4
Alternative Approaches for Estimating Food Availability: International and Domestic
This chapter summarizes the third session of the workshop, which focused on alternative approaches for estimating international and domestic food availability. Jay Variyam (Economic Research Service [ERS]) moderated the session. The first speaker was Josef Schmidhuber (Food and Agriculture Organization [FAO]), who described FAO’s approach for estimating food availability. Klaus Grünberger (FAO) then described FAO’s approach to comparison and reconciliation of food consumption from household surveys and food balance sheets. Aylin Kumcu (ERS) described potential availability of alternative data sources, including scanner data. The fourth speaker, Alanna Moshfegh (Agricultural Research Service [ARS]), talked about her group’s work to disaggregate food mixtures in nutrition data. The final sections in this chapter present a summary of open discussion between panelists and the audience, followed by a facilitated discussion conducted by steering committee chair Mary Muth.
STATEMENT OF JOSEF SCHMIDHUBER FAO’S APPROACH FOR ESTIMATING FOOD AVAILABILITY
Schmidhuber explained that the FAO food balance sheet system is currently in transition to update conversion factors and extraction rates and, importantly, to develop new imputation methods for all elements (variables) of its balances. In tandem, the questionnaires sent to FAO member countries are being improved and new technologies to ease data
collection, processing, and dissemination are being adopted to harness operating efficiencies and ultimately reduce costs.
He said FAO currently publishes comparable food balance sheets with data from 1961 to 2011 for 185 countries. The most recent estimates extend the series to 2011 and were published in May 2014; a new statistical working system will allow FAO to generate preliminary data up to 2013 by the end of 2014. He said FAO plans to provide open and easy access to the data. He noted some improvement in dissemination with the FAO’s FAOSTAT version 3 website.1 The food balance sheets database is a four-dimensional cube, and the new system pivots the data in ways that make them easy to access. He said FAO’s key goal is to promote statistics for evidence-based policy making. FAO uses the food balance sheets to prepare food security indicators for prevalence of undernourishment, indicators of food adequacy, and so on. They have also established simple balance sheets for early warning purposes and worked on tracking dietary patterns.
He said FAO tries to serve the needs of its many internal and external clients. One key type of external client uses the data in economic models. For example, the International Food Policy Research Institute (IFPRI) uses the FAO commodity balances, including the food balance sheets, in its IMPACT model. The same holds for the FAO/OECD Aglink/Cosimo model,2 and FAO’s long-term projections to 2050 (FAO’s global perspectives studies).3
FAO is pursuing a two-pronged approach to improve the balance sheets. One prong is to make sure countries have available the best possible data collection methods. FAO is moving into the next World Census on Agriculture (WCA), called the WCA2020 round, starting in 2015. The WCA2020 will make numerous recommendations to improve data coverage and quality and to help countries access new and efficient data collection methodologies. Other programs to improve data collection methods include the Global Strategy to Improve Rural and Agricultural Statistics (GS), Agricultural Market Information System (AMIS), and, in part, CountrySTAT.4 Second, FAO is trying to improve imputation methods and plans to roll them out to member countries. He noted that much of the food utilization data that FAO requests from countries are not provided, forcing FAO to impute them.
________________
1See http://faostat3.fao.org/faostat-gateway/go/to/home/E [July 2014].
2For information on the model, see http://www.oecd.org/site/oecd-faoagriculturaloutlook/oecd-faoagriculturaloutlook-tools.htm [July 2014].
3For information on these studies, see http://www.fao.org/waicent/faoinfo/economic/esd/gstudies.htm [July 2014].
4For more information, see http://www.fao.org/economic/ess/ess-capacity/countrystathome/en/ [July 2014].
Schmidhuber said FAO is also trying to tap into new technology for data collection. For example, FAO has an application programming interface (API) with the United Nations Statistics Division (UNSD) to get trade data automatically and seamlessly. FAO is trying to establish the same technology with member countries, even though there is greater reluctance to engage in such automated processes at the level of individual countries. Schmidhuber noted that FAO does not generally collect data directly within a country, instead relying on data already collected by the country. The only notable exception is a cooperative agreement with Gallup called Voices of the Hungry.5
Because of this reliance on secondary data, FAO is trying to work with countries to improve their data collection methodologies through a number of efforts, including, as noted above, GS, WCA, AMIS, and, in part, CountrySTAT. Such efforts also include work on experimental designs to measure food waste at different stages of the value chain.
Another FAO effort, he said, has been to ensure that member countries use internationally accepted standards to report their data, using classifications such as the Harmonized System (HS) or the Central Product Classification (CPC). FAO also tries to make sure member countries compile data in a comparable manner so the agency can be assured that it has comparable food balance sheets. Before overwriting official data, FAO goes back to member countries, asking them to verify their submissions. FAO undertakes extensive quality checks to ensure there are no logical errors, transcription errors, or shifts in decimal points.
He noted the generic FAO balance system is similar to that described earlier for the Food Availability Data System (FADS). For grains, however, FAO includes use for food, feed, seed, waste, and other uses in separate columns, whereas a U.S. commodity spreadsheet typically has fewer end-use columns. Like the U.S. system, the FAO system is set up to ensure that domestic supply is equal to domestic utilization, and total supply is equal to total utilization. He noted that few countries report all end-uses. Most countries report trade and production, some also report stocks, but all are estimates with varying levels of accuracy. This means that there is a considerable need to impute missing data, he explained.
FAO is currently reviewing and revising their imputation approaches, which Schmidhuber highlighted, starting with a description of the new feed use imputation system. He explained that few countries have official feed use estimates based on feed surveys; instead, most countries, including the United States, compute the estimate as a residual. The new FAO feed use imputation method links feed use estimates to four fac-
________________
5For more information, see http://www.fao.org/economic/ess/ess-fs/voices/en/ [July 2014].
tors: number of livestock, composition of the herd, feeding intensity (the share of compounding concentrate feed and the total ration), and feeding efficiency.
