In observational studies, cholesterol intake has been associated with impaired glucose metabolism( Reference Feskens and Kromhout 1 ) and type 2 diabetes risk( Reference Meyer, Kushi and Jacobs 2 , Reference Salmeron, Hu and Manson 3 ). Preliminary results from small randomised-controlled trials have shown that adding eggs, an important source of dietary cholesterol, to the diet improved insulin sensitivity( Reference Blesso, Andersen and Barona 4 ) and atherogenic lipoprotein profile( Reference Blesso, Andersen and Barona 4 , Reference Mutungi, Waters and Ratliff 5 ) and decreased inflammatory markers( Reference Ratliff, Mutungi and Puglisi 6 , Reference Andersen, Lee and Blesso 7 ). Furthermore, observational data support the role of circulating small and dense LDL and HDL particles( Reference Hodge, Jenkins and English 8 ) and inflammation( Reference Wang, Bao and Liu 9 ) on diabetes risk. However, other nutrients found in eggs (e.g. choline) could also play a role in diabetes risk( Reference Tang, Wang and Levison 10 – Reference Gao, Xu and Jiang 12 ). In contrast, a previous meta-analysis of epidemiological studies concluded that egg consumption may increase diabetes risk( Reference Shin, Xun and Nakamura 13 ). Owing to the absence of randomised trials directly assessing the effect of egg consumption on type 2 diabetes, we conducted a dose–response meta-analysis of prospective cohort studies to quantify the association between habitual egg intake and risk of type 2 diabetes.
Methods
Search strategy
We followed standard criteria for conducting and reporting meta-analyses of observational studies( Reference Stroup, Berlin and Morton 14 ). We searched for all prospective cohort studies that evaluated egg consumption and risk of diabetes mellitus in adults from the earliest available online indexing through June 2015. We conducted our study search without language restrictions on MEDLINE (egg[tw] or eggs[tiab] or ‘animal food’[tiab]) and (‘diabetes mellitus’[mesh] or diabetes[tiab]), EMBASE and EBSCOhost; we reviewed related articles, hand searched reference lists and directly contacted the authors. The search key words were ‘eggs’, ‘animal food’ and ‘diabetes mellitus’. One investigator (M. T.) screened titles and abstracts, and two investigators (M. T., M. L.) independently and in duplicate reviewed full texts of potentially relevant articles and assessed study eligibility. We included studies that provided multivariable-adjusted risk estimates. We resolved differences by consensus. A priori we excluded ecological and cross-sectional analyses, case–control studies, commentaries, general reviews and case reports (Fig. 1). When duplicate studies were present, we chose the most recent publication. We also excluded studies reporting only crude risk estimates. After screening titles and abstracts of 416 articles and reviewing thirty-one full-texts, we extracted study characteristics and data in duplicate from ten eligible studies( Reference Montonen, Jarvinen and Heliovaara 15 – Reference Virtanen, Mursu and Tuomainen 24 ) (Fig. 1) with a total of thirteen estimates. Thus, twenty-one studies were excluded because they were reviews (n 2)( Reference Jarrett 25 , Reference Djoussé 26 ), cross-sectional (n 4)( Reference Agrawal and Ebrahim 27 – Reference Shi, Yuan and Zhang 30 ) or case–control (n 1)( Reference Radzeviciene and Ostrauskas 31 )studies; assessed overall dietary patterns (n 4)( Reference Batis, Mendez and Sotres-Alvarez 32 – Reference Yu, Zhang and Xiang 35 ) or nutrient intakes (n 2)( Reference Silva Ton, das Gracas de Almeida and de Morais Cardoso 36 , Reference Houston, Ding and Lee 37 ); included participants with prevalent disease (n 4)( Reference Pearce, Clifton and Noakes 38 – Reference Horwath and Worsley 41 ) or gestational diabetes (n 1)( Reference Qiu, Frederick and Zhang 42 ) or other outcome different from diabetes (n 2)( Reference Kuriki, Tokudome and Tajima 43 , Reference Nicklas, O’Neil and Fulgoni 44 ); or because of duplicate study population (n 1)( Reference Ericson, Sonestedt and Gullberg 45 ).
Flow diagram of the study selection process.
