IFPRI Discussion Paper 02361

September 2025 

The Future of Food Demand 

A Global Meta-Analysis and Projections of Income and Price Elasticities 

Maxime Roche 

Andrew Comstock 

Olivier Ecker 

Foresight and Policy Modeling Unit 



INTERNATIONAL FOOD POLICY RESEARCH INSTITUTE 

The International Food Policy Research Institute (IFPRI), a CGIAR Research Center established in 1975, 

provides research-based policy solutions to sustainably reduce poverty and end hunger and malnutrition. 

IFPRI’s strategic research aims to foster a climate-resilient and sustainable food supply; promote healthy 

diets and nutrition for all; build inclusive and efficient markets, trade systems, and food industries; 

transform agricultural and rural economies; and strengthen institutions and governance. Gender is 

integrated in all the Institute’s work. Partnerships, communications, capacity strengthening, and data and 

knowledge management are essential components to translate IFPRI’s research from action to impact. 

The Institute’s regional and country programs play a critical role in responding to demand for food policy 

research and in delivering holistic support for country-led development. IFPRI collaborates with partners 

around the world.  

AUTHORS 

Maxime Roche (m.roche21@imperial.ac.uk) is a PhD Candidate in the Centre for Health Economics & 

Policy Innovation, Department of Economics & Public Policy, Imperial Business School, Imperial 

College London, London, United Kingdom.  

Andrew Comstock (a.comstock@cgiar.org) is a Senior Research Analyst in the Foresight and Policy 

Modeling (FPM) Unit, International Food Policy Research Institute (IFPRI), Washington, DC. 

Olivier Ecker (o.ecker@cgiar.org) is a Senior Research Fellow in IFPRI's FPM Unit, International Food 

Policy Research Institute (IFPRI), Washington, DC.   

Notices 

1 IFPRI Discussion Papers contain preliminary material and research results and are circulated in order to stimulate discussion and 

critical comment. They have not been subject to a formal external review via IFPRI’s Publications Review Committee. Any opinions 

stated herein are those of the author(s) and are not necessarily representative of or endorsed by IFPRI.  

2 The boundaries and names shown and the designations used on the map(s) herein do not imply official endorsement or 

acceptance by the International Food Policy Research Institute (IFPRI) or its partners and contributors. 

3 Copyright remains with the authors. The authors are free to proceed, without further IFPRI permission, to publish this paper, or 

any revised version of it, in outlets such as journals, books, and other publications. 

mailto:m.roche21@imperial.ac.uk
mailto:a.comstock@cgiar.org
mailto:o.ecker@cgiar.org


iii 

ABSTRACT 

Understanding how food demand responds to household income and price changes is essential for 

anticipating global food needs and designing effective food policies. Yet existing elasticity estimates vary 

widely due to differences in data, estimation methods, and study settings. This study aims to assess how 

empirical choices influence elasticity values and examine theory-based predictions related to income 

growth and inequality, urbanization, and demographic change. It provides the most comprehensive global 

systematic literature review and meta-analysis of income and price elasticities to date, compiling over 

13,000 elasticity estimates from 215 peer-reviewed studies published between 1974 and 2022. We estimate 

two-level random effects meta-regressions and use the results to generate predicted elasticities for nine food 

groups by world region. While most data and methodological choices have little effect on price elasticity 

estimates, income elasticities are influenced by factors such as demand model type, use of conditional 

specifications, and the choice of expenditure measure. We find empirical support for Engel’s Law but only 

partial support for Bennett’s Law. Income elasticities are positively associated with urbanization, 

particularly in lower-income countries, and negatively associated with population aging. By projecting 

income elasticities through 2050 under alternative Shared Socioeconomic Pathways, we show that ignoring 

structural shifts in sociodemographics can yield meaningfully different estimates of future food demand. 

Keywords:  Food demand, income elasticity, price elasticity, systematic review, meta-analysis, projection 



iv 

ACKNOWLEDGMENTS 

This work was undertaken as part of the CGIAR Science Program on Policy Innovations and the CGIAR 

Research Initiatives on National Policies and Strategies and Foresight. We would like to thank all funders 

who supported this research through their contributions to the CGIAR Trust Fund: www.cgiar.org/funders. 

We thank Shaonan Wang for his research assistance in conducting the systematic literature review and 

compiling the food demand elasticity dataset. 

http://www.cgiar.org/funders


1 Introduction

The theoretical foundations of consumer demand trace back to David Ricardo’s early 19th-century work

on taxation and income distribution (Ricardo, 1821), which highlighted how purchasing power influences

consumption behavior. This was a precursor to modern concepts of income elasticity of demand, i.e., the

proportional response of quantity demanded to a change in real income. Later, Marshall (1890)’s Principles of

Economics formalized the relationship between prices, marginal utility, and the quantity of goods demanded.

Central to this framework was the notion of price elasticity of demand, i.e., the proportional response of

quantity demanded to a change in price, which remains a cornerstone of both neoclassical demand theory

and empirical welfare analysis.

Estimated elasticities of demand play a critical role in applied economics. First, they reflect revealed

consumer preferences and provide critical information about how people’s consumption is expected to

change when household income or the prices of food and non-food goods and services change, as a result

of economic growth or policy interventions. Economic theory has long offered structured predictions on

the food consumption effects of income growth and demographic shifts. The first formulated relationship

between food consumption and household income is Engel’s Law, which posits that, as income increases, the

proportion spent on food decreases (Engel, 1857). Formulated much later, Bennett’s Law describes a typical

pattern of food consumption changes in the context of agricultural and economic development. It postulates

that, with rising income, household consumption shifts away from calorie-dense starchy staple foods toward

more diverse and nutrient-rich foods (Bennett, 1941). Second, estimated elasticities of demand are essential

inputs in a wide range of economic models for a variety of analytical purposes, including models for business

strategy development, market and trade analysis, and partial general equilibrium models, to assess the

impacts of economic shocks and policies on development outcomes. One of the most immediate purposes

of such models, especially in a developing context, is projecting future food demand in populations. While

it is widely accepted that demand will continue to rise over the coming decades, especially in low- and

middle-income countries (LMIC), the rate at which this growth will occur, and the associated patterns of

typical consumption shifts between distinct food groups, are open questions.

Despite a large empirical literature estimating income and price elasticities of food demand across

countries, time periods, and food groups, reported values vary widely. This heterogeneity reflects differences

in data sources, demand models, estimation strategies, and contextual settings. However, the extent to

1



which data and methodological choices systematically shape elasticity estimates remains poorly understood.

