Subscribe to Blog via Email
January 2021 S M T W T F S « Jan 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
If we saw that headline we might just laugh. But a few weeks ago the media was abuzz with the opposite conclusion. A Minnesota station said ‘U’ Study: Breakfast Could Decrease Diabetes Risk saying the study “links the meal to a decrease in risk for type 2 diabetes” based on a study at the University of Minnesota. Star Tribute’s headline said Breakfast cuts risk of diabetes, U study finds. Minnpost.org did a better job saying “Skipping breakfast associated with higher risk of obesity and diabetes, U of M study finds.” What they added was a discussion of the limitations of the study saying
“Now for those caveats. First, this was an observational study, which means it can only show a correlation between two things (in this case, breakfast and a lower risk of certain medical conditions), not a cause-and-effect. Other “hidden” factors, not controlled for in the study, could explain the results.”
That is important. What they have observed is A is correlated with B. What makes the media with the help of the authors of the paper is: A causes B. But with a correlation study it can just as easily be said the B is correlated with A and therefore B causes A. Thus my title. Both conclusions are unjustified.
The headlines were based on a paper published in the Diabetes Journal titled “Breakfast Frequency and Development of Metabolic Risk.”
A reading of the paper provides a different view of the results. That actual estimate of the hazard ratio for diabetes was 0.81 with a confidence interval from 0.63 to 1.05. Thus a claim that diabetes risk is lower for those eating breakfast cannot be made under the usual rules of classical statistical testing. Failing to find an overall effect the authors then looked at breakdowns by sex and race. There they found some effects. However this is a risky statistical approach as the analysis has entered the realm of multiple comparisons and the confidence intervals need to reflect this. When the overall effect is 0.81 it is a virtual certainty that the rate for one sex group will be lower than 0.81 and one will be higher. The calculation of the confidence intervals and the associated statistical testing must take into consideration that two estimates are being looked at – not just one estimate at a time.
So where did the claim that diabetes had a lower risk come from if the results are not in the paper. I suspect is came out of the subgroup estimates. But the University of Minnesota web posting on the paper also clearly implies the link, saying “A study by University of Minnesota School of Public Health researchers has found consuming breakfast daily, regardless of diet quality, is strongly associated with a reduced risk of developing type 2 diabetes.” This statement, at least to my way of thinking, goes beyond what is provided in the analysis in the published paper.
A second issues in this work is that of conflict of interest. The research was funded in part by General Mills. The authors stated “they had absolutely no hand in any part of the paper or in the design of it. Like you, they’re reading it for the first time today.” This is good but it does not completely solve the conflict of interest problem. Think in terms of what is usually referred to as publication bias. A company can fund such studies. If the results a negative the results are not likely to be published. There is not harm to the company. If the results are positive the results get published. The company can than say the research shows that you should eat breakfast. That in turn will benefit the company. The company does not need to have hand in the actual research. So funding such studies is a no risk proposition for the companies involved. Innocent or guilty, there is no way around this problems for the companies involved. This is one reason why a common advice is to look first at research that has no funding involvement from parties likely to benefit from the results.