Data source used:

A literature search using the Pubmed database was performed.

Cancer: For cancer in relation to vegetables & fruits, the following combination of keywords was used:

(vegetable OR vegetables OR fruit OR fruits OR alfalfa OR apple OR apples OR apricot OR apricots OR artichoke OR artichokes OR asparagus OR aubergine OR avocado OR avocados OR bamboo OR banana OR bananas OR bean OR beans OR beet OR beetroot OR beets OR berries OR berry OR broccoli OR cabbage OR cabbages OR cantaloupe OR cantaloupes OR cauliflower OR carrot OR carrots OR celery OR cherries OR cherry OR chicory OR chili OR chilli OR citrus OR coconut OR coconuts OR coleslaw OR corn OR cruciferae OR cucumber OR cucumbers OR currant OR currants OR dates OR eggplant OR eggplants OR endive OR figs OR garlic OR gherkins OR grape OR grapes OR grapefruit OR grapefruits OR greens OR kale OR kiwi OR kohlrabi OR leek OR leeks OR legume OR legumes OR lemon OR lemons OR lentil OR lentils OR lettuce OR lime OR limes OR maize OR mandarin OR mandarins OR mango OR mangos OR melon OR melons OR mushroom OR mushrooms OR nectarine OR nectarines OR okra OR onion OR onions OR oranges OR papaya OR parsley OR parsnips OR pea OR peas OR peach OR peaches OR pear OR pears OR pepper OR peppers OR pickle OR pickles OR pineapple OR pineapples OR plum OR plums OR pomegranate OR potato OR potatoes OR prune OR prunes OR quince OR radish OR radishes OR raisin OR raisins OR raspberries OR rhubarb OR salad OR salads OR sauerkraut OR scallion OR scallions OR shallot OR shallots OR seaweed OR soy OR soya OR soyfoods OR spinach OR sprout OR sprouts OR squash OR strawberries OR strawberry OR tangerine OR tangerines OR tempeh OR tofu OR tomato OR tomatoe OR tomatoes turnip OR watercress OR watermelon OR watermelons OR yams OR zucchini) AND cancer

A search using the words: "AND (neoplasm OR neoplasms)" instead of the word "cancer" resulted in 0 additional relevant articles.

End points other than cancer: For all other end points, such as CVD, in relation to vegetables & fruits, the following combination of keywords was used:

(vegetable OR vegetables OR fruit OR fruits OR alfalfa OR apple OR apples OR apricot OR apricots OR artichoke OR artichokes OR asparagus OR aubergine OR avocado OR avocados OR bamboo OR banana OR bananas OR bean OR beans OR beet OR beetroot OR beets OR berries OR berry OR broccoli OR cabbage OR cabbages OR cantaloupe OR cantaloupes OR cauliflower OR carrot OR carrots OR celery OR cherries OR cherry OR chicory OR chili OR chilli OR coconut OR coconuts OR coleslaw OR corn OR cucumber OR cucumbers OR currant OR currants OR eggplant OR eggplants OR endive OR figs OR garlic OR gherkins OR grape OR grapes OR grapefruit OR grapefruits OR greens OR kale OR kiwi OR kohlrabi OR leek OR leeks OR legume OR legumes OR lemon OR lemons OR lentil OR lentils OR lettuce OR lime OR limes OR maize OR mandarin OR mandarins OR mango OR mangos OR melon OR melons OR mushroom OR mushrooms OR nectarine OR nectarines OR okra OR onion OR onions OR oranges OR papaya OR parsley OR parsnips OR pea OR peas OR peach OR peaches OR pear OR pears OR pepper OR peppers OR pickle OR pickles OR pineapple OR pineapples OR plum OR plums OR pomegranate OR potato OR potatoes OR prune OR prunes OR quince OR radish OR radishes OR raisin OR raisins OR raspberries OR rhubarb OR salad OR salads OR sauerkraut OR scallion OR scallions OR shallot OR shallots OR seaweed OR soy OR soya OR soyfoods OR spinach OR sprout OR sprouts OR squash OR strawberries OR strawberry OR tangerine OR tangerines OR tempeh OR tofu OR tomato OR tomatoe OR tomatoes turnip OR watercress OR watermelon OR watermelons OR yams OR zucchini) AND (prospective OR cohort OR follow-up OR longitudinal)