This paper asks: 1) Do empirical choices, including data characteristics and estimation methods, explain

the observed variation in food demand elasticities? 2) Are theoretical predictions regarding the role of

sociodemographic factors empirically supported? 3) How is food demand expected to vary with projected

sociodemographic changes?

To answer these questions, we first conduct a global systematic literature review of empirical food demand

estimation studies. We compile over 13,000 income and price elasticity estimates extracted from 216 peer-

reviewed studies covering five continents, along with detailed information on their data characteristics and

estimation methods. The meta-sample data is merged with country-level sociodemographic indicators.

Second, we perform a meta-regression analysis using a two-level random intercept model to quantify how

methodological choices, data characteristics, and sociodemographic factors contribute to the variation in

reported elasticities. Population-weighted average predicted income and price elasticities are estimated

for nine food groups by world region as of 2021. Third, we use these estimated relationships to project

global food demand through 2050, allowing elasticities to evolve in line with scenario-based projections of

sociodemographic characteristics.

Our meta-analysis synthesizes demand elasticities from the literature and offers a framework for inter-

preting the wide variation in published estimates. It quantitatively assesses how specific empirical choices

and evolving sociodemographic contexts shape these demand parameters that are crucial for policy analysis

and market modeling. The strength of this approach is its ability to generate robust, generalizable meta-

estimates. The trade-off is a loss of context-specific detail. Our predicted elasticities should therefore be seen

as a standardized reference value, to be refined with local data when possible. Lastly, the projection exercise

demonstrates how structural sociodemographic shifts can alter consumption trajectories and illustrate the

limitations of assuming static demand elasticities in long-term modeling.

The meta-regression results indicate that some estimation methods help explain empirical differences in

elasticity estimates. Income elasticities estimated using conditional demand1 tend to yield higher values,

1Conditional elasticities measure the demand response for a specific food product or group under the assumption that certain
factors, such as total expenditure on food or the consumption of related products, remain fixed. These elasticities capture substitution
effects within a specific category of food but do not account for shifts in consumption across broader food or non-food categories.
In contrast, unconditional elasticities measure demand responses without such restrictions, reflecting adjustments across all food
and non-food items. In our meta-sample and meta-regression analysis, both conditional and unconditional elasticities are included.
However, we incorporate a dummy variable to distinguish between them, as they represent different scopes of demand response and
should not be treated as equivalent measures.

2



while those estimated using household income rather than total expenditure to measure income elasticities, or

using a quadratic almost ideal demand system (QUAIDS), tend to yield lower values. As expected, Hicksian

(compensated) price elasticities are larger in magnitude than Marshallian (uncompensated) ones.2

In line with Engel’s Law, we find that income elasticities decline as gross domestic product (GDP)

per capita rises. This effect is more pronounced in higher urbanized settings. However, evidence for

Bennett’s is limited. While income elasticities for food away from home (FAFH) rise with real income

—consistent with their discretionary nature —the decline in income elasticities for grains & starchy staples

is not significantly steeper than that for total food, and income elasticities for animal-sourced foods (ASF) do

not rise significantly with real income. Income inequality shows little explanatory power in shaping elasticity

estimates. Urbanization is associated with higher income elasticities across most food groups, particularly

in LMIC contexts, likely reflecting increased market access and dietary diversification. This pattern reverses

for FAFH, suggesting early adoption and saturation effects in urban settings. Population aging is linked to

lower income elasticities, consistent with reduced caloric needs and more stable dietary habits among older

individuals.

Our population-weighted average predicted income and price elasticities are broadly consistent with

established findings in the literature, highlighting significant regional and food-group variation. Estimated

income elasticities are the highest in Africa and for more discretionary food groups, such as ASF (0.98)

and FAFH (1.29). They are the lowest for grains & starchy staples (0.59), which form the bulk of diets

in many LMIC. Demand for necessities with fewer close substitutes, such as grains & starchy staples

(−0.58) and oils & fats (−0.51), is less price elastic, whereas FAFH demand is the most price-responsive

(−1.00). Disaggregation reveals greater price responsiveness at finer food group levels, likely reflecting

higher substitution potential.

A key finding is that income elasticities are not static. They evolve with sociodemographic characteristics

like income growth, urbanization, and population aging. Our projections illustrate this dynamism. For

instance, under the ”Middle of the road” scenario of the Shared Socioeconomic Pathways, reflecting a

2Uncompensated (Marshallian) elasticities measure the effect of a price change for one good while keeping household income
and the prices of other goods constant, thereby reflecting both income and substitution effects. In contrast, compensated (Hicksian)
elasticities measure the effect of a price change for one good while holding consumer utility constant, assuming that any price change
is offset by household income adjustments to maintain the same utility level, thus excluding income effects. Both are included in
our meta-sample and meta-regression analysis. However, we include a dummy to distinguish between the two types as they capture
different demand responses to price changes and cannot be treated as equivalent.

3



continuation of historical trends as well as moderate progress towards sustainable development goals, we

project that food demand in Africa will remain highly responsive to household income growth. In contrast,

demand responsiveness to income for most non-discretionary foods in other regions is projected to decline

as they approach consumption saturation and face population aging.

Using these evolving elasticities leads to significantly different conclusions about future global food

demand compared to assuming today’s elasticities remain constant. The most striking impact is on ASF,

where our demand projection for 2050 is 10.7% lower than what a static assumption would suggest. While

we abstract from price changes and do not intend to provide a definitive forecast, our projection results

highlight the critical importance of incorporating dynamic sociodemographic shifts into long-term food

system modeling.

Many elasticity values used in global food models today can be traced back to estimates developed

before 1990, such as those embedded in the U.S. Department of Agriculture (USDA)’s Static World Policy

Simulation model (Roningen et al., 1991). Integrated global databases of demand elasticities are rare and

often outdated. For example, the USDA Commodity and Food Elasticities database,3 which covers elasticity

estimates from 1979 to 2005, is no longer maintained. Similarly, the International Evidence on Food

Consumption Patterns report, which provides estimated income and price elasticities across more than 100

countries using International Comparison Program data, was last updated in 2011 (Muhammad et al., 2011).

Recently, the International Food Policy Research Institute (IFPRI) has started to close this gap by periodically

estimating income and price elasticities for a wide range of developing countries using harmonized food

demand system estimations and representative household survey data as they become available (Income

and Price Elasticities of Food Demand (E-FooD) dataset) (Ecker and Comstock, 2021). Our contribution

complements these efforts. We compile the most comprehensive and up-to-date global dataset of income and

price elasticity estimates published in the peer-reviewed literature (1974-2022). Our scope extends beyond

LMIC, enabling meta-analytic identification of patterns in elasticity variation across settings, data, methods,

and food types.