Data extraction
We extracted data using a standard electronic form independently and in duplicate by two investigators (M. T., M. L.). Information included the first author’s name, contact information, year of publication, number of years the study was performed, study name, study location, population (age, sex, race, exclusion criteria and sample size), mean age and standard deviation at baseline, duration of follow-up, exposure assessment, egg consumption categories, outcome definition, outcome ascertainment, covariates adjusted for, number of participants, person-years, number of events and adjusted risk estimates and 95 % CI. When the mean age of the total study population was unavailable, we calculated the weighted mean age and weighted standard deviation based on exposure categories’ specific information. We assumed one serving of egg to be equivalent to 50 g. We gave preference to multivariable estimates from models with the greatest control for potential confounders. Hazard ratios and OR were assumed to approximate risk ratios (RR). We assigned studies a degree of covariate adjustment: minimal (socio-demographic covariates), adequate (socio-demographic plus either other risk factors or dietary variables) and optimal (socio-demographic plus risk factors and dietary variables). Issues regarding missing data or definitions were resolved by direct contact with authors.
We assessed study quality based on the degree of covariate adjustment and the Newcastle–Ottawa quality assessment scale for observational studies in meta-analyses( Reference Wells, Shea and O’Connell 46 ). This scale for observational studies in meta-analyses grants a maximum of 9 points to each study: a maximum of 1 point for each item within the selection (representativeness of the exposed cohort, selection of the non-exposed cohort, ascertainment of exposure, demonstration that outcome of interest was not present at start of study) and outcome categories (assessment of outcome, length of follow-up, adequacy of follow-up) and a maximum of 2 points for the comparability category based on the design or analysis. We assigned scores of 0–6 for low-quality and 7–9 for high-quality studies based on the distribution of scores among studies. Differences in quality assessment scores between investigators were unusual and were resolved by consensus.
Statistical analyses
We conducted random-effects, dose–response regressions by using generalised least squares for trend estimation (one-stage estimation)( Reference Greenland and Longnecker 47 ). We assumed hazard ratios and OR-approximated RR. Covariance was fit with the use of total numbers of cases plus non-cases for studies that reported OR or person-years for studies that reported RR or hazard ratios, at each level of exposure. For completeness, we also performed two-stage estimation: separate generalised least squares models for trend were evaluated for each study to derive study-specific, log-linear dose–responses (log RR), and then each study-specific log RR was pooled in a second generalised least squares model for trend. We pre-specified one-stage estimation as our primary outcome because it uses all available estimates in each study yielding a better estimate of the variance–covariance matrix relative to the two-stage estimation. We tested the between-study heterogeneity with goodness of fit (χ 2). When exploring dietary factors in relation to diabetes, BMI is commonly included in models to adjust for confounders. However, this variable could also be considered an intermediate in the causal pathway. Therefore, we conducted sensitivity analyses including (when available) estimates from models that were not adjusted for BMI. We explored a priori potential sources of heterogeneity by using meta-regression (sex, study location (USA v. Europe/Asia), study quality (Newcastle–Ottawa score <7 or ≥7), years of follow-up (<15, ≥15), mean age (<50, ≥50 years) and method for assessing dietary intakes (FFQ v. other methods)). We constructed funnel plots for visual inspection of publication bias and evaluated statistically the bias using Begg’s test( Reference Begg and Mazumdar 48 ). Finally, we stratified the data by sex, study location and quality score. Analyses were performed using Stata 11.2 for Mac (StataCorp LP), with two-tailed α<0·05. Analytical code used is provided in the online Supplementary Material, and databases and documentation are available as supplemental digital content.
Results
We identified ten studies with thirteen different estimates that included 251 213 individuals (173 463 women and 77 750 men) and 12 156 incident cases of type 2 diabetes. The studies were conducted in the USA (n 4), Asia (n 1) and Europe (n 5). Age ranged from 38 to 95 years, and the median daily egg consumption ranged from 0 to 1·1 eggs across studies (Table 1). Nearly all studies adequately adjusted for important diabetes risk factors including age, sex, BMI, smoking status, alcohol use, physical activity and dietary factors.
Characteristics of participants in included studies of egg consumption and type 2 diabetes risk
FMCHES, Finnish Mobile Clinic Health Examination Survey; AMS, Adventist Mortality Study; AHS, Adventist Health Study; PHS, Physicians’ Health Study; WHS, Women’s Health Study; CHS, Cardiovascular Health Study; SUN, Seguimiento Universidad de Navarra; JPHC, Japan Public Health Center-based Prospective; JHS, Jackson Heart Study; MDC, Malmö Diet and Cancer Cohort; E3N, Etude Epidémiologique auprès des femmes de la Mutuelle Générale de l’Education Nationale; KIHD, Kuopio Ischaemic Heart Disease Risk Factor Study.
