Bad diets, heart disease & deaths
* A study was published on April 22nd, 2021 which reviewed the association between diet and cardiovascular disease (CVD) and deaths (from CVD or any cause).
* It was different to typical epidemiological studies as it tried to look at dietary patterns rather than single foods (e.g., meat) as usually happens.
* The study used a statistical technique that relied on assumptions about certain nutrients. The researchers assumed that energy density, saturated fat and free sugars were bad and that fiber was good. The statistical technique then identified the foods that met these assumptions.
* Two dietary patterns were identified. Dietary pattern 1 was characterised by high intakes of chocolate and confectionery, butter and low-fiber bread, and low intakes of fresh fruit, vegetables, and high-fiber breakfast cereals. Dietary pattern 2 was characterised by a higher intakes of sugar-sweetened beverages (SSBs), fruit juice, and table sugar/preserves.
* Dietary patterns 1 and 2 are both junk food diets. It was claimed that people whose diet was most like dietary pattern 1 were up to 40% more likely to have a heart incident or death during the study.
* The usual three flaws of epidemiological studies applied to this study (association not causation, relative not absolute risk, the healthy person confounder).
* There were additional flaws of this study: i) the assumptions about nutrients biased outcomes; ii) the healthy person was not fully adjusted for; and iii) the idea of a ‘bread/butter & jam’ diet might describe the British diet of 50 years ago, but it’s not a typical diet anywhere today.
* The absolute risk differences were tiny. Surprisingly small in fact. The dietary patterns were so bad that it would have been reasonable to expect far greater association with disease. The biggest finding was how small the impact actually was.
Just when you thought nutritional epidemiology had nowhere left to go, it’s gone somewhere else. A paper was published on April 22nd, 2021, called “Associations between dietary patterns and the incidence of total and fatal cardiovascular disease and all-cause mortality in 116,806 individuals from the UK Biobank: a prospective cohort study” (Ref 1).
The rationale for the approach was explained in the abstract. “Traditionally, studies investigating diet and health associations have focused on single nutrients.” This study aimed to identify dietary patterns and to examine the association of these with cardiovascular disease (CVD), CVD deaths and deaths from any cause.
The concept is sound. It never made sense to focus on one food or nutrient to examine associations, when we all eat whole diets. We have had decades of articles claiming, “red meat will kill you” and “whole grains will save you” and the food intake table has then revealed that the red meat/whole grain intake was 15g a day, or similar, and the study seems farcical before it starts. When we eat 2,000 – 2,500 calories (on average) a day, trying to claim that the 50 or so calories from one food will kill or save us has always seemed an imaginary stretch.
This study aimed to look at whole dietary patterns, rather than single foods. The method used was called Reduced Rank Regression (RRR). The paper explained, “RRR is a data-dimension reduction technique that aims to identify the combination of food groups that explain the maximum amount of variation in a set of response variables (e.g., nutrients) hypothesized to be on the causal pathway between diet and health outcomes.” (We’ll come back to what this means for this paper).
In terms of novelty, the introduction reported that – to the best of the researcher’s knowledge, only six other (longitudinal population) studies have reviewed CVD risk and/or all-cause mortality using this RRR technique. They were in smaller populations and none focused on the UK. This was thought to be the first large, UK, study to use this approach.
This study used UK Biobank data (as our meat study did a couple of weeks ago (Ref 2)). UK Biobank is a population study that recruited 502,536 participants aged 37 to 73 between 2006 and 2010 and then followed them for a number of years. Full details of the protocol can be seen here (Ref 3). The dietary information for Biobank was gathered using a web-based 24-hour dietary assessment tool called Oxford WebQ (Ref 4). The questionnaire was collected at baseline and on up to four further occasions.
This particular paper used data from participants who completed a dietary assessment on two or more occasions. The paper reported that “The RRR analysis was used to identify dietary patterns that could explain the maximum variation in a set of nutrient response variables hypothesized to be on the causal pathway between predictors (food groups) and outcomes (CVD and all-cause mortality events).” The key word there is “hypothesized.” This is where things went wrong. Instead of taking all dietary data and looking for associations between dietary patterns and disease, assumptions were made about the dietary patterns.
The word prejudice means to pre-judge. The authors pre-judged what harmful dietary patterns would be. They selected four factors “which contribute to excess energy intake”:
– energy density (defined as kJ/gram – this is a proxy for high energy intake. It takes the amount of energy (kJ) divided by the total food weight (g) excluding beverages because of their disproportionate impact on weight).
– saturated fat (SFA) (as a % of total energy).
– free sugars (as a % of total energy); and
– fiber density (defined as g/MJ – this takes total dietary fiber intake in grams divided by total daily energy intake in kJ and then multiplies it by 1,000).