Given the heterogeneity in data sources and estimation methods, meta-analytic synthesis can offer more

robust and policy-relevant elasticity estimates for use in general and partial equilibrium models. To our

knowledge, we present the first predicted income elasticities by world region for nine disaggregated food

3U.S. Department of Agriculture, Economic Research Service. Commodity and Food Elasticities (Accessed 22 January 2025).

4

https://www.ers.usda.gov/data-products/commodity-and-food-elasticities/


groups, defined in alignment with global dietary guidelines.4 Our study also updates Green et al. (2013)’s

global meta-estimates for price elasticities.5

Previous studies have investigated the influence of the choice of data and methods on elasticity estimates

using meta-regression analysis, including Gallet (2010) and Bouyssou et al. (2024) for ASF, Chen et al.

(2016) and Colen et al. (2018) for all types of food in China and Africa, respectively, and globally for

price elasticities only by Cornelsen et al. (2016). Our analysis covers a longer and more recent period and

incorporates more studies and estimates. This allows us to include additional variables characterizing the

data and estimation methods, and interact disaggregated food groups with socio-demographic variables in our

meta-regressions. This enhances our understanding of demand heterogeneity and improves the specificity

of predicted elasticities across food groups.

Our study contributes to the empirical literature testing theory-based demand predictions. Support for

Engel’s Law corroborates earlier empirical evidence (Banks et al., 1997; Cranfield et al., 1998). The weaker

patterns in support for Bennett’s Law align with recent findings from Bellemare et al. (2024), who leverage

data from randomized evaluations of five conditional cash transfer programs in Mexico, Nicaragua, the

Philippines, and Uganda. They show that with additional income, poor households tend to shift from coarse

to fine staples and coarse staples to protein, but find no evidence of substitution from staples to protein overall,

nor fine staples to protein. A limitation of our analysis is that we cannot assess substitution patterns within

narrowly defined food groups (e.g., between coarse and fine cereals, or by quality tiers within a category).

This reflects a common constraint in meta-analyses, where food group definitions vary across primary studies

and results must be harmonized to a common set of aggregated categories. The relationship between income

distribution and food demand has received limited empirical attention. Cirera and Masset (2010) review

the theoretical literature and argue that within-country income inequality is unlikely to substantially affect

aggregate food demand. Our findings that Gini coefficients have no statistically significant effect on demand

elasticities provide empirical support for their conclusion.

Finally, we contribute to the literature projecting global food demand by emphasizing the importance of

4Other meta-analyses of income elasticities have either covered only a specific region, e.g., Africa (Colen et al., 2018), or specific
food groups, e.g., animal-sourced foods (Bouyssou et al., 2024), or focused on income-calorie and income-nutrient elasticities (Bouis
and Haddad, 1992; Zhou and Yu, 2014). In addition to animal-sourced foods, Bouyssou et al. (2024) provides predicted elasticities
for four food groups (fats, grains, fruits vegetables pulses and tubers, and other foods). These groups are overly aggregated, limiting
their use as parameters in general or partial equilibrium models.

5To our knowledge, Green et al. (2013) represents the latest and only available study to date.

5



accounting for the impact of shifts in sociodemographic structure on future demand elasticities (Fukase and

Martin, 2020; Gouel and Guimbard, 2019). Our projected elasticities, along with our meta-sample dataset

and meta-elasticity estimates, provide a globally harmonized, empirically grounded resource to support

improved food demand projections and enhanced food system and policy modeling.

The remainder of this paper is organized as follows. Section 2 describes the systematic review protocol

and the meta-sample. Section 3 outlines the empirical strategy, including the meta-regression framework,

the estimation of predicted elasticities, and the projection of global food demand to 2050. Section 4 presents

the main results along with sensitivity tests. Section 5 discusses the findings and their implications and

concludes.

2 Systematic literature review and descriptive statistics

We built our meta-sample following the PRISMA guidelines (Page et al., 2021a,b). Candidate primary

studies were identified by conducting an exact keyword search for publication titles, abstracts, and keywords

using the terms “food” and “demand”, along with “elasticities” or “elasticity” through the ISI Web of

Science platform and Google Scholar. Search engines considered in other food demand meta-analyses, such

as AgEcon, EconLit, Medline, or RePEc (Green et al., 2013; Chen et al., 2016; Colen et al., 2018), led to

similar results. Only peer-reviewed journal articles were considered. Database searches were performed as

of 27 September 2022. The time frame of the studies was restricted by the earliest available year for the

Science Citation Index on the Web of Science platform and encompassed 1974 to 2022. After removing

duplicates, 557 candidate studies were identified.

Two members of the research team independently conducted a two-step screening process. We considered

a paper eligible if both authors agreed to include it. In the case of a disagreement, a third author conducted a

full-text review and made the final decision. The first step of the screening process consisted of excluding non-

English text studies.6 The second step screened the abstracts for the provision of own-calculated estimates

6Page et al. (2021c) suggest including non-English text studies to minimize the risk of bias. However, our search revealed less
than 1% of non-English text studies.

6



 

Records identified from 

database searching  

(n=557) 

Records screened 

(n=254) 

Full text articles assessed for 

eligibility 

(n=212) 

Study included in qualitative 

synthesis 

(n=215) 

Study included in quantitative 

synthesis (meta-analysis) – 

price elasticities 

(n=199) 

Records excluded 

(n=303) 

Full-text articles 

excluded  

(n=42) 

Targeted search  

(n=3) 

Study included in quantitative 

synthesis (meta-analysis) – 

expenditure elasticities 

(n=186) 

Figure 1. PRISMA flow diagram for selection of primary studies.

of food demand elasticities, rather than predicted or projected. This process resulted in 254 included studies.

At this stage, a final screening step took place to remove 42 papers that we deemed ineligible for not reporting

the sample size or not containing enough information to infer it.7 Finally, we supplemented our meta-sample

with three studies that we had pre-identified from prior reviews and our expertise.8 The final sample covers

215 articles. Figure 1 illustrates this process and Appendix A lists the primary studies included in the

meta-sample.

For each study, we collected the names of the first four authors, title, year, and journal of publication.

Regarding the estimates of interest, we included the income elasticity of demand and the own-price elasticity

7Out of these 42 papers, Harding and Lovenheim (2017) was specifically not included because its sample size (34.2 million
observations, approximately 30 times larger than the second largest sample size in the meta-sample) would significantly drive the
meta-regression results given our approach using the square root of the sample size as sampling weight to address heteroskedasticity.
A sensitivity meta-regression analysis including this study yields consistent results and is available upon request.

