Diet synergies and mortality—a population-based case–control study of 32 462 Hong Kong Chinese older adults

Abstract

Background Food and drink are not consumed in isolation and can have complimentary effects enhancing or blocking the overall uptake of nutrients. We investigated how combinations of foods, drinks, and smoking affected mortality.

Method Adjusted logistic regression was used to assess the joint effect of healthy foods, less healthy foods, smoking, and alcohol use on mortality in a case–control study of all Chinese adults aged 60 or over who died in 1998; 21 494 dead cases (81% of all registered deaths) and 10 968 live controls were included.

Results There was a significant trend of increasing all-cause mortality risk with decreasing healthy food consumption (P < 0.001), and the increase in risk was significantly steeper for people with high intakes of less healthy food (P for interaction <0.001). There was a steeper risk from increasing less healthy food intake in ever-smokers and people not drinking tea regularly (P < 0.001), while the J-shaped relationship between alcohol and mortality differed in shape with level of less healthy food intake.

Conclusion Intake of some dietary items may modify the effect of others. An analysis framework explicitly recognizing complementary and potentially synergistic effects of food, drinks, and smoking could enhance our understanding of dietary epidemiology.

Introduction

A reductionist approach has been very successful in determining the relationship between diet and disease. However, specific nutritional items are not consumed in isolation, and the absorption, bioavailability, and action of dietary constituents may depend on what else is ingested. These potential synergies have mostly been investigated from the perspective of providing adequate nutrition to identify foods that should be eaten together, such as the protein complementation from bean and cereal combinations.¹ Others have examined how food combinations may block or promote the uptake of nutrients, such as iron or calcium.^2–4 Potential synergies have less often been investigated from the perspective of over-nutrition or chronic disease prevention to investigate whether there are any foods that affect the uptake of nutrients, which may be harmful in excess, or if there are foods which eaten in combination may prevent disease. Nevertheless, there is some evidence for such synergies. Tea may moderate lipid metabolism.^5–8 Certain nutrient combinations, such as tea and soy, are particularly effective in suppressing some types of tumour growth^9–11 or lipid oxidation.¹² The role of diet may also vary for smokers.^13,14 Thus, there is increasing recognition of the need for a synergistic approach^15–17 although no consensus on how to investigate synergies in epidemiological studies.

In assessing the overall effects of diet, studies have generally focused on data driven dietary patterns or theoretically derived diet scores.¹⁸ Data derived dietary patterns are typically obtained using factor analysis or cluster analysis. Factor analysis is premised on exposing underlying dimensionality, implicitly assuming the existence of independent orthogonal factors. Cluster analysis is designed to categorize subjects discretely. Although both approaches have uncovered elements of a healthy diet, they are predicated on linear combinations of dietary items, which will not specifically reveal their joint or synergistic effects nor provide insight into whether particular patterns are healthy because of the constituent items or combinations thereof. Empirically in this study, we explored the factor structure of dietary components using principal components analysis with varimax rotation but the resulting 2-factor solution grouped fish and meat consumption together (fruits, vegetables, soy, and dairy was the other factor grouping) and thus would have precluded the examination of our main research question of dietary synergy between a priori healthy and unhealthy food items. On the other hand, diet scores are typically uni-dimensional giving a ranking according to overall healthiness of the diet. However, diet scores have been operationalized in terms of two dimensions,¹⁹ giving the possibility of examining joint effects. As we aimed to investigate potentially synergistic effects of different dietary components, we followed the two-dimensional food score components of Michels et al.¹⁹ Each person receives a recommended food score (RFS), based on consumption of pre-defined healthy foods, such as fruits and vegetables, and a not recommended food score (NRFS) based on consumption of pre-defined less healthy foods, such as red meat. We examined the effect of RFS and NRFS in relation to all-cause mortality, using data from a large population-based case–control study in a homogeneous Asian population. Additionally, we considered their joint effect and whether the effect of less healthy food (NFRS group) or healthy food (RFS group) varied with tea consumption, alcohol use, or tobacco consumption.

