On PURE

The PURE macronutrients studies were published in the Lancet journals today and the headlines/commentaries are reminding us that everything we thought we think we were told we knew about nutrition is wrong/misguided, etc. Below is my non-epidemiologist's run down of what happened in PURE.

A couple papers came out related to PURE, but the one causing the most buzz is the relationship of the macronutrients to mortality. With a median follow up of 7.4 years, 5796 people died and 4784 had a major cardiovascular event (stroke, MCI). The paper modeled the impacts of self reported dietary carbohydrate, total fat, protein, monounsaturated (MUFA), saturated (SFA), and polyunsaturated (PUFA) fatty acid intakes on cardiovascular (CVD), non-CVD and total mortality; all macros were represented as a percentage of total self reported energy intakes and reported/analyzed in quintiles (energy intakes between 500-5000kcals/day were considered plausible..). All dietary data was determined by a single food frequency questionnaire (FFQ) at baseline. The mortality modeling for each of the macronutrients adjusted for age, sex, education, waist-to-hip ratio, smoking, physical activity (assessed by questionnaire), diabetes, urban and rural locations , centre, geographical regions, and energy intake (self-reported). In a subset of sensitivity analyses on carbohydrate and saturated fat intakes relationship to total mortality, the investigators adjust for proxies of SES such as household income/wealth, education and economic level of the country.

The results of this PURE analysis found that carbohydrate intake was associated with a higher risk of mortality, and non-cardiovascular diseases. As the fat-carb seesaw tends to go, higher intakes of fat were associated with lower risks of total mortality, and non-cardiovascular disease mortality (common causes of non-CVD mortality were cancer and respiratory diseases, except in Africa where respiratory and infectious diseases were the most common causes). Total protein and animal proteins were also associated with lower risks of total mortality. When looking within fatty acids, saturated fats were associated with lower risk of total mortality, stroke, and non-cardiovascular mortality. Similar findings for total mortality and non-cardiovascular mortality were found for MUFA and PUFA. These associations typically held up to comparisons of Asian vs Non Asian countries, though no report of non-Asian/African countries combined is given (these are the highest CHO intake countries). Lastly, replacement modeling suggested that replacing CHO with SFA lowered the risk of stroke, and replacing CHO with PUFA lowered the risk of non cardiovascular disease mortality. 

A few notes on interpreting this:

1) On Carbs: The findings that have been highlighted are the impacts of high carbohydrate intake on mortality. The associations between carbohydrate intake and total/non-CVD mortality are largely confined to comparisons between quintiles 5 and 1; this is comparing individuals with a median of 77.2% of their daily intakes from carbohydrates to 46.4%. One model with WHR included also finds a marginally significant effect of quintile 4 compared to quintile 1. The models for total fats and total mortality/non-CVD mortality are much more consistent across quintiles (confidence intervals don't intersect 1).  Essentially, at face value, these data suggest that very high intakes of CHO are associated with a modest increase in total mortality and non-CVD mortality whereas moderate increases in median fat intakes across quintiles (Q1=10.6% fat, HR=1; Q2=18% fat, HR=.9, Q5=35% fat, HR=.77) reduce total mortality. 

Of note here, the top 5 sources of carbohydrate intake in these populations are largely refined sources (white bread/rice, sugary drinks, cakes, fruit/fruit juices). There are some odd-ish sources for some countries, such as tea (also a top 5 source of protein and saturated fat for some populations), so i'm assuming these populations are consuming highly sweetened teas with added dairy. No information about the fiber content of the diet, or adjustment for fiber intakes is given though this should theoretically have been available from the FFQ data. It's a bit odd to see fats broken down by class but no further classification of carbohydrate. For all of the fan fair about these results showing that 'low fat diet' recommendations were 'wrong', these data don't actually test whether the recommended low fat diet rich in high fiber carbohydrates was beneficial. Further analysis looking at food sources of less refined carbohydrates (fruits, vegetables and legumes) with similar quality data, were also published in the Lancet in a different analysis. These effects for very high CHO intakes are driven primarily by non-North American/non-European populations, given the relatively small cohort of individuals consuming that high a percentage of energy from carbohydrates; that hasn't stopped American news from making clickbaity headlines about carbohydrates though. Carbohydrates also didn't give much to talk about with regard to cardiovascular disease outcomes either, so again, a bit odd to see American news headlines referencing long standing debates about carbs and fats in the CVD realm in relation to this study. 

