Current Diets

Overview

The model represents current consumption patterns by combining three intake datasets — GDD-IA, GBD, and NHANES (for the USA) — with item-level food supply data from FAOSTAT for within-group disaggregation. The pipeline produces a single per-country, per-food baseline diet whose mass basis is aligned with what the model’s food bus delivers after applying food loss and waste. The baseline diet serves several roles:

• Health impact assessment: dietary risk exposure for the burden of disease attributable to current diets.

• Optimization reference: comparison point for optimized diets and, optionally, an equality constraint when enforce_baseline_diet is enabled.

• Calibration: anchors the consumer-utility piecewise blocks (Consumer Values) and the production-stability L1 calibration (Calibration).

Population-weighted mean food group consumption (g/person/day) by UN M49 macro-region, showing how dietary patterns vary across world regions.

Global population-weighted mean consumption (g/person/day) broken down by individual foods within each food group.

Data Sources

Global Dietary Database — Integrated Assessment (GDD-IA)

• Provider: Marco Springmann (University of Oxford / UCL). GDD-IA combines the Global Dietary Database (GDD) survey-based intake estimates with FAOSTAT Food Balance Sheets and applies a multi-source caloric-intake normalisation procedure to produce consistent per-country food and energy intake estimates.

• Status: Pending publication; available upon personal request from Marco Springmann. Will be re-licensed under CC-BY-NC on release.

• Coverage: ~185 countries, per-country mean dietary intake at the reference year, reported in parallel grams/day and kcal/day for every food category.

• Role: Primary source of per-country food-group totals for all food groups except the GBD-anchored risk groups (see below).

Global Burden of Disease (GBD) 2019 dietary risk exposure

• Provider: Institute for Health Metrics and Evaluation (IHME) [Brauer2024]

• Coverage: country-level mean intake (g/day) for the GBD dietary risk factors, adults 25+.

• Role: anchors the risk-factor food groups (fruits, vegetables, whole_grains, legumes, nuts_seeds, red_meat) so the baseline lines up with the same exposure basis the GBD relative-risk functions were calibrated against.

NHANES — What We Eat in America / FPED

• Provider: USDA ARS / CDC NHANES

• Coverage: United States; population-mean intake per food group derived from the FPED demographic table.

• Role: USA-only override for every food group it covers.

FAOSTAT FBS + QCL

• Provider: FAO Statistics Division

• Role: Item-level supply (FBS) drives within-group disaggregation of food-group totals into per-food consumption. Production statistics (QCL) resolve shared FBS items (e.g. several millet species under one FBS code) and weight module-pool projections (see Step 2: Within-group food shares). FBS supply also serves as the anchor source for the foods in diet.fbs_override_foods (meats, eggs, yam, coffee, cocoa).

Weight Conventions

GDD-IA reports intake “as consumed” (cooked weight for cereals and meats, fresh weight for fruits and vegetables). The pipeline derives the food-group mass values that downstream rules consume in the model’s basis, so no further conversion is needed when reading dietary_intake.csv:

• For most groups (cereals, vegetables, fruits, nuts/seeds, oil, sugar, legumes, poultry, eggs) the IA-reported grams are already close enough to the model basis to be passed through as-is.

• Red meat is inflated from cooked to raw retail mass by the configured diet.gdd_ia.cooked_to_raw factor (default 1.43, i.e. 1/0.7).

• Dairy mass is derived from energy at cow-milk density (0.607 kcal/g) so the value is on a strict cow-milk-equivalent basis. All dairy subcategories reported by GDD-IA (fluid milk, yoghurt, cheese, condensed/evaporated, ice cream, butter, cream) are pooled by energy before the conversion.

GBD exposure is converted to the model basis at load time via diet.source_basis and diet.weight_conversion (cooked→dry for whole_grains and legumes at 0.45 and 0.40; cooked→fresh for red_meat at 1.43). NHANES values are intake-based and pass through unchanged.

Units in the merged dietary_intake.csv distinguish g/day (fresh wt) from g/day (milk equiv) for dairy and g/day (refined sugar eq) for sugar.