8The three studies added at this stage are: Zhen et al. (2014), McCullough et al. (2022), and Headey et al. (2023).

7



of demand.9 There was significant heterogeneity in the aggregation of food types for these estimates.

Figure B1 illustrates our re-aggregation scheme. It consists of nine food groups (Level 1) as typically defined

by dietary guidelines (Herforth et al., 2019). These include: animal-sourced foods (ASF); beans, lentils, peas

& soy; condiments & sweeteners; fruits, vegetables, and nuts (FV&N); grains & starchy staples; processed

meals & snacks; oils & fats; food away from home (FAFH); and all beverages. Further disaggregation for

some groups is reported in our database (Level 2 and Level 3). These are not included in the meta-regression

analyses, the predicted elasticities, or the projections presented in this paper, as the main analysis focuses on

Level 1 food groups.

In the database, we gathered variables describing the data used to estimate elasticities. These include

the sample size, the structure of the data (i.e., cross-sectional, time series, or panel data), the level of

representativeness (e.g., national, subnational), and the country and the year (or multiple years) in which

the data were collected. We further computed the average year of data collection. Many studies report

elasticity estimates by subsample (e.g., rural vs. urban, over a given subsample period). Thus, we compiled

the subsample information and estimated the corresponding sample size (if missing) for each elasticity

observation.

We constructed several model variables to capture key features of each elasticity estimate. These include

the functional form (e.g., single-equation or demand system), the specific demand model used (e.g., AIDS,

QUAIDS), the number of estimation stages, the type of demand (conditional or unconditional), the income

measure (total expenditure or household income), and whether the price elasticity estimate is compensated or

uncompensated. We coded whether the specification controlled for demographics to account for preference

heterogeneity. In addition, we identified whether the estimation addressed common empirical challenges

such as censored demand and endogeneity in prices or total expenditure.

To enrich this dataset, we added external information on countries’ sociodemographic characteristics.

We retrieved data on per capita GDP, population by age, and the share of the population living in urban areas

from the World Bank.10 National-level Gini indices were drawn from the World Institute for Development

Economics Research to capture income inequality.11 These external variables were merged with each

9This analysis does not include cross-price elasticities. In this study, price elasticities only refer to own-price elasticities.
10World Bank. World Development Indicators, 2024 (Accessed 18 September 2024).
11United Nations University, World Institute for Development Economics Research. The World Income Inequality Database -

WIID, Companion dataset, 2025 (Accessed 20 May 2025).

8

https://databank.worldbank.org/source/world-development-indicators
https://www.wider.unu.edu/database/world-income-inequality-database-wiid##WIIDcomp
https://www.wider.unu.edu/database/world-income-inequality-database-wiid##WIIDcomp


observation for the average year of data collection linked with the elasticity estimate.12

The final meta-sample covers 57 countries and includes 6,572 income elasticity estimates and 6,701 price

elasticity estimates. Figure B2 shows the geographic distribution of studies included in the meta-sample.

The United States (46 studies) and China (26 studies) are the most represented countries. Central Africa,

North Africa, West Africa (except Nigeria), the Middle East, and Eastern Europe are underrepresented. In

the rest of the analysis, countries are grouped into four regions: Africa, the Americas, Asia-Oceania, and

Europe.13

Figure 2 presents the density distributions for the elasticity estimates by food group.14 For income

elasticities, the shaded areas indicate if the goods are estimated to be inferior (negative), normal (between

zero and one), or luxury goods (above one). The mean value of the distribution is between zero and one for

all food groups. Unsurprisingly, it is closest to one for the most discretionary or ‘luxurious’ food groups,

i.e., ASF (0.97) and FAFH (0.97). It is lowest for grains & starchy staples (0.59), which are considered

necessities in many settings. Regarding price elasticities, most estimates fall within the [−2, 0] interval, with

predominantly price-inelastic demand (i.e., estimates greater than −1). The mean value of the distribution

is below −1 only for FAFH and beverages (i.e., price-elastic demand), which represent mostly non-essential

goods. It is the lowest in absolute value for oils & fats (−0.62). It should be noted that the price elasticity

density functions include both uncompensated and compensated elasticities. The former reflects both income

and substitution effects, while the latter isolates the substitution effect by holding utility constant.

Table 1 summarizes the coverage of our meta-sample by food group and type of demand elasticity. The

ASF group represents 37.6% of income elasticity observations and 36.7% of price elasticity observations,

followed by FV&N with 19.0% and 17.8%, respectively, and grains & starchy staples with 16.9% and

17.2%, respectively. The least represented food groups are All foods with 1.4% and 1.1% and FAFH with

12The average year of data collection was calculated by truncating the average value between the first and the last year of data
collection in the original study. Two studies lacked information on the period of study data. For these, we assumed the average year
of data collection to be 5 years before the publication year of the study. For countries for which the GDP per capita series started
after 1960 (the earliest average year of data collection in the meta-sample), we imputed the value from the first non-missing year.
This was the case for 7.1% and 5.8% of income and price elasticity observations, respectively. We followed the same strategy to
impute missing values for the Gini index (27.0% and 32.2% of income and price elasticity observations, respectively). Sensitivity
meta-regression analyses using only non-imputed data as well as other sources for Gini indices, including the World Bank’s Poverty
and Inequality Platform and World Development Indicators (Accessed 16 April 2025), yield consistent results and are available
upon request.

13This grouping is selected due to the limited number of income elasticity observations in the Oceania continent (32 observations)
and the South American continent (68 observations) in the meta-sample.

14Table B2 additionally presents descriptive statistics for the elasticity estimates for Level 2 and Level 3 food groups.

9

https://pip.worldbank.org/home
https://pip.worldbank.org/home
https://databank.worldbank.org/source/world-development-indicators


0 1 2

ASF

0 1 2

Beans
lentils

peas & soy

0 1 2

Condiments
& sweeteners

0 1 2

F V & N

0 1 2

Grains &
starchy
staples

0 1 2

Processed
meals &
snacks

0 1 2

Oils & fats

0 1 2

FAFH

0 1 2

All
beverages

Income
elasticity

-2 -1 0 -2 -1 0 -2 -1 0 -2 -1 0 -2 -1 0 -2 -1 0 -2 -1 0 -2 -1 0 -2 -1 0

Price
elasticity

Figure 2. Density distributions of the elasticities by food group. Based on the full sample. The vertical
dotted line represents the mean value of the distribution. Price elasticity density functions include both
uncompensated and compensated elasticities. ASF: animal-sourced foods; FAFH: food away from home;
FV&N: fruits, vegetables, and nuts.

