USDA FoodData Central holds five databases, not one

FoodData Central is five databases behind one search box

USDA FoodData Central is not one database. One search box sits in front of five collections, built in different ways for different jobs, and on screen a row from one looks like a row from any other. Foundation Foods and SR Legacy hold analysed whole foods. Branded Foods holds what manufacturers print on labels, FNDDS holds the food codes behind the national dietary surveys, and Experimental Foods holds research data. For a client's plan, which collection a number came from matters more than the number.

FoodData Central provides a web-based data system that is composed of five distinct types of food and nutrient composition data, each with a unique purpose and acquired with different approaches.

The data type documentation sets the five side by side. This is the short version of that table.

Foundation Foods is the analysed data, and it is the smallest set

Foundation Foods is the one data type built to show its working. A food profile carries the number of samples behind it, where they were collected, when, which analytical method was used, and for crops the genotype and the production practices. USDA added a Limit of Quantification field to it, so a value too small to measure with confidence reads as a less than value instead of a zero.

That depth costs time, which is why the set is small and grows slowly. Two releases a year, in April and October. If you want the most defensible number for a raw ingredient, this is where to look first, and if the food is not there yet, you fall back.

SR Legacy stopped in April 2018 and still runs most of the world's apps

SR Legacy was the primary food composition database in the United States for decades. USDA describes its values as derived from analyses, imputations and the published literature, and says those values have supplied most other public and private food composition databases. The last release was April 2018, and USDA states it will not be updated again.

That is worth holding in your head the next time two apps disagree. A tracking app built on a 2018 snapshot and a coach's plan built on this year's Foundation Foods release are both quoting USDA, honestly, and neither is quoting the same USDA. The other reasons behind that gap are in why two apps give the same meal different calories.

Branded Foods is label data, and it moves every month

Branded Foods comes from a public-private partnership: food industry data providers send the label information, and USDA standardises how it is presented. The data type is refreshed monthly through the API, and the bulk downloads are regenerated every six months.

So a Branded Foods row is a label, with everything a label is. The manufacturer chose the serving size, the values were rounded to the increments the labelling rules allow, and the recipe behind the barcode can change without the barcode changing. Useful for the tub of yoghurt a client actually buys. Wrong for "chicken breast".

FNDDS exists to code survey answers, not to portion a client's dinner

The Food and Nutrient Database for Dietary Studies holds the foods reported in What We Eat in America, the dietary intake part of NHANES. It is released every two years, and its values come from Foundation Foods and SR Legacy rather than from new analysis.

FNDDS carries portion weights, which is why it turns up in apps. It is a reasonable source for a mixed dish a client describes in words, such as a chicken salad sandwich, because somebody has already decided what goes in one. It is a poor source for an ingredient you are going to weigh.

The calories are calculated from the macros, not measured

No one burns a portion of each food in a calorimeter. USDA's FAQ says most FoodData Central energy values are calculated with the Atwater general factors of 4, 9 and 4 for protein, fat and carbohydrate, and that carbohydrate by difference is what remains after water, protein, fat, ash and alcohol are subtracted from 100. Some foods carry a second energy value calculated with Atwater specific factors alongside it.

Those two systems do not agree. The FAO's 2003 report on food energy sets the general factors at 4.0 kcal/g for protein, 9.0 for fat, 4.0 for carbohydrate and 7.0 for alcohol, and reports that the specific factors run from 2.44 to 4.36 kcal/g for protein and, for carbohydrate, from 2.70 kcal/g in lemon and lime juices to 4.16 kcal/g in polished rice.

Application of general factors to the mixed diet common in the United States resulted in values that were on average about 5 percent higher than those obtained with specific factors.

For a few foods, including snap beans, cabbage and lemons, the same report puts the gap between the two systems at 20 to 38 percent. A calorie figure is an estimate with a method attached, and the method is a choice somebody made before you saw the number.

Pick one data type per food and write the choice down

Our rule: build a client's whole-food plan on Foundation Foods and SR Legacy, and reach for Branded Foods only for the packets the client actually buys. Mixing them inside one meal is how a plan ends up with a protein total nobody can reproduce: the rice from SR Legacy, the chicken from a supermarket's own label, the olive oil from a brand that reports per 15 mL.

The rule that survives a busy week is one line per food in your own list: the food, the data type, the FDC ID. When a client questions a number in eight weeks, you can open the same row you opened the first time. Peas & Pans writes that row for you on an imported recipe: every ingredient is matched to USDA data, and the match is recorded rather than remembered. To see that match on a recipe of your own without an account, paste the ingredients into the free recipe nutrition calculator.

FoodData Central will not resize a portion for you either. USDA says so plainly: there is no way inside it to change a portion size and its nutrient values automatically. Values are per 100 g unless the profile says otherwise, and the arithmetic down to a 140 g serving is yours.

The API allows 1,000 requests an hour, and the data are public domain

If you or a developer you hire wants the data programmatically, the FoodData Central API has two endpoints, one to search for foods and one to fetch the details of a food. Anyone may use it, with a data.gov key in every request. The default limit is 1,000 requests per hour per IP address, and going over it returns HTTP 429 and blocks the key for an hour.

The data themselves are in the public domain under CC0 1.0, with no permission needed. USDA asks that you list FoodData Central as the source and, when you can, tell it what you built. That makes it the one nutrition database you can build a client-facing document on without a licence conversation, which is a large part of why almost every app you have used is quoting it back to you.