Each egg per day was associated with a 13 % higher risk of incident type 2 diabetes (one-stage estimation RR 1·13; 95 % CI 1·04, 1·22; P heterogeneity<0·001) (Fig. 2). In contrast, secondary results from the two-stage estimation (which does not use all available information) were null (RR 1·07; 95 % CI 0·93, 1·24). The results from a sensitivity analysis that used estimates that did not include BMI, a potential intermediate, were qualitatively the same (online Supplementary Fig. S1). Using meta-regression, we explored sources of heterogeneity: sex (P=0·84), mean age (P=0·15), study location (P=0·03), study quality (P=0·18), years of follow-up (P=0·46) and method for assessing dietary intakes (P=0·20). When we stratified the analyses by study location, we observed that in the studies conducted in the USA an egg per day was associated with a 47 % higher risk of type 2 diabetes (RR 1·47; 95 % CI 1·32, 1·64), whereas the association for studies conducted elsewhere was null (Table 2). Moreover, the association for high-quality studies was null (RR 0·94; 95 % CI 0·74, 1·19). We found no evidence of publication bias on the funnel plot or Begg’s test (P=0·46) (Fig. 3).
Risk ratios (RR) of incident type 2 diabetes per egg per day. One-stage overall pooled dose–response estimate and 95 % CI are boxed. 
and 
Study-specific dose–response and 95 % CI; 
, pooled dose–response; 
, 95 % CI combining each study specific dose–response (two-stage); FMCHES, Finnish Mobile Clinic Health Examination Survey; AHS, Adventist Health Study; AMS, Adventist Mortality Study; WHS, Women’s Health Study; PHS, Physicians’ Health Study; CHS, Cardiovascular Health Study; SUN, Seguimiento Universidad de Navarra; JPHC, Japan Public Health Center-based Prospective; JHS, Jackson Heart Study; MDC, Malmö Diet and Cancer Cohort; E3N, Etude Epidémiologique auprès des femmes de la Mutuelle Générale de l’Education Nationale; KIHD, Kuopio Ischemic Heart Disease Risk Factor Study.
Funnel plot for graphical assessment of potential publication bias.
Consumption and risk of type 2 diabetes stratified by sex, study location and study quality (Risk ratios (RR) and 95 % confidence intervals)
Discussion
In a dose–response meta-analysis of ten publications of prospective studies using thirteen estimates, we observed a direct association between egg consumption and type 2 diabetes. We found evidence that results may be driven in part by studies conducted in the USA and by studies of a lower quality.
In animal studies, cholesterol intake has been associated with impaired glucose metabolism( Reference Adamopoulos, Papamichael and Zampelas 49 ) and inflammation( Reference Lewis, Kirk and McDonald 50 ). However, these studies generally use a very high dose of cholesterol, thus potentially limiting the applicability of results to humans. Eggs are important contributors of dietary cholesterol, raising concerns that egg consumption may affect cardiovascular health and diabetes risk. However, there is no clear relationship between dietary cholesterol consumption and serum cholesterol( Reference Fuller, Sainsbury and Caterson 39 ), although there seems to be significant heterogeneity in the response to cholesterol intake. In addition to genetic factors( Reference Ordovas, Lopez-Miranda and Mata 51 ), for example, obesity and insulin resistance appear to affect cholesterol absorption( Reference Knopp, Retzlaff and Fish 52 , Reference Tannock, O’Brien and Knopp 53 ). In addition, experimental studies in humans have shown that increased egg intake has rather had a beneficial impact on several risk factors for type 2 diabetes, such as insulin resistance( Reference Blesso, Andersen and Barona 4 ), inflammation( Reference Ratliff, Mutungi and Puglisi 6 , Reference Andersen, Lee and Blesso 7 , Reference Blesso, Andersen and Barona 54 ) and lipid particle size( Reference Blesso, Andersen and Barona 4 , Reference Mutungi, Waters and Ratliff 5 ). However, a previous meta-analysis of prospective cohorts concluded that individuals who ate an egg or more per day had a 42 % higher risk of diabetes compared with individuals who never consumed eggs( Reference Shin, Xun and Nakamura 13 ). This meta-analysis was based on five cohorts, included US studies only and dose–response was not assessed( Reference Vang, Singh and Lee 16 – Reference Djoussé, Kamineni and Nelson 18 ).
