘controlled’, i.e., a relationship between exposure A and outcome B was significant, but then after adjusting for third variable C, the relationship was no longer significant. In this example, C is statistically ‘adjusted’ in order to have a more true estimate of the effects of the exposure on the outcome. It is a means to control variables, other factors, that might influence the results unless they were accounted for. Different statistical tests can be used to achieve this.
Let’s say, for example, that there is a relationship observed between coffee consumption and heart disease. However, in the data on the participants, it also appears that coffee consumption related to smoking rates. The researchers then adjust for smoking, and the relationship between coffee consumption and heart disease is no longer evident statistically. This indicates that smoking was acting as a confounder in the relationship observed, and implicates smoking as a more likely factor related to the outcome.
However, over-adjustment is also possible, and this can weaken a true relationship. Over-adjustment occurs when the variable adjusted is in fact part of a chain of causation between the exposure and outcome. In our previous example, coffee and smoking may correlate as a behaviour, but coffee consumption is not the cause of smoking. However, let’s now say we’re looking at the relationship between saturated fat and heart disease. We know that the main factor saturated fat influences is LDL-cholesterol levels, and increased LDL-cholesterol causes atherosclerosis. This is a causal chain: saturated fat>LDL>heart disease. If we adjust for LDL in this scenario, we remove a variable in the causal chain, and there is no longer an association evident. If LDL was not adjusted for, the relationship would be stronger. This is where over-adjustment can obscure true relationships.