GDD-IA to Food Group Mapping

GDD-IA’s food categories are mapped onto the model’s food groups in workflow/scripts/prepare_gdd_ia_dietary_intake.py. The mapping covers every food group the model uses (fruits, vegetables, starchy_vegetable, legumes, nuts_seeds, oil, sugar, grain, whole_grains, red_meat, poultry, dairy, eggs, plus stimulants for downstream tea/coffee handling). Categories that are out of scope for the model (alcohol, seafood, spices, rendered animal fats, miscellaneous “other”) are excluded from food-group totals but their energy is tracked separately for the kcal-normalisation step described below. Refined and whole-grain mass are tracked separately so cereals can be split between the model’s grain and whole_grains groups; plantain is routed to starchy_vegetable; and all red-meat subcategories (including processed) are folded into red_meat so the consumption side stays consistent with FAOSTAT slaughter-volume animal production.

A more detailed category-level mapping will be added once GDD-IA is published.

Country Coverage

GDD-IA covers ~185 countries. For a handful of territories without separate IA estimates the pipeline copies values from a configured proxy. The built-in proxies live in workflow/scripts/prepare_gdd_ia_dietary_intake.py and can be extended via diet.gdd_ia.country_proxies in the config:

Missing country	Proxy	Rationale
Afghanistan (AFG)	Iran (IRN)	Persian/Pashtun dietary similarity.
American Samoa (ASM)	Samoa (WSM)	Pacific islands; geographic proximity.
Brunei (BRN)	Malaysia (MYS)	Regional similarity.
Bhutan (BTN)	Nepal (NPL)	Himalayan diet.
Eritrea (ERI)	Ethiopia (ETH)	Existing convention.
Equatorial Guinea (GNQ)	Cameroon (CMR)	Central African neighbour.
French Guiana (GUF)	France (FRA)	French overseas territory.
Palestine (PSE)	Jordan (JOR)	Regional similarity.
Puerto Rico (PRI)	United States (USA)	US territory.
Somalia (SOM)	Ethiopia (ETH)	Existing convention.
South Sudan (SSD)	Sudan (SDN)	Regional and historical ties.
Taiwan (TWN)	China (CHN)	Regional similarity.

Data Processing

The diet pipeline runs in three preparation stages followed by the baseline-diet estimation:

1. Prepare GDD-IA (prepare_gdd_ia_dietary_intake): reads the parallel grams and kcal CSVs, maps GDD-IA’s food categories to the model’s food groups, derives the per-food-group mass in model basis (pooling all dairy subcategories by energy, applying the cooked-to- raw meat inflation), and emits two files:

   • gdd_ia_dietary_intake.csv — per-(country, food group) intake (g/day) at age = All ages.

   • gdd_ia_kcal_target.csv — per-country kcal accounting: the total dietary energy, the out-of-scope subtotal, the in-scope target (total minus out-of-scope), and the refined / whole-grain cereal energy split. Consumed by the cereal residual fix and the kcal-normalisation step in estimate_baseline_diet.

2. Prepare NHANES (prepare_nhanes_dietary_intake): parses the USDA FPED demographic-table PDF for the configured cycle and emits USA-only per-food-group intake with the FAOSTAT butter top-up, cured-meat fold, and fruit-juice projection (see Data Sources for the FPED specifics).

3. Merge sources (merge_dietary_sources): NHANES overrides GDD-IA for the (country, item) pairs it covers; the merged file dietary_intake.csv is the input to estimate_baseline_diet.

The GBD risk-exposure data is processed independently by prepare_gbd_food_group_intake into gbd_food_group_intake.csv and is read directly by estimate_baseline_diet for the GBD-anchored groups.

Output Format

dietary_intake.csv:

unit,item,country,age,year,value
g/day (milk equiv),dairy,USA,All ages,2018,...
g/day (fresh wt),fruits,USA,All ages,2018,...
...

• unit: g/day (fresh wt), g/day (milk equiv) (dairy), or g/day (refined sugar eq) (sugar).

• item: food group name.

• country: ISO 3166-1 alpha-3 code.

