Critiquing `All-Cause and Cause-Specific Mortality with Low-Carbohydrate Diets' by Mohsen Mazidi, Niki Katsiki, Dimitri P. Mikhailidis, Naveed Sattar and Maciej Banach

Link to the ungated distillation of the article by the first three authorsLink to the article itself (gated) — *Link to the ungated distillation of the article by the first three authors*
*Link to the article itself (gated)*

I am fortunate to have free access to the full text of most medical and nutritional studies through the wonderful website for the University of Colorado Boulder’s Norlin library. So let me contribute to the nutritional debate by telling what I learned from digging into the full article “Lower carbohydrate diets and all-cause and cause-specific mortality: a population-based cohort study and pooling of prospective studies” behind the ungated distillation by the authors Mohsen Mazidi, Niki Katsiki, Dimitri P Mikhailidis, Naveed Sattar and Maciej Banach, which is supertitled “Higher risk of all-cause and cause-specific mortality with low-carbohydrate diets.”

The authors divided people into four groups based on a score that gave “low carbohydrate diet” points for having a diet low in percentage of calories from carbohydrates, points for being high in percentage of calories from fat and points for being high in percentage of protein. The coefficients are not transparent. They write:

]Consumption of carbohydrates was scored from 10 (lowest consumption) to 0 (highest consumption), whereas protein and fat intake were scored from 0 (lowest consumption) to 10 (highest consumption).

A key point I want to emphasize is that they are not testing the seeming effects of a low-carb, high-fat diet, but the effects of a low-carb, high-fat and high-protein diet. In particular, their strongest results are for the fourth quartile, which has dramatically higher protein as well as dramatically higher fat than the other quartiles. They write:

Participants were stratified into quartiles, based on LCD score:

Q1: median LCD score of 12, 367 g carbohydrates/day, 77 g protein/day, 73 g fat/day [reference]
Q2: median LCD score of 15, 245 g carbohydrates/day, 69 g protein/day, 65 g fat/day
Q3: median LCD score of 18, 205 g carbohydrates/day, 72 g protein/day, 70 g fat/day
Q4: median LCD score of 21, 214 g carbohydrates/day, 103 g protein/day, 105 g fat/day

The distillation has a striking graph:

The colored lines with the small circles show the point estimates and 95% confidence intervals for risk ratios implied by the regression coefficients in a multivariable Cox proportional hazards model. The risk ratios are for each of the other quartiles compared to the first quartile, which is very high carb.

The Possible Dangers of a High-Protein Diet. I am not at all surprised that the high-protein diet indicated by the red line might be dangerous. I have written about the possible danger that too much protein might promote cancer in these two blog posts:

The results in the study by Mohsen Mazidi, Niki Katsiki, Dimitri P Mikhailidis, Naveed Sattar and Maciej Banach raises the question of whether protein creates a risk for stroke and heart disease as well. Unfortunately, it is hard to distinguish in an observational study like this between the effects of dietary fat and the effects of dietary protein since dietary fat and dietary protein are highly correlated. My suspicion, partly coming from just my love of being contrarian, is that dietary protein has too positive a reputation and dietary fat too negative a reputation. That is, I think that dietary fat is often getting blamed for the harm caused by dietary protein, especially animal protein.

Controlling for Calories Consumed Changes the Interpretation Dramatically, in a Way the Authors Do Not Recognize or Acknowledge. Leaving the fourth quartile results undisputed as a possible warning about high-protein diets (with other besides me free to dispute them), let’s turn to the light and dark blue lines showing the estimated risk ratios for the second and third quartiles relative to the first. If you look carefully at what I have copied out above, you can see something strange. The second and third quartiles have a median consumption of all three macronutrients that is lower than in the first quartile. Why would eating less fat, less protein and less carbs lead to higher mortality? It isn’t as if these folks are starving. The answer is that the multivariable Cox model controls for total number of calories consumed. The authors write in the full paper:

… we had two different models: Model 1 adjusted for age, sex, race, education, marital status, poverty to income ratio, total energy intake, physical activity, smoking, and alcohol consumption; and Model 2 adjustment for Model 1 plus body mass index (BMI), waist circumference, hypertension, serum cholesterol and diabetes.

