All others things being equal

Regression analysis is often interpreted as being the effect of changing one variable while holding all other factors constant. Sadly, when working with complex human behaviors, like in nutritional epidemiology, factors rarely stay constant. A change in one parameter can shift other parameters making inference difficult. A very good example of this was reviewed by Travis Saunders of Obesity Panacea.

Now obviously randomized experiments allow an estimate of the average causal effect of an intervention but they are expensive to run and these types of research questions raise grave issues of equiposie and adherence in the design of these experiments!