Obesity and mortality: challenging the conventional wisdom, part II: Categorisation is fatal

by Alan Cohen

This post is the second in a series critiquing the much-publicized recent study by Flegal et al. in JAMA, which claims that there is a lower risk of mortality among overweight people than normal weight people. As in my last post, my main beef is that our standard categories of BMI are a problem, but here I argue against categorizing altogether. In fact, I think there is a good argument that researchers categorizing BMI is literally fatal for many people.

How so? Well, once we put someone in a bin such as “obese” or “normal,” we have a mental tendency to accept the categories as something real, something objective. But they are not. These cut-offs are arbitrary, chosen mostly because they are nice round numbers. This can be seen easily on a histogram of 53,000 BMIs taken from the NHANES study (American adults age 20+, roughly 1998-2010):

NHANES BMI distThis figure makes it clear that there is no real substantive difference between someone with a BMI of 24.9 and someone with 25.1 – at least no more than there is between 24.7 and 24.9. Yet a person with a BMI of 25.1 is classified as “overweight” and may be counseled on what diets or exercise regimes to follow, whereas someone with a BMI of 24.9 will be told they are perfectly healthy and sent home. (True, some physicians will take into account how close one is to the threshold, but not always or perfectly.)

So why are categories used? Basically, for medical tradition and simplicity. Doctors have always liked categories: you either have a disease or you don’t. This is why psychologists tie themselves in knots with each new version of the DSM (Diagnostic and Statistical Manual) trying to precisely define what constitutes ADHD or schizophrenia or Asperger’s syndrome, all conditions that fall on a continuum. In addition, this is often practical: doctors have to make a decision to prescribe a treatment or not, and they can’t always choose to halfway prescribe the treatment, so they need some clear guideline for when to make such a yes-no decision. As a result, researchers who publish in medical journals are essentially obliged to divvy things up into arbitrary groupings, even when the underlying reality provides no support for these categories.

From a traditional statistical perspective, this means a little bit of information is lost. When the analyses are conducted, we have removed the fact that an individual had a BMI of 32.4 and have simply classified her as obese (grade 1). This loss of information means that we have less statistical power, and therefore that we have less chance to detect a real association between variables. It’s not ideal, but it may not be the end of the world, especially if we nonetheless did detect an association. In fact, we may underestimate how certain we should be about such associations as a result of using categories.

However, this traditional perspective misses some additional risks of using categories, and the Flegal et al study is a perfect example of why. To explain where the problem comes from, it is first necessary to describe how BMI actually associates with mortality risk, something I think can actually be established relatively well from previous studies, with a few caveats. I will mostly skip the caveats for now for the sake of simplicity, except to say that the same BMI can mean different things in different individuals depending on their genetic background, body type, sex, and other factors.

What we know about changing mortality risk with BMI and age

In general, mortality risk is high with very high BMI or with very low BMI, with some sort of optimal level of minimal risk. The risk will increase more quickly as BMI gets too low than when it gets too high, so that the final risk curve will look something like this:

BMI -mort risk 2dWhy is this? Well, we all know that being obese can increase the risk of various diseases (heart disease, diabetes, etc.). But if someone is too skinny, there is also risk: risk that the person is skinny because she has cancer, or because he is anorexic. A very skinny person also has less energy stored, and thus may be less able to survive a severe disease such as cancer or pneumonia. So it makes sense that the minimal risk is at an intermediate weight. But the curve cannot be symmetrical: it is possible to have a BMI of 60, which is 35 higher than a typical BMI of around 25, but it is not possible to have a BMI of -10, which is 35 lower than 25. In fact, based on some quick calculations, it’s probably not really possible to have a BMI much below 10 and still be alive: for me at 6 feet tall (184 cm), a BMI of 10 would imply weighing 73 lbs (33 kg). So the curve must be asymmetrical: risk must increase faster at the low end than at the high end, even if there are not many people with very low BMIs. The exact functional form of the curve is not necessarily known; here I have used a quadratic, which should be close enough, though I suspect that there is less difference in risk close to the optimal BMI than shown here.

So it’s pretty clear, both for theoretical reasons and based on the many studies available, that the general curve follows the figure above. But there is one other factor which is clearly important, and which has been ignored in most studies: all of this depends heavily on age. A 90-year old who has a low BMI (17, let’s say) is at very high risk; a 30-year old with a similarly low BMI is at much lower risk, even relative to the minimum-risk BMI for their age group. Being underweight in the elderly is often a sign of cancer, wasting, frailty, and other concerns that tend to precede death. A 30-year old with a low BMI may be anorexic, or may just have a skinny body-type; neither of these factors is very likely to lead to mortality. On the other hand, a very obese 30-year old may have a low absolute risk of mortality, but the risks from cardiovascular disease and diabetes may make the person orders of magnitude more at risk than a normal-weight 30 year old; having the same amount of excess weight in old age may make less difference, both because the risk is higher for normal-weight people as well, and because the extra fat reserves can be useful in some cases.