One combined search term was used for these variables in relation to all end points:

(acrylamide OR bacon OR beef OR "dietary cholesterol" OR "dietary fat" OR fish OR ham OR hamburger OR heterocyclic OR hotdog OR hotdogs OR meat OR meats OR pork OR poultry OR salami OR sandwich OR sandwiches OR "saturated fat" OR sausage OR sausages OR seafood OR seafoods OR steak OR steaks OR veal OR vegetarian OR vegetarianism OR vegetarians OR butter OR cheese OR cream OR "dietary calcium" OR egg OR eggs OR lactose OR margarine OR milk OR dairy OR yoghurt) AND (prospective OR cohort OR follow-up OR longitudinal)

One combined search term was used for these variables in relation to all end points:

("dietary cholesterol" OR "dietary fat" OR "omega 3" OR "omega 6" OR "omega 9" OR "unsaturated fat" OR "unsaturated fats" OR "unsaturated fatty" OR "saturated fats" OR "saturated fatty" OR "polyunsaturated fat" OR "polyunsaturated fats" OR "polyunsaturated fatty" OR "monounsaturated fat" OR "monounsaturated fats" OR "monounsaturated fatty" OR "trans fat" OR "trans fats" OR "trans fatty" OR "linoleic acid" OR "octadecadienoic acid" OR "linolenic acid" OR "octadecatrienoic acid" OR "oleic acid" OR "octadecenoic acid" OR "eicosapentaenoic acid" OR "docosahexaenoic acid" OR "stearic acid" OR "palmitic acid" OR "myristic acid" OR "lauric acid" OR "butyric acid" OR "elaidic acid") AND (prospective OR cohort OR follow-up OR longitudinal)

Inclusion/exclusion criteria for articles:

Inclusion criteria:
1) Consumption of a dietary variable.
2) Endpoint: cancer risk, disease progression, or cancer mortality/survival risk.
3) Prospective studies (cohort or nested case-control design).
4) The full text article was published in English. Articles excluded because of language restrictions are defined in the related abstracts.

Exclusion criterium:
Data was excluded if risk among cases stratified by genetic polymorphisms was examined instead of risk among a total population. Since the practical use for this kind of information is extremely low.
Example: The article Red meat intake, CYP2E1 genetic polymorphisms, and colorectal cancer risk. only provides information about the relationship between 2 functional polymorphisms and their modifying effects on the association between diet and cancer. Whereas no information is shown about risk of diet for all cases vs controls.

The way data from the articles was used:

Tables: The systematic reviews contain simple tables on the main page defining the author, cohort name, amount of cases, and Relative Risk.
Possible reproducibility is a key factor in a systematic review. For this reason it is important to extract data in a transparent manner. This is why - for all dietary variables - also larger "extended tables" were created to include raw data from the articles using predefined methods for certain variables, described here:


1) FOOD GROUPS.
Data was extracted of all total food groups (e.g., vegetables), subgroups/botanical families (e.g., cruciferous vegetables), and specific food items (e.g., broccoli) related to the food group(s) defined in the review of interest. A definition of the dietary variables was added if directly available from the related publication.
If more than 1 article published data from a cohort about the same dietary variable, all data was added to the extended tables. But data from the most recent publication was chosen to be included for judging the evidence (unless a previous version included a substantially larger part of subjects & cases from the cohort).
Data about all different units of consumption was added. E.g., when results were published in tertiles, RR's for all tertiles were added. Definitions of these consumption units were added if directly available from the related publication.

