Food composition database development for between country comparisons

Nutritional assessment by diet analysis is a two-stepped process. The first step is the evaluation of food consumption, and the second the conversion of food into nutrient intake. To do this we need a food composition database, which lists the mean nutritional values for a given food portion. We then multiply food intake by the mean nutrient content of that amount of food (obtained from the food composition database) [1]. As most professionals conducting nutritional assessments are primarily concerned with the evaluation of food intake, a large part of the literature on nutritional assessment focuses on minimizing errors at this step. However, errors and discrepancies may arise in nutrient estimation from the food composition database, arising mainly from the assays used in nutrient estimation and the sampling procedure and date of the test foods [2]. Between-country comparisons are particularly prone to error when different food composition tables have been used to estimate nutrient intake. In recognition of these difficulties there are ongoing efforts since 1984 to standardize food composition databases over the world [2]. This is a continuous process because of increased global trade in foods, changes in fortification policies, development of newer assays for nutrient estimation, and addition of new foods in the global diet.

Over the last 40 years, global trade of foods grew in both developed and developing countries [3]. Import of foodstuffs in developing and developed counties grew by 115% and 45% respectively. Since the late 1990s, the least developed countries have become the major net importers of agriculture products [3]. Foods traded internationally include cereals, edible oils, animal products, sugar, fruits, vegetables, nuts, and coffee. Global food trade has also affected North American food consumption. Imports to America between 1980 and 1997 rose for fish from 45% to 62%, for fresh fruits from 24% to 34% and vegetables from 5% to 10% [3].

We have started recruiting participants into the Prospective Urban and Rural Epidemiologic Study (PURE), an investigation that evaluates the societal, familial, and personal determinants of diet and their relation with obesity and chronic disease globally. When recruitment is complete we anticipate participation from about 14 countries and approximately 120,000 persons. We are using the food frequency questionnaire (FFQ) to evaluate diet in this study as described earlier [4]. The use of a multi-country food composition database that would provide comparable data across these countries is a critical part of our nutrient assessment. The objective of this paper is to describe our process and rationale for the selection and adaptation of food composition tables to make cross-country nutrient comparisons.

The last time the local food composition databases were partially updated ranged from 3–15 years ago, as compared with one year for USDA's SR18 (Table 1). The number of food items ranged from 201–2000, and maximum number of nutrients from 16–33 for the local food composition databases as compared with 7,146 food items and 136 nutrients for USDA's SR18. The local food composition databases did not provide detailed information about how the test foods were sampled, or the assays that were used for nutrient estimation, as compared with detailed documentation that was available for the USDA nutrient database.

Table 2 contains a list of the different types of uncooked rice in the USDA nutrient database, and the nutrient values for uncooked rice from the respective local food composition tables. At the end of the table there is the result of applying our proposed algorithm to select the rice in the USDA nutrient database that is closest in nutrient content to the local rice.

The nutrient content of four recipes of stew (100 g) varied depending upon the food composition table that was used in the estimation in all the countries that we examined (Table 3). For instance, carbohydrate intake estimated from the PURE-USDA database for Argentina was 11.0 g versus 12.5 g estimated using the same data but the PURE-USDA database for Zimbabwe (Table 3). There were differences in micronutrient and macronutrient estimations depending upon the choice of the database.

The intraclass correlations using FFQ data comparing nutrient intake from stew estimated using two nutrient databases both derived from the USDA nutrient database were generally good, but sometimes varied for macronutrients and micronutrients (Table 4). For instance the intraclass correlation coefficient for folate in Argentina was 0.78, for protein in Brazil it was 0.54, and for fat in UAE it was 0.79 (Table 3).

Our strategy to improve between-country comparisons is to use a common food composition database from which to derive nutrient estimates of foods for all countries. The need to have a common food composition database was felt in an analysis of food composition databases of countries participating in the European Prospective Investigation of Cancer and Nutrition (EPIC) study [19]. Even though the nutrient estimates obtained for most nutrients were comparable some were not [20, 21]. Nutrient estimates from FFQ data from Chile using American and British food composition tables had good overall agreement; the agreement was excellent for macronutrients (ICC for energy = 0.98, protein 0.98, carbohydrate 0.97), but less for few micronutrients (ICC = Iron 0.86, Zinc 0.91) [22].

