|According to Harvard, this truck has saved more lives than an ambulance.|
There are a number of methodological flaws in all these studies, and they are worth highlighting.
Firstly, the authors have combined two somewhat heterogenous cohort studies, previously published separately, and which present different findings, into what they now call one cohort.
Another way of describing this method is to say that they have cherry-picked two studies to put together. There are other studies that they could have combined with HPfS, or with NHS, to dilute or amplify their results. Of course they chose these studies because they are in charge of both of them, but nonetheless this is probably a unique proceeding.
Secondly, the results are now presented as person-years. This creates a larger number which looks impressive, but obscures the actual n= in each result.
Thirdly, the validity of the data is more questionable than the authors admit. Respondents were asked to estimate how many times they had eaten listed foods on average in the past year. The only verification seems to have been a comparison between a sample of the respondents completing both the FFQ and a 7-day food diary.
"In each FFQ, participants were asked how often, on average, they consumed a standardized portion size of each food during the previous year."
"The Spearman correlation coefficient of intake assessed by the FFQs and 7-day dietary record was 0.56 for animal protein and 0.66 for plant protein."
A Spearman correlation of 1 would have meant that the results were identical. 0.56 may be considered "high validity" in diet epidemiology, but wouldn't be accepted at the vehicle testing station. The results are meant to estimate 365 days not 7 days, so this comparison was incomplete.
So incomplete that the NHS cohort (the female half of this population) has reported eating 1,500 kcal/day on average for many years by the FFQ system.
"Among participants who returned baseline questionnaires, we excluded those who had a history of cancer (except nonmelanoma skin cancer), CVD, or diabetes at baseline, left more than 10 items blank on the baseline FFQ in the NHS and more than 70 items blank in the HPFS, or reported implausible energy intake levels (under 500 or over 3500 kcal/d for women, or under 800 or over 4200 kcal/d for men)."
This seems to state that respondents who seriously under- or over-stated energy intake were still included in the two studies.
Those are objections that pertain to the studies as a whole, but what of the specific findings of this study?
"Of the 131 342 participants, 85 013 were women (64.7%) and 46 329 were men (35.3%) (mean [SD] age, 49  years). The median protein intake, as assessed by percentage of energy, was 14% for animal protein (5th-95th percentile, 9%-22%) and 4% for plant protein (5th-95th percentile, 2%-6%). After adjusting for major lifestyle and dietary risk factors, animal protein intake was weakly associated with higher mortality, particularly cardiovascular mortality (HR, 1.08 per 10% energy increment; 95% CI, 1.01-1.16; P for trend = .04), whereas plant protein was associated with lower mortality (HR, 0.90 per 3% energy increment; 95% CI, 0.86-0.95). These associations were confined to participants with at least 1 unhealthy lifestyle factor based on smoking, heavy alcohol intake, overweight or obesity, and physical inactivity, but not evident among those without any of these risk factors. Replacing animal protein of various origins with plant protein was associated with lower mortality. In particular, the HRs for all-cause mortality were 0.66 (95% CI, 0.59-0.75) when 3% of energy from plant protein was substituted for an equivalent amount of protein from processed red meat, 0.88 (95% CI, 0.84-0.92) from unprocessed red meat, and 0.81 (95% CI, 0.75-0.88) from egg.
There are two things that should jump out here. The first is that intakes of animal protein and plant protein differ by a factor of 3. Most people on LCHF and paleo diets are eating more plant protein than the people in NHS and HPFS cohorts. For the people in the lowest quintile of plant protein, this supplied 2.6% of energy. That's consistent with bread and processed meat being the main sources of plant protein. (wheat is 14% protein, most cheap commercial sausages contain wheat and soy protein. I'm not sure if Song et al factored this latter into their analysis).
The comparison between high and low plant protein intake is between median 2.6%E (about 10 grams of protein for NHS) and 6.6%E (about 25 grams). 25 grams is associated with less mortality than 10g. Neither amount is sufficient to sustain life.
In the animal protein stakes, median of lowest quintile is 8.9%E and highest is 20%E, and this is a range of protein intake consistent with life.
We're not really comparing like with like.
The second thing that jumps out is this:
"These associations were confined to participants with at least 1 unhealthy lifestyle factor based on smoking, heavy alcohol intake, overweight or obesity, and physical inactivity, but not evident among those without any of these risk factors."
This screams "residual confounding". If your associations disappear when you minimise confounding variables, you probably haven't measured or adjusted for these properly.
To their credit, Song et al do recognise this;
"First, given the remaining variation of health behaviors across protein intake categories in the unhealthy-lifestyle group, residual confounding from lifestyle factors may contribute to the observed protein-mortality associations. However, our results are robust to adjustment for a wide spectrum of potential confounders and the propensity score. "
This seems to be saying that because they performed adjustments, and this produced consistent results, therefore those results are likely to be correct.
However, in the last paper by this group based on the exact same data sets, there was evidence of residual confounding, in the form of a positive correlation between respiratory disease mortality and saturated fat (HR 1.56; 95% CI, 1.30-1.87). Saturated fat consumption was associated with a higher incidence of smoking, but this had been adjusted for.
This finding was described as "novel", because it had no support in the literature. Respiratory disease mortality is usually associated with smoking (which was controlled for) and other air quality factors (passive smoking and traffic proximity) which were not.
There are two possible explanations for this correlation.
Either saturated fat strongly increases respiratory mortality via unknown mechanisms which are only operative in doctors and nurses living in the USA in 1980-2012, or,
Doctors and nurses living in the USA between 1980 and 2012, years of strong anti-smoking campaigns and adjusted insurance premiums, are more likely to underreport smoking than the other, non-medical populations in other cohort studies.
I leave it to you to judge which of these explanations is more ontologically parsimonious.
[Edit 18/04/17; Wang et al have confirmed that the low SFA quintile was 15 years older than the high SFA quintile, due to their cumulative method and changes made by health-conscious individuals. This introduces a larger possibility of both unmeasured confounding and over-adjustment than suggested by baseline data. Individuals have moved between quintiles since baseline, rendering baseline data useless.]
Let us, for arguments sake, take the results at face value; there is no harm from eating extra animal protein from mixed sources instead of carbohydrate, especially if you don't eat commercial crap, and some benefit from eating plant protein (probably from its richer, higher fat sources, as only these will supply extra protein in replacement for carbohydrate).
Imagine a diet where you replace carbohydrate from wheat flour with protein from almond flour. Why, such a diet will reduce your chances of dying, according to Harvard!