2.1% and 0.5%, respectively. Table 1 also provides a summary of the descriptive statistics for the elasticity

estimates. There are considerable differences in the magnitude of the income and price elasticities across

regions. The average income elasticity is the smallest in the Americas, while it is the largest in Europe.

The average price elasticity is the smallest in absolute terms in Asia-Oceania, while it is the largest in the

Americas. The Americas is the least represented region in the meta-sample.15 More than three out of four

elasticity observations are based on data collected this century. The large majority of estimates are based on

a subsample of nationally representative data and cross-sectional household-level data and use a quadratic

almost ideal demand system (QUAIDS, 48% for income elasticity and 44% for price elasticity) or an almost

ideal demand system (AIDS, 23% and 31%, respectively). The meta-sample is approximately evenly split

between conditional and unconditional demand estimates (49% and 59% of income and price elasticity

15South America and Oceania are the least represented subregions in the meta-sample.

10



estimates are conditional, respectively) and between censored and uncensored demand (40% and 46% of

income and price elasticity estimates are derived from censored demand, respectively). Most observations

are estimated using a single-stage model and do not account for expenditure or price endogeneity, but control

for demographics. In most cases, household income is proxied using total expenditure (74%) and price

elasticities are uncompensated (78%).

Table 2 additionally presents descriptive statistics for the continuous data variables. Sample size varies

greatly, ranging from only five to more than 1.1 million observations. Studies based on homescan or retail

scanner panel datasets tend to have significantly larger sample sizes than studies based on time series or

cross-sectional data. Regarding external variables, we observe a significant dispersion in GDP per capita,

ranging from $419 to $51,457,16 and in the Gini index, ranging from 20% to 67%. The mean of the share of

the population below 15 years old is 28% and 8% above 64 years old. On average, approximately half of the

population in the sample countries lives in an urban area with a population dependency ratio of 0.6.

16GDP per capita in constant 2005 purchasing power parity international dollars to match the GDP per capita data used in Shared
Socioeconomic Pathways projections (Dellink et al., 2017; Crespo Cuaresma, 2017; Leimbach et al., 2017).

11



Income elasticity Price elasticity

N Share Mean Sd N Share Mean Sd

Food group
All foods 89 0.014 0.739 0.296 72 0.011 -0.685 0.209
ASF 2,468 0.376 0.966 0.633 2,452 0.366 -0.899 0.752
Beans lentils peas & soy 166 0.025 0.824 0.686 171 0.026 -0.932 0.898
Condiments & sweeteners 263 0.040 0.615 0.759 293 0.044 -0.722 0.655
F V & N 1,247 0.190 0.835 0.491 1,191 0.178 -0.801 0.415
Grains & starchy staples 1,110 0.169 0.594 0.594 1,151 0.172 -0.804 0.740
Processed meals & snacks 204 0.031 0.608 0.775 309 0.046 -0.933 0.641
Oils & fats 420 0.064 0.743 0.437 463 0.069 -0.626 0.508
FAFH 137 0.021 0.966 0.836 33 0.005 -1.130 0.669
All beverages 468 0.071 0.798 0.488 566 0.084 -1.081 0.625
Region
Africa 1,643 0.250 0.850 0.568 1,680 0.251 -0.886 0.545
Americas 783 0.119 0.546 0.712 1,116 0.167 -0.893 0.855
Asia-Oceania 2,615 0.398 0.859 0.656 2,482 0.370 -0.813 0.729
Europe 1,531 0.233 0.863 0.473 1,423 0.212 -0.863 0.534
Sample type
All 1,904 0.290 0.732 0.688 2,412 0.360 -0.810 0.738
Subsample 4,668 0.710 0.856 0.574 4,289 0.640 -0.881 0.634
Structure of data
Cross-sec. & TS 4,787 0.728 0.854 0.583 4,692 0.700 -0.852 0.712
Panel data 1,785 0.272 0.730 0.676 2,009 0.300 -0.863 0.575
Data representativeness
National 5,297 0.806 0.833 0.606 5,805 0.866 -0.864 0.662
Subnational 1,275 0.194 0.766 0.634 896 0.134 -0.800 0.745
Data period
Pre-2000 1,516 0.231 0.706 0.715 1,534 0.229 -0.706 0.835
2001-2010 2,644 0.402 0.794 0.604 2,704 0.404 -0.930 0.628
2011-2020 2,412 0.367 0.921 0.531 2,463 0.368 -0.866 0.592
Type of conditionality
Conditional 3,234 0.492 0.970 0.575 3,978 0.594 -0.862 0.701
Unconditional 3,338 0.508 0.676 0.613 2,723 0.406 -0.845 0.633
Demand model
Single-equation 801 0.122 0.581 0.501 258 0.039 -0.793 0.860
AIDS type 1,478 0.225 0.883 0.580 2,106 0.314 -0.891 0.782
QUAIDS type 3,119 0.475 0.930 0.624 2,969 0.443 -0.837 0.585
Other 1,174 0.179 0.613 0.591 1,368 0.204 -0.852 0.633
Censored demand
No 3,922 0.597 0.783 0.627 3,590 0.536 -0.811 0.639
Yes 2,650 0.403 0.876 0.586 3,111 0.464 -0.906 0.710
Number of estimation stages
Single-stage 4,065 0.619 0.849 0.589 4,224 0.630 -0.855 0.705
Multi-stage 2,507 0.381 0.774 0.646 2,477 0.370 -0.855 0.619
Account for exp. endogeneity
No 5,013 0.763 0.790 0.575 5,276 0.787 -0.841 0.677
Yes 1,559 0.237 0.916 0.710 1,425 0.213 -0.909 0.663
Account for price endogeneity
No 5,857 0.891 0.812 0.631 5,818 0.868 -0.826 0.702
Yes 715 0.109 0.885 0.428 883 0.132 -1.049 0.398
Control demographics
No 1,452 0.221 0.499 0.629 1,093 0.163 -0.683 0.822
Yes 5,120 0.779 0.911 0.576 5,608 0.837 -0.889 0.636
Income elasticity measure
Income 1,737 0.264 0.708 0.581 - - - -
Total expenditure 4,835 0.736 0.861 0.618 - - - -

12



Income elasticity Price elasticity

N Share Mean Sd N Share Mean Sd

Uncompensated or compensated demand
Compensated - - - - 1,487 0.222 -0.771 0.647
Uncompensated - - - - 5,214 0.778 -0.879 0.680

Total 6,572 1.000 0.820 0.612 6,701 1.000 -0.855 0.674

Table 1. Descriptive statistics of data and model variables included in the meta-sample. Subnational
also includes non-representative data. TS: time series. N: number. AIDS: Almost Ideal Demand System;
QUAIDS: Quadratic Almost Ideal Demand System.