To my mind, this doesn’t give a low-carbohydrate diet a fair chance. The main harm I see from carbohydrates (especially easily-digested carbohydrates) is that they make you hungry so you eat more total calories. That story certainly matches what is happening with the median consumption numbers in the first quartile: carbohydrate grams are much higher in the first quartile without fat or protein grams being any lower—indeed, fat and protein grams are somewhat higher. If most of the harm of a high-carb diet is that people end up eating too much and the benefit of a low-carb diet is that people end up eating less, controlling for total calorie consumption in the regression slices out one of the main mechanisms through which low-carb diets are helpful. Note that Model 2 goes even further in this direction, by controlling for body mass index, which thereby insures that any benefit of a low-carb diet that operates through weight loss is sliced out. And without other analyses, that means that effects of low-carb diets that operate through reducing appetite and through weight loss are ignored.

It is important to have a multivariable model, because the first quartile has lower alcohol consumption and less smoking, but controlling for something that might be a key intermediate causal variable totally changes the interpretation of the results. Let me point to one specific possibility to make the issue more vivid. One type of low-carb diet that would still leave me hungry so that I’d be likely to eat a lot would be cutting back drastically on vegetables and whole grains while keeping my sugar consumption at full throttle. So when the regression looks at people whose carb consumption is low but who are eating a lot of calories, it might be focusing on people whose carb consumption is almost entirely unhealthy carbs that keep appetite up, with the healthy carbs cut out.

The Author’s Meta-Analysis of Other Studies also Makes High-Protein Look Bad. In addition to their own regressions, the authors to a meta-analysis of regressions by other authors. In their report on the meta-analysis, they appropriately emphasize the “high-protein” aspect of the story:

Low-carbohydrate/high protein diet mortality

There was a significant association between LC/HP and overall mortality [RR 1.16, 1.07–1.26, P < 0.001, n = 5 studies, (no heterogeneity, I2 = 17.6, P = 0.825), Supplementary material online, Figure S2], as well as a positive correlation between LC/HP and CVD mortality (RR 1.35, 1.07–1.69, P < 0.001, n = 5 studies, Supplementary material online, Figure S3), with minimal evidence of heterogeneity (I2 = 21.5, P = 0.736). In contrast, a significant trend between LC/HP and cancer mortality was observed (RR 1.03, 0.99–1.07, P = 0.084, n = 3 studies, Supplementary material online, Figure S4), but with modest of heterogeneity, (I2= 57.3, P = 0.036).

Conclusion. A general point is that authors in the nutrition area are not always good at interpreting their own results. Economists are rewarded professionally for finding the flaws in other researchers’ regression designs and interpretations. So we get good at finding such chinks in the armor. I think everyone would get a much more accurate sense of what the evidence about nutrition really says if more economists got involved in thinking about that evidence. I could personally be mistaken quite easily. But if many economists were scrutinizing nutrition studies, collectively they would add a lot to understanding in this area.

Don’t miss my other posts on diet and health:

I. The Basics

II. Sugar as a Slow Poison

III. Anti-Cancer Eating

IV. Eating Tips

V. Calories In/Calories Out

VI. Other Health Issues

VII. Wonkish

VIII. Debates about Particular Foods and about Exercise

IX. Gary Taubes

X. Twitter Discussions

XI. On My Interest in Diet and Health

See the last section of "Five Books That Have Changed My Life" and the podcast "Miles Kimball Explains to Tracy Alloway and Joe Weisenthal Why Losing Weight Is Like Defeating Inflation." If you want to know how I got interested in diet and health and fighting obesity and a little more about my own experience with weight gain and weight loss, see “Diana Kimball: Listening Creates Possibilities” and my post "A Barycentric Autobiography. I defend the ability of economists like me to make a contribution to understanding diet and health in “On the Epistemology of Diet and Health: Miles Refuses to `Stay in His Lane’.”