In fact, we can clearly see that people do gain weight with age (as we all know). Here is the same data from above, broken down into two age groups:

NHANES BMI dist by ageThere are several important differences between the distributions of BMI in the younger and older groups here. First, in the younger group the peak is in the normal category, whereas in the older group it is in overweight, and there are many fewer people at the low end of normal. Second, in the older group there are fewer people in the underweight category. And third, in the older group there are fewer people who are very obese (BMI of 45+), probably because the very obese people do not survive to age 60.

Having traveled to many countries, I’ve noticed that the tendency for people to gain weight with age is not just in the US, where it might be considered a sign of unhealthy eating habits, but also in places like Japan and Italy. Here, for example, are the BMIs of a sample of older Italians:

inchianti BMI histoMy interpretation of this is that the ideal weight for good health changes with age, and there is substantial data now to support this: in the elderly, you are at lower risk if you weigh a bit more. Many of the guidelines, however, are based on weights for young people. In fact, we should expect the curves of risk to be different in young and old people, something like this:

BMI -mort risk 2d by ageAs some of you may have noticed, I have been categorizing people by age (young versus old) despite complaining about categorization in principle. This is because I think we are accustomed to think this way, and so it is easier to understand my point with these categories. But we can just as easily make age a continuous variable (as it should truly be considered) and represent risk of mortality by BMI and age together like this:

bmi-age-mort risk relThis figure shows mortality risk by BMI and age, relative to the minimum possible risk for a given age (the white band at the bottom). The minimum-risk BMI increases with age. Relative to this minimum, the risk of being underweight increases sharply, and the risk of being overweight declines. Of course, we can also look at absolute mortality risk (represented here on a log-scale for easier viewing):

bmi-age-mort risk absIn this case, we see that mortality increases with age quite generally, as we know to be the case. This is true for people at all BMIs. However, the absolute risk increases more slowly for people in higher weight classes, and increases more quickly for those in lower weight classes.

I want to emphasize here that these curves and surfaces are hypothetical: they are based strongly on what we know, but not on actual measurements. The precise shape could vary, but the general principles are almost certainly correct. (The histograms above are actual measurements.)

Why this changes our interpretation of Flegal et al.

OK, that was a long introduction to understanding how mortality risk changes with BMI and age. (We could but won’t add in other factors such as sex, genetic background, and diet.) If we accept this summary as the underlying reality, why would we observe that mortality risk is lowest in overweight individuals, as reported by Flegal et al.?

The simplest answer is that we have poorly defined our weight categories, and that what we call “overweight” is not truly overweight, as I argued in my previous post. But the more subtle and important answer is that the old system of classification might be approximately correct for a person in her twenties, but is the wrong classification for an older individual.

Many of the studies used in Flegal et al. (which combined results of 140+ studies) were performed exclusively on older individuals, and the rest had mixed ages. Absolute mortality risk is much higher in older individuals, almost regardless of BMI, so the older individuals (and the studies on older individuals) will have a highly disproportionate influence on the results. This means that even though some younger individuals were included in the study, we can pretty much interpret the results as mortality risk by BMI in the elderly. And in this case, it is not at all surprising that the lowest-risk category is those we call “overweight”. It is already known that a bit of excess weight can be protective in the elderly, and that being underweight is a particular risk.

In other words, our traditional categories are not very meaningful because they depend too much on age and other factors.

Why categories are harmful

At the beginning, I said that categorizing BMI could actually be fatal. So far I have shown why it is not justified statistically or conceptually, but not why it is harmful. To understand the harm, think about what happens when a study like Flegal et al. comes out in one of the world’s premier medical journals. The media pick up on it, and inevitably the results are reported to the public without any of the nuances in the orginal study. The message is simple: mortality risk is lower for overweight people.

What this means is that many younger people will worry just a little less about being overweight. A 25-year old with a BMI of 24.3 may change to 24.7, for example. This change may not make much difference in his mortality risk at age 25 (which is low in any case), but will carry over to later ages, when he will be more at risk to gain weight beyond the optimal, increasing his risk of heart disease and diabetes 40 years later. Some individuals may not change their behavior, and few will make major changes, but many people making small changes could have a huge effect at the population level down the road.

At the same time, the fact that we have used these rigid (literally one-size-fits-all) categories for so long means that elderly people are assessing their weight by standards that should only apply to young people. Many elderly people who are at a healthy weight for their age may continue to try to lose weight because they perceive themselves as overweight based on the standard classification. In so doing, they risk disrupting their physiology and increasing their risk of mortality due to being underweight.

So what is the solution to all this? I would argue that focusing on weight is a red herring. Stop worrying about weight altogether, and worry about eating well. It is pretty hard for weight to be a problem for someone eating a healthy diet, but even someone with the “ideal” BMI can be profoundly unhealthy (and unaware of it) if she is eating poorly. By focusing on weight, we have distracted ourselves from the real problem, which is diet. Most people who eat a healthy, balanced diet will tend to gain weight slowly with age. This is natural and good, and helps keep mortality risk down at older ages.

What do you think? Has classifying people into weight categories been as harmful as I suggest? Leave your thoughts in the comments section…