Subjects and methods

The ethnic Chinese in Hong Kong, comprising 95% of the total population, are a culturally homogeneous group with a common cooking style. The staple food is white rice, accounting for ∼80% of cereal consumption,²⁰ with a correspondingly low consumption of wheat and potatoes. The main cooking oils are corn oil, peanut oil, and safflower oil.²⁰ Typically, the main meals (i.e. lunch and dinner) consist of rice or noodles, vegetables and meat, fish or tofu, possibly followed by some dessert, such as fruit. Chinese meals are usually served as a selection of separate dishes to be shared around the table. Although eating habits are changing with increasing Westernization, this is less likely to be the case for older people.

The LIMOR (LIfestyle and MORtality) study is a population-based case–control study of adult deaths among ethnic Chinese residents in Hong Kong between mid-December 1997 and mid-January 1999. The study was originally designed to examine the effect of smoking as a risk factor for mortality.²¹ It captured 81% of all deaths registered during the study period. Information on demographic characteristics (age, sex, education, and type of job) and health behaviours (smoking, use of alcohol, leisure exercise, and dietary habits) 10 years before the death of the case was collected from the person registering the death, who was usually one of the more educated members of the deceased's family. The same information was also obtained for a living person, other than the informant, either the decedents' spouse or a person, preferably aged 60 or above, whose habits 10 years ago the informant was most familiar with. When the informant was the spouse an eligible control was seldom obtained, therefore there were more cases than controls, especially in the older age-groups, and our method tended to select female controls (since wives generally survive their husbands). Information about both cases and controls was obtained from a third-party, who was usually an adult child of the subject, often living in close, multi-generational Chinese families. Proxy reports were chosen first for expediency, second because of the collectivist values of Chinese society, and third because proxies have been shown to be capable of providing reliable answers to simple questions,^22,23 particularly in Hong Kong.²⁴ In addition proxy reports of food intake are potentially preferable because self-report dietary data validates poorly against associated bio-markers²⁵ probably because of socio-cultural pressures to present oneself as eating a ‘correct’ diet.^26,27 Issues of self-presentation are unlikely to be so strong when reporting on another person. Proxy reports of food intake in a similar setting, i.e. adult reports on other family members, such as children, validated better against bio-markers.²⁸ Thus, simple food frequency questions were asked about the consumption of nutritionally significant foods groups, which are commonly eaten in Hong Kong, vary between individuals,²⁰ and were thought to be related to disease risk at the time of the study. Simple food frequency questions on food groups, such as fruits and vegetables, have been found to relate to relevant bio-markers²⁹ and have been validated in developed Chinese societies.³⁰ The food groups chosen were fresh fruits, fresh vegetables, soy products, fish (excluding salted fish), Chinese tea, dairy products (specified as milk, cheese, and ice cream), and meat. Potatoes, bread, and wheat products were not included because of relatively low consumption, while assessing oil use would have been unlikely to produce useful data and would have been potentially less relevant when the main oils used are unsaturated. Consumption was specified as ‘not at all’, ‘monthly less than once’, ‘monthly 1–3 times’, ‘weekly 1–3 times’, and ‘weekly at least 4 times’. These data are suitable for ranking the intake of certain foods, but do not enable us to assess overall dietary content, the intake of specific nutrients, or total dietary intake; as such they are useful for the assessment of food groups or dietary patterns, but not of food constituents. As a reliability check, repeat interviews were conducted by telephone on a random sample of 235 cases and 106 proxy controls, 3 weeks after initial interview on average. The percentage agreement to within one category for case and controls, respectively, were 92 and 97% for fruit, 98 and 100% for vegetables, 90 and 93% for soy products, 98 and 99% for fish, 90 and 91% for tea, 70 and 66% for dairy products, and 99 and 98% for meat.