Schmidhuber said many developing countries are moving from a backyard livestock system to a family-based livestock system, and eventually to an industrialized system. To capture this, the FAO model estimates the amount of feed needed to support the food requirements of a country’s herd. This estimate is then mapped into feed availability and allocated to livestock based on country-specific feeding practices, which allows for a number of consistency checks. For example, it is possible to ensure that grain use is consistent with changes in feeding intensity, as well as the number and types of animals to be fed.
He observed that industrial use of agricultural products was fairly small when oil was cheap, but this situation changed dramatically with higher oil prices. FAO has observed that more vegetable oils and cereals are used either for biodiesel production or ethanol and that more agricultural products are used to produce paints, detergents, and starch-based products. Higher energy prices not only made agricultural products competitive for the biofuels market (not ignoring the help of subsidies, he noted), but also traditional or nonexisting industrial use became competitive as agricultural raw materials became more competitive with traditional sources of energy and synthetic raw material. FAO is trying to collect as much information as possible on these new uses from its member countries. Many countries provide information on production of biofuels, and some have information on paints, detergents, and starches. For the rest, FAO imputes based on an economic model.
Schmidhuber said FAO is also reviewing and updating its classification system. In the old approach, the food balance sheets were classified using the FAOSTAT Commodity List (FCL) that was developed decades ago. However, the FCL largely remained static, so few countries use it today. One of the consequences has been a sometime arbitrary conversion process between data provided by countries and the FAO classification system. In the future, the FCL will be replaced with the UN CPC Version 2.1, expanded and the HS 2012, where available.6 The CPC and HS have a high level of communicability, he said, allowing complete one-to-one
________________
6These classification systems and their evolution are described at http://www.fao.org/fileadmin/templates/ess/ess_test_folder/Workshops_Events/APCAS_24/Documents_and_ppt/APCAS-12-INF5-International_Classification_of_Agri_Commodities.pdf. This is a summary of a meeting to discuss classification systems that was held in October 2012. It also lists other FAO websites such as the page on classifications and standards: see http://www.fao.org/economic/ess/ess-standards/en/ [September 2014] and the page on FAOSTAT commodities and inputs, see http://www.fao.org/economic/ess/ess-standards/commodity/en/ [September 2014].
commodity mapping between the two systems. This will eliminate mapping errors, reduce the country response burden, and increase international comparability.
FAO publishes balances for more than 80 primary commodities plus 10 commodity groups. Underlying these primary product balances is a vast array of processed products, which are converted into primary equivalents using extraction rates, conversion factors, and a detailed structure of the processing chain. The conversion and aggregation process is referred to as “standardization.” FAO has developed standardization commodity trees, flow diagrams that trace the flow of a primary agricultural commodity along the value chain from raw agricultural commodity to processed products, to facilitate this process. Commodity trees for many commodities can be found on the FAO website.7 For example, he stated that quantifying wheat supply is quite complex, because wheat is processed into flour, bran, and germ. He noted one of the key wheat products is bran and it goes into many other products such as breakfast cereals and feed. In addition, wheat flour can go into cereal, bread, pasta, pastry, or starch and gluten. There may be subsequent processing levels as well. FAO estimates the amounts of wheat included in imports and exports of processed products and includes them in the wheat balance sheets. FAO has mapped these multi-ingredient products back into primary products via commodity trees.
Schmidhuber provided a detailed illustration of the standardization process using millet, a product with one of the simplest commodity trees because millet is not consumed directly and is used for human consumption only as millet flour and bran. He described an FAO balance sheet for Niger in 1981, illustrated in Figure 4-1. Like FADS, FAO starts with the basic supply and disposition balance in primary products. In the first row of Figure 4-1 labeled millet, the columns show raw millet grain production (P), imports (I), exports (X), and stock change (dSt) as well as estimates for amounts used for feed, for seed, waste, and other uses (O_Use). The column showing millet processed into food (Food Proc.) is the residual.
In the line in Figure 4-1 for flour, FAO uses an exogenous extraction rate—in this case, 0.7—to estimate the fraction of millet flour that is extracted from millet grain. Thirty percent is bran or is lost in the milling process. The product of the amount of millet grain used for food and 0.7 is an estimate for the amount of millet flour produced as shown in red in the row labeled flour. In the third line, FAO shows imports of millet flour, and in the fourth line, in green, shows the negative of exports of
________________
7See http://www.fao.org/fileadmin/templates/ess/documents/methodology/tcf.pdf [September 2014].
FIGURE 4-1 Example of Food and Agriculture Organization (FAO) balance sheet for millet in Niger, 1981.
NOTE: dSt = stock change, Food Proc. = processed into food, I = imports, O_Use = other uses, P = production, X = exports.
SOURCE: Prepared by J. Schmidhuber for presentation at the workshop. Based on FAO food balance sheet (FBS).
millet flour, also in green, to get the total amount of millet flour available for consumption. Creating a sub-balance at the level of millet flour offers the possibility of using a commonly available, region-specific calorie conversion factor for millet flour (3.52) to estimate the calorie content of the flour produced, imported and exported. These estimates are shown in the column labeled Calories, with the last row of that column showing the calorie content of the flour that was available for consumption. Calorie conversion factors are seldom reliably available for commodities such as millet grain.
For purposes of presentation in the food balance sheet (FBS), the results for millet flour imports and exports are converted back to units of raw millet by dividing by 0.7 (as shown in the row conversion back to primary millet). The last two rows of Figure 4-1 show the standardized FBS for millet that includes imports and exports of both millet grain and its primary product, millet flour.
STATEMENT OF KLAUS GRUNBERGER COMPARING AND RECONCILING FOOD CONSUMPTION FROM HOUSEHOLD SURVEYS AND FOOD BALANCE SHEETS
Grünberger described his current project at FAO to develop an approach for using food consumption results from household surveys to improve food balance sheets, pointing to Conforti and Grünberger (forthcoming) for additional details. FAO developed a model that adjusts the food group shares in the food balance sheets by reconciling them with the information gathered from survey data.
