Substitutions of animal-based foods consumption and risk of hypertension: a cohort study

This study assessed the association between the consumption of animal-based foods (ABFs) and hypertension, a recognized risk factor for cardiovascular disease. The adverse effects of red and processed meat (RPM) consumption and the beneficial effects of dairy products consumption and other ABFs have been discovered separately; however, the constrained nature of food intake was typically ignored. We aimed to assess the substitution effects between RPM and other ABFs using compositional transformation substitution analysis and further compared this analysis with the traditional substitution analysis method. Methods We followed 5394 Chinese adults aged 18-60 years at baseline in the China Health and Nutrition Survey from 2004 to 2011. Food consumption was assessed using a combination of individual-based consecutive 24-h recalls and household-based food weighing approaches. Both traditional substitution analysis and substitution analysis based on compositional transformation were used to assess substitution effects.

food consumption into account. This is a flexible approach to estimate substitution effects under different substitution patterns, to obtain personalized estimation effects based on different consumptions of ABFs at the individual level, and to provide an individualized dietary recommendation.

Background
Hypertension is a major risk factor for cardiovascular diseases and is the leading risk for premature deaths worldwide, contributing to more than 10 million deaths in 2016 1 . From 1975 to 2015, the number of adults with hypertension increased from 594 million to 1.13 billion and is predicted to grow to nearly 1.56 billion by 2025 2, 3 . As the prevalence of hypertension escalates, preventive interventions are urgently needed. Diet is considered to be a critical and amendable factor for hypertension 4 . Accordingly, understanding how dietary components impact on hypertension is most important in constructing preventive strategies from public health perspectives.
Animal-based foods (ABFs), especially meat, are a major component of dietary patterns. Red meat represents 58% of all meat consumed in the USA and its consumption continues to rise 5 . In China, the consumption of ABFs keeps increasing in line with the economic growth, especially regarding meat consumption which has leaped 18 percent, from 64 million to 78 million (metric) tons from 2007 to 2012 6 . As a rich source of protein, minerals and vitamins, ABFs have attracted considerable attention regarding hypertension prevention 7 . A significant amount of evidence from observational studies has shown that a higher RPM intake is associated with an increased risk of hypertension 7-10 .
Meanwhile, the beneficial effects of other ABFs, such as dairy products and eggs, have also been probed 11, 12 . As people do not consume food in isolation, current dietary interventions no longer focus on single ABFs but advocate for the substitution of unhealthy ABFs with diverse healthy foods of animal origin 13 . Therefore, it is essential to explore the substitution effects of healthy and unhealthy ABFs on hypertension.
The traditional substitution analysis (TSA) is based on regression modeling, including the sum consumption of interested foods and all foods, except one, as explanatory variables 14 . However, these variables are exactly correlated and cannot independently and correctly predict the substitution effects 15 . Besides, TSA can only predict one-to-one substitution effect theoretically without considering the individual dietary differences 16 . Thus, whether the predictive results of TSA can correctly and practically guide a reasonable diet behavior at the individual level is questionable.
Another substitution method based on compositional data (CoDa) analysis, which was originally used in the time allocation for different behaviors, can be used in nutritional substitution analysis 17, 18 . Daily individual food consumption, just like time allocation, is essentially finite data and one food consumption cannot be shifted without compensatory changes in the others. Thus, data regarding food consumption has a constrained property and exists in the constrained simplex space with the definition of operations differing from those in the real Euclidean space 19, 20 . The CoDa analysis that can work well with constrained data by translating data into real space through the application of isometric log-ratio (ilr) coordinate transformation is becoming the new approach to deal with dietary data 21, 22 .
In this study, the compositional transformation substitution analysis (CoTSA), which combines CoDa with the notion of replacement, is first proposed in food substitution. We used CoTSA to implement substitution of RPM and other ABFs and explored the association between different types of replacements and the risk of hypertension. We further compared the findings obtained from the CoTSA with the results from the TSA.