• age: All ages for GDD-IA rows; NHANES uses the configured diet.baseline_age literal (the FPED single population-mean row).

• year: reference year.

• value: mean daily intake in grams per person, in model basis.

Baseline Diet Estimation

The dietary intake stage produces food-group-level totals. The optimization model operates at the level of individual foods, so workflow/scripts/estimate_baseline_diet.py disaggregates the totals into per-(country, food) consumption estimates and applies a small number of consistency fixes:

Step 1: Food group totals

For groups in health.risk_factors (currently fruits, vegetables, whole_grains, legumes, nuts_seeds, red_meat) the per-country total is taken from GBD when GBD reports a value and falls back to the merged GDD-IA/NHANES value otherwise. GBD strictly takes precedence on these groups — no averaging — so the baseline is on the same intake basis the GBD relative-risk functions are calibrated against. All other groups use GDD-IA (or NHANES for the USA).

GBD exposure is converted to the model’s basis at load time, per food-group, using diet.source_basis plus per-(source, country, food_group) overrides from data/curated/diet_source_basis_overrides.csv and the conversion tables in diet.weight_conversion. The script also logs cross-validation metrics: median and range of the GDD-IA/GBD ratio across countries for every risk group, and GBD’s milk exposure as a cross-check on the dairy total.

Step 1b: Cereal residual fix

GBD’s whole_grains risk factor is defined narrowly (dry whole-grain flour). GDD-IA’s whole_grains is broader (any product with substantial whole-grain content). When Step 1 anchors whole_grains to GBD, ~250 kcal/day of cereal energy can disappear from the country’s cereal budget. To preserve the cereal energy budget, the deficit is reassigned to refined ``grain``:

\[ \begin{align}\begin{aligned}\text{deficit\_kcal} = (\text{kcal}_{\text{whole\_grains}}^{\text{IA}} + \text{kcal}_{\text{grain}}^{\text{IA}}) - g_{\text{whole, anchored}} \cdot k_{\text{whole, model}}\\\text{new}\ g_{\text{grain}} = \max(0, \text{deficit\_kcal}) / k_{\text{grain, model}}\end{aligned}\end{align} \]

The IA cereal kcal pool comes from gdd_ia_kcal_target.csv (basis- aware), not from nutrition.csv per-group averages.

Step 1c: Anchor-aware kcal normalisation

For each country, the unanchored groups are scaled by a single multiplicative factor so that total kcal across all groups lands on the in-scope dietary-energy target from gdd_ia_kcal_target.csv (total energy minus the out-of-scope subtotal). GBD-anchored groups and the refined-grain residual from Step 1b are held fixed. The factor is clipped to [0.1, 5.0] to guard against pathological values; the mean, std, and range of the factor across countries are logged.

Step 2: Within-group food shares

Once food-group totals are set, the algorithm determines how to distribute each total across its constituent foods using FAOSTAT FBS item-level supply. The shares pipeline (in build_within_group_shares) covers four resolution patterns:

Direct (1:1) mapping. A food that is the unique claimant of its FBS item receives 100 % of that item’s supply.

Shared FBS item, QCL-resolved. When several foods share an FBS item (e.g. cowpea, chickpea, gram, phaseolus-bean and pigeon-pea all map to FBS 2546 “Beans”), country-level FAOSTAT QCL production data splits the shared supply between QCL buckets, and within a bucket the default is an equal split. Two cases use explicit within-bucket overrides:

• pearl-millet / foxtail-millet (both QCL “Millet”): a fixed 80 / 20 global split based on literature production shares.

• dairy / dairy-buffalo (both FBS 2848): QCL items 882 and 951 resolve the cow / buffalo split. The split is then post-processed by cap_buffalo_share_at_production to cap each country’s buffalo share at its domestic buffalo production (buffalo milk has very limited international trade), with any excess share reassigned to cow dairy. Without the cap, GBD-anchored dairy intake exceeds domestic milk production in buffalo-heavy importers (PAK is the textbook case) and the production-share split over-allocates buffalo demand, surfacing as unrelievable buffalo shortage at solve.

