Cheesecake makes you fat, but correlation is not causation

By Peter Dearden 26/03/2014

Julia Horsfield

I was one of the happy people rejoicing in new gastronomic possibilities after hearing that eating saturated fats may not cause heart disease after all.

Yay! I never could bring myself to opt for that trim latte. Maybe I can even ditch the Olivani in favour of butter. But, as my nutritional friends point out, it’s all how you look at the data. Unfortunately, we mustn’t get too excited, and the best dietary advice is still to stick with omega-3 polyunsaturated fats as part of a balanced diet.

Fair enough, it’s hard not to agree. But I’m an experimental biologist, inordinately interested in the nuts and bolts of how things work. I like cause-and-effect, or as my sons would tell you, ‘consequences’. Stuff you can measure and be unequivocally convinced by. I’ve always been faintly uncomfortable with the nature of the research that encourages folk to burst forth with dietary advice.

In a nutshell, dietary studies look at what people eat, and then determine what happens to them, health-wise, over a period of time. Given enough people, enough time, and some fancy statistics, the studies often conclude that what the people were eating was causative of their health outcomes.

This more than certainly true for simple case studies. For example, eating half a cheesecake a day will make you fatter, worse luck. However, it seems that when the dietary input being measured is just one factor among many (like the type of fat), then the outcome may not be simple. The saturated fat debate outlined above is a good example of where different kinds of analyses show different outcomes, even using the same datasets.

How to establish cause-and-effect? Good old-fashioned biochemistry can sometimes determine the how different foods affect our bodies. Eat enough carrots and you will actually turn orange because of persistence of the carotene pigment (however, I’m told you have to eat an awful lot of carrots to mimic a spray-tan). But eating fat (whether saturated or unsaturated) in large enough amounts to noticeably affect your biochemistry is probably going to be very, very bad for you regardless. Biochemistry is in itself a complex thing, and may not establish causation.

Perhaps the most convincing cause-and-effect relationship with obesity in humans comes from genetics. It is true: our genetic makeup determines our body weight to some extent. Geneticists have shown, using large populations of people, that small variations in DNA sequence are associated with obesity. That is to say, people with a particular version of a DNA sequence are more inclined to be overweight than people who do not have that particular version. The stats are very convincing, and often the variations are within or near genes that could affect human bodyweight. But, like the dietary studies, this is still correlation. Surely, since we’re talking about genes here, and genes actually do stuff, correlation should equate to causation, right?

Not necessarily. DNA sequence variants near a gene called FTO are significantly associated with obesity. People with the risky version of the DNA sequence were on average 3 kg heavier than those with the non-risk version. So, scientists naturally assumed that FTO, being the nearest gene to the genetic variants, was responsible for the increased weight in humans with the risk version. They went so far as to suggest FTO is an ‘obesity gene’ in humans. The only fly in the crème brulee was, there was no correlation between the actual amount of FTO in humans and their body weight. Therefore genetic correlation didn’t entirely equate to causation, in the case of FTO.

But there may be an answer, after all. A highly significant new study in Nature (20 March 2014) shows that the genetic variants do not affect the function of the nearest gene, FTO, but rather, one that’s almost a million base pairs away – IRX3.

The piece of DNA that harbours the genetic variants communicates over this immense distance to the IRX3 gene, completely bypassing the nearest candidate, FTO. This communication happens by the creation of a large DNA loop that brings the piece of variant DNA, (known as an ‘enhancer’, because its role is to enhance gene expression), into contact with the IRX3 gene. And the snuggling up of the enhancer with the IRX3 gene really does seem to change IRX3’s function.

And when the scientists looked at how the IRX3 gene works in humans, they found that the amount of protein made by the IRX3 gene significantly correlates with obesity – more IRX3 usually means you are fatter.

The scientists then turned to mice to prove that it is the function of the IRX3 gene, rather than FTO, that is responsible for the weight changes observed in human populations. Mice lacking the IRX3 gene were 30% thinner than their littermates, and didn’t put on weight when fed the mouse equivalent of cheesecake.

Therefore IRX3 is now the best smoking gun candidate for an obesity gene, even though the genetic changes in humans are actually much closer to the FTO gene.

Only this large collection of functional evidence could convincingly place IRX3 as an important gene associated with human obesity. The collective evidence included mapping of human genetic variants, enhancer function and DNA looping analysis, and mouse models (complete with fat analysis).

It’s tempting to infer cause and effect from all sorts of human studies, including dietary studies. But this excellent genetic example is a good reminder to all that correlation is not causation, and we still have to figure out how things really work before we can assign causation.


0 Responses to “Cheesecake makes you fat, but correlation is not causation”

  • Correlation is a very poor statistical tool to assess causation or even association, and there are many well established statistical tools for assessing whether an exposure and outcome are associated without resorting to “fancy statistics”.

    In trying to understand the role of foods or nutrients in health and disease, nutrition researchers and epidemiologists practice “evidence-based nutrition”. Evidence-based nutrition has developed in parallel with evidence-based medicine. Individual studies are evaluated and ranked according to the type and quality of the study. Ecological association studies and cross-sectional studies which examine correlations at one point in time between a nutritional exposure and a health outcome are considered the weakest level of evidence; they considered hypothesis-generating studies. Meta-analyses of randomised controlled intervention studies are the highest level of evidence; inferring causality. When making nutritional recommendations we look for convincing evidence of causality from high quality randomised controlled interventions and meta-analyses of these types of studies.

    The reviews referred to in our blog (see link above) are meta-analyses of both prospective cohort studies/quasi controlled studies (providing “probable” evidence of causality) and randomised controlled trials (providing “convincing” evidence of causality). These are not simple correlation studies. The trick to interpreting such studies is to understand the questions they attempt to answer. This is dependent on appropriate selection of studies which individually address the same question of interest.

    Nevertheless, genetic research has huge potential to contribute to our understanding of the effects of nutritional exposures on health outcomes and this will help to refine dietary recommendations for both individuals and populations. So geneticists please come and talk to us nutritionists about opportunities to collaborate.

  • hey there 🙂 thanks for your comment.
    I guess my point was ‘inferring causality’ is not the same as proving causality, which applies to genetics as much as anything else (as illustrated by FTO/IRX3). It’s not a field-specific thing. To *prove* causality (at least in my field), you need multiple lines of evidence, which usually means doing more than one *type* of study (ie functional work in addition to an RCT selection), and including mechanism. I am usually not able to publish my own work without providing a mechanism for HOW something affects something else. It’s not enough that changing something DOES affect something else, since the connecting points in between are unknown. Happy to collaborate any time!

  • I see your point now.

    In an applied science like nutrition understanding the the finer molecular details doesn’t often enhance our understanding of the roles of foods and nutrients in human health as much as we would hope. This is probably because of the incomprehensibly complex interactions between an individual’s genes, the environment, behaviour and their highly variable diets etc.

    However we can see macro level patterns and effects that show how different diets/nutrients influence the risk of developing disease, and we can show with a high degree of certainty in controlled intervention studies how nutrients, foods or even dietary patterns differentially affect disease risk biomarkers (and sometimes even disease states – though this is much more difficult to test).