He noted food balance sheets look at food at the macro level, while household surveys provide micro data. These two types of estimates are not perfectly comparable because household surveys usually measure consumption by considering incremental expenditures. For the most part, the surveys provide estimates of the economic value of food that needs to be transformed into calories obtained from purchases. An additional challenge faced by FAO, he said, is that household surveys in different countries are quite heterogeneous, with differences in timing and content. Some surveys have a recall period of seven days while others have a recall period of one month.
In contrast, the food balance sheets are homogenous because they are an FAO product. He described the challenges with using household survey data. FAO has processed 64 household surveys from 52 countries, primarily low- and middle-income countries. They calculated average consumption from the survey data and categorized the food items into 17 food groups. Then, they compared the consumption pattern indicated by the surveys and by the food balance sheet.
FIGURE 4-2 Total calories in national household surveys and food balance sheets.
NOTE: FBS = food balance sheets, NHS = national household surveys.
SOURCE: Prepared by K. Grünberger for presentation at the workshop. Based on Food and Agriculture Organization data.
He showed a comparison of the total calorie consumption levels from the two sources (see Figure 4-2). On the Y axis is the calorie consumption level from the household surveys, and on the X axis is the calorie availability from the food balance sheets. The dotted line is the equality line. There is a correlation, he said, but also a lot of variability, pointing to an interesting pattern that with increasing food balance sheet calories, the household survey calories are lower. This is consistent with older studies that showed that this difference is significantly correlated with income.
Grünberger then showed Figure 4-3 to illustrate the relative difference between calories measured from household surveys and calories measured from food balance sheets against gross domestic product (GDP) per capita. This graph shows that the difference is not significant in low-income countries. However, with increasing income, the differences increase. He noted even after controlling for some confounding factors, such as the ratio of foods consumed away from home, correlated with GDP, the negative correlation still holds. The difference may be because food balance sheets are too high or because household surveys are too
FIGURE 4-3 Differences in calorie measurements against gross domestic product per capita.
FBS = food balance sheets, NHS = national household surveys.
SOURCE: Prepared by K. Grünberger for presentation at the workshop. Based on Food and Agriculture Organization data.
low. For example, it is difficult to assess food away from home because food type and quantity are rarely specified in the surveys, but measured only with economic value.
Although the data tend to indicate that food balance sheets might overestimate food availability, such as due to underestimation of losses at the retail level, household surveys may underestimate food processed in the hospitality sector, he said. Food wasted in restaurant kitchens, for example, is not included in the household surveys that only include already-prepared dishes. He pointed to a systematic pattern in the differences.
Terming the data noisy, Grünberger said he switched to an analysis of consumption patterns. To do this, he calculated the caloric shares of the food items and calculated the contribution of food groups to the total consumption. He said categorization of the household surveys and the food balance sheet is quite similar, so they were easy to match. Seventeen food groups were developed for comparison.
He displayed Figure 4-4, with 16 plots (for descriptive purposes, tree nuts are combined with oil crops in a single-item group) showing the calorie share from household surveys to the calorie share from the food balance sheets for each food group. He noted a correlation for cereal that does not show up for all other products. For stimulants, there is no apparent correlation. There are some items that might be biased in household survey data such as alcoholic beverages, stimulants, and spices, but these
FIGURE 4-4 Calorie shares by food group.
NOTE: Tree nuts are combined with oil crops in a single-item group. FBS = food balance sheets, NHS = national household surveys.
SOURCE: Prepared by K. Grünberger for presentation at the workshop. Based on Food and Agriculture Organization data.
three groups are of minor importance in terms of their contribution to total calories. Seven food groups were excluded from the analysis: alcoholic beverages, animal fats, miscellaneous food, oil crops and tree nuts, spices, stimulants, and vegetable oils.
He then described the model used to adjust the food group shares of the balance sheet data using the household survey data. Box 4-1 shows the important part of the model, he said. Its objective function was the Kullback-Leibler divergence, a cross-entropy measure that allows for measurement errors in both data sources. It is a divergence measure between old and new shares, where the new shares are better in line with household survey shares. The objective function was minimized subject to stochastic constraints. The error term was added to incorporate the
BOX 4-1
Model Specification
Model to adjust food group shares of balance sheet data using household survey data.
Minimize the objective function (Kullback-Leibler divergence)
subject to the stochastic constraint
s.t. Fi = Ni + εi, ∀i
where the variables Fi, i and N i are the shares of food group i of the updated FBS, the old FBS and the NHS, respectively
εi is an error term with a discrete i,l of L dimensions and respective probabilities wi,l. Probabilities are going to be updated as well.
The model finds updated FBS food group shares that meet the stochastic constraint (2), and at the same time are ‘close’ to the old FBS.
SOURCE: Prepared by K. Grünberger for presentation at the workshop.
known noise in the household surveys, and it was developed as a discrete distribution based on the error structure of the differences between the food balance sheet and household data that were shown in the correlation plots.
According to Grünberger, the objective function can be viewed as a penalty function that searches for updated shares that are close to the old food balance sheets and simultaneously meet the stochastic constraint, which is defined by the household survey shares and the error term. The noise term determines whether household survey shares or food balance sheet shares are favored. If the household survey is regarded as reliable, the result will be closer to the household survey share.
Grünberger next displayed Figure 4-5 to summarize the results of the comparisons and reconciliations. The approach does not change total food balance sheet (FBS) overall calories. Instead, changes were made to the shares of the food groups to bring them more in line with those from the national household surveys (NHS). For each food group, the
FIGURE 4-5 Mean differences between shares of the updated FBS/NHS and old FBS. (See text for explanation of boxes.)
NOTE: FBS = food balance sheets, NHS = national household surveys.