Statistics

The outcome considered is all-cause mortality. Based on the RFS/NRFS scoring system¹⁹ and the likely constituents of food groups in Hong Kong in 1998,²⁰ the food groups were classed as RFS (fresh fruits, fresh vegetables, soy, and fish) or NRFS (meat and dairy products). In Hong Kong in 1995, most meat consumed was red meat (pork, beef, lamb, organ meat).³¹ Dairy products include ice cream, cheese, and milk; most milk consumed at the time was full fat.²⁰ Vegetables are mainly green vegetables, such as Chinese cabbage, Chinese kale, lettuce, and broccoli.²⁰ Fruits include apples, oranges, bananas, and a wide range of tropical fruits, such as melon, papaya, and mango.²⁰ Subjects were given a score according to how often they ate foods in these groups. The score for each item was weighted according to the proportion of days on which it was consumed, e.g. 2/7 for 1–3 times a week. The weighted scores were summed for RFS and NRFS foods, and divided into groups representing a ranking of food scores for the controls, such that there were at least three groups for each dimension (to investigate dose–response relationships) and we avoided having >40% of controls in the highest or lowest group.

Chinese tea is tea without milk often drunk with meals. In Hong Kong 90% of the tea imported at the time of exposure ascertainment in 1988 was black tea.³² This is in contrast to Japan, where tea is mainly green. Tea was categorized on frequency into less than weekly, 1–3/week, and 4 or more times per week.

On the basis of the potential benefits of light drinking, alcohol was categorized as never, fewer than 8 drinks per week, 8 or more drinks per week, and ex-drinkers, based on the frequency of alcohol use, typical consumption per occasion, and type of alcohol. A drink was defined as equivalent to 12 g of ethanol. Smokers were classified into current smokers, ex-smokers, and never-smokers. Ex-smokers and ex-drinkers are a heterogeneous group who might have changed their habits as a health protection measure or as a response to ill health. We excluded ex-smokers and ex-drinkers when investigating the potential moderating effect of smoking or use of alcohol, thus in a mainly older sample we are comparing surviving and possibly healthier smokers and alcohol consumers with lifelong abstainers, which will make our findings conservative.

Unconditional logistic regression adjusted for potential confounding factors was used to assess the effect of NRFS and RFS group independently on all-cause mortality. The joint association of NRFS and RFS with mortality was examined from (i) the significance of interaction terms within a multiplicative model and (ii) the effect of variables representing the combination of RFS and NRFS groups. Additionally, we considered in a similar way whether the effect of NRFS or RFS group varied with tea, smoking, and use of alcohol. Men and women, and all ages were considered together as there is little evidence that diet has a different effect on men and women or by age-group. Potential confounding factors considered were age (in 5 year age-bands), sex, leisure-time exercise (<1/month or ≥1/month), educational level (none, primary, secondary/tertiary), type of longest held job (none, sedentary, light/medium, or heavy manual), type of housing (public/hut/shared, self-owned, or other), and where appropriate, other food score, tea, smoking (current smoker, ex-smoker, or never-smoker), and alcohol history (never, <4/weeks, 4+/week or ex-drinker). The nature of the data collection made it impossible to collect body mass index but that should be correlated with diet and physical activity.

To minimize the potential for reverse causality and to obtain more conservative results, the cases who had been in very poor health or unable to go outdoors alone more than 10 years previously were excluded, as these people might have changed their diet in response to their health problems. Consistent with the control selection and to ensure an adequate number of controls representative of the population in terms of socioeconomic status, we restricted our analysis to cases and controls aged 60 or over. The project received ethics approval from the Ethics Committee of the Faculty of Medicine, University of Hong Kong.

Results

After excluding all the under-60s (n = 6386) and cases in very poor health or housebound 10 years previously (n = 1395) there were 21 494 cases and 10 968 proxy controls. Consistent with Hong Kong mortality patterns, cancer was the leading cause of death (33.7%), followed by cardiovascular disease (27.8%), respiratory diseases (excluding lung cancer) (21.4%), and other (17%). The cases and controls had similar birthplaces, education, type of housing, jobs, and ever use of alcohol (Table 1). The mean (SD) age of the cases and controls were 73.6 (8.1) and 75.2 (8.4) years for men but 71.0 (7.3) and 79.7 (9.2) years, respectively for women. These differences in age were controlled for automatically in the analyses. As reported previously, a higher proportion of the cases were ever smokers and rarely took exercise.^33,34 In keeping with the traditional diet in a wealthy, coastal Chinese city³¹ most subjects ate meat (90%), fish (90%), vegetables (95%), and fruit (80%) at least 4 times a week, but fewer ate soy products (21%) or dairy products (16%) as frequently. Healthy food scores for the controls created three RFS groups, with 26% in the group eating least healthy food, 55% in the middle group and 19% in the high group. Less healthy food scores had a bimodal distribution and fell most evenly into four NRFS groups with 38% in the lowest group, 19% in the next group, 29% in the next, and 14% in the highest group. The proportion of people eating high amounts of healthy food was greatest in the controls in both age-groups, but there was little difference in less healthy foods between the cases and controls (Table 1). Smokers and drinkers were less likely to be in the high healthy or high less healthy food groups, while regular tea drinkers were more likely to be in the high healthy and high unhealthy food groups.