Some of the author names are familiar and I know that former obesity Tsar, Susan Jebb, for example believes that people are obese because they eat too much (energy density); saturated fat is bad; free sugars are bad (she is a strong supporter of the fizzy drinks tax and has co-authored papers about this); and she’s a big fan of fruit, vegetables, wholegrains and thus fiber.
Participants were assigned a ‘z-score’ for their dietary pattern. This quantified how much their reported dietary intake reflected each dietary pattern relative to other participants. An increasing intake of food groups having positive factor loadings increased the dietary pattern z-score, while an increasing intake of food groups with negative factor loadings decreased the dietary pattern z-score.
A total of 116,806 participants were included in all analyses after exclusions. People were excluded for a number of reasons: almost 300,000 people in Biobank didn’t provide any dietary data! Approximately 85,000 were excluded for only having one dietary questionnaire. Other exclusions were CVD having occurred before baseline assessment, missing data, implausible energy intake and other valid reasons.
The paper reported that the RRR analysis identified two major dietary patterns that could explain the greatest amount of shared variation in the response variables (energy density, saturated fat, free sugars and fiber density) (43% for dietary pattern 1 and 20% for dietary pattern 2). Dietary pattern 1 was characterised by high intakes of chocolate and confectionery, butter and low-fiber bread, and low intakes of fresh fruit, vegetables, and high-fiber breakfast cereals. Dietary pattern 2 was characterised by a higher intakes of sugar-sweetened beverages (SSBs), fruit juice, and table sugar/preserves.
What they have done here is as follows. First they have judged energy dense foods and those high in saturated fat and/or free sugars as bad and foods high in fiber as good (call these variables). They have then used a statistical technique called RRR to identify the combinations of food groups that best capture these chosen variables. Table S4 in the Supplementary File reports how well the dietary patterns capture the chosen variables. Dietary pattern 1 is the combination of foods that best captures high energy density and low fiber density – it’s moderately OK as a representation of high saturated fat and high free sugars. Dietary pattern 2 is much better for free sugars, useless for energy density and fiber density and moderately OK for saturated fat. However, taken together, these two dietary patterns capture what the researchers set out to look for.
Both dietary patterns are fictious. Neither say much about intake of the healthiest foods: meat; fish; eggs; and dairy products. I would expect people who consume chocolate and confectionery to consume fizzy drinks, but apparently those are different types of people.
Over an average of 4.9 years of follow-up, 4,245 cases of total CVD, 838 cases of fatal CVD, and 3,629 deaths from any cause occurred. This was during 907,431 person years of follow-up. The researchers reported a positive linear association between dietary pattern 1 and the risk of total CVD (risk ratio 1.07), fatal CVD (risk ratio 1.07) and all-cause mortality (risk ratio 1.08). These risk ratios were for each standard deviation increase in dietary pattern 1 z-score, which is a large difference (Ref 5). The researchers reported a nonlinear association between dietary pattern 2 and the risk of total CVD (risk ratio 1.02), fatal CVD (risk ratio 1.02) and all-cause mortality (risk ratio) 1.01 – the latter was not statistically significant and so should be ignored (Ref 6).
Figure 2 in the paper split people into five groups (quintiles). It compared the fifth of people consuming a diet most like dietary pattern 1 with the fifth of people consuming a diet least like dietary pattern 1. It claimed that the risk ratio for those in the top group for total CVD incidence was 1.40 compared to those in the bottom group. This was reported in the media as “Two distinct dietary patterns featuring fatty and sugary favourites and which appeared to be linked to a higher risk of heart disease and death emerged from the food diaries. Those whose diets were most closely aligned to one of these patterns were up to 40 per cent more at risk” (Ref 7).
The three flaws of epidemiology
These still apply whether or not we’re looking at an individual food or a dietary pattern:
1) This is association not causation.
The top 10 foods that defined dietary pattern 1 were: chocolate and confectionery; butter and other animal fat spreads; low-fiber bread; table sugars and preserves; grain-based desserts; sugary drinks; high fat cheese; crisps and savoury snacks; alcohol; and milk-based desserts (think ice cream). With the exception of butter and cheese, that is all junk food. Leave butter out of this – the average person in the UK consumed 33g of butter per week (fewer than 5 grams a day) in 2018/2019 (Ref 8). A smudge on one’s fingernail of butter a day is not killing us. I am not surprised that the other foods are associated with CVD and deaths. In fact…
2) This is relative not absolute risk.