SOURCE: Prepared by K. Grünberger for presentation at the workshop. Based on Food and Agriculture Organization data.
height of the overall box shows the mean difference across countries between NHS and FBS, while the height of the dark box shows the mean difference across countries between the adjusted balance sheet and the original balance sheet. Figure 4-5 shows the adjusted share is between the household survey share and the food balance share for cereals, meat and offal, sugar products, and eggs. For these products, the approach is like a compromise. For roots and tubers, fruits, and vegetables, the adjusted share is close to the household survey share. Shares of cereals increased, while shares of roots and tubers decreased. Shares of vegetables increased, while shares of fruits decreased.
Grünberger called the approach promising, although it has limitations and could be improved. The procedure provides one way to inform the pattern of consumption in food balance sheets using information from household surveys. The analysis can be used to cross-validate household surveys and the food balance sheet, and to detect potential errors. Ultimately, he said, FAO hopes that the approach could be a tool to improve data for the measurement of undernourishment and to show why it is important, at both the level of consumption and distribution.
STATEMENT OF AYLIN KUMCU POTENTIAL USES OF SCANNER DATA AND OTHER DATA RESOURCES
Kumcu described four data sources that might be considered to improve FADS. The first series is data on food expenditures from the
Census Bureau and published by ERS. Annual data are currently available from 1928 to 2011. The second series is data from Information Resources Incorporated (IRI) that includes retail scanner data as well as household survey data. Both datasets are available from 2008 to 2011. Third, she noted the Food Acquisition and Purchase Survey (FoodAPS) has recently been released by ERS for 2012. The data are based on a household survey. Finally, the Census Bureau’s Economic Census of Manufacturers has been done every 5 years, with the most recent in 2012.
She explained that the ERS food expenditure series8 measures the total value of all food and beverages purchased by consumers in the United States. It is supposed to measure the expenditures both for consumption at home and away from home. The data represent the entire population, including the institutionalized population, as well as taxes, tips, and so on. The data are available in total, with no breakdown by commodity, although some detail is provided through tables. She suggested that the tables of greatest interest might be total expenditures for alcoholic beverages; food expenditures by source of funds (consumers, home production, government, business); per capita food expenditures; sales of food away from home by type of outlet; and sales of meals and snacks away from home by type of outlet.
Next, she explained that the proprietary IRI data include weekly retail scanner data9 that come from selected retailers across the United States (not all food retailers), including grocery stores, supercenters, and convenience stores. Data are available by store, Uniform Product Code10 (UPC), sales quantities, and cost. Not all variables are available for all retailers. She noted the IRI retail data have advantages and disadvantages. The data include all sales, not just to consumers. Some types of outlets are not included, nor are all U.S. retailers included. Some chains are missing, and there are no private-label data for some retailers. She said one of the advantages of the 2012 data is that they include random weight information, for example, for fruits and vegetables sold in bulk.
A second source from IRI is the consumer network,11 similar to the Nielsen Homescan approach. IRI has a sample of more than 60,000 U.S.
________________
8For more information, see http://www.ers.usda.gov/data-products/food-expenditures.aspx [July 2014].
9For more information, see http://www.iriworldwide.com/SolutionsandServices/Detail.aspx?ProductID=181 [July 2014].
10The UPC provides information on nutrition facts from the label and random weight. For UPC-coded perishable products, it includes weight/volume. With the UPC information, ERS can break out information about products to get down to the commodities. For prepared food and some canned food, more work would be needed to use the information.
11See http://www.iriworldwide.com/SolutionsandServices/Detail.aspx?ProductID=180 [July 2014.]
TABLE 4-1 Summary of Pros and Cons of Selected Data Sources
Data Source | Pros | Cons |
ERS Food Expenditure Series (1929-2011+) |
• Nationally representative • FAH and FAFH • Entire food system: retail, household, gov, business, etc. |
• No quantities • No commodity detail |
IRI InfoScan: Retail (2008-2012+) |
• Includes quantity • Detailed food groups |
• Not nationally representative • FAH only; retail only • Some limits to PL, RW |
IRI Consumer Network: HH (2008-2012+) |
• Nationally representative • Includes quantity • Detailed food groups |
• FAH only • No quantities for RW • Consumers only |
USDA (ERS) FoodAPS (2012) |
• Nationally representative • FAH and FAFH • Includes quantity • Detailed food groups |
• Consumers only • Coverage: 50 PSUs |
Economic Census 1902-2012+ |
• Nationally representative • FAH and FAFH • Some food groups, FAH |
• No quantities • No commodity detail, FAFH • Not annual (every 5 years) |
NOTE: FAH = food at home, FAFH = food away from home, IRI = Information Resources Incorporated, PL = private label, RW = random weight, PSU = primary sampling units.
SOURCE: Prepared by A. Kumcu for presentation at the workshop.
households that scan all their food purchases. Projection factors allow aggregation to the U.S. level. Data are available daily, by UPC, price, and quantity. The advantage is that the data can be weighted up to be representative of food consumption at home for the United States.
The third data source described by Kumcu is USDA’s FoodAPS National Household Food Acquisition and Purchase Survey.12 It was a nationally representative sample survey conducted in 2012, with over-sampling of Supplemental Nutrition Assistance Program (SNAP) and low-income households, and a sample size of almost 5,000 households. The data were collected using a seven-day recall by day for all food con-
________________
12See http://www.ers.usda.gov/data-products/foodaps-national-household-foodacquisition-and-purchase-survey.aspx [July 2014.]
TABLE 4-2 Summary of Food Group Availability in Selected Data Sources
Food Groups | ERS Food Expenditures | IRS InfoScan | IRI Consumer Network | ERS FoodAPS | Economic Census |
Meat, poultry, and fish | No | Yes | No RW | Yes | FAH, no FAFH |
Dairy products | No | Yes | Yes | Yes | FAH, no FAFH |
Eggs | No | Yes | Yes | Yes | No |
Fats and oils | No | Yes | Yes | Yes | No |
Fruits | No | Yes | No RW | Yes | No |
Citrus |
No | Yes | No RW | Yes | No |
Noncitrus |
No | Yes | No RW | Yes | No |
Legumes, nuts, and soy | No | Yes | No RW | Yes | No |
Vegetables | No | Yes | No RW | Yes | No |
White potatoes |
No | Yes | No RW | Yes | No |
Dark green, deep yellow |
No | Yes | No RW | Yes | No |
Other vegetables |
No | Yes | No RW | Yes | No |
Grain products | No | Yes | Yes | Yes | Some FAH |
Sugars and sweeteners | No | Yes | Yes | Yes | Some FAH |
Miscellaneous | No | Yes | Yes | Yes | Some FAH |
NOTE: FAFH = food away from home, FAH = food at home, IRI = Information Resources Incorporated, RW = random weight.