Study population
The China Health and Nutrition Survey (CHNS) is a large-scale cohort study that was launched in 1989 and has finished eight follow-up panels up until 2011. Furthermore, the CHNS is a prospective, ongoing study that explores the socioeconomic factors, demographic changes, and nutritional behaviors that affect the health status in China. A multi-stage, random-cluster sampling procedure was employed to enroll samples in each of the 9 provinces where the total population accounts for roughly 50% of the Chinese population and represents populations from rural, urban and suburban areas. Details concerning the study project and design have been presented in the cohort profile 23 .
The study was approved by the institutional review committees of the University of North Carolina at Chapel Hill and the Chinese Institute of Nutrition and Food Safety, China Center for Disease Control and Prevention. All participants provided an informed consent form.
A total of 8240 volunteers aged 18-60 years old were included between 2004 to 2011 on the ground that the participant retention maintained was higher than 80% across all surveys after 2004 23, 24 .
Participants with hypertension, diabetes, myocardial infarction, stroke, and cancer, and women who were pregnant at baseline, were not included (n = 2070). Eighty-one participants with missing data on smoke and job status and BMI value at baseline were excluded, as well as 726 vegans, leaving 5394 participants for our final analysis (Fig. 1).

Covariates data collection
Baseline questionnaires provided information about age, gender, residential areas, marital status, educational level, working status, smoking status, and alcohol consumption. Residential areas were divided into urban or rural areas. Marital status was grouped into never, married, divorced, widowed, or separated. Educational attainment was classified into two levels according to a seven-category primary question: middle school degree or below (never, graduation from primary or middle school), high school degree or above (high school, technical or vocational, university or college, master's degree or higher). Working status was classified into jobless, employees, and retirees. Smoking status (never, former, or current) and alcohol consumption (never, current drinking) were also documented in the questionnaire. The weekly time spent on performing physical activities, being sedentary, and lying in bed, were collected. Body mass index (BMI) was calculated by dividing weight in kilograms by height in square meters.

Dietary assessment
In the CHNS, dietary surveys were assessed over three consecutive 24-h recalls at the individual level in combination with weighing inventory at the household level over the same three day period 25, 26 .
The consecutive three days were randomly allocated from Monday to Sunday and were equally balanced across the seven days of the week for each sampling unit. At the individual level, all food consumed at home and away from home each day was recorded with food models and pictures by trained interviewers. At the household level, food consumption changes in inventory from the beginning to the end of each day were measured with a weighing scale. By comparing the average daily dietary intake (g/day) calculated from the household survey and 24-hour recall, significant discrepancies were found and solved by revisiting at both levels. One serving of ABFs was defined as one Liang, which is the commonly used unit in China and is equivalent to 50 g or 50 ml. The classification of food groups, such as ABFs, including RPM, poultry, seafood, dairy products, and eggs was based on the dietary consumption data using the Chinese Food Composition Tables 2 7 .

Ascertainment of incident hypertension
Arterial blood pressure was measured on the right arm three times using a mercury sphygmomanometer with the cuff maintained at the heart level in the sitting position after at least a 10 minutes rest. An average of three blood pressure measurements was used for hypertension ascertainment. Hypertension ascertainment was based on the follow-up survey in 2011 by meeting at least one of the following criteria: (1) a systolic blood pressure ≥ 140 mmHg; (2) a diastolic blood pressure ≥ 90 mmHg; (3) a self-reported diagnosis or new diagnosis of hypertension by a physician; (4) a self-reported antihypertensive treatment.