Income elasticity Price elasticity

min mean max sd min mean max sd

Sample size 5 16,720 1,152,528 84,437 10 31,595 1,152,528 108,022
GDP per capita (2005 PPP) 419 12,411 46,778 12,753 419 14,230 51,457 13,925
Urbanization (%) 0.150 0.495 1.000 0.217 0.131 0.529 1.000 0.231
Population >64 yo 0.015 0.083 0.211 0.051 0.015 0.084 0.211 0.052
Population <15 yo 0.135 0.283 0.508 0.109 0.135 0.289 0.508 0.111
Population dependency ratio 0.355 0.598 1.122 0.194 0.355 0.614 1.122 0.193
Gini index 0.199 0.382 0.670 0.053 0.220 0.385 0.670 0.057

Table 2. Descriptive statistics of the sociodemographic and continuous variables. PPP: purchasing
power parity international dollars. yo: years old.

13



3 Estimation and projection methods

3.1 Meta-regression

We estimate separate meta-regressions to analyze the relationship between household income and food

demand and between prices and food demand. The model can be described as follows:

𝜖𝑖,𝑠, 𝑗 = 𝛽0 +
∑︁
𝑚

𝛽𝑚𝐹𝑚,𝑖 +
∑︁
𝑛

𝛽𝑛𝐷𝑛,𝑠, 𝑗 +
∑︁
𝑝

𝛽𝑝𝑀𝑝,𝑖,𝑠, 𝑗 +
∑︁
𝑞

𝛽𝑞𝑆𝑞,𝑠, 𝑗

+
∑︁
𝑚

∑︁
𝑞

𝛽𝑚,𝑞𝐹𝑚,𝑖 × 𝑆𝑞,𝑠, 𝑗 + 𝛾
1√︁
𝑁𝑠, 𝑗

+ 𝜐 𝑗 + 𝜈𝑠 + 𝜀𝑖,𝑠, 𝑗 (1)

𝜖𝑖,𝑠, 𝑗 is the elasticity estimate for food item 𝑖 for sample 𝑠 in study 𝑗 . 𝐹𝑚,𝑖 , 𝐷𝑛,𝑠, 𝑗 , 𝑀𝑝,𝑖,𝑠, 𝑗 , and 𝑆𝑞,𝑠, 𝑗

are the 𝑚 food group dummies (Level 1) and the 𝑛, 𝑝, and 𝑞 variables that contain the data, model, and

sociodemographic attributes for estimate 𝜖𝑖,𝑠, 𝑗 . 𝛽𝑚, 𝛽𝑛, 𝛽𝑝, 𝛽𝑞, and 𝛽𝑚,𝑞 represent the coefficients of interest

and quantify the impact of each of these attributes. 𝜀𝑖,𝑠, 𝑗 is the error term.17

This two-level random intercept model is estimated via maximum likelihood. It accounts for random

effects at two hierarchical levels: the study 𝜐 𝑗 (the higher level) and the sample 𝜈𝑠 (the lower level). By doing

so, it allows for further random variation within samples (e.g., national, rural only, subsample period). This

model is preferred over pooled OLS as it controls for possible study-specific and sample-specific unobserved

heterogeneity (i.e., it accounts for between- and within-study variation). This specification is also favored

relative to fixed effects. A fixed effects estimator would control for within-study unobserved heterogeneity.

However, many covariates in our model only exhibit between-study variation, thus random effects are more

appropriate.

We perform Egger et al. (1997)’s test for publication bias involving regressing elasticity estimates on an

inverse indicator of their precision. Publication bias may arise given the likely preference for statistically

significant and theory-consistent empirical results among research outputs (Stanley and Doucouliagos, 2014).

The first-best indicator of precision is the standard errors of elasticity estimates. However, only around one-

third of papers reported standard errors, and half reported the significance levels of estimates. Solely

including studies reporting standard errors or significance levels may bias the meta-sample, as these studies

17It was not possible to include additional interaction terms (e.g., between food groups and regions) due to singularity issues. In
addition, interaction terms consume degrees of freedom (i.e., number of studies minus the number of parameters estimated), leaving
fewer to assess variability across studies in our two-level random intercept model.

14



could be more likely to report statistically significant estimates. As a consequence, we use the inverse square

root of the sample size as an indicator of study precision, relying on the correlation between sample sizes

and t-statistics.18 The Egger test shows the presence of publication bias in our meta-sample (Table B3).

As a result, we include the inverse square root of the sample size 1√
𝑁𝑠, 𝑗

as an additional regressor in our

meta-regression model (Stanley and Doucouliagos, 2014).

We allow for heteroskedasticity using robust clustered standard errors at the study level. This addresses

potential violations of the independence assumption, as estimates within the same study are likely correlated.

In addition, we apply the square root of the sample size (as a proxy for study standard errors) as a sampling

weight to address heteroskedasticity arising from the variance of the elasticity estimates differing across

studies and between samples (Card and Krueger, 1995). Table B4 shows the relevance of this approach as

larger sample sizes are a strong predictor of a higher significance level for elasticity estimates.

The variables included in the main meta-regression specification are summarized in Table B1. We

include a control variable indicating whether the elasticity estimate is derived from the full sample or a

subsample. This accounts for systematic differences in consumption behavior and elasticities that may arise

due to sample composition. For instance, elasticities estimated from more homogeneous subsamples may

differ in magnitude and direction compared to those based on nationally representative samples. Most

elasticity estimates in the meta-sample are based on nationally representative data (81% for income and

87% for price elasticities) (Table 1). Owing to this limited variation, we omit the national vs. subnational

distinction from the main specification. While the data period may capture global dietary preference shifts,

this variable is excluded due to multicollinearity with other covariates.19

Regarding the model variables, we emphasize key structural choices such as whether a single-equation

model or a demand system is used, the type of demand system, and the type of conditionality (conditional

vs. unconditional demand), as these reflect decisions from the econometrician. We also distinguish between

income elasticities estimated using total expenditure as a proxy for household income and between compen-

sated and uncompensated price elasticities, as they capture different behavioral responses to price changes.

Failing to distinguish between the latter would conflate conceptually distinct measures and introduce unnec-

18Table B4 shows that larger sample sizes are a strong predictor of a higher significance level for elasticity estimates in our
meta-sample.

19This is because certain model choices, access to new data sources (e.g., scanner data), or sociodemographic indicators (e.g.,
per capita GDP) may evolve over time.