Indexing vegetables or fruits into botanical families or vegetable/fruit subgroups:
Some articles specify data about certain botanical families. Depending on the FFQ from the cohort, any number of specific vegetables and/or fruits are indexed into these families.
-When a) no definition of a botanical family is given, or b) the definition of a dietary variable included > 1 specific vegetable and/or fruit, data from this variable is indexed under the specific botanical family it belongs to.
-When the definition of the botanical family includes only 1 vegetable or fruit item, data from this variable is indexed under this specific vegetable or fruit item.
-Data about gramineae is added to the variable "corn", unless it came from an Asian cohort.
-Data about musaceae is added to "bananas", and data about vitaceae is added to "grapes", since these are the only fruit items indexed under these families.

For the systematic reviews on this page the following index was used. Names for subgroups in this index overrule the ones chosen by authors if a difference occurred.

• Allium vegetables: garlic, leek, onions.
• Chenopodiaceae: beetroot or beet, chard greens/Swiss chard, spinach.
• Citrus fruit (rutaceae): grapefruit, mandarins, oranges, tangerines.
• Compositae: endive, lettuce.
• Convolvulaceae: sweet potatoes, yams.
• Cruciferous vegetables (cruciferae/brassica vegetables): broccoli, Brussels sprouts, cabbage, coleslaw, cauliflower, kale, mustard greens, rutabaga, sauerkraut)
• Cucurbitaceae: cantaloupe, cucumber, gherkins, squash, watermelon, zucchini/courgette.
• Gramineae: bamboo, corn/maize.
• Green leafy vegetables: beet greens, borage, cabbages, chard, chicory, chingensai, chives, collards, dandelion greens, endive, escarole, garden rocket, garland chrysanthemums, Jew's mallow, kale, lettuce, mugwort, mustard greens, parsley, spinach, seaweed, thistle, turnip greens, watercress.
• Green vegetables: containing any amount of items from the group of green leafy vegetables & any amount of other green vegetable items, such as broccoli, Brussels sprouts, green peppers, or legumes.
• Green-yellow vegetables: containing any amount of items from both the green vegetables and yellow vegetables group.
• Legumes (leguminosae/pulses): alfalfa, beans, lentils, peas, soy.
• Musaceae: bananas.
• Root vegetables: beetroot, carrots, celeriac, ginger, parsnip, radish, rutabaga, salsify, swedes, turnip.
• Rosaceae: apples/pears, apricot, peaches, plums, prunes, strawberries.
• Solanaceae: aubergine/eggplant, potatoes, peppers, tomatoes.
• Umbelliferae: carrots, celery.
• Vitaceae: grapes, raisins.
• Yellow fruits: apples, apricots, bananas, melons, oranges, peaches.
• Yellow vegetables: carrots, pumpkin, squash, sweet potatoes, red peppers, tomatoes, yams.

2) CANCER RISK.
Data of dietary variables was extracted regarding the relationship with cancer (incuding disease stage if data is available), disease progression, cancer mortality, or cancer survival.
Data about modifying effects on risk by potential confounders (e.g., age, sex, obesity, physical activity, menopausal status, hormone replacement therapy, ethnicity, and smoking) was added as well.

3) RELATIVE RISK AS A CATEGORIZED-, OR CONTINUOUS VARIABLE.
When RRs were available for associations evaluated both as a categorized variable (increasing units of consumption), and as a continuous variable (for an increment of X g or servings/day), the categorized variable was chosen to be included in the review. The categorized variable allows the possibility to define information about a) possible tresshold effects, and b) J, U, or other-shaped effects.
In addition, Relative Risks for an increase per 2 units of consumption may not reflect predicted Relative Risks based on increases per 1 unit of consumption, which complicates translating data to recommendations for individuals. For example, Chan JM (2006) found a significant association with prostate cancer progression risk of 4 dietary variables based on an increase of 1 serving/day. When this risk was based on an increase of 2 servings/day, all 4 associations almost disappeared.