SOURCE: Prepared by A. Kumcu for presentation at the workshop.
sumed at home (UPC/food product detail) and away from home for all members of the household.
One advantage of FoodAPS, she said, is that it includes food away from home, as well as food that households produced themselves in gardens. It includes UPC-level information for purchased food so it could be
used to break out commodity groups; it also has receipts for food away from home.
The fourth source of data Kumcu described was the Census Bureau’s Economic Census,13 collected every 5 years since about 1902. It provides information about sales of food by category at supermarkets and warehouse clubs and supercenters. The Economic Census provides information about the composition of food used at home, though it might also include sales that were not for direct consumption. Sales categories include groceries and other food for consumption off premises;14 meals, snacks, and nonalcoholic beverages prepared for immediate consumption; and meals, snacks, and nonalcoholic beverages prepared for catered events.
The Economic Census also provides information about sales by full-service restaurants. While there is no information by commodity, it includes meals and snacks for immediate consumption off the premises; groceries and other food for consumption off the premises; meals and snacks served at a table by a server; meals and snacks dispensed without table service for consumption on the premises; and meals and snacks dispensed through drive-in service.
Kumcu summarized highlights of the pros and cons of the four data sources and their possible relevance for FADS (see Table 4-1). She noted it is unlikely any of these sources would provide comprehensive information, but they may be useful. She also provided Table 4-2 as a summary of whether each source has information by food group.
STATEMENT OF ALANNA MOSHFEGH DISAGGREGATION OF FOOD MIXTURES IN NUTRITION DATA
Moshfegh described special databases developed to go from individual consumption of processed food back to the food supply in terms of commodities. In particular, this requires use of a disaggregation process for foods and beverages.
First, Moshfegh described the dietary intake portion of the National Health and Nutrition Examination Survey (NHANES), which is the basis of the two special purpose databases described later in her presentation. NHANES is a nationally representative sample of about 5,000 individuals of all ages, conducted annually beginning in 1999. Interviews are conducted every day of the week, with two 24-hour dietary recalls. The first
________________
13See http://www.census.gov/econ/census [July 2014].
14Meat, fish, seafood, and poultry, including prepackaged meats; produce, including fresh and packaged fruits and vegetables; frozen foods; dairy products; bakery products baked on premises; bakery products not baked on the premises; delicatessen items; soft drinks; candy.
is administered by an in-person interview, the second by telephone 3 to 10 days later. The data from 2011-2012 were released in 2014.
Moshfegh said the USDA’s automated multiple-pass method is used for collecting the data. The questionnaire goes through a series of steps to elicit information about the foods and beverages that Americans eat. She said many items are mixture foods or multi-ingredient items, and that is how the data are collected. For example, NHANES would find out a respondent ate a slice of pepperoni pizza, then whether it was thick or thin, whether it contained extra cheese—with a probe for the size of the slice. She noted because the majority of the foods people eat are mixture items, ARS needs to convert them back to nutrient and commodity levels.
ARS developed the Food and Nutrient Database for Dietary Studies (FNDDS) to identify the nutrient profile for 8,000 foods/beverages. The USDA National Nutrient Database for Standard Reference is the basis of the 65 nutrient/component values. This database is defined to support specialized research and policy needs and has been funded by other federal agencies.
Moshfegh said that, in the FNDDS, about one-third of the items are simple, single-ingredient foods, such as types of milks by fat level and flavorings, apples, and so on. However, two-thirds of the items in the database are mixture foods. Examples of commodities and some of the multi-ingredient foods that contain them include apples in pies and apple jacks, tuna filet in tuna salad and tuna casserole, and eggs in cake and cookies. ARS releases the FNDDS every 2 years in concert with the 2-year releases of NHANES data in What We Eat in America (WWEIA). ARS is continuing to expand the number of food items in the database to reflect differences in foods that are being consumed.
She then discussed two special purpose databases developed primarily in support of other agencies. The Food Intakes Converted to Retail Commodities Database (FICRCD) is designed to translate the foods and beverages in FNDDS into 65 food commodities at the retail level, as defined by ERS. Two versions of the database have been prepared, one based on the 1994-1998 NHANES, the other based on the 1999-2008 NHANES. FICRCD translates all of the items in FNDDS into its 65 food commodities, and ARS worked in collaboration with ERS to develop the list.
Moshfegh showed Box 4-2, which lists food categories on the far left and the 65 commodities within those categories. ARS used foods as reported and converted them to the retail amount of a commodity. This provides the capability to look at consumption of commodities that appear in multi-ingredient foods by individuals. She said, for example, one category is apples from fruit. Another is apples from juice. ARS disaggregated all of the different sources of foods that have apples in them.