Table 2 shows the effect of NRFS and RFS on all-cause mortality in three models, adjusting for an increasing number of confounding factors. There was no effect of NRFS group in the first two models, which adjusted for age, sex, and education (Model 1) and then also for housing type, job type, and leisure-time exercise (Model 2). However, the effect of NRFS group changed when the RFS group, tea frequency, alcohol use, and smoking status were entered into the model (Model 3). In the fully adjusted model there was a linear dose–response relationship between higher NRFS group and increased risk of mortality. In contrast, a higher RFS group was associated with lower mortality in all three models, with consistent odds ratios and clear dose–response relationships in each model. Deaths from injury and poisoning did not show the same pattern of relationships with NRFS or RFS groups. In the fully adjusted model (Model 3) there was no evidence that the effect of NRFS or RFS group varied with sex (P for interaction >0.5 in both cases—data not shown). The relationships in the fully adjusted model were essentially unchanged when we restricted the sample to the potentially most reliable informants, i.e. those living with the case or control 10 years ago during the time of exposure ascertainment (Model 4), or those reported on by their adult children (data not shown). Subsequent analyses (as presented in the following tables) gave similar results for all the subjects, or the subjects living with the informant at the time of exposure ascertainment, or for the subjects reported on by their adult children, so results for all subjects are presented.

The effect of NRFS group varied with RFS group (Table 3), and this effect persisted within analyses stratified by sex or age-group (60–74 and 75 or over) (data not shown). Table 3 also shows the joint effect of NRFS group and RFS group on all-cause mortality, with high RFS group and low NRFS group as the reference category. In the lowest RFS group, there was a clear dose–response relationship for an increased risk of mortality with increasing NRFS group. In the middle RFS group there was also a clear dose–response relationship for an increased risk of mortality with increasing NRFS group. However, in the high RFS group there was no clear relationship between increasing NRFS group and increasing risk. Within NRFS and RFS combinations, the highest risk was observed for the combination of highest NRFS group and the lowest RFS group. The effect of low RFS group was more marked in the higher NRFS groups than in the low NRFS group. In the highest NRFS group there was more than double the risk between the low RFS group and the high RFS group while in the lowest NRFS group the difference was smaller.

Similarly the effect of NRFS group also varied with tea drinking frequency, smoking status, and alcohol use (Table 4). Table 4 also shows the joint effect of NRFS group and tea drinking, NRFS group and smoking, and NRFS group and alcohol use on all-cause mortality. In the lower two tea drinking groups, there was an increasing risk of mortality with increasing NRFS group. However, in the high tea drinking group, although there was still a trend for increasing risk with increasing NRFS group, the gradient was attenuated. Within NRFS and tea combinations, the highest risk was observed for the lowest tea frequency and highest NRFS group. The effect of infrequent tea drinking was more marked in the higher two NRFS groups than the lower two NRFS group.

For current smokers there was a clear trend for higher NRFS group to be associated with greater risk. However, for never-smokers there was only a small increase in risk in the higher NRFS, and the trend was not significant.

There was a clear trend for increased risk with higher NRFS group for the never-drinkers, there was also a trend for a increased risk with higher NRFS group in the light drinkers (<8 units per week), however there was no clear trend for drinking >8 units per week. The U-shaped relationship between alcohol use and mortality was very clear in the lower two NRFS group. In the medium high NRFS group the relationship appeared more J-shaped, and in the highest NRFS group it appears that never drinking is most risky. However additional analysis of alcohol use within the highest NRFS group found that there was a U-shaped relationship that became evident at higher levels of alcohol use, such that there was almost the same risk for never-drinkers and people drinking 21 or more units per week (data not shown). Thus NRFS group appeared to modify the shape of the U, rather than change the nature of the relationship.