Module-pool projection. For food groups whose modelled foods share a GAEZ RES06 supply-side module, the demand-side attribution pools all module-aligned FBS codes (both explicit and “Other”-style residuals) and splits the pool across the modelled foods. Pooling matches the supply-side attribution by construction: each modelled food’s supply comes from FAOSTAT direct area plus a share of the module’s residual raster area, so routing the explicit FBS supply (e.g. onion FBS 2602) to one food on the demand side while supply spreads it across the module would produce systematic within-group slack.

Each pool sub-projection carries a share_method that decides how to allocate the pool across its modelled foods:

• "blend" — country/global production-share blend \(s_{c,f} = w \cdot s_{c,f}^{\mathrm{country}} + (1-w) \cdot s_f^{\mathrm{global}}\) over FAOSTAT crop production (currently \(w=0.7\) for all pools using this method).

• "frt_attribution" — per-(country, crop) shares read directly from the supply-side frt_area_attribution.csv (target_production_tonnes column), so the demand-side within-pool split mirrors the supply-side FRT attribution exactly. Used for the fruits FRT pool, where the supply side intentionally uses area-share (not production-share) weighting to avoid over-attributing residual area to high-yield fruits; the blend method on the demand side would drift from that choice.

Food group	Pooled FBS items	Projection foods	Share method
vegetables	2602 “Onions”, 2605 “Vegetables, Other”	onion, cabbage, carrot	blend
starchy_vegetable	2534 “Roots, Other”	potato, sweet-potato, yam, cassava	blend
nuts_seeds	2551 “Nuts and products”	groundnut, sesame-seed, coconut, sunflower-seed	blend
fruits (BAN sub-projection)	2616 “Plantains”	banana only	blend
fruits (FRT sub-projection)	2611–2614 (citrus), 2617 “Apples”, 2618 “Pineapples”, 2619 “Dates”, 2625 “Fruits, other”	citrus, mango, watermelon, apple	frt_attribution

For fruits the projection is split into two sub-projections so that the demand-side attribution mirrors the GAEZ RES06 module split on the supply side: banana and plantain share the GAEZ BAN raster (plantain supply is therefore projected onto banana exclusively), while citrus / mango / watermelon / apple jointly absorb the FRT raster plus CROPGRIDS-backed apple.

Tomato (FBS 2601) and individually-itemised starchy vegetables and nuts retain their explicit FBS supply in addition to any pool share they receive: each has its own GAEZ raster on the supply side and its own FBS item, so the explicit-route is already symmetric.

Equal split fallback. Where total FBS supply for a food group is zero in a country, foods within the group are assigned equal shares.

Note

The within-group share computation weights each food’s FBS supply by its edible portion (FAO edible_portion_coefficient looked up via data/curated/foods.csv -> fao_edible_portion.csv) before normalising. The GDD-IA group totals are on an edible-mass basis while FBS supply is reported on a fresh-whole-commodity basis, so splitting an edible-mass total by fresh-mass weights would over-allocate intake to low-edible-portion foods (plantain 0.59, watermelon 0.62, citrus 0.72). The weighting is applied in both the direct-supply branch and the pooled-projection branch of build_within_group_shares; in pooled projections the edible portion follows the recipient food (e.g. plantain FBS supply in the BAN sub-projection is redistributed to banana using banana’s edible portion).

Step 3: Per-food consumption

Per-food consumption is the product of the food-group total (post Steps 1b and 1c) and the within-group share:

\[c_{i,f} = T_{i,g(f)} \cdot s_{i,f}\]

As a validation check, the within-group sums are verified to match the group totals to within 0.1 g/day (excluding foods that will be replaced by FBS overrides in Step 4).