15



essary heterogeneity into the meta-regressions. Other potentially secondary model characteristics —such

as the data structure (e.g., panel data, time series, cross-sectional data), whether the study accounts for

censored demand, the number of estimation stages, whether it corrects for endogeneity in total expenditure

or prices, and whether it includes demographic controls —are excluded from the main model specification

for parsimony and because of their limited variation across studies (Table 1). We show the robustness of our

meta-regression results to the inclusion of these secondary data and model variables in Table D1.

Per capita GDP is used to examine how real income affects food consumption. It allows us to empirically

test Engel’s Law, which states that as household income increases, the share of income spent on food declines

(Engel, 1857). It also enables us to test Bennett’s Law, which suggests that rising household income leads

to a shift away from coarse, calorie-dense staples toward finer staples, animal-sourced foods, and more

diverse diets (Bennett, 1941). Our main model specification includes the level of urbanization, as it may

influence food consumption patterns through differences in the type of labor and income, the opportunity

cost of time, or proximity to agricultural production and marketplaces (Regmi and Dyck, 2001). Information

on demographics can also be indicative of the caloric and nutritional needs or food preferences of different

age groups. Demographic changes represent a significant source of aggregation bias (Blundell and Stoker,

2005). In demand theory, the aggregation problem refers to the capacity to transition from micro- to

macro-economics and represent all households’ demand at one point in time. Given the significant global

population aging trend (Lutz et al., 2008), we specifically include the share of the population aged 65 or

above. Life-cycle theory predicts consumption smoothing and declining marginal propensities to consume

with age (Gourinchas and Parker, 2002). Older individuals are expected to have more stable and habitual

consumption patterns, which could induce flatter Engel curves. Finally, our main model specification

includes geographical regions. This is motivated by the significant heterogeneity in estimates observed

between regions.20

We further explore alternative specifications. First, we control for income inequality using the Gini

index, as the distribution of income may influence aggregate food demand patterns and elasticities when

Engel curves are non-linear (Cirera and Masset, 2010). In such settings, an increase in inequality can lead

20Table B5 shows t-test statistics for equality of elasticity estimates by region. Meta-regression results are presented before
and after the introduction of regional dummies, as they may be correlated with other sociodemographic variables (Table C2 and
Table C3). Due to a lack of observations when splitting the sample, the stratified meta-regression models by country income group
(Table C4) and for ASF, FV&N, and grains & starchy staples are not including regions (Table C5, Table C6, and Table C7).

16



to a reduction in total food demand, as poorer households typically have higher marginal propensities to

consume food than richer peers. Second, we include the share of the population below 15 years of age. This

variable may proxy household size, as a higher proportion of children typically implies larger households,

which can give rise to economies of scale in food consumption. Third, we examine whether demographic

structure affects demand through the population dependency ratio, given its potential association with income

levels, saving behavior, and household consumption dynamics. We define this ratio as the combined share

of individuals below 15 and above 64 years old.

Using the meta-regression results from our main specification, we estimate predicted elasticities for Level

1 food groups by region. Data and model covariates are set to their sample means. For socio-economic

variables, the values for each region are the weighted average of the 2021 values across the region’s countries.

Population is used as weight.21

3.2 Projections

This is the first review to provide meta-estimates to systematically quantify the relation between household

income and demand for all main food types at the global level. We additionally shed light on differences

in this relation between regions, income, demographics, and urbanization levels. To demonstrate the value

of the elasticities presented in this study and the importance of these heterogeneities, we estimate projected

income elasticities up to 2050 based on the Shared Socioeconomic Pathways (SSPs) (Equation (2)) and

compare values with constant predicted elasticities as of 2021 (O’Neill et al., 2014). SSPs outline a range

of five plausible pathways for societal development, grounded in hypotheses about the key societal factors

that most significantly influence the challenges of climate change mitigation and adaptation.22 They consist

of projected socioeconomic global changes, including GDP (Dellink et al., 2017; Crespo Cuaresma, 2017;

Leimbach et al., 2017), urbanization (Jiang and O’Neill, 2017), total population, and population by age group

21We estimate the predicted elasticities’ standard errors using the following expression: 𝑠𝑒𝐸 =
[
XVX′ + 𝑠𝑑2

𝜈 + 𝑠𝑑2
𝜐 + 𝑠𝑑2

𝜀

] 1
2 . X

represents the row vector of covariate means, including a one for the intercept. V is the variance-covariance matrix of the estimated
coefficients. 𝑠𝑑𝜈 and 𝑠𝑑𝜐 are the standard deviations of the random effects. 𝑠𝑑𝜀 is the standard deviation of the residuals. By
including the off-diagonal elements of V (covariances) in the computation of the standard errors for predicted elasticity estimates,
we account for the correlation between regression coefficients, improving the precision of the standard error estimates.

22The five pathways are: SSP1: Sustainability (“taking the green road”); SSP2: “Middle of the road”; SSP3: Regional rivalry
(“a rocky road”); SSP4: Inequality (“a road divided”); and SSP5: Fossil-fueled development (“taking the highway”) (O’Neill et al.,
2014).

17



(KC and Lutz, 2017).23

𝜖𝑖,𝑟 ,𝑡 = 𝛽0 +
∑︁
𝑚

𝛽𝑚𝐹𝑚,𝑖 +
∑︁
𝑛

𝛽𝑛𝐷𝑛 +
∑︁
𝑝

𝛽𝑝𝑀 𝑝,𝑖 +
∑︁
𝑞

𝛽𝑞𝑆
𝑟 ,𝑡

𝑞 +
∑︁
𝑚

∑︁
𝑞

𝛽𝑚,𝑞𝐹𝑚,𝑖 × 𝑆
𝑟 ,𝑡

𝑞 + 𝛾̂
1√︁
𝑁

(2)

where 𝐷 and 𝑀 represent the sample mean of the data and methods variables, respectively. 𝑁 is the

sample mean of the sample size. 𝑆
𝑟 ,𝑡

is the population-weighted average for the sociodemographic variables

for region 𝑟 , based on the projected population for each country at time 𝑡.

Using our projected income elasticity estimates, we simulate global food demand up to 2050. Baseline

national-level food consumption is proxied using the Food and Agriculture Organization (FAO) data on food

availability in 2021.24 This dataset provides the annual total food supply (in tons) for each country and has

been shown to correlate with other indicators of food intake and health outcomes (Naska et al., 2008). We

focus on the following food groups: ASF; beans, lentils, peas & soy; condiments & sweeteners; FV&N;

grains & starchy staples; and oils & fats.25 Several models of food and agricultural markets have been used

in the literature to project global trends, such as IFPRI’s IMPACT (International Model for Policy Analysis

of Agricultural Commodities and Trade) or the OECD-FAO AGLINK-COSIMO model (Rosegrant and The

IMPACT Development Team, 2012; OECD and Food and Agriculture Organization, 2022). Generally, these

models assume that demand elasticities remain constant over time. In this application, we demonstrate that

allowing demand parameters to vary with projected socio-economic variables may significantly influence

projection results.