BOX 4-2
Food Intakes Converted to Retail Commodities Database (FICRCD) Commodities by Category, 2003-2008
FICRCD Commodities by Category | |||
Category (no. of commodities) |
Commodities | ||
Dairy Products (10) | Total dairy products Total fluid milk Fluid whole milk Fluid 2% milk Fluid 1% milk |
Fluid skim milk Butter Cheese Yogurt Other dairy products |
|
Fats and Oils (5) | Total fats and oils Margarine Salad & cooking oils |
Shortening (includes industrial shortenings) Other oils |
|
Fruits (14) | Total fruit Total apples Apples from fruit Apples from juice Bananas Berries Grapes |
Melons Total oranges Oranges from fruit Oranges from juice Other citrus fruits Stone fruits Tropical fruits |
|
Grains (5) | Total grains Corn flour & meal Oats & oat flour |
Rice (dry or uncooked) Wheat flour |
|
Meat, Poultry, Fish & Eggs (10) | Total meat, poultry, & fish Total meat Beef Pork Total poultry |
Chicken Turkey Finfish & shellfish Eggs, shell included Eggs, without shell |
|
Nuts (tree nuts & peanuts) (3) | Total nuts Peanuts |
Tree nuts | |
Sweeteners, caloric (1) | Total caloric sweeteners | ||
Vegetables, Dry Beans & Legumes (17) |
Total vegetables
Total brassica (cruciferous) vegetables Broccoli & cauliflowerCarrots Celery Cucumbers Green peas Total leafy vegetables Lettuce (head & leaf) |
Onions Peppers (bell & non-bell) Tomatoes Sweet corn Total roots & tubers Potatoes Snap beans (string beans) Dry beans and peas (legumes) |
SOURCE: Available: http://www.ars.usda.gov/SP2UserFiles/Place/12355000/pdf/ficrcd/FICRCD_2003_08_factsheet.pdf [September 2014].
Moshfegh described the process of preparing the disaggregation, starting from the food category and a determination of whether the category needed disaggregation. ARS has compiled recipes for foods in each category. Recipes may be within FNDDS, from Internet sites, product labels, or the profile of the ingredients from similar foods and beverages. The recipes are used to determine ingredients (commodities) as well as their amounts and proportions. She said conversion factors are applied, if necessary. Food/beverage ingredients are characterized by predefined characteristics/criteria.
She gave the disaggregation for a tuna noodle casserole as an example. Their recipe calls for light tuna fish canned in oil (drained), egg noodles (cooked), fluid milk, margarine (80 percent fat), and white flour. Further disaggregation would be for the tuna into tuna fish, soybean oil, and salt, and the egg noodles into egg noodles (dry), and further into the amount of whole eggs and raw wheat flour.
She noted some foods and beverages may appear simple, but still need to be disaggregated. As examples, canned pineapple in pineapple juice needs to be disaggregated into the amount of pineapple and amount of juice, while buttered popcorn needs to be disaggregated into popcorn and butter.
She described the second special purpose database, the Food Patterns Equivalents Database (FPED),15 which is available for 2005-2006, 2007-2008, and 2009-2010. It is developed using the same translation process for the foods and beverages in FNDDS, but it uses a different criteria. FPED translates foods and beverages into the 32 USDA food pattern groups that have been defined by USDA’s Center for Nutrition Policy and Promotion (CNPP) based on the 2010 Dietary Guidelines for Americans. She acknowledged the efforts of Susan Krebs-Smith, whom Moshfegh said served as their champion in developing the FPED. Moshfegh noted some of the 37 food pattern components of FPED (see Box 4-3) are similar to those in FICRCD, but in FPED, none was taken back to the retail level. She pointed out that the last items on the list are oils, solid fats, added sugar, and alcoholic drinks.
Moshfegh described another new product based on WWEIA that groups the 8,000 food and beverages in FNDDS into 150 unique categories. These are called the WWEIA food categories. It is to be used for analyzing consumed foods and beverages. There is no disaggregation, but, for example, one can look at milk and how it is consumed separately by fat content and flavorings.
She closed by noting all the ARS databases, including FICRCD and
________________
15The Food Patterns Equivalents Ingredient Database (FPID) is provided by ARS as part of the FPED release.
BOX 4-3
Food Patterns Components in the Food Patterns Equivalents Ingredient Database (FPID) and the Food Patterns Equivalents Database (FPED)
37 Food Patterns Components in FPID and FPED
Main Components | FPID/FPED Components |
Fruit |
1 Total fruit |
2 Citrus, melons, and berries |
|
3 Other fruits |
|
4 Fruit juice |
|
Vegetables |
5 Total vegetables |
6 Dark green vegetables |
|
7 Total red and orange vegetables |
|
8 Tomatoes |
|
9 Other red and orange vegetables (excludes tomatoes) |
|
10 Total starchy vegetables |
|
11 Potatoes (white potatoes) |
|
12 Other starchy vegetables (excludes white potatoes) |
|
13 Other vegetables |
|
14 Beans and peas computed as vegetables |
|
Grains |
15 Total grains |
16 Whole grains |
|
17 Refined grains |
|
Protein Foods |
18 Total protein foods |
19 Total meat, poultry, and seafood |
|
20 Meat (beef, veal, pork, lamb, game) |
|
21 Cured meat (frankfurters, sausage, corned beef, and luncheon meat made from beef, pork, poultry) |
|
22 Organ meat (from beef, veal, pork, lamb, game, poultry) |
|
23 Poultry (chicken, turkey, other fowl) |
|
24 Seafood high in n-3 fatty acids |
|
25 Seafood low in n-3 fatty acids |
|
26 Eggs |
|
27 Soybean products (excludes calcium-fortified soy milk and immature soybeans) |
|
28 Nuts and seeds |
|
29 Beans and peas computed as protein foods |
|
|
|
Dairy |
30 Total dairy (milk, yogurt, cheese, whey) |
31 Milk (includes calcium-fortified soy milk) |
|
32 Yogurt |
|
33 Cheese |
|
Oils |
34 Oils |
Solid Fats |
35 Solid fats |
Added Sugars |
36 Added sugars |
Alcoholic Drinks |
37 Alcoholic drinks |
SOURCE: Available: http://www.ars.usda.gov/SP2UserFiles/Place/12355000/pdf/fped/FPED_2009_10_Fact_Sheet.pdf [September 2014].
FPED, as well as documentation for them, are available on the ARS website.16
OPEN DISCUSSION
Laurian Unnevehr asked Grünberger about Figure 4-3. She said the graph shows the relative difference between the food balance sheet and the household survey data gets larger as GDP increases, in accordance with expectations. She asked whether his analysis would be feasible in a country like the United States and whether a similar analysis might be done comparing NHANES versus the food balance sheet over multiple years. She asked whether the analysis might provide some insights into U.S. loss and waste estimates. Grünberger agreed it would be an interesting strategy for estimating food waste at the household level.