Step 4: FBS supply overrides

For foods listed in diet.fbs_override_foods the Step-3 estimate is replaced with an FBS-supply-anchored intake. The override formula is

\[c_{i,f} = \frac{S_{i,f} \cdot \sigma_{i,f} \cdot r_f \cdot 1000}{365} \cdot (1 - w_{i,g(f)})\]

where

• \(S_{i,f}\) is the FAOSTAT FBS supply (kg/capita/year) for the food’s FBS items (carcass-weight for meat);

• \(\sigma_{i,f}\) is the within-FBS-item share (1.0 unless several override foods share an FBS code, in which case the supply is split between them by country-level QCL production weights — e.g. dairy / dairy-buffalo both map to FBS 2848);

• \(r_f\) is the carcass-to-retail factor for meat (0.67 cattle, 0.73 pig, 0.66 sheep, 0.60 chicken; 1.0 for non-meat foods);

• \(w_{i,g(f)}\) is the country- and group-level consumer-waste fraction from processing/{name}/food_loss_waste.csv.

Note that the override deducts only consumer waste, not supply-chain loss: the FAOSTAT FBS “Food supply” element is already net of production-side losses (production − feed − seed − processing − other − losses = food). The \((1-w)\) factor lands the override on the same post-FLW intake basis the model’s food_processing and animal_production links deliver after applying their FLW multipliers, so the diet mass-balances against the food bus.

Why yam needs an override

GDD-IA / GBD starchy-vegetable intake for sub-Saharan Africa is well below FAOSTAT food supply (e.g. Nigeria: GBD ≈ 70 g/day vs. FBS ≈ 700 g/day for starchy vegetables). Because yam production is almost entirely concentrated in West Africa, the within-group underestimate translates directly into a ~10× underestimate of yam demand. The within-group shares are correct — the problem is in the group total — so overriding yam consumption with FBS supply ensures the model’s demand matches observed food availability.

Why animal products use FBS, not survey intake

For meats, poultry, and eggs the per-food intake is anchored to FAOSTAT FBS supply rather than the survey-disaggregated group total. Three reasons:

1. Survey bias on socially significant foods. Self-reported food intake systematically over-reports red meat against slaughter- volume supply in many populations. GDD-IA harmonises survey data but does not reconcile against production. The combined intake total for red meat sat ~24 Mt/yr above what total world supply (production net of feed/non-food/exports, after post-loss and consumer waste) can deliver — physically impossible — and previously inflated the calibrated animal_feed_l1_cost ninefold because the production-stability calibration was forced to fight intake-derived consumer values that were structurally above supply.

2. Trade is handled implicitly. FBS supply per country already encodes production + imports − exports − feed − seed − non-food − stock_changes, so country-level diet automatically reflects observed importer/exporter patterns. The model’s trade hubs then only have to reproduce the observed FAOSTAT trade flows at solve time, instead of resolving a mismatch via expensive feed-deviation L1 penalties.

3. Same FAOSTAT backbone as production. Baseline animal production is built from QCL element 5510 with the shared weight_conversion.carcass_to_fresh table applied. FBS aggregates the same QCL primary commodities at carcass-weight balance level. Anchoring both sides to FAOSTAT removes a class of unit/source mismatches that otherwise surfaces as residual slack after solve.

Dairy is intentionally excluded from the override list. Its food_loss_waste convention is non-standard — the curated dairy override sets loss_fraction=0 and waste_fraction=0.30, where the 30 % lumps in non-food uses of raw milk (calf feed, processing, industrial) plus retail and consumer waste, because the model does not have an explicit non-food milk outlet. Under that convention the GDD-IA-based dairy total happens to mass-balance against the production-side QCL × 0.7 delivered to the food bus. Switching dairy to an FBS override would break that balance.

Output

processing/{name}/baseline_diet.csv has one row per (country, food):

Column	Description
country	ISO 3166-1 alpha-3 country code.
food	Model food name (e.g. banana, rice-white, cowpea).
food_group	Food group to which the food belongs.
consumption_g_per_day_intake	Estimated daily consumption in grams per person, on post-loss, post-waste consumer-eaten intake basis — the same basis the food bus delivers after the build_model FLW multiplier (see Weight bases for animal products).

Rows are sorted by (country, food_group, food).

Downstream Uses

• Baseline diet enforcement: when config.validation.enforce_baseline_diet is true, the solver adds per-food, per-country equality constraints on food consumption links.