The effect of RFS group did not vary with smoking status (P = 0.94) or use of alcohol (P = 0.36), but did vary with tea drinking (Table 5). Table 5 also shows the joint effect of tea drinking and RFS group on mortality. In the lowest tea-drinking group, there was a trend for an increase in mortality with decreasing RFS group. In the highest tea drinking group there was a strong trend for increasing mortality with lower RFS group. There was also a strong trend for increasing mortality in the higher two RFS groups as tea drinking decreased. However, for the low RFS group the effect of tea drinking was less clear.

Discussion

In common with other research showing that the consumption of fruits and vegetables is protective against major causes of death such as cancer³⁵ and heart disease,³⁶ we have shown that consumption of healthy foods including fruits and vegetables is associated with lower all-cause mortality. Consistent with a previous study using similar dietary dimensions, we found that in an Asian population higher use of healthy foods had a protective effect, but higher use of less healthy foods had a smaller impact on mortality.¹⁹ In addition, we extended the analysis to demonstrate in our population that use of less healthy foods did not have a consistent impact on mortality, varying with healthy food use, tea drinking, possibly alcohol use, and smoking status. Our finding that smoking may magnify the harm of a less healthy diet is consistent with a previous ecological study that suggested the effect of smoking on lung cancer might vary with fat intake.³⁷ Smokers may also have different diets.³⁸ Although the best strategy for reducing the harm of smoking is cessation, a better diet would be beneficial. We found that the effect of high intake of less healthy food was partially offset by a high intake of healthy food, frequent tea drinking, and possibly moderate alcohol consumption. However at lower intakes of less healthy foods use of healthy foods or tea was less relevant. Finally it appeared that healthy food and tea combined decreased the risk of mortality more than either singly. Statistical interaction does not imply biological synergy but there are several mechanisms by which the effect of some foods could be moderated by the intake of other items, for example where some items specifically neutralize less healthy foods or where healthy foods complement other items.

Many healthy foods are plant based and so provide fibre. Fibre alters the physical properties of the food being digested, by increasing bulk, owing to its own volume and the resulting increased bacterial mass, which decreases transit time. Both the faster transit and the greater volume decrease duration of mechanical contact between potentially harmful substances, such as carcinogens from cooked meat, and the intestine wall, and absorption is also decreased.³⁹ However, faster transit time and decreased absorption would be of little value in the absence of harmful substances and could even lead to deficiencies. Bowel movements have been investigated in relation to colorectal cancer with somewhat mixed results.^40–43 However, bowel movements alone cannot approximate the complex process of food digestion and passage within the body. Hitherto, nutritional epidemiology has mainly been concerned with intake, not length of time in the body, internal flora, or absorption, so we know little about these potentially moderating effects, which deserve greater consideration.

Plant matter also has chemical properties that may inhibit the formation of harmful substances or neutralize them. Fruits and vegetables may inhibit nitrosation⁴⁴; associated fibre may absorb hydrophobic carcinogens.⁴⁵ Many plants contain anti-nutrients, which prevent absorption of potentially harmful substances, for example plant sterols reduce cholesterol absorption, particularly in people with high fat intakes,^46,47 while yams reduce fat absorption in mice.⁴⁸ Both tea and alcohol may share some of the neutralizing qualities of other plant-based foods. Black tea inhibits nitrosation⁴⁹ and has a wide range of anti-mutagenic properties.⁵⁰ Tea may enhance insulin action,⁵¹ while moderate alcohol intake enhances insulin sensitivity.⁵² Tea and alcohol may moderate lipid absorption or metabolism^7,53 with particular benefit for high fat consumers.⁵⁴ However, these neutralizing effects are not always advantageous, for example alcohol also disrupts folate uptake and metabolism, such that alcohol users need higher folate intakes to avoid increased mortality, while folate intakes make little difference for non-users.⁵⁵ The mechanism for synergist effects of tea and some plant compounds are not yet clear, but could be due to a complementation process. Building on findings of synergies dietary items could be considered within a paradigm of potentially positive and negative qualities based on their biological action. An item whose action is mainly neutralizing might have a harmful effect in the absence of its partner, and a more obviously beneficial effect in those consuming large amounts of its partner, and possibly a U-shaped relationship overall, as for alcohol. Recent studies that found increased mortality in those taking large doses of vitamin E or antioxidant supplements^56,57 would be consistent with this paradigm. Examining the effect of dietary items with respect to doses and other intakes, as in this study, may help unravel these complex relationships and reconcile some of the conflicting findings in nutritional epidemiology, which could be due to different levels of intakes of other items between groups studied.