Unnevehr further questioned whether the analysis would hold since the gap is likely to be larger for the United States. Grünberger replied the challenge associated with the gap is not related to the analysis of shares. He said the reason he moved away from analyzing levels to analyzing shares is that shares are robust to differences in levels, and the data in shares were more correlated than the data in levels. While an analyst would need to consider these issues, he said, the suggestion to use the U.S. data to estimate waste is excellent. The Statistics Division of FAO is primarily focusing on losses before food reaches households, he said, because food balance sheets do not consider consumer waste, and the division is currently producing updated estimates of preconsumption food losses. Once the latter are updated, he said, the division may want to focus on waste at consumption by using nutrition surveys.
Suzanne Thornsbury (ERS) asked Grünberger to comment on the observation that while differences are getting larger for higher-income countries, they are also becoming more variable. Grünberger pointed to Figure 4-3, observing variability is also high for low-income countries. This contradicts the intuitive assumption that suggests variability is larger in richer countries where the production chain is more complex. He observed that FAO did not use the data in levels because the data were so noisy. He said one step forward might be to categorize surveys to see if that would help in explaining differences, then decide whether differences are due to loss or lack of information on foods consumed away from home.
Krebs-Smith noted that Moshfegh’s presentation about the FICRCD database is particularly relevant because it illustrates a way to draw link-
________________
16See http://www.ars.usda.gov/ba/bhnrc/fsrg [July 2014] for links to all databases and documentation.
ages between the survey data and the balance sheet data, or using one data source to cross-check the other. She suggested the disaggregation approach Moshfegh described for the FICRCD might be useful in dealing with imports and exports of composite foods in the FADS data.
Kai Robertson (World Resources Institute) asked Schmidhuber about the waste column in Figure 4-1. Schmidhuber responded that the column covers waste only at the primary level. It might be useful, he said, to identify losses at the primary level, the retail level, and the consumer level. He explained that the conversion factor in Figure 4-1 is just the primary extraction rate and has nothing to do with losses. Robertson asked whether waste is a function of the conversion factors. Schmidhuber responded in the negative, saying that the 0.7 is a technical conversion factor: one kilogram of millet typically has 70 percent flour, and the rest is bran and germ.
Morvarid Bagherzadeh asked how optimistic others are that the discrepancies between the food balances and the household surveys can be explained. She asked whether it might be possible to generalize to a country from modeled comparisons of survey and balance sheet results in other countries, suggesting maybe there are categories of countries for which this is possible. She observed that household surveys are resource-intensive and are difficult for many developing countries. Grünberger responded that the model is country-specific at the moment, and he has not thought about the proposed generalization. Countries are used in the analysis only if they have both balance sheet and survey information, and only low- and middle-income countries are included. He noted it would be interesting to extend this research to developed countries, to see how the relationship with GDP changes.
Schmidhuber stated one of the purposes of the analysis was to have a comparison between household surveys and food balance sheets and to use it to calibrate the food balance sheets. The calibration would be used in updated balance sheets until such time as additional household survey data become available. He said FAO is looking at the comparison commodity by commodity and item by item. In some cases, country-specific information is important; in other cases, some generalizations are possible. As an example of a country-specific result, he said, in Cambodia, 8 kilos of sugar are available in the food balance sheet and 4 kilos in the household surveys. In this case, the household surveys would not be used for calibration because the food balance sheets appear to be more accurate. Cambodia imports nearly all of its sugar. The household survey reports sugar consumed as sugar, but does not include sugar consumed in the form of processed products. In other cases, such as Mexico for maize (corn), the household surveys could be used to recalibrate the food bal-
ance sheets because there is reason to believe that the household surveys are the more accurate information.
FACILITATED DISCUSSION WITH PANELISTS AND AUDIENCE
Muth next moderated the discussion, asking the panel members and audience to consider this session in light of the question posed earlier by ERS Administrator Mary Bohman (see the open discussion in Chapter 2) in which she asked participants to identify high-priority improvements for ERS data. Muth asked panelists to consider how ERS might use the data and approaches they presented to improve the food availability (FA) data. (She noted that the loss-adjusted food availability [LAFA] data would be discussed in the next workshop session.) She also asked panelists and the audience to suggest their personal one or two highest priority research ideas for ERS to consider for the FA data.
Sarah Nusser noted that the FAO analysis calibrating the food balance sheet and the survey data together was interesting, and asked about any promise for that approach within ERS. Jay Variyam (ERS) questioned whether the different purposes of the FAO and ERS balance sheets have an impact on methods that might be used. He stated that FADS has been used for meta-analysis and other purposes, while the FAO data are used to understand food safety and availability in many countries. He stated that, to the extent the FAO data are based on models, he thinks their use is limited, while the U.S. FA data come from balance sheets and represent production estimates and, hence, can be used in forward modeling. He expressed a need for caution in adopting the use of models.
Nusser observed that she views a balance sheet as a model about how the system works, and it is perfectly closed. She asked whether Variyam is concerned about imputed values in the balance sheet and how that might impact further analysis. Variyam said he disagreed that a balance sheet is a model. Mark Jekanowski stated he also thinks a balance sheet is not a model, but acknowledged this may be a semantic difference. He said he views a balance sheet more as an accounting framework, one that accounts for all of the different types of supply and use. To him, this differs from a modeling approach to estimating food availability directly from other sources, which he acknowledged may be a subtle difference. Nusser agreed that a balance sheet is an integrative framework with a set of constraints, which is why she thinks of it as a model. It is expected that production, imports, and beginning stocks add to supply, and, in theory, they do. However, estimates are needed in all the cells, some are more accurate than others, and key items are calculated as residuals. She said she views balance sheets as incorporating model assumptions.