Reference: Chan JM, Holick CN, Leitzmann MF, Rimm EB, Willet WC, Stampfer MJ. Diet after diagnosis and the risk of prostate cancer progression, recurrence, and death (United States). Cancer Causes Control. 2006 Mar;17(2):199-208. Abstract

4) AMOUNT OF CANCER CASES.
Often, data about the amount of cancer cases related to dietary variables was less than the total amount of cases from the cohort as specified in the abstract. This is due to missing values from incomplete Food Frequency Questionnaires. Data about the amount of cases was collected in the following descending order:
a) Data about the amount of cases was extracted from tables added by the authors to provide information about the dietary variables of interest.
b) If no direct data about the amount of cases for a dietary variable was provided, It was chosen to use the amount of cases - as specified by the author - from the food group the dietary variable belonged to. And to add a "?" symbol, indicating that the true amount of cases may be less.
c) If no direct data about the amount of cases for a dietary variable was provided, and no data was was provided by the author about amounts of cases from total food groups, It was chosen to use the total amount of cases from the cohort. And to add a "?" symbol, indicating that the true amount of cases may be less.

5) SIGNIFICANT OR NONSIGNIFICANT EFFECTS (RISK OR TREND).
a) Definition of an effect.
Background: When an author of an article speaks of a significant effect, mostly he is talking about the trend instead of the Relative Risk. But a trend does not consider the possibility of different effects at different levels of consumption, such as the ones found between alcohol consumption and heart disease.
An effect can be found in two ways:

• The Relative Risk can show an effect. E.g., RR = 2.02 (95% CI = 1.05-3.87) excludes the "1" in the 95% CI. Both boundaries are either above- or below 1, and therefore the RR is significantly different from 1.
• The trend can show an effect. E.g., RR = 1.49 (95% CI = 0.88-2.52; P = 0.008) includes the 1 in the 95% CI through it's lower boundary (0.88). Therefore, the RR is not significantly different from 1, but the trend is significant (< 0.05).

This site: For the systematic reviews on this page, levels of evidence will be created based on the (non)significant effects found in the baseline articles. Effects in the reviews will be included if found as an RR, and/or as a trend.
b) Definition of significance.
Background: Though recent articles seem to use the same definition for a "significant association" (based on P-value for trend), no consistent definition can be found for the term "nonsignificant association". In various articles, the term "nonsignificant association" was used for fairly strong associations without a dose response, or for weak associations with a dose response, but with a P-value sometimes exceeding 0.20 or even 0.40. Also, some authors will define an RR of 0.90 (95% CI = 0.72-1.14) as a nonsignificant protective effect - even when the P-value is > 0.20 - while most will speak of "no association".

This site: For the reviews on this site, associations will be defined as "increased risk" or "protective effect". An addition is given (in parenthesis) to make clear if the association was linked to the risk, or the trend.
Any of the following definitions is required for an effect:

• A significant effect:
  1) The 95% CI does not embrace a RR of one.
  2) P (for trend) = ≤ 0.05.
  3) The article provides no P-value, but uses the term "significant" in relation with the association.
• A nonsignificant effect:
  1) One boundary from the 95% CI embraces the one, and the other boundary is ≥ 10% different from the one (e.g., 95% CI = 0.90-1.00, or 1.00-1.21 would indicate a nonsignificant protective effect, and a nonsignificantly increased risk, respectively).
  2) P (for trend) = > 0.5, and ≤ 0.1.
  3) The article provides no P-value, but used the term "nonsignificant" in relation with the association.

Note: Definitions 1), and 2) were the primary criteria for defining an effect. And therefore overruled definition 3 if contradictory.

