Absolute vs. relative
Sometimes one makes more sense than the other
Absolute vs. relative – most people understand these two inseparable concepts. In everyday life, it is usually clear when absolute judgments are required and when such judgments are not very helpful. I love this Economics joke: Mister A. greets Mister B. with “How do you do?” “Relative to what?” is the counter-question with which Mister B. answers. That is funny, or at least I find it so. Something as innocuous as a “How do you do?” should not require an anchor, like “Compared to a man on death row, or somebody who just won the lottery?” Some other comparisons, on the other hand, should be relative rather than absolute, otherwise they are not meaningful. Comparing the GDPs of Liechtenstein and China without considering the population of both countries is not very helpful, and a per capita adjustment is much more enlightening.
But back to absolute judgments. These sometimes concern you as an individual for maximum meaning, but are skillfully and deliberately short-circuited into relative verdicts, often for commerce. My hoarding dad comes to mind. He is a smart guy and good with numbers, but like everybody falling for a bargain, he is easily suckered into considerations of rate, i.e., per item price, when he really should be looking at the absolute total. “Ten for the price of 5” is not really a good bargain if you are in genuine need of zero.
We can take this example and translate it to epidemiological studies where multiple risk factors are tallied for large cohorts of people with a precise observation of the incidence of certain diseases. Now, some abstractions are often made to present relative statements and one can get lost quickly. First, often the outcome measure presented is relative risk, i.e., risk in relation to a reference group’s risk either at the top or bottom end of the spectrum of the risk factor or exposure of interest. Imagine the risk factor meat consumption and let us pick the bottom quintile of meat consumption as the reference group against which relative risk is judged. You might see a finding like the following: “the risk of disease X for the highest quintile of meat consumption was more than double than that of the lowest quintile with a relative risk ratio RR=2.1 and confidence interval for RR of [1.1, 5.3].” This is doubly relative, and can mislead in two ways: first, the computation of quintiles has removed all absolute information. You cannot easily work out how much actual meat the top and bottom quintile is consuming per week without going back to the text. Second, there is the relative risk itself. The baseline anchor for the lowest-quintile meat-eaters is important here too. A doubling of disease risk sounds alarming indeed. But doubling from what? A baseline risk of 0.01%? Or 10%? Obviously, the latter scenario will give much more cause for alarm than the former.
But there is another absolute vs. relative confusion here, this time cutting back in the other direction. The confusion is lurking in the way intelligent lay people are sometimes processing the reports of epidemiological studies with the best of intentions; it cannot be blamed on the authors of such studies, but it is helpful to keep in mind that these studies usually employ a variety of “nuisance variables” to make sure that the influence of the risk factor of interest is isolated appropriately. For instance, it could be that meat eaters are less healthy in general, for reasons that have nothing to do with meat eating per se. Maybe they smoke too, or are generally inactive; maybe they are not very health conscious in general, and do not use healthcare providers as often. To cleanly get at the pure effect of meat eating, everything else being equal, these nuisance variables play a critical role. To attribute a unique effect to meat consumption it needs to surpass what is already explained by the nuisance variables. This is obviously a very important step for knowledge gain and might be followed up with more detailed studies on the bench, looking at the mechanisms by which meat consumption influences disease risk. In clinical trials, one can ensure by design that possible confounders are uncorrelated with the intervention of interest. For observational studies, we can take the less kosher alternative of statistical adjustments.
However, the valid research perspective “Does meat consumption influence disease risk above all else?” is often not most relevant for deriving practical health advice. One might rather ask “What is the most effective thing I can do in attenuating disease risk?” This is inherently a relative question since it involves comparison of effect sizes across different risk factors. Ideally, we would minimize all risk factors, but real life is messy, and we only have limited potential for behavior modification. So, while cutting your meat consumption might be beneficial, maybe the resulting effect on disease risk is trivial compared to your giving up smoking and starting to walk half an hour every day?
To get a handle on such relative assessments with sufficient rigor, a lot more work is required involving the exact nuisance variables whose influence the authors took pains to remove: consulting the paper to obtain their regression weights and associated standard deviations is necessary, before doing some calculations. This is too involved for most people who might just consult an abstract of the study results. If the nuisance variables are “silent” then the corresponding regression weights might not be reported at all. Further, the most crucial nuisance variables might be missing altogether. So even with good stats chops, the most important relative assessments might not be possible.
There you have it. I have no real solution, but at least it raises your awareness, both for epidemiology and real life.