Milk and dairy consumption and risk of cardiovascular diseases and all-cause mortality: dose–response meta-analysis of prospective cohort studies

With a growing number of prospective cohort studies, an updated dose–response meta-analysis of milk and dairy products with all-cause mortality, coronary heart disease (CHD) or cardiovascular disease (CVD) have been conducted. PubMed, Embase and Scopus were searched for articles published up to September 2016. Random-effect meta-analyses with summarised dose–response data were performed for total (high-fat/low-fat) dairy, milk, fermented dairy, cheese and yogurt. Non-linear associations were investigated using the spine models and heterogeneity by subgroup analyses. A total of 29 cohort studies were available for meta-analysis, with 938,465 participants and 93,158 mortality, 28,419 CHD and 25,416 CVD cases. No associations were found for total (high-fat/low-fat) dairy, and milk with the health outcomes of mortality, CHD or CVD. Inverse associations were found between total fermented dairy (included sour milk products, cheese or yogurt; per 20 g/day) with mortality (RR 0.98, 95% CI 0.97–0.99; I2 = 94.4%) and CVD risk (RR 0.98, 95% CI 0.97–0.99; I2 = 87.5%). Further analyses of individual fermented dairy of cheese and yogurt showed cheese to have a 2% lower risk of CVD (RR 0.98, 95% CI 0.95–1.00; I2 = 82.6%) per 10 g/day, but not yogurt. All of these marginally inverse associations of totally fermented dairy and cheese were attenuated in sensitivity analyses by removing one large Swedish study. This meta-analysis combining data from 29 prospective cohort studies demonstrated neutral associations between dairy products and cardiovascular and all-cause mortality. For future studies it is important to investigate in more detail how dairy products can be replaced by other foods. Electronic supplementary material The online version of this article (doi:10.1007/s10654-017-0243-1) contains supplementary material, which is available to authorized users.


Introduction
Cardiovascular disease (CVD) is the leading cause of mortality and disability worldwide [1]. Together with smoking, obesity and inactivity, diet is considered to be one of the most important prevention strategies for CVD Electronic supplementary material The online version of this article (doi:10.1007/s10654-017-0243-1) contains supplementary material, which is available to authorized users. [2]. Milk and dairy foods have been recommended in most dietary guidelines around the world, but the association of milk or dairy food consumption with CVD is still controversial [3,4]. An earlier meta-analysis [5] which included 17 prospective cohort studies showed that milk intake was not associated with total mortality or CHD mortality, but there was a borderline significant inverse association with CVD mortality based on limited studies. There were not enough data to examine the effects of other dairy products or milk fat content. Since then, further prospective cohort studies have been published. For example, one recent Swedish publication with two large Swedish cohorts [6] reported that higher milk consumption was associated with a doubling of mortality risk including CVD mortality in the cohort of women. Since this paper was published in 2014, there has been mounting debate from different researchers regarding its seemingly contradictory results [7,8]. This has caused new uncertainty about the effects of milk and dairy intake on human health. Recently, new meta-analyses of dairy consumption and risk of stroke [9], butter and risk of CVD, diabetes and mortality [10] have been published, showing predominantly neutral or marginally beneficial associations for all dairy products. Therefore, we conducted a comprehensive dose-response meta-analysis to examine linear and non-linear associations between milk and dairy products with all-cause mortality, CHD and CVD events using existing prospective cohort studies of adequate quality.