Contrasting results from intervention studies and observational studies should take into account design limitations. Conducting randomised trials to evaluate the effects of foods can be challenging because of costs, difficulties in blinding individuals and non-compliance because of the length of time that is necessary to observe incidence of outcomes. For example, only short-term intervention studies on egg consumption using intermediate risk markers are feasible but do not necessarily reflect diabetes risk. Thus, in the absence of trials that use diabetes as the outcome, one strategy is to infer these effects from long-term prospective cohort studies. Besides the fact that these studies may not adequately reflect the question of the potential effect of altering food consumption, two key limitations are the potential for residual confounding and for misclassification of the exposure. Our results are based on individual studies where the potential for these two major limitations is always present, and therefore the results should be interpreted with caution.
The observation that the association could be driven by studies conducted in the USA may reflect the possibility that egg consumption may be confounded by behaviours or dietary habits associated with diabetes risk that are common in this population. For example, in the US studies, egg intake is often associated with smoking( Reference Djoussé, Gaziano and Buring 17 , Reference Djoussé, Kamineni and Nelson 18 , Reference Djoussé, Petrone and Hickson 21 ) or lower physical activity( Reference Djoussé, Gaziano and Buring 17 ) or higher intake of red meat( Reference Djoussé, Gaziano and Buring 17 , Reference Djoussé, Kamineni and Nelson 18 , Reference Djoussé, Petrone and Hickson 21 ), whereas this is generally not observed in studies outside the USA( Reference Zazpe, Beunza and Bes-Rastrollo 19 , Reference Virtanen, Mursu and Tuomainen 24 ). However, although one study in France did find such associations with egg intake, it still reported null findings for the relation of egg intake and type 2 diabetes( Reference Lajous, Bijon and Fagherazzi 23 ). Food preparation methods (e.g. boiled or fried eggs, whole eggs or only egg whites) or concurrent consumption of other foods that may increase diabetes risk (e.g. home fries, bacon) may also account for a part of the differences, but such information is not available in these studies. Our results are consistent with a recently published meta-analysis( Reference Djoussé, Khawaja and Gaziano 55 ), and the conclusion that the association was not present in non-US populations is strengthened by the inclusion in the current meta-analysis of an additional study from Finland. The importance of adequate designs, robust ascertainment of exposure and outcome and collection of information on potential confounding factors with as much detail as possible is further underscored by the observation that better quality studies were less likely to find an association between egg consumption and diabetes risk. We classified four studies as low quality mainly because diabetes was self-reported and the follow-up rate was inadequate or not described( Reference Vang, Singh and Lee 16 , Reference Djoussé, Gaziano and Buring 17 , Reference Kurotani, Nanri and Goto 20 ). It is difficult to interpret why results of these studies differ from high-quality studies. However, three of these studies were conducted in the USA.
The possibility that the observed differences across populations are the result of underlying biological mechanisms is still present. Intestinal microbiota may vary across populations and there is evidence that intestinal flora affects the production of trimethylamine-N-oxide from dietary phosphatidylcholine (egg yolks are important contributors)( Reference Tang, Wang and Levison 10 ). In animal studies, this metabolite appears to play a role in glucose metabolism( Reference Gao, Xu and Jiang 12 ). Thus, there is a need to further study this association between egg consumption and glucose metabolism across populations.
Our meta-analysis has several strengths. We systematically reviewed multiple databases for all prospective studies on egg consumption and diabetes risk. We contacted authors directly when clarifications of findings or additional data were necessary, thus minimising potential misclassification and publication bias. We performed study inclusion/exclusion and data extraction in duplicate and independently. We explicitly assessed dose–response rather than carrying out simple categorical comparisons using generalised least squares models for trend estimation.
Our results suggest that the association of habitual consumption of eggs and incidence of type 2 diabetes observed in prospective studies may be restricted to studies conducted in the USA. In the absence of a clear biological mechanism, the possibility that the observed relation may be due to residual confounding by dietary behaviours or food preparation methods restricted to this population is always present.
Acknowledgements
The authors of the present study acknowledge the authors of studies included in this analysis who provided additional information.
This analysis was supported by the National Council of Science and Technology (CONACyT, Mexico), the University of Eastern Finland, National Institute of Public Health and the Bernard Lown Fund for Cardiovascular Health.
M. L. and J. K. V. conceived the study. M. L. and M. T. extracted data and contacted authors. M. T. conducted the analysis. M. L., M. T. and J. K. V. wrote and revised the manuscript.
M. L. has an investigator-initiated research grant from AstraZeneca. Other authors have no disclosures.
Supplementary material
For supplementary material/s referred to in this article, please visit http://dx.doi.org/doi:10.1017/S000711451600146X
