Dear all. As you may have seen, in response to my request for a journal to publish a critique of PURE, Carl Heneghan, editor of BMJ EBM, has kindly agreed to review a submission on this.
So could I ask members of the List to provide me with any comments on PURE, both positive and negative, which I will put together in a collective review. All significant contributors will be included as authors.
I would like to focus the critique largely on the methodology of PURE rather than on the relationship between diet and CVD. I believe PURE is seriously flawed with regard to addressing questions
of causality, however I am open to both positive and negative comments.
James McCormack and David Nunan have already offered some comments (thanks guys) and I would like to hear more.
James, you point out that the authors found similar associations across populations, which I think you thought was a possible strength whereas I thought it was consistent with universal confounding.
He suggests it is largely confounded by socioeconomic deprivation which is strongly associated with diet, other behaviours, medical care and actually just about anything you can think of. He also points out the very important
issue that many epidemiologists and statisticians miss about the downsides of multivariate adjustment modelling when investigating causation. In the PURE carbohydrates and disease papers published last year, the expected univariate association between increasing
fruit and vegetable intake and lower mortality, changed into the surprising conclusion that you can eat too much fruit and vegetables, after multivariate adjustment.
I do large-scale cohort studies of CVD, but for prediction not causation research and I am very aware of the strange things that multivariate analyses do to the associations between CVD risk factors and CVD outcomes. Fortunately
in prediction research, a confounder is not a problem - it is just another predictor - however it is a major headache for causation research. In a multivariable analyses, variables in effect ‘fight for their place in the model’ and the final model can be
as much to do with the degree of measurement error of the different variables as with any causal association. Multivariate models like those in PURE are just a nightmare to make sense of.
Cheers Rod
Hi Rod,
We’d be more than happy to see such a paper submitted to BMJ Evidence-Based Medicine. Whether we’re reputable, or not, I’ll leave that to you and others to decide, but we are not shy of publishing controversial pieces and they do get considerable attention.
On the wider issue, I agree, the plethora of observational (perhaps epidemic would be a better term) research that infers causation is concerning and adding little to overall decision making. The growth of big data, the availability of data sets and
the media reaction means we see more of this poor quality evidence.
Are we going backwards not forwards in our evolution of evidence?
All the best Carl
Sent from my iPhone
On 24 Nov 2018, at 02:31, Rod Jackson <[log in to unmask]> wrote:
Dear all. For a great example of almost certain substantial confounding (not to mention measurement error and cross-over) please critique the latest PURE Study analyses in this week’s Lancet. Why do journals continue to publish
papers that could be used in introductory Epidemiology classes as ideal examples of how to introduce confounding into an analysis? The paper is controversial and comes from a well-known clinical trialist who has an international reputation for undertaking
high quality randomised controlled trials, a study type designed specifically to minimise confounding. Yet this paper describes an analysis of the association between dairy food consumption and disease in people from countries as diverse as Canada and China
and Sweden and Zimbabwe. This mix of a food group that can be expensive and difficult to store without refrigeration, a disease group that is easy and quick to modify and a participant group from extremely diverse populations provides an ideal environment
for confounding. I don’t know how it would be possible to confidently control for confounding in the association between dairy food and disease in just one of these populations, let alone across such diverse populations? And using the same adjustment factors
in all populations. Who eats dairy products in Canada and Sweden? Are they likely to be similar to those who eat the same foods in China and Zimbabwe or Bangladesh? Higher consumption of total dairy products was associated with lower total mortality, lower
non CVD mortality and lower CVD mortality - appeared to be good for everything! Yet, when examining more specific outcomes, there was no protective association observed for the only specific atherosclerotic outcome - myocardial infarction - and the strongest
protective association was for stroke - a large proportion of which would be haemorrhagic rather than atherosclerotic, given the large proportion of participants from low and middle income countries.
Does anyone on this list know of a reputable and frequently read journal that would accept a substantial critique of PURE? I have only scraped the surface of the necessary critique. It needs more than a letter to the Lancet editor.
Regards Rod Jackson
Professor of CVD Epidemiology
University of Auckland, New Zealand
########################################################################
To unsubscribe from the EVIDENCE-BASED-HEALTH list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=EVIDENCE-BASED-HEALTH&A=1