Literature search and study selection
This review was conducted based on guidelines of Metaanalysis of Observational Studies in Epidemiology [11]. Prospective cohort studies published up to Sep 2016 (without language restriction) were searched using PubMed, Embase, and Scopus database, the query syntax of searching is shown in the Supplemental Methods (see search strategy). After excluding duplicates and based on titles and abstracts, we excluded studies on animals, baseline age B18 years, or populations with prior CVD, diabetes, or any other chronic diseases. Eligible studies were selected by using predefined inclusion criteria of prospective cohort studies, healthy populations and original articles on the association of milk and dairy intake and allcause mortality, CHD or CVD. In addition, supplementary hand searching of reference lists of previous reviews or meta-analyses was conducted. Of 59 eligible full articles, 29 articles [6, met the inclusion criteria (see Fig. 1). Several authors or coworkers provided additional data for this meta-analysis [14,16,19,23,27,28,32,34,37,40].
Data extraction and quality assessment Data were extracted from published articles by using a structured extraction form, which included descriptive characteristics of the study, range of intake, median intake, number of participants, number of mortalities, CHD or CVD cases, person-years at risk, and relative risk (RR) with 95% CI for each unit of dairy intake. For studies that reported results from different multivariable-adjusted models, the model with the most confounding factors was extracted for the meta-analysis. If dairy intake was presented in servings or times per period of time [12-20, 22, 23, 34-36, 39], we converted the portion size into grams per day by using standard units of 244 g for milk (585 g for 1 pint of milk); 244 g for yoghurt and 40 g for cheese [41,42]. One serving of total dairy, high-fat dairy and low-fat dairy was taken to be 200 g, similar to our previous meta-analysis [5]. When studies reported country specific conversion factors, these were used to calculate intake as g/day [26,29,30].
In some studies the mean intakes of dairy categories were not reported, in which case we calculated the mean value by using the lower and upper limit. For open-ended upper limits of intake, the same range as the lower category was applied. The categories of dairy types were defined in accordance with the definition in the original articles (Supplemental Table 2).
Two independent reviewers determined the quality of the 29 studies based on the Newcastle-Ottawa quality assessment scale (NOS, Supplemental Methods) [43]. By evaluation of selection, comparability and outcome, the rating system scores studies from 0 (highest degree of bias) to 9 (lowest degree of bias). Additionally we investigated the funding sources of all of the eligible studies. The four categories of funding were recorded as industry, partial funded by industry, research institution and unknown.

Statistical analysis
Meta-analyses of each dairy type were performed if the number of studies was three or more. Splined variables were generated by MKSPLINE in STATA version 13.0 to determine the most appropriate knot points of nonlinear associations from goodness-of-fit tests and Chi square statistics. Spine analysis and dose-response generalised least-square trend (GLST) meta-analysis were applied for the further analysis of linear or nonlinear associations. Incremental dose-response RRs were derived from the random-effects meta-regression trend estimation of summarised dose-response data. Ding's spaghetti plot was used to present the shapes of the association within individual studies, as described previously [44]. Forest plots were created to assess the linear dose-response slopes and corresponding 95% CI across relative studies with increments of 200 g/day for total, high-fat, and low-fat dairy; 244 g/day for milk; 20 g/day for total fermented dairy (includes cheese, yogurt and soured milk products); 10 g/day for cheese; 50 g/day for yogurt. Sensitivity analysis was based on linear dose-response slopes by excluding one study population at a time.
To explore heterogeneity between studies, I-squared was calculated from Cochrane Q test [45]. In addition, subgroup analyses were performed providing that at least 6 study populations were available by age (B50 years, [50 years), follow-up duration (B10 years, [10 years), gender (men, women, both men and women), continent, confounding factors (whether analyses were or were not adjusted for the following 7 confounders age, sex, smoking, alcohol, body mass index (BMI), physical activity, food energy intake), BMI (B25 kg/m 2 , [25 kg/m 2 ) and Newcastle-Ottawa quality score\or C7. When number of the examined studies C10, potential publication bias was assessed by means of the Eggers test [46] and symmetry of the funnel plot. All of the statistical analyses were performed in STATA version 13.0 (StataCorp. College Station, Texas, USA). Two-sided P values \0.05 were considered as statistically significant.

Results
Overviews of key characteristics of the 29 prospective cohort studies are shown in Table 1. The included participants of each dairy exposure data on all-cause mortality, CHD or CVD are presented in Table 2. A total of 783,989 participants, 93,158 mortality cases, 28,419 CHD and 25,416 CVD were included in the analysis. There were 3 studies conducted in Asia (Japan and Taiwan) [28,35,39], 2 studies in Australia [24,29], 7 in the United States [12, 14-16, 19, 22, 34] and the remaining 17 studies in Europe. A total of 6 studies presented sex-specific results, 3 studies were in men [18][19][20] and 3 in women [15,16,30]. There was one study [12] with missing data on age and 4 studies with missing BMI data [12,21,33,36]. The estimated mean age was 57 years (range 34-80 years) and mean value of BMI was 25.4 kg/m 2 (range 22.3-27.1 kg/ m 2 ). The duration of follow-up ranged from 5 to 25 years, with a mean follow-up of 13 years. Study characteristics of each dairy intake category by outcomes are shown in Table 2. Results of quality assessment are shown in the Supplemental Table 1, with 18 studies scoring C7. All of the studies were funded by a research institute except one study [13] without funding information, thus sub-group analysis was not conducted by funding source. There was no evidence of publication bias in the meta-analyses of milk or dairy consumption with different health outcomes .