2) On Fat Type: The fatty acid data is pretty much a let down for anyone who thinks the effects of dietary fats on lipids has a meaningful impact on cardiovascular disease, given the very limited effects of these on CVD events/mortality that were observed in the primary analysis. However, the data are not too surprising when you look at the paper released on the effects of fats on lipids and blood pressure from the PURE study. Whereas controlled feeding studies consistently show that PUFA (largely n-6's) and MUFA reduce LDL-C and ApoB concentrations, particularly when replacing SFA and carbohydrates, the PURE analyses found positive associations with all fat subclasses and LDL; SFA and PUFAs (reported food sources suggest these were also largely n-6s though no quantitative analysis is done) were also positively associated with ApoB, a common marker of all atherogenic lipoproteins. Whereas every randomized controlled trial to date using non-hydrogenated/non TFA containing PUFAs yield lower concentrations of LDL and ApoB relative to the other macronutrients, the PURE cohort sees an association with higher LDL/ApoB. MUFAs were associated with higher ApoA-1 in PURE as they are in trials, but had a neutral effect on ApoB in PURE, whereas they lower ApoB in trials. PUFAs were also associated with higher ApoA-1 in the PURE cohort, whereas trial evidence hasn't shown an impact of PUFA on ApoA-1. Given the observational nature of the study, there's very likely some confounding going on. The only likely mechanism by which PUFAs would be associated with higher ApoB is if they were partially hydrogenated, but PURE does not report trans fatty acid intakes. It's more likely that additional factors which impact lipoprotein levels are coming into play that are unmeasured in the cohorts. 

The study authors make a big deal about their results not supporting guidelines to reduce saturated fatty acid intakes; apart from the fact that the lipid data lends itself to concerns that these international estimates of saturated fat intakes serve more as markers of unmeasured confounding factors, guidelines stress the importance of replacing saturated fats with unsaturated fats, primarily PUFA. This replacement modeling was not performed in this study, oddly, given all of the handwaving about running counter to dietary guidelines.

Apart from the lipid surrogate data, the effects of saturates on stroke are interesting, but are marginally statistically significant (p=.0498).

3) On Multiple Comparisons: Speaking of p=.0498, like many observational cohorts, the results of this study must take into account the very many correlations that were drawn. Correcting for multiple tests would eliminate some of the observed positive associations. If you're thinking of making causal inferences from this study, keep that in mind.

4) On Effect Sizes: The effect sizes, regardless of statistical significance, are quite small. If taken at face value, eating a ton of refined carbohydrates (literally the devil if you ask Professor Google) at 77.2% of calories resulted in 1.36 HR relative to quintile 1 consuming 46.4% of kcals from CHO. That's a pretty low HR when thinking in terms of whether or not we could make causal inferences from this study; this is not the type of risk increase that gets most evidence based medicine advocates all riled up and ready to make causal inferences.

5) On PURE's Rationale: Speaking of our ability to make causal inferences, I can't not comment on the rationale for the study. The goal of epidemiology is to assess variations in an exposure and its relationship to some outcome. Ideally, you want to compare populations that are relatively homogeneous in their general lifestyle and other environmental exposures, and measure confounding factors that might influence the outcome, while capturing some decent variability in your exposure of interest. With variability in the exposure, and variability in the outcome, while keeping mostly everything else fixed, you can estimate an effect size and draw some clear hypotheses/maybe make causal inferences. This is a particularly necessary approach for examining the relationship of diet with chronic diseases, given the numerous factors that influence causal disease pathways (i.e. many dietary and non-dietary factors impact LDL particles/ApoB to influence CVD risk). In many modern nutrition cohorts, like those out of Harvard and Framingham, we look within a population at the relationship between dietary exposures and outcomes. While all cohorts in nutrition have their problems, particularly relating to self reported dietary intakes (macronutrients, kcals), a strength of these cohorts is that they look at the associations between nutritional exposures and disease in somewhat similar people. Even then, we still have concerns about confounded relationships. PURE took the potential for confounding to another level. It compares individual level dietary data with disease outcomes across a broad range of countries and geographical regions. While the study crudely attempts to model the impacts of socioeconomic status and account for geographic region, it's the epitome of using epidemiology as a blunt force tool to get at the relationship between macronutrients and disease. In this paper, the investigators could have examined the relationships only within countries/regions and disease outcomes (admittedly limiting their power) to compare more 'like' individuals, but the authors did not do this (at least in this paper). This lacking analysis makes it rather odd that they focus so much on comparing their results to other cohorts where the individuals are all from the same country (table 10 in the supplement even compares to Harvard cohorts and EPIC).