When we compared two food composition databases, both derived from the USDA nutrient database, we found good overall agreement, but there were several notable exceptions (Tables 3, 4). This difference was there even though both tables were derived from the same version of the USDA nutrient database but only differed in the food that was selected. These differences in estimates therefore do not reflect the assay or method of test food selection, but likely true variation in nutrient content of the foods, or the different methods of cooking in the case of mixed dishes. This underscores the need to select the appropriate food from the USDA nutrient database. We did this by choosing the food in the USDA nutrient database that was most similar to the local food with respect to nutrients of interest. Moreover, the selection algorithm is reproducible, and we consistently applied it to develop the food composition databases for all the countries we studied. Any residual errors in nutrient content of foods resulting from this process were probably similar in all the country databases.

The errors that arise in nutrient estimation as a result of nutrient databases are systematic [23]. For instance, if an analytical method that underestimated fiber in foods was used in a food composition database, nutrient estimation using that nutrient database will consistently underestimate fiber. If different food composition databases were used to estimate nutrients in different countries the systematic errors would be different, and in all likelihood vary in magnitude and direction for different nutrients. Using different food composition databases for different countries would increase errors.

Even if food composition tables are based on sound principles, there are differences in estimates of nutrients depending upon which one is used for nutrient estimation. The advantage of using a common food composition database is that the errors are consistent between the countries, hence making data more comparable. For instance, if fiber intake is underestimated because of the method used to determine fiber content in the food composition database, it is equally underestimated in all the countries. This would impact estimates of absolute intake but probably not the relative ranking by fiber intake.

With increasing globalization the food available in different countries is not all grown in the region [3] and this is a potential source of errors in nutrient estimation. For instance, using the composition of an orange grown in Argentina for nutrient estimation would be erroneous if it were imported from elsewhere. To address this issue we consulted regional food composition databases and obtained the nutrient composition of a food commonly used in that country and matched it up with the food with the closest nutrient composition on the USDA list. Shai et al used a similar approach to develop a food composition database in Israel [24]. Choosing the food from the food USDA nutrient database that most closely resembles local foods would further reduce errors in nutrient intake estimation between countries. Another reason why our approach is less likely to be prone to error is the consistent method we used in estimating the nutrient content of mixed dishes. We measured dry matter and took into account retention and loss of nutrients as a result of food preparation. Many food composition tables did not consider yield and retention factors. Finally, local food composition tables do not have a comprehensive list of foods or nutrients; when these are missing the nutrient value from the missing food or nutrient are frequently taken to be zero resulting in error, when in fact they are missing.

Our approach also has some limitations. First, there are a fixed number of foods for which we have comparable food data. We therefore cannot use the common food composition database in its present form for the analysis of 24-hour recalls or food diaries but we will expand this list when we do study to validate our FFQs. However, all the foods on our FFQs are found in the table and it thus can be used to estimate nutrients from data generated with those FFQs and the tables can be updated to accommodate more foods. Second, the use of average portion sizes and categories of intake to estimate intake frequency are potential sources of errors in estimation using the FFQ, but these have been shown to be small [22]. Third, most importantly the errors are consistent [23], and thus not likely to influence the relative ranking of the persons by nutrient intake, and increasing the between country comparisons. Fourth, the USDA estimates may not be completely accurate [25], or the test foods in the USDA nutrient database may be dissimilar from the local foods [25]. Fifth, we cannot validate our method. To do that we would need to have a gold standard. In this case the gold standard would be nutrient estimates from country specific food composition tables, which would be prohibitively expensive to develop. However, the method has face validity, is reproducible, and removes potential errors arising from different assays to estimate nutrient content and food selection. Last, it is still possible that there may be large discrepancies in nutrient content of foods between the USDA database and what is found in the field because of manufacturing practices. For instance, an analysis of dietary sources of alpha-linolenic acid (from solid cooking oil) in Iran revealed very high trans fat contents [26], not accounted for by the USDA database. To correct for such anomalies it is necessary to periodically sample and chemically analyze selected commonly consumed foods.

We have prepared a food composition database primarily based on the USDA food composition database, from which to estimate nutrients using FFQ data from different countries. By using this method we have ensured that the units of measurement, sampling scheme for foods, and assays used for nutrient estimation are consistent and current, and have taken into account local conditions. Using this common metric for nutrient assessment will likely reduce differential errors arising from nutrient estimation and make the between country comparisons more valid.