Total fermented dairy, cheese and yogurt
Total fermented dairy intake (weighted median intake 77 g/day, 19 populations, 11 studies) was non-linearly and marginally associated with lower mortality risk, with a RR of 0.98 (95% CI 0.97-0.99) per 20 g/day but with high heterogeneity (I 2 = 94.4%, P \ 0.001; Fig. 2). In sensitivity analysis, by excluding the Swedish study [6] of women's results for cheese, I 2 was reduced to 45.2% (P = 0.02), with RR of 1.00 (95% CI 0.99-1.00). Similarly, total fermented dairy intake (17 populations, 9 studies) was non-linearly and modestly associated with a 2% lower CVD risk per 20 g/day (RR 0.98, 95% CI 0.97-0.99) (Fig. 3). Significant heterogeneity was present  Relative risk Fig. 2 Relative risk of all-cause mortality for an increment of 20 g/day of fermented dairy intake. Squares represent study-specific RR. Square areas are proportional to the overall specific-study weight to the overall meta-analysis. Horizontal lines represent 95% CIs.

Discussion
This meta-analysis combining data from 29 prospective cohort studies showed there were no associations between total dairy, high-and low-fat dairy, milk and the health outcomes including all-cause mortality, CHD or CVD. The modest inverse associations of total fermented dairy were found with all-cause mortality and CVD, but not CHD. By examining different types of fermented food in relation to  CVD, we found marginally inverse association with cheese but not yogurt. However, further sensitivity tests showed the inverse associations of fermented dairy and cheese with all-cause mortality or CVD disappeared after removing the study of Michaelsson et al. [6]. No associations were found between total dairy and milk consumption with all-cause mortality, CHD or CVD in the current study, which is in agreement with several metaanalyses [47,48]. Larsson et al. [47] reported neutral associations of dairy and milk consumption with mortality or CVD mortality. Mullie et al. [48] reported neutral associations of milk consumption with all-cause mortality or CHD. In addition, the current study is in agreement with a recently published review [49] which indicated neutral associations between the consumption of total dairy and risk of CHD or CVD. Results of sub-group analyses showed the inverse associations were observed between total dairy intake and CHD, or the association between milk consumption and CVD when studies did not adjust for major confounders. Thus, confounders included in statistical analyses in prospective studies have substantial effects on the final findings and conclusions. Furthermore, inverse associations were also found in sub-groups of studies defined by mean age (B50, [50 years) or BMI ([25 kg/ m 2 ) of the associations between total, high-fat, low-fat dairy and milk with risk of all-cause mortality, CHD or CVD, which indicated the findings and conclusions were also affected by characteristics of the study populations within different studies.
Three US prospective cohort studies described by Chen et al. [50] showed a substantially lower risk of CVD when animal fats, including dairy fat, were replaced by unsaturated fats. Recently, UK National Health Service (NHS) has recommended low-fat milk and dairy products as healthy choices [51]. However, in the current study, highfat and low-fat dairy consumption were investigated separately and no substitution models replacing high by lowfat dairy products were carried out. We found no significant associations between high-and low-fat dairy and all-cause mortality, CHD or CVD. This supports two previous metaanalyses [5,52] which also reported no association of high or low-fat dairy and CHD. Furthermore, beneficial effects of high-fat dairy foods on human health were reported by a cross-sectional study [53], which showed an inverse association of full-fat dairy food and the metabolic syndrome. In addition, another US study [54], which reviewed cross-  sectional and prospective cohort studies, showed that 11 of the 16 studies identified that population with higher full-fat dairy intake had less adiposity. It is also noteworthy that butter as a high fat dairy food containing 80% fat [55], a recent meta-analysis on the effects of butter [10] showed that whilst consumption was weakly associated with allcause mortality (per 14 g/day: RR 1.01, 95% CI 1.00-1.03), there was no significant association with CHD, CVD or stroke and there was an inverse association with incidence of diabetes (RR 0.96, 95% CI 0.93-0.99). Therefore, the effect of dairy fat on CVD is complex and may be influenced by the nature of the fat containing food vehicle, which needs confirmation in further studies. Despite their fat content and composition, milk and dairy products are naturally rich in various minerals (e.g. calcium, potassium), protein and vitamins (e.g. vitamin A and vitamin B 12 ) [56]. Nutrients including calcium, potassium and magnesium have been suggested to be associated with lower risk of stroke [57,58]. Short-term human intervention studies [59,60] also indicated that subjects who have high-fat diets enriched with dairy minerals or calcium have significantly lower total cholesterol and LDL-cholesterol levels than those on a control diet. This may explain in part why total dairy consumption has a neutral role in terms of the effect on health outcomes.