Mark Denbaly (ERS) asked whether the underlying data for balance
sheets are survey-based. Jekanowski replied that the underlying data come from NASS and Census Bureau surveys and are estimates. Denbaly went on to say when estimates are added up, the result is a variable that has some probability distribution. Jekanowski again agreed, saying that in reality, the estimates might not add up perfectly. He added that there is potential for the data to be measuring things that are slightly different than assumed. However, they are the best data available and are put in a common framework with assumptions.
Muth reminded the audience about the interest in being able to break out food at home and food away from home in the LAFA data. She wondered whether the other data sources discussed might be used to come up with a disaggregation at the commodity level that could be used to estimate food at home and food away from home. Moshfegh replied that the FICRCD is linked to the data from NHANES. For every food and beverage reported in NHANES, a lot of information is available. Data are available for the time the individual started eating, whether the food was consumed at home or away from home, and the source of the food. This level of detail might be useful to estimate amounts of food consumed at home versus amounts of food consumed away from home.
Muth asked how the 65 commodities in FICRCD map to the 200 or so commodities in FADS. Moshfegh replied some are mapped very cleanly, some less so. Denbaly added that when the project with ARS started, ERS asked for the 200 commodities that are in FADS. Unfortunately, data were limited, but he is not sure why only 65 categories were chosen. Moshfegh explained that some of the commodities are not frequently consumed, according to the survey data. The commodity-level detail in FADS is food available for the whole country. NHANES surveys 5,000 people in a year in a country with a population of about 315 million. For that reason, ARS did not think that the data were robust enough to drill down to all of the smaller commodities.
Kumcu noted that other efforts might fill in some data gaps. For example, different groups have been working with ERS and the Food and Nutrition Service (FNS) of USDA to map UPC products to the FNDDS groups or the MyPyramid equivalents.
Muth asked whether the FoodAPS data will provide estimates for food acquisition at home versus away from home, saying those estimates might support an estimate in the FADS system. Denbaly replied that FoodAPS does have the designation of at home versus away from home. He observed, however, that measuring quantities of food away from home is difficult, although they are trying to do so.
Krebs-Smith asked about mapping of UPC codes to some commodities or the food pattern equivalents. Denbaly stated that CNPP, ERS, and ARS will meet to launch a project to link up the IRI data to FNDDS, which
would result in obtaining more information on price and quantity measures for food availability. He went on to say that theoretically speaking, it could also be done for FoodAPS, but this is not currently planned. Krebs-Smith asked about plans to link a commercially available database to FNDDS. She noted that a branded foods database is planned to be available soon. Denbaly expressed hope that this will also be part of the effort.
Schmidhuber stated that FAO is interested in continuing work comparing household surveys with food balance sheets. FAO is going through the comparison commodity by commodity to understand the strengths and weaknesses of the two data sources. He said the question is when it is better to use household surveys versus food balance sheets, noting the answer can be country-specific. More generally, he said, it is important to do comparisons. He asked whether the various U.S. sources support a total calorie availability comparison, calling it a useful benchmark. He said he would like to see a comparison of the U.S. balance sheets with household surveys.
He noted that most of the surveys available to FAO are household income and expenditure surveys representing food available at the retail level. Food consumption surveys would permit a much better evaluation of waste, he said, as the difference between food purchased and food consumed would be waste or loss. He noted that an international effort to sponsor more food consumption surveys is something FAO could support.
FAO has been using the National Nutrient Database for Standard Reference, he said. However, one of its limitations for FAO use is its inclusion of fortified food. When FAO used it on aggregated commodity balances at the process level, they noticed that micronutrient deficiencies in developing countries went away, which was an artifact of applying the nutrient content of fortified foods. He suggested a similar database, but without fortification, because it would serve as a default database for nutrient factors where they do not have country-specific nutrient tables.
Muth related a question that came in through the workshop webcast from Lisa Johnson (North Carolina State University) about whether supply and availability data would change drastically by including the portion that is unharvested, especially for the fruit and vegetable commodities. Unharvested amounts are estimated to be enormous, Johnson said; since growers report what leaves the farm to NASS, they could perhaps also report what is left. Jekanowski replied that for many commodities, NASS reports both planted and harvested acreage, and abandonment estimates are available for these commodities. However, no reason is given for that abandonment, and it is not clear whether it is appropriate to consider abandonment as a loss. Thornsbury said planted and harvested acres and a production number are typically reported on NASS surveys
for fruits and vegetables. The production quantities are used in the food balance sheets. NASS has national surveys that provide the production estimates, which are calibrated to the Census of Agriculture every five years. NASS does not collect data about abandonment in the field.
Muth noted one reason for abandonment might be a crop failure. She said the difference between planted and harvested acres could give an upper-bound estimate of farm loss. Thornsbury agreed, but also suggested other reasons for abandonment. For fruits and vegetables, it could also be economic abandonment. If too much product is available, prices are driven down, and there may not be enough value in the crop to pay for the harvest. Jekanowski noted one other complication: if not harvested, the yield is unknown. Assumptions would be needed to come up with loss estimates.
Helen Jensen said some abandonment has to do with food safety. Farmers may find a part of their fields has contamination, so they rope it off and do not harvest it. She said she was not sure whether that would be considered loss. Thornsbury elaborated that now, if a crop is not harvested, it is not captured in the supply and use table. There are many reasons why food might not be harvested, but none is captured in the supply and use tables.
Muth asked the audience about their ideas concerning high-priority projects for improving the FA data.
Krebs-Smith suggested the notion of abandonment and the reasons for abandonment might be good to capture. She said data that illuminate any unknowns along the way would be helpful for various types of analyses, especially at a time of concern about sustainability and feeding a hungry world, where food waste is a big issue.
Moshfegh suggested another type of data that may be of interest for international uses is food loss in transportation across countries. Inside a country, there are ways to capture transportation losses. However, for international trade, there seems to be no information at all, although perhaps insurance companies have data. Schmidhuber responded that trade statistics databases may provide the information because they include the exports of reporting countries and the imports of partner countries. Combining these quantities and values may provide an estimate of loss. He noted, however, that there is a lot of noise in the data, and the trade data are more reliable for developed than developing countries.