Strengths

Our study is based on an ethnically homogeneous, East Asian population. Our study was large enough to have sufficient power to investigate interactions. We were able to adjust for potentially confounding factors, such as physical activity and socioeconomic status, using several indicators, i.e. job type for longest held occupation, education, and type of housing. Finally our study provides a new way of analysing dietary patterns that could be used by others to investigate the joint effect of dietary items within a framework based on complementary and antagonistic effects of foods and drinks.

Weaknesses

First, we are relying on 10 year recall of dietary patterns, which may be inaccurate, however other studies have found that remote diet recalled from 10 years earlier may be as reliable as recent dietary recall.⁵⁸ Random recall error would bias the result towards the null. We did not find that NFRS and RFS had any relationship with deaths from injury and poisoning, which should be biologically unrelated. Second, it is possible that our informants ascribed some form of socially undesirable or unhealthy eating pattern to the people who had died (cases) but not to the controls, although it would have needed to be a complex pattern ascribing varying intakes of healthy foods with different levels of other foods. However, the dietary items chosen are largely neutral in Chinese tradition, where war and famine occurred within recent memory. Dietary health promotion was not a priority in a hospital-focused health service in laissez faire colonial Hong Kong in the 1990s. At the time of data collection interest was focused on dietary fat and salt,^31,59 rather than on specific foods. In addition in Chinese society communitarian values and a Confucian respect for authority may have contributed towards informants providing quality responses to a survey introduced as helping ‘all Hong Kong people to lead more healthy lives’. Finally, the main focus of LIMOR was to assess the effects of active and passive smoking on mortality, therefore there was little reason to suspect such a strategic bias on the part of informants or research assistants conducting the interviews. Third, we constructed crude food groups from a limited number of food items, however our results were fairly resilient to changing the definitions of the less healthy group, such as collapsing down to two groups, or of excluding fish from the healthy food score (data not shown). Fourth, the two dietary components are a very broad categorization that may not correspond to dietary recommendations for all groups; however this does not detract from findings about synergies. Fifth, we do not have total food intake. We do not know if the effect of these dietary patterns occurred because overall people ate less of the less healthy foods as they ate more of the healthy foods, or because some nutrients are blocking others, or because the response is different to some combinations, for example if a mixture of food may reduce the insulin response with associated benefits.⁶⁰ However, we found similar modifications for tea drinking, which should not affect food intake, and the correlation between NFRS and RFS is relatively low (0.28), which suggests it is more simply an amount producing the joint effect.

Conclusion

Our analysis of mortality in a large case–control study in an Asian population has found similar results to other studies on the importance of eating healthy foods. However, our analysis is strongly suggestive that a framework explicitly recognizing synergistic effects of healthy foods, tea drinking, and alcohol use on less healthy foods and of complementary effects could enhance our understanding of nutritional epidemiology and possibly reconcile some of the conflicting findings. We have also demonstrated that a simple way of exploring potential dietary synergies can produce useful results.

The LIMOR study was supported by Hong Kong Health Services Research Committee (631012), Hong Kong Council on Smoking and Health. We thank the Immigration Department of the Government of the Hong Kong Special Administrative Region for their help with data collection. All the authors jointly planned the study. C.M.S. did the analysis and wrote the first draft, all authors made substantial contributions to the final draft. None of the authors had any financial interest.

References