The current study also showed total fermented dairy and cheese intake to be marginally inversely associated with mortality and CVD risk, respectively, and large heterogeneity was present. However, by removing the study of Michaelsson et al. [6], heterogeneity of the associations of total fermented dairy and mortality or CVD, cheese and mortality or CVD were markedly reduced. Also, the marginally inversely associations were disappeared. To our knowledge, the present study is the first dairy meta-analysis to include the large Swedish cohort results [6]. The markedly reduced heterogeneity after removing the results of the Swedish female cohort [6] indicated the heterogeneous nature of the Swedish study, which may be related to the diet and lifestyle characteristics of the study participants, as they had a relatively low education level (80 and 70% for women and men were educated for B9 years, respectively), also the highest milk drinkers had highest percentage of smokers and those living alone.
Cheese consumption based on 11 populations was found to be modestly and inversely associated with CVD risk, with a 2% lower risk of CVD per 10 g/day of cheese, however, the significant association disappeared after removing the study of Michaelsson et al. [6]. Compared with other meta-analyses on cheese, Alexander et al. [4] has reported 11% lower risk of CVD per 35 g/day (95% CI 0.78-1.01), while Chen et al. [61] presented 10% lower risk of CHD per 50 g/day (95% CI 0.84-0.95). However, the analysis of the associations between cheese and CVD in studies of Alexander et al. [4] and Chen et al. [61] were based on 3 and 8 populations, respectively, which was less than our current study of 11 populations.
Furthermore, total fermented dairy and cheese were modestly inversely associated with risk of CVD but not CHD in the current meta-analysis, so perhaps both dairy types play a role in reducing the risk of stroke. This is supported by the evidence of another recent meta-analysis [9], which found a 9% lower risk of stroke (RR 0.91, 95% CI 0.82-1.01) associated with higher total fermented dairy intake and a 3% lower risk of stroke (RR 0.97, 95% CI 0.94-1.01) with higher cheese consumption, although none of these associations were statistically significant. As there was limited information of the different sub-types of the CVD events, the understanding of the association of fermented dairy products with varied CVD types remains unclear. In addition, unlike the result for cheese, the association of yogurt with disease outcomes was neutral. However, a previous review of randomised trials suggested that yogurt is associated with lower risk of CVD [62]. Our null results for yogurt intake and CVD may be due to the limited number of participants from only 3 populations. In addition, a very recent meta-analysis showed a 14% lower risk of type 2 diabetes for 80 g/day yogurt intake (RR 0.86, 95% CI 0.83-0.90) based on 11 prospective cohort studies [63].
The mechanism of the beneficial association of fermented dairy products and reduced CVD risk and mortality is uncertain. Evidence from randomised controlled trials suggests that the reason, at least in part, may be an effect of the food matrix reducing lipid absorption and short chain fatty acids produced by the bacteria in the large intestine [64]. Moreover, omics-techniques have suggested that some of the beneficial effects of cheese can be accounted for by microbial fermentation producing short chain fatty acids such as butyrate [65].
Strengths of our study include the use of dose-response meta-analysis, the inclusion of more studies than in previous meta-analyses and the consideration of examination the individual dairy products separately such as dairy products in terms of fat content (high-fat, low-fat) or processing method (fermented or non-fermented). However, investigation of total dairy or total fermented dairy consumption with disease outcomes by combining dairy foods, high and low-fat dairy foods, solids and liquids, simply adding these up is a limitation which should be addressed in future studies by collecting and analyzing more detailed data. In addition, limitations of the study include sub-group analyses that lack statistical power, such as for Asian studies and effects of gender. We have 9 studies with scores of 7 or less by using the Newcastle-Ottawa Scale (NOS) [43]. Study quality could explain some heterogeneity but not all. For example, NOS scores of all studies containing high-fat dairy or low-fat dairy were C7, which could have resulted in lower heterogeneity for those analyses. Furthermore, residual confounding is a limitation of prospective cohort studies. The background diet should be taken into account in the statistical analyses as major confounders, which was done in 15 out of 29 cohort studies. Comparisons of dairy products with other foods in replacement models were not possible from the available data. The neutral risks of dairy products with mortality and CVD risk could be because of replacement by other foods, for example, those with high intake of dairy products may consume less sugar sweetened beverages which could lead to lower CVD mortality [66] or consume more processed meat which could lead to higher CVD risks [67,68]. For future studies it is important to investigate in more detail how dairy products can be replaced by other foods.

Conclusions
The current meta-analysis of 29 prospective cohort studies suggested neutral associations of total, high and low-fat dairy, milk and yogurt with risk of all-cause mortality, CHD and CVD. In addition, a possible role of fermented dairy was found in CVD prevention, but the result was driven by a single study.