… what does surprise me is how tiny the risk differences are. Even taking the biggest claim – for total CVD having a (relative) risk ratio of 1.4 when the highest quintile is compared with the lowest – this is still way off any likelihood of being causal. The absolute risk is minute. With 907,431 person years and 4,245 CVD events, the event rate is 0.47% for the study. A 40% difference on this would be an event rate of 0.39% vs 0.55%. As I often say, who cares?
3) The healthy person confounder.
The usual healthy person confounder was reported in the paper. Dietary pattern 1 was associated with a higher proportion of men, current smokers, less physical activity, higher prevalence of obesity and hypertension. Dietary pattern 2 was associated with a higher proportion of men again and less physical activity again. However, interestingly, this pattern was also associated with a lower prevalence of obesity and diabetes.
Added flaws with this paper
As well as the three flaws that are typical of all epidemiological studies, there were some additional ones with this study:
1) The assumptions undermined the objectivity of this study from the outset. By making assumptions up front – that free sugars, saturated fat, and energy density are bad and fiber is good – the researchers got what they were looking for.
2) The paper does not appear to have adjusted for the full healthy person confounder. As a general tip, papers report what they adjusted for – not what they didn’t adjust for. So, if the paper doesn’t report that they adjusted for, say, obesity, then they didn’t. This paper reported adjusting as follows: “All analyses were stratified by sex and region (Scotland, Wales, England) and adjusted for ethnicity, a measure of deprivation (Townsend), education group, smoking status, physical activity, energy intake, and menopause status.”
It didn’t, therefore, adjust for obesity, hypertension, or diabetes. Obesity rates were 24% in the top group for dietary pattern 1 vs. 16% in the bottom group. Diabetes rates were 5.3% in the bottom group of dietary pattern 2 and 2.6% in the top group. Confounders were not consistent, therefore or fully adjusted for.
3) The paper reported a number of limitations of the study as follows. The usual unreliability of dietary recall questionnaires. The CVD risk factors were only measured at baseline, which was one to four years before the dietary data collection. This is significant, as someone could have been diagnosed with CVD before the dietary questionnaire and changed their diet as a result. The paper reported another limitation of the study as “dietary patterns may only be applicable to a UK population, since the combination of foods is likely to be culturally specific.” Bread, butter and jam – yes – quintessentially British! Oh, and British pensioners, not young people.
This was an interesting paper to review for the new approach and this is an approach likely to be repeated in the future. The media takeaway message was “’British diet’ staples like white bread, butter, jam, fruit juice and chocolate are sending us to an early grave, researchers warn.” My takeaway messages were – if you assume that sugar, fat and energy are bad and fiber is good and then go looking for foods that are high in sugar, fat and energy and low in fiber, you will find them. Primarily you will find them in junk food and that junk food was, unsurprisingly, associated with heart disease and deaths. Surprisingly, the association of a trolley load of junk food was not as strongly associated with bad outcomes as I would have expected. This echoes the findings of the French NutriNet-Santé study, which found that diets high in ultra-processed foods were not as strongly associated with cardiovascular disease as we might expect (Ref 9). Bad diets are not good, but they might not be as bad as we think.
Until the next time
All the best – Zoë
Ref 1: Gao et al. Associations between dietary patterns and the incidence of total and fatal cardiovascular disease and all-cause mortality in 116,806 individuals from the UK Biobank: a prospective cohort study. BMC Medicine. April 2021. https://bmcmedicine.biomedcentral.com/articles/10.1186/s12916-021-01958-x
Ref 2: https://www.zoeharcombe.com/2021/04/meat-disease-again/
Ref 3: http://www.ukbiobank.ac.uk/wpcontent/uploads/2011/11/UK-Biobank-Protocol.pdf
Ref 4: Liu et al. Development and evaluation of the Oxford WebQ, a low-cost, web-based method for assessment of previous 24 h dietary intakes in large-scale prospective studies. Public Health Nutr. 2011.
Ref 5: The risk ratios with confidence intervals were total CVD (hazard ratio 1.07, 95% confidence interval 1.04 to 1.09), fatal CVD (1.07, 1.02 to 1.13), and all-cause mortality (1.08, 1.05 to 1.11).
Ref 6: The risk ratios with confidence intervals were total CVD (1.02, 1.01 to 1.03), fatal CVD (1.02, 1.01 to 1.04), and all-cause mortality (1.01, 1.00 to 1.03); the latter being non-significant.
Ref 7: https://www.dailymail.co.uk/news/article-9497923/British-diet-white-bread-butter-chocolate-sending-early-grave-researchers-say.html
Ref 8: https://www.statista.com/statistics/281189/weekly-consumption-of-butter-in-the-united-kingdom-uk/
Ref 9: https://www.zoeharcombe.com/2019/07/processed-food-cvd-nova/