6) ADJUSTMENT FOR CONFOUNDERS.
a) Whole foods: In general, RRs were used from the model which adjusted for the largest number of confounders. However, if a model was found which additionally adjusted for nutrients-only, it was chosen not to use RRs from this model. This might imply that an effect from a nutrient can be equally found among all dietary sources of this nutrient. It was decided that this is an assumption which does not reflect current knowledge about nutrients. These levels of adjustments were also chosen for nutrients found to reflect consumption of whole foods, and who were therefore included in the analysis of whole foods, such as dairy calcium.
b) Nutrients: The model was chosen which adjusted for the largest number of confounders, including other nutrients.

Defining the level of consumption at which an effect was found.

Dealing with different units of measure for vegetables & fruits.
When evidence suggested a possible association, results were put into graphics. Results often were defined in grams or servings, but some authors published results in units (= grams/servings)/(X kcal./KJ)/day. The following conversions were used to switch between different units of measure:

• -One serving of total vegetables = 77 g. One serving of total fruit = 80 g (1).
  -Serving sizes from specific fruit/vegetables items were derived from the USDA (2).
  -No attempt was made to convert different units of measures for vegetables/fruits subgroups, but graphics of these variables were created when results were defined in identical measures for units of consumption.
• One serving of total vegetables or total fruit = 0.5 cup.
• When units of consumption were defined in consumption/1000 kcal./day, it was assumed that the average intake of energy can be found at the level of 2,500 kcal among men and 2,000 kcal among women.

Dealing with different units of measure for milk.

• One cup or glass of milk = 245 ml.
• One pint of milk = 0.568 l.

Results put into graphics.
An attempt was made to find out if evidence could be found for a possible association at certain levels of consumption. Graphics were made to see if significant results from different cohorts overlapped at any given level of consumption. Two additional criteria were created to enable quantification of results. Results were included in graphics if:

• Results were published as a categorized variable, and in at least 3 units of consumption. Excluding study results published as a continuous variable.
• Results were published as consumption of a specific amount (grams, servings, or cups) over a given time period. Excluding study results published in frequency/time period, or without a definition for different units of consumption.

|References:
1) He FJ, Nowson CA, Lucas M, MacGregor GA. Increased consumption of fruit and vegetables is related to a reduced risk of coronary heart disease: meta-analysis of cohort studies. J Hum Hypertens. 2007 Sep;21(9):717-28. Abstract
2) USDA Nutrient Data Labatory. Link|

Calculating the effect size: Average RR

Background: Often, authors from meta-analysis' use a model to calculate the RR from RR's of individual studies which are given a certain % weight (World Cancer Research Fund. 2007; Skeaff CM. 2009; Soedamah-Muthu SS. 2011). This % weight is then equal to the power of this individual study to contribute to the RR from the overall meta-analysis. An example of this type of analysis can be seen in the table below. The table shows the association between an increment of 2% of total energy from trans fat, and CHD (Skeaff CM. 2009).

The first association examined (CHD death) most clearly shows the possible implication of giving individual studies a specific weight, based on a chosen model. The RR after meta-analysis is 1.21 (0.89-1.65), and the RR is calculated from 2 studies (cohorts). The size of this effect is mostly driven by results from the "Zutphen Elderly Study", which has been given a 73.56% weight to contribute to the RR. But when we take a look at the specifics from both cohorts, we see something unexpected: The amount of CHD deaths in the "Zutphen Elderly Study" is only 49 out of 278 total events (17.6% of all cases), and the amount subjects included in this cohort is only 667 out of 44,424 subjects (1.5% of all subjects).
How can a cohort contributing to only a small minority of both the amount of events, and the amount of subjects, be given so much power to contribute to the RR from the overall meta-analysis? This is partly driven by the sample size. A larger % of subjects within the "Zutphen Elderly Study" died from CHD (49 out of 667 = 7.3%).
Giving % weight to individual studies in this way is a highly debatable way of creating meta-analysis'. Both the cohort size and the follow-up period of the individual cohorts should be taken into account when a meta-analysis is done. And the combination of the cohort size and follow-up period is reflected in the amount of cases in a cohort. Simpified: a cohort of 1,000 subjects followed up for 20 years will create approximately the same amount of cases as an identical cohort of 10,000 subjects followed up for 2 years (taking age into account, which itself is a risk factor). In this example, both cohorts will provide the same amount of person-years of follow-up.
In the example above, a more reliable way to give the individual cohorts a % weight, would be to divide the weight based on the % of cases at which the individual cohorts, contribute to the total amount of cases. This type of calculation also takes into account the sample size. This would give the "Zutphen Elderly Study" a 17.6% weight, instead of 73.56%.