𝑓𝑖,𝑟 ,𝑡 =
𝑓𝑖,𝑟 ,2021

𝑝𝑟 ,2021
× 𝑝𝑟 ,𝑡 × 𝑔

𝜖𝑖,𝑟,𝑡
𝑟 ,𝑡−2021 (3)

In this projection exercise, GDP per capita growth (𝑔𝑟 ,𝑡−2021, as the ratio of GDP per capita in future year

𝑡 over GDP per capita in 2021) is used to proxy changes in real income at the population level. 𝑝𝑟 ,𝑡 is the

23The SSP projection data is extracted from the International Institute for Applied Systems Analysis (IIASA)’s SSP Database
- Version 2.0, 2018 (Accessed 18 October 2024). When various models (IIASA, OECD, and PIK) are available for a given SSP
scenario, year, and outcome, we take the average between the available models.

24Food and Agriculture Organization. FAOSTAT. Food balance sheets, 2022 (Accessed 13 January 2025).
25FAO’s Food balance sheet data is not available for processed meals & snacks, FAFH, and all beverages. Table C15 provides

details for regrouping FAO commodities.

18

https://tntcat.iiasa.ac.at/SspDb/dsd?Action=htmlpage&page=10
https://tntcat.iiasa.ac.at/SspDb/dsd?Action=htmlpage&page=10
https://www.fao.org/faostat/en/##data/FBS


regional population for year 𝑡. 𝑓𝑖,𝑟 ,𝑡 and 𝜖𝑖,𝑟 ,𝑡 represent the estimated projected demand and the estimated

projected income elasticity for food group 𝑖 in region 𝑟 in future year 𝑡, respectively. While prices are an

important determinant of food demand, we abstract from price changes. This is motivated by our focus on

long-run structural shifts in income-driven demand and by the considerable uncertainty surrounding long-

term global food price forecasts. The goal is not to predict market-equilibrium outcomes, but to highlight

how projected changes in sociodemographic and elasticity structures may shape future consumption patterns.

4 Results

4.1 Meta-regression

Table 3 provides the estimated coefficients for the meta-regression analysis for each elasticity type.26,27 Our

results indicate that using a demand system rather than a single-equation model does not significantly affect

elasticity estimates. However, the specific type of demand system does matter. In particular, QUAIDS-type

models are associated with lower income elasticity estimates compared to AIDS-type models. This likely

reflects QUAIDS’ ability to flexibly capture nonlinearities in Engel curves, particularly the diminishing

marginal responsiveness of food demand at higher income levels, as motivated by Banks et al. (1997).

In contrast, alternative demand systems grouped under “Other” (e.g., exact affine stone index (EASI),

translog, Rotterdam) do not exhibit systematic differences relative to AIDS-type models.28 Income elasticities

estimated conditional on food demand or the demand for a higher-aggregate food category are associated with

higher values. This is consistent with economic intuition, where spending patterns respond more strongly

26Table C1 displays the meta-regression analysis results with all coefficients, including the food group dummies and their
interactions with sociodemographic variables. Table C2 and Table C3 present the meta-regression analysis results with stepwise
introduction of variables.

27The conditional R2 represents the proportion of variance explained by both the independent variables and the random effects.

It is estimated as follows: 𝑅2 =
𝑠𝑑2

𝑓
+𝑠𝑑2

𝜈+𝑠𝑑2
𝜐

𝑠𝑑2
𝑓
+𝑠𝑑2

𝜈+𝑠𝑑2
𝜐+𝑠𝑑2

𝜀

. 𝑠𝑑2
𝑓

represents the variance explained by the fixed effects (i.e., the independent

variables, without considering random effects), 𝑠𝑑2
𝜈 and 𝑠𝑑2

𝜐 the between-group variances (i.e., variance explained by the random
effects), and 𝑠𝑑2

𝜀 the residual (within-group) variance.
28In robustness checks, we further disaggregated the “Other” category to isolate EASI-type models, which represent higher-order

flexibility systems than QUAIDS (Lewbel and Pendakur, 2009). We found a negative, though not statistically significant, difference
in income elasticity estimates between EASI and AIDS models; however, this comparison may suffer from limited statistical power,
as only 7.5% of income elasticity observations in our sample are based on EASI-type models. Results are available upon request.

19



once a baseline demand for food or broad categories is already established. Using household income rather

than total expenditure to measure income elasticities leads to lower estimates. This corroborates prior

evidence suggesting that household income is measured with greater error, introducing attenuation bias, and

may poorly reflect the permanent consumption capacity of households compared to total expenditure (Bhalla,

1979). Column (2) of Table D1 provides results for a sensitivity analysis in which we add the extra data

and model variables discussed in Section 3.1 and presented in Table B1. Findings indicate no statistically

significant effects on income elasticity estimates.

In Table C4, we disaggregate the sample by country income group, distinguishing between LMIC and

high-income countries (HIC).29 The results suggest that measuring income elasticities using household

income instead of total expenditure has a negative effect in LMIC but a positive effect in HIC. In LMIC,

household income is more volatile, frequently under-reported, or missing in household surveys, making

total expenditure a more reliable proxy for permanent household income (Carletto et al., 2021). Rural

households in these settings often smooth consumption through in-kind transfers, remittances, or savings

(Combes and Ebeke, 2011). Consequently, using reported household income introduces measurement error

that attenuates elasticity estimates. Given that food represents a large share of household budgets in LMIC,

this results in steeper Engel curves when using expenditure. In contrast, household income tends to be

more stable and accurately reported in HIC, while food constitutes a smaller share of the household budget.

Expenditure may be less sensitive to short-term income fluctuations due to access to credit, habit persistence,

and intertemporal smoothing. In this context, income variation is more likely to reflect permanent differences

in living standards, which may lead to higher estimated elasticities. As a result, Engel curves estimated using

household income may be steeper.

29Country income classifications are based on gross national income (GNI) per capita thresholds defined by the World Bank in
1987—the first year such rankings were published—closely aligning with the start of our meta-sample. Source: World Bank, GNI
per capita Operational Guidelines & Analytical Classifications (OGHIST), 2023 (Accessed 23 February 2025). LMIC include low-,
lower-middle-, and upper-