Statistical analysis
The potential confounding covariates were examined by the cross-tabulation of dichotomous ABFs.
The mean ± standard deviation or percentage was presented to describe continuous or categorical variables appropriately. The general linear model and Pearson χ 2 test were performed for examining differences between groups for continuous and categorical variables, respectively. The Pearson correlation coefficient was used to detect the baseline correlations among the different ABFs components. Based on the multivariable Cox proportional hazards model, two substitution analysis methods were conducted to explore the association between substituting RPM for another or all other ABFs, with the risk of hypertension. The proportional hazards assumption was not violated by observing the -ln(-ln) survival plots.
Pol, Dry, Egg, Sea and ABF represented the consumption of poultry, dairy products, eggs, seafood, and total ABFs, respectively. For the TSA, the effects on outcome f(x) of substituting RPM for other factors, like Pol, could be estimated by taking RPM out of the model that includes Pol, Dry, Egg, Sea, and ABF as the following formula: β was the estimated regression coefficient. The multivariate model was adjusted for w e on behalf of the covariates of age, gender, BMI, marriage, urban location, education degree, working status, smoking, alcohol consumption, time spent on physical activity, sedentary and lying on bed, consumption of vegetable, fruit, nuts and legumes, whole grain, sweetened beverages and sodium.
The hazard ratio (HR) and its 95% confidence intervals (CIs) were reported. The HR for β 1 was explained as the risk of hypertension by replacing one serving of RPM with the same amount of poultry.
For the CoTSA, the 5 categories of ABFs had a restricted sample space of 5-simplex which was not compatible with operations in real space without transformations. We obtained the compositional values of ABFs for individuals by an operation called closure, which divided out a set of ABFs by its total. These compositional values were actually proportions that adding up to 1. Based on the compositional values, the isometric log-ratio (ilr) coordinates constructed by the sequential binary partition (SBP) process were used to accomplish the transformation. The SBP was based on a combination of a priori information of interest and a data-driven approach of the principal balance.
The ilr coordinates pertaining to the above SBP were defined as the following: The coordinates were independent of each other. When substituting RPM for other all ABFs, RPM was changed to [(1+r) RPM], and the remaining components were equally reduced through multiplying by (1-s), where rRPM = s(Pol + Dry + Egg + Sea). Accordingly, z i turned into the following coordinates: Based on the Cox proportional hazards model with ilr coordinates, HR was given as, , and the 95% CI of HR one−for−many was The Similarly, other substitution patterns could be derived based on this method.
Data cleansing and consolidation were completed using SAS software version 9.4 (SAS Institute Inc., Cary, NC, USA). Statistical tests were performed using R software (version 3.6.1). Statistically significant meant a P<0.05 and the 95%CI of HR dose not contain 1. All P values were 2-tailed.

Baseline characteristic distribution
Data from 5394 participants were analyzed whom 1267 were newly diagnosise hypertensives. The mean follow-up period was 6.2±0.02 years. The baseline characteristics were summarized according to the dichotomous consumption of ABFs (Table 1). The main consumption of ABFs was RPM, with an average of 1.68 servings per day. Participants with higher consumption of all five ABFs were more likely to have a higher educational level, to spend more time sitting and less time doing physical activity, and to consume more fruit. Males, drinkers, and individuals who consumed more sugarsweetened beverages, also tended to consume more meat and dairy products. Higher consumption of RPM meant more vegetable intake and less whole grains intake. Participants with higher consumption of poultry consumed fewer vegetables, nuts and legumes, as opposed to those who consumed more dairy products. Contrary to individuals with a higher intake of eggs, participants with a higher seafood intake ate less whole grain. Participants, who consumed more RPM, seafood, dairy products, and eggs, consumed more sodium.

TSA
The TSA model was used to predict the associations between changes (serving/day) in specific ABFs components and risk of hypertension (see Table 2). A meaningful substitution effect was found between RPM and eggs. Model A suggested that replacing 1 serving/day of RPM with eggs were associated with a 14.8% lower risk of hypertension (HR = 0.852, 95%CI: 0.781-0.934). As expected, the opposite association was seen when eggs was dropped out of the model (Model E). The line where risk of hypertension varies with the amount of substitution is shown in Fig. 2(A). The slopes of lines are different between different ABFs substitute patterns. With the maximum slope of 0.3%, substituting RPM with eggs was associated with the greatest benefit in reducing the risk of hypertension.
However, there was a strong correlation between individual ABFs and total ABFs, especially between RPM and total ABFs consumption (r = 0.64), which implies the multi-collinearity (Table 3).

CoTSA
The regression results with ilr coordinates are presented by modeling a multivariable Cox proportional hazards regression (Table 4). Coordinates with RPM in the numerator were a significant predictor of hypertension. The first ilr-coordinate that contained all the information, regarding RPM relative to the remaining ABFs, was a significant predictor of hypertension (HR = 1.015, 95%CI: 1.003-1.027).
The means of the compositional values were 50%, 23%, 17%, 7%, and 3% for RPM, eggs, seafood, poultry, and dairy products, respectively. This implied that the average consumption proportion of different ABFs varied, and the ranges of substitution between different ABFs were dissimilar ( Fig. 2(B)). Since the mean consumption proportion of RPM was 50%, the substitution curve between RPM and other ABFs was plotted in a symmetrical range, while the substitution curve between RPM and another ABF was drawn in an asymmetrical range. with dairy products, while the amount of substitution was 64% for participant B. However, the risks estimated by TSA were consistent regardless of the differences in baseline ABFs consumption.
We further estimated the substitution effects between RPM and poultry, seafood, eggs, and all other ABFs in participant A and participant B respectively (Fig. 4). Whether based on the average or individual (Participant A and participant B) consumption of ABFs, substituting dairy products for RPM was associated with a greatest benefit in reducing the risk of hypertension. This implies that, to achieve the same HR, the lowest amount of substitution was required between RPM and dairy products. For instance, the risk of hypertension for participant B could be reduced by 5% (HR = 0.95) most effectively by replacing approximately 62% of RPM with dairy products, compared to the replacement of about 65% RPM by all other ABFs, and 68% by eggs, poultry, and seafood (Fig. 4), while for the average consumption of ABFs the amounts of replacement were 43%, 47%,and 48% respectively ( Fig. 2(B)).

