Vinay Prasad Tweetorial: nutritional epidemiology

Vinay Prasad @VinayPrasadMD a year ago, 34 tweets, 10 min Read on Twitter

Copied & pasted via Thread Reader:

A couple weeks ago I said these HARSH words about a recent Lancet study & the media coverage where the authors argued more than 5-7 drinks a wk (100g/wk) was too many
I took heat
Well, I meant it then, and I mean it now. 
And here is the TWEETORIAL on this paper/ nutritional epi


First, let me prove to you the AUTHORS, not the journalists, the PR officers, etc, but the AUTHORS are guilty of this interpretation
From the paper:

Here, in the press release, the AUTHOR & funder are digging deeper

Consuming more than five drinks a week could shorten your life: Even moderate alcohol drinking linked to heart and circulatory diseases, study findsRegularly drinking more than the recommended UK guidelines for alcohol could take years off your life, according to new research. The study shows that drinking more alcohol is associated with a highe…

The media runs wild, and I get increasingly annoyed

For many readers of this genre of utter rubbish nutritional epidemiology research, we feel as if the news is engaging in this

I am curious, if you get annoyed by this news, who do you blame?There are 2 main problems with these kinds of studies; First confounding, meaning that the people who drink >100g/wk are different than those who drink less, and these differences, not the alcohol drive the outcome; the second is multiplicityLet me talk about multiplicity
For common exposures/ outcomes (e.g. aspirin, chocolate, tea, beer) & (cancer, death), there are thousands of interested researchers w retrospective data & stats softwareIn such situations, the published literature may not reflect the totality of published analyses, but just the analyses that slipped past the human & publication selective reporting biases & which serve little more than opinion pollJohn Ioannidis, B B, and @chiragjp wrote an amazing paper to show how analytical plan can alter the relationship btw exposure & outcome

Their insight as to realize that each published paper adjusts for some variables (age, sex, race), or others (age, SES, heart disease), what if you adjust for all variables (they picked 13 common ones) and all combos

By doing this, they get clouds of associations
Basically, showing, by covariate selection (picking what to adjust for), you can get huge variations in effect & sometimes, effects in either direction (+ or -)
Most associations were null however

When you consider multiplicity in this field, you get a total disastrous situation where most published & reported analyses may be little more than those that tell us what we wish were true, and nothing resembling the truth.Ioannidis pushed this in a related paper & showed we can find 40/50 random foods are linked to cancer in all sorts of directions
This finding is of course, implausible aka stupid

Looking at the Z scores of these relationships reveals people hug the 1.96 aka significance threshold, which also stinks, and suggests the published literature is little more than an opinion poll or self fulfilling prophecy on these topics.

This is part of the reason John puts nutritional epi studies at the bottom of the septic tank of science in his paper

Why Most Published Research Findings Are FalsePublished research findings are sometimes refuted by subsequent evidence, says Ioannidis, with ensuing confusion and disappointment.

Got interrupted. more parts coming and then this paper is getting destroyedPart 2: IT is not often that observational nutritional science claims are tested in corresponding RCTs, but over the years they have occurred. After enough occur, one can perform a concordance study– asking how often they agreeThe last time I read a concordance study was this one from the AHRQ in 2007 on these 34 topics

The first finding was that concordance was limited

On just 6 topics did they agree

But subsequent trials eroded that for 4 topics, and by 2013 (when I looked last) the ability of observational nutrition studies to say anything of value occurred in at most 2 topics.

Now, back to the Lancet paper 
BIG DATA, right?

Put aside the fact that the studies are of vastly different periods of time, and that alcohol use was documented over a limited period of time, and that most of the data is self reported, and HR extrapolated for Years life lost estimates that is rampant BSConsider the simple fact that the authors are putting a lot of stock in this estimate

Specifically they put a lot of stock in this estimate at 100g, which is borderline INSANE given that the data certainly has residual confounding (aka there are other factors about the Etoh users not the Etoh driving the outcome)
(last part coming)

The last point that reveals the data is surely confounded is the association between type of beverage and all cause mortality
Wine essentially adjusts for unmeasured socioeconomics here

Or if you truly believe this garbage field, then why not mention this data that might have a different public health implication — actually encouraging a daily drink

The hardest conclusion to accept here is that when it comes to common nutritional exposures– tea, coffee, chocolate, alcohol– we may have to make decisions the same way we decide how often to go to the bathroom or movies, i.e. using common sense and not bad epiAnd that if governmental/health authorities wish to retain credibility, they should make recommendations only when STRONG data exist, and that would most certainly not be the case for this (or similar) recommendationsFinal poll to see if you read this

Was any of this helpfulAnd, now, who is innocent for the media coverage of this foolish paper1 more question: was I wrong to say this study was BS?Looks like this thread made the news, and Ioannidis feels similarly

University of Twitter? Scientists give impromptu lecture critiquing nutrition research | CBC NewsA media frenzy seems to accompany any study showing that a common food or beverage is hazardous to our health. But some scientists squirm when they see those headlines. “The science is almost always …

Missing some Tweet in this thread?

Leave a Reply