Today’s (16/02/2018) newspapers (e.g. The Guardian) and other media (e.g. BBC) carry headlines about eating “ultra-processed foods” being linked to higher cancer risk. These headlines are in response to a paper published in the BMJ by a group of mainly French scientists. In my opinion this paper and the media coverage it has generated epitomizes what is wrong with much of current nutrition research. It uses weak associations from dubious and flawed data to reinforce existing prejudices (prejudices which I partly share).
This study found that people who eat more of what is termed “ultra-processed foods” have a higher overall risk of developing cancer. Whatever their opinion about the merits of “ultra-processed foods”, most nutritionists would have predicted this outcome because the people who eat lots of these “ultra-processed foods” are likely to be very different from those who eat little of them. For example those who ate a lot of these foods were more likely to smoke, were less active, ate more calories and were more likely to be taking oral contraceptives. Of course, the authors did collect data about many other dietary, lifestyle and socioeconomic characteristics of the participants and used sophisticated mathematical models to make multiple “corrections” for these and quite a few other so-called confounding variables. There is, however, no statistical magic wand that precisely corrects for all the potential confounders. In several cases, the information on confounding variables is a relatively crude categorisation, for example:
- Categorisation of physical activity levels into low, medium and high
- Categorization of education level into three categories – less than high school degree, less or more than two years after high school degree
- Smokers were just divided into current smokers or never/former smokers.
The less precise the information about confounders is, then the less reliable the correction process is and, of course, one can never be sure that all of the potential confounders have been corrected for. The dietary assessment process is also relatively crude (two or more online 24h recall questionnaires), as dietary assessments of free-living people always are. The sample was a self-selecting one and was not representative of the French population e.g. it was predominantly female (78%) and tended to be better educated, more health conscious, to eat less “ultra-processed foods” and they had substantially lower rates of cancer than the French population as a whole.
The size of the cohort (c105000) makes this sound like an impressively large study but the statistical analysis of course compares rates of cancer in different categories. The overall rate of cancer in the study group was 79 cases per 10,000 people per year c.f. 97 cases per 10,000 per year in the general French population (comparison standardised for age and sex).
Epidemiological studies have been important in the discovery of important links between lifestyle and behavioural factors and disease risk e.g. smoking and lung cancer, sleeping position and cot death, asbestos exposure and the normally rare, lung condition mesothelioma. However, the risk differences seen between exposed and unexposed in such examples is large; certainly more than doubled and often more than 10 times. When one is dealing with such large differences in risk then it is difficult to believe that such associations are not cause and effect and are an artefact caused by some unexpected or improperly corrected confounding variable. In this paper, the French authors suggest that increased cancer risk is 12% (or between 6 and 18%) for every 10% increase in use of “ultra-processed foods”. Given the cancer rate in this cohort, this amounts to just a few extra cases per 10,000 people per year; it now becomes much more credible, likely even, to suggest that this difference could be explained by the necessarily crude methodology and any flaws in the correction process. They may be measuring the bias in their study.
I would also suggest that even if we accept the findings at face value they are of very limited practical value for public health guidance. The category of “ultra-processed foods” is a very broad and diverse one, it includes:
- Mass produced packaged breads and buns
- Sweet or savoury snacks
- Industrialised confectionery and desserts including many dairy products
- Sodas and sweetened drinks including calorie-free versions
- Meat products which have been preserved with anything other than salt e.g. bacon, ham, sausages etc.
- Instant noodles and soups
- “Ultra-processed fruits and vegetables”
- Ready meals including chilled and frozen.
They list a large range of processes and additions which would make any food “ultra-processed”.
The authors suggest many theoretical mechanisms by which “ultra-processed foods” might increase cancer risk. These range from the traditional low in vitamins and fibre but high in fat, saturated fat, salt and added sugar, through deliberate food additives to contaminants from packaging or produced during processing. Only some of these “potential hazards” will apply to each of the categories above.
The foods categorized as “ultra-processed” make up a substantial proportion of the food eaten by many (most?) people living in an industrialised country like France or the UK. How are public health advisers, let alone ordinary consumers, supposed to interpret this weak, blanket-association with one type of disease into practical dietary change? Should consumers be advised to reduce or avoid a huge number of foods because some of them may be weakly associated with an increased cancer risk? A return to a bygone age where most people bought largely fresh ingredients to make home-cooked food or bought bread and other freshly processed foods from a local artisans seems like a pretty unlikely prospect, certainly for the typical French or British person working relatively long hours to generate a limited income.
The authors call for other large-scale observational studies with different populations and settings. I cannot see how this would substantially improve the usefulness of these findings. A weak association with a cohort of half a million is still a weak association and unless it is extremely well-resourced then it seems more likely to amplify the problems mentioned rather than reduce them; well resourced studies, designed specifically to test links between diet and cancer like the major EPIC project (European Prospective Investigation into Cancer and Nutrition) may use better methods of assessing diet and potential confounding variables
They also call for further studies to try to discriminate between the various potential causes of any association between “ultra-processed foods” and cancer e.g. between nutritional factors, contaminants or additives. Investigation of many of these factors has already been ongoing for decades so exactly what further studies do the authors think are justified by their findings?
What these authors have found is a weak association between the total consumption of a huge category of foods and a slightly increased cancer risk. I can see no practically useful conclusions that can be drawn from this work even if one accepts the findings at face value. I can foresee it generating a steady stream of related studies that will soak up research resources and increase the mountain of published papers that do not really add to our scientific understanding or our ability to give clear and consistent advice about dietary improvement. On the positive side, these future papers will add extra lines to the curricula vitae and enhance the careers of their authors and help fill the mass of available space in some of the tens of thousands of scientific journals.
Ioannidis and Trepanowski (2018) acknowledge the importance of diet to good health but say that the influence of individuals foods and nutrients is ambiguously tiny or non-existent. They go on to say that:
“Substantial reliance on observational data for which causal inference is notoriously difficult also limits the clarifying ability of nutrition science”