Main findings
In this prospective cohort of Chinese adults, for TSA models, a decrease in RPM consumption and a simultaneous increase in the consumption of eggs were associated with a lower risk of hypertension.
For CoTSA models, substituting RPM for other ABFs with the equivalent consumption, including one-toone and one-to-many substitution, were associated with a higher risk of hypertension.
Systematic reviews and meta-analyses of prospective cohort studies have pointed out an adverse association of RPM consumption and a beneficial association of the consumption of seafood, eggs and dairy products, with hypertension 28, 29 . For instance, a recent meta-analysis that summarized 10 prospective cohorts, suggested that the highest level of red meat consumption was associated with a 22% higher risk of hypertension, while the highest level of egg consumption was associated with a 21% lower risk of hypertension, both compared to its lowest consumption level 28 . Furthermore, a systematic review supported favorable associations between dairy product consumption and risk of The TSA, that deals with modelling all ABFs except one of them, is limited to assessing only estimates for the replacement of one ABFs (e.g., dairy products) for another (e.g., RPM). Whereas, CoTSA breaks through this limitation and flexibly deals with the different substitution combinations within ABFs, such as the one-to-one replacement between RPM and another ABFs, or the one-to-many substitution between RPM and the total amount of ABFs considered in our current study, or the substitution between any ABFs which researchers are interested in.
Since the substitution effect is estimated based on equivalent consumption replacement, the ABFs should be replaced with a reasonable amount of consumption 16 . In this cohort, the average consumption of RPM was three times more than that of eggs (1. substituting dairy and its products for all RPM would bring the greatest benefit of decreasing risk of hypertension. Based on current evidences, the high content of potassium and calcium in dairy products has been associated with lower blood pressure 40, 41 . Moreover, substituting eggs, poultry and seafood for RPM were associated with a reduced risk of hypertension, respectively. However, the evidence on these foods is limited and need to be further confirmed 28 .

Limitations
Our study has drawn on a large prospective cohort study and proposed a flexible and efficient approach of CoTSA to assess the substitution effects of ABFs consumption. However, this study has several limitations. Although we have comprehensively controlled demographic, lifestyle, and dietary factors, we cannot ensure the causality of the observed association since the residual confounding factors, especially unmeasured confounders cannot be completely excluded. There are inherent disadvantages of CoTSA in that non-zero elements are required because of the log-transforms. We deal with essential zeros with random values between a tenth and a half of the smallest positive value, using the modified EM algorithm before transformation 15 . The participants we studied come from China, which may limit the generalizability of the results to populations from other countries.

Conclusions
In the results of CoTSA, replacing RPM with healthier ABFs was associated with a lower risk of hypertension, moreover, in the case of constant substitution, substituting dairy products for RPM was associated with a greatest benefit in reducing the risk of hypertension. While the TSA has only found that substituting eggs for RPM was associated with a lower risk of hypertension, our analysis provided

Consent for publication
Not applicable

Availability of data and materials
The datasets generated and/or analyzed during the current study are available in the China Health  Values are linear correlation coefficients among the different ABFs consumptions at baseline, p<0.05 indicates statistical significance.     The HR varies with the change in consumption of substitutes between RPM and dairy products with the method of TSA (the dotted line) and CoTSA (the solid line). The thin solid line represents participant A whose consumption proportions of RPM and dairy products were 10% and 45%; The thick solid line represents participant A whose consumption proportions of RPM and dairy products were 70% and 7%. The 95% CI of HR calculated by CoTSA within the horizontal dashed line range (L and U) contains 1.

Supplementary Files
This is a list of supplementary files associated with this preprint. Click to download. STROBE_checklist_cohort.doc