Thise site: It was decided that the strength of the effect should be a criterium for linking the results to the different levels of evidence.
The "average RR" is the average of RR's from all different cohorts for any variable. And this was adjusted for the number of cases per cohort, which is a reflection of both the cohort size, and the follow-up period.
This means that the number of cases per cohort was multiplied by the RR, creating a "total RR" per cohort. The total RR's for all cohorts were added up, and then divided by the total number of cases from all cohorts.

|References:
Skeaff CM. Dietary fat and coronary heart disease: summary of evidence from prospective cohort and randomised controlled trials. Ann Nutr Metab. 2009;55(1-3):173-201. Link.
Soedamah-Muthu SS. Milk and dairy consumption and incidence of cardiovascular diseases and all-cause mortality: dose-response meta-analysis of prospective cohort studies. Am J Clin Nutr. 2011 Jan;93(1):158-71. Link.
World Cancer Research Fund/American Institute for Cancer Research. Food, Nutrition, Physical Activity, and the Prevention of Cancer: a Global Perspective. Washington DC: AICR, 2007. Link.|

Judging the evidence:

Motivation for criteria used to create the evidence model.
Background: In 2007, The World Cancer Research Fund (WCRF) published the largest systematic literature review about the relation between diet (+ other factors) and cancer ever. The goal of this report was to review all the relevant research, using the most meticulous methods, in order to generate a comprehensive series of recommendations on food, nutrition, and physical activity, designed to reduce the risk of cancer and suitable for all societies.
The WCRF found that the best evidence does not come from any one type of scientific investigation. It comes from a combination of different types of epidemiological and other studies, supported by evidence of plausible biological mechanisms. Still, for all 3 main levels of evidence, results from ≥ 2 independent cohorts studies were required (or ≥ 5 case-control studies for the lower 2 main levels of evidence).

Results from systematic reviews examining the relation between dietary items and health outcomes are completely based on the choice which data to include/exclude as "relevant" for the review. Fairly often, the author of an article provided no relative risk for an association, but only stated that "no (significant) association was found" in a given cohort. Both systematic reviews and meta-analysis' often do not mention "results" from these articles, or perhaps were not able to find them from the literature search. It is obvious that such statements can not be included to calculate the relative risk in a meta-analysis.
Still, this type of data should be included somehow when the evidence for an association is judged. For some associations examined, authors from a fairly large amount of cohorts (≥ 5) did not provide a RR, but only stated that some type of association was found. Excluding this type of data in an analysis can seriously bias results in any given direction. Results without a RR can not be included in a meta-analysis, but there are 2 ways to adjust for this type of data:

• A relative risk of 1.00 can be given to cohorts in which "no (significant) association was found". This will always attenuate the effect size from a meta-analysis.
• Grading the evidence is based on the consistency at which (non)significant effects were found. This will not influence the effect size, but can direct the way in which data from a meta-analysis is presented. E.g., inconclusive-, weak-, strong-, or very strong evidence.