• Within-group ratio fixing: when config.food_groups.fix_within_group_ratios is true, foods within each group are constrained to keep their baseline proportions while group totals may vary.

• Piecewise consumer utility calibration: baseline per-food consumption and baseline food-equality duals together calibrate results/{name}/consumer_values/utility_blocks.csv (Consumer Values).

• Health impact assessment: baseline consumption feeds the population-attributable fraction calculation (Health Impacts).

Workflow Integration

Snakemake rules (see workflow/rules/diet.smk):

• prepare_gdd_ia_dietary_intake

• prepare_nhanes_dietary_intake

• merge_dietary_sources

• prepare_gbd_food_group_intake

• prepare_faostat_fbs_items

• prepare_food_loss_waste

• estimate_baseline_diet

• validate_baseline_diet and compare_baseline_diet_to_gbd (consistency checks)

Input data:

• data/manually_downloaded/GDD-IA-intake_grams_{baseline_year}.csv

• data/manually_downloaded/GDD-IA-intake_kcals_{baseline_year}.csv

• data/manually_downloaded/IHME_GBD_2019_DIET_RISK_1990_2019_DATA/*.csv

• data/downloads/usda_fped/Table_1_FPED_MaleFemale_{cycle}.pdf

• FAOSTAT FBS and QCL (auto-fetched via the FAOSTAT bulk API)

Curated data files:

File	Purpose
data/curated/faostat_food_item_map.csv	Maps model foods to FAOSTAT FBS item codes for within-group share calculation.
data/curated/faostat_food_qcl_resolution.csv	Maps foods sharing an FBS item to QCL production codes for disambiguation.
data/curated/food_groups.csv	Food → food-group mapping.
data/curated/food_basis.csv	Per-food native mass basis (dry / fresh / cooked / milk-equiv).
data/curated/diet_source_basis_overrides.csv	Per-(source, country, food_group) basis overrides for the cross-source conversion.
data/curated/nhanes_fped_mapping.csv	FPED column → model food-group mapping and unit-conversion factors.
data/curated/food_loss_waste_overrides.csv	Per-(country, food_group) loss/waste overrides feeding food_loss_waste.csv.

Configuration parameters:

• config.countries — list of countries.

• config.food_groups.included — food groups to process.

• config.baseline_year — reference year for GDD-IA and GBD.

• config.diet.baseline_age — age label written to NHANES rows (default "All ages").

• config.diet.fbs_override_foods — foods anchored to FBS supply. See Why animal products use FBS.

• config.diet.source_basis and config.diet.weight_conversion — per-source native bases and conversion tables.

• config.diet.gdd_ia.cooked_to_raw — per-group cooked→raw inflation factors for GDD-IA (currently red_meat: 1.43).

• config.diet.gdd_ia.country_proxies — extra proxies beyond the defaults in prepare_gdd_ia_dietary_intake.py.

• config.diet.nhanes.cycle and .reference_year — FPED release.

• config.health.risk_factors — drives which food groups are anchored to GBD in Step 1.

• config.byproducts — foods excluded from share calculation.

Output:

• processing/{name}/gdd_ia_dietary_intake.csv — GDD-IA group-level intake.

• processing/{name}/gdd_ia_kcal_target.csv — per-country kcal accounting (total dietary energy, out-of-scope subtotal, in-scope target, refined / whole-grain cereal energy split).

• processing/{name}/nhanes_dietary_intake.csv — USA NHANES override.

• processing/{name}/dietary_intake.csv — merged GDD-IA + NHANES.

• processing/{name}/gbd_food_group_intake.csv — GBD exposure.

• processing/{name}/baseline_diet.csv — per-food, per-country baseline diet.

Scripts:

• workflow/scripts/prepare_gdd_ia_dietary_intake.py

• workflow/scripts/prepare_nhanes_dietary_intake.py

• workflow/scripts/merge_dietary_sources.py

• workflow/scripts/prepare_gbd_food_group_intake.py

• workflow/scripts/estimate_baseline_diet.py

• workflow/scripts/diet/food_group_projection.py — within-group pooled-projection helpers (FBS-code pools, production-share blends).