6) On the Models: Given this high possibility of confounding from unmeasured variables when comparing individuals living across countries, the modeling presented in the paper appears relatively simple. Given the variability in developed and developing countries included in PURE, you'd imagine that the models would at least include variables such as medication or alcohol use, but they don't report doing so (but do report collecting this additional information).  Additional nutrition analyses could've been useful as well, such as including multiple macronutrients in one model; for example, the associations between animal protein and mortality and saturated fat and mortality are quite similar; HRs for total protein and total mortality/non CVD mortality are .88 and .85, HRs for saturated fat and total mortality/non CVD mortality are .86 and .86. Adjustments for multiple macronutrients in the same model would've been potentially informative; are animal protein and saturated fat (both largely of animal origin) so highly correlated in PURE that their effects are indistinguishable, and are just serving as markers of animal food intakes?  In most modern cohorts within countries, we tend to see robust efforts to adjust for additional nutritional factors, such as the Alternative Healthy Eating Index, a marker of overall nutrient rich diets; PURE could've adjusted for micronutrients, especially given that highly refined carbohydrate rich diets are probably low/deficient in multiple micronutrients (which can impact some of the common causes of mortality, such as respiratory/infectious diseases). Additional adjustments with surrogates such as lipids or blood pressure would've been potentially informative as well, to see if the effects of macronutrients on mortality are mediated by these measures. Alas, the paper on fatty acids and lipids/blood pressure only estimates the effects on mortality and surrogates aren't included in the actual mortality models. 

7) On all of the data: I would've enjoyed full macronutrient and energy breakdowns by country and region in the supplement. We get the table below for carbohydrates, which essentially confirms for us that the relationship between very high carbohydrate intakes and mortality are driven primarily by individual, often poorer, countries. It would've been nice to see average kcal intake as well, given the huge variation in plausible intakes. I strongly believe we should've gotten a table of the top 5 causes of mortality across the countries as well so we could attempt to understand how plausible the relationships seen are. Total mortality is certainly a useful metric but it requires context to interpret as well.

It's important to keep in mind that all of this data comes from a single FFQ at baseline and that many of these FFQs were developed for this study and validated on small cohorts. Additionally, as Erik Arnesan pointed out to me, for many of these developing countries (i.e. Bangladesh), it's likely that developing food systems and nutrition transitions would alter their food intakes across time.

Conclusions: Overall, PURE is a huge cohort and hopefully will add some more interesting analyses as the results come out. The current macronutrients and mortality data is less compelling to me, for reasons discussed above. There are many more analyses that can come from this cohort and I look forward to seeing them. As of now, it's hard to look at this PURE publication, where individual level data are compared, and not just see macronutrients look a lot like proxies for unmeasured variation related to wealth, SES, and access to healthcare, as well as a very broadly nutrient poor. Less developed countries who happen to eat a ton of carbohydrates having higher total/nonCVD mortality doesn't feel like a super novel finding, but it certainly fuels headlines; there's potential biological plausibility in the finding, particularly if the individuals live in high infection risk environments and have very low intakes of vitamins/minerals/PUFAs which play a role in the immune response but aren't found in high amounts in refined carbs (vitamin A/beta carotene, zinc, iron, linoleic acid), but these lack a lot of relevance for developed countries; they certainly don't warrant a reduction in concerns about the impact of diet on cardiovascular disease in the developed world. 

At the end of the day, if you believe the data at face value, they confirm that super high intakes of refined carbohydrates are not a great idea, and add to the 'which is worse' nutrient war related to refined carbohydrates vs saturated fats. In this analysis, saturates win out, but the limited comparisons of fats to one another and the discordant fatty acid- lipid surrogate data between controlled trials and this data leave me not jumping to many conclusions or ready to throw away all prospective cohorts and randomized controlled trials performed to date.

For other takes, see sciencemediacentre's expert reaction , Nutrition Wonk's coverage of the study, Pauli Ohukainen's twitter rant.

If you love the nutrition gossip, the PURE study investigators have been talking up their findings for a while now, with some lovely accusations of bias from nutrition scientists and how PURE, an observational study with one time point of self reported dietary information, is upending all nutrition recs. It also doesn't take very long on snapbird to see a libertarian 'the government/establishments have been horribly wrong on diet' (and climate change...) vibe on some of the lead author's twitter(s); the discussion within the papers and the lack of robust analyses testing current dietary guidelines correlate interestingly.