This site: Judging of evidence was inspired by the WCRF model. Of course, the methodology for the systematic reviews on this site can not compete with the methodology from the WCRF, since reviews were created solely by results from cohort studies. Still, levels of evidence were created with similar goals. For the highest level of evidence this was defined as: "To be robust enough to be highly unlikely to be modified in the foreseeable future as new evidence accumulates".
To reach this goal within limits of a single study design, some criteria were created to judge the evidence, being consistency of effects, and the strength of the effect. Factors taken into account were: a) the amount of cohorts, b) the size of the cohorts, and c) a criterium incorporating both variables "cohort size" and "follow-up time" in a single variable, namely the amount of cases.
All types of results were included to grade the evidence for an association. Results without a RR were not included to calculate the effect size through the average RR, but were included to grade the level of evidence for an association.

Levels of evidence.
Background: Levels of evidence were created, based on consistency of effects, and the strenght to withstand possible opposite findings from current ongoing cohorts within the next couple of years.
A lot of cohorts are currently providing information about diet & various health outcomes. Some cohorts are of very large size: 200,000 to 1,000,000 subjects (1-4). Each of these single cohorts may - on their own - provide information based on an amount of disease cases so large, that this may compensate for associations found by the current combined results of all cohorts for a large amount of disease outcomes!

This site: Evidence is divided into 3 main levels (possible, probably, convincing) which are based on:

• The amount of cohorts in which an effect was found.
• The size of the cohorts in which an effect was found.
• The strength of the effect.

In addition, nonsignificant associations can be judged as suggestive evidence.

Cohort size.
It was decided that cohorts of small size should have less power to influence the judgement of evidence than cohorts of moderate-large size, and that results from cohorts of very large size would be necessary to provide sufficient strenght for the higher levels of evidence.
Cohort sizes were defined af follows:

• Very small size: < 5,000 subjects.
• Small size: ≥5,000-< 20,000 subjects.
• Moderate size: ≥ 20,000-< 60,000 subjects.
• Large size: ≥ 60,000-< 150,000 subjects.
• Very large size: ≥ 150,000 subjects.

Strenght of the effect.
The strength of the effect is divided into 3 levels:

• Weak effect: RR = ≥ 0.91, or ≤ 1.10.
• Moderate size effect: RR = ≥ 0.83-< 0.91, or > 1.10- ≤ 1.20.
• Strong effect: RR = < 0.83, or > 1,20.

In general, any evidence requires:
Little heterogeneity between study results. Defined as: a) no significant association in the opposite direction in a cohort of moderate-large size (≥ 20,000 subjects), and b) no nonsignificant association in the opposite direction in an amount of cases covering ≥ 10% of the total amount of cases from all cohorts combined.
The 4 levels of evidence are described here. Criteria differ in consistency of effects, and the strength of the effect. For analysis stratified by sex, requirements for cohort size are divided by 2.

1) Suggestive evidence.
This level of evidence stands for an interesting finding in more than one study:

• A significant association in ≥ 2 cohorts of any size (including ≥ 33% of all cases), or
• A nonsignificant association in ≥ 4 cohorts of any size (including ≥ 33% of all cases).

2) Possible evidence.
This level of evidence stands for a significant association in more than one study of ≥ moderate size:

• A significant association found in at least 2 cohorts of moderate-large size (≥ 20,000 subjects each), and including ≥ 50% of all cases.

3) Probable evidence.
This level of evidence stands for fairly consistent findings from multiple studies.
Both following findings are required to carry the lavel of "probable evidence":

• A significant association found in at least 4 cohort studies, including at least one cohort of very large size (≥ 150,000 subjects), and including ≥ 50% of all cases.
• The average RR for all cases shows a moderate size/strong effect.

4) Convincing evidence.
This level of evidence stands for consistent findings from multiple studies, some of which are of very large size.
Both following findings are required to carry the label of "convincing evidence":

• A significant association found in at least 6 cohort studies, including at least two cohorts of very large size (≥ 150,000 subjects), and including ≥ 67% of all cases.
• The average RR for all cases shows a strong effect.

References:
1) The Million Women Study.
2) The NIH-AARP Diet & Health Study.
3) The EPIC Study.
4) The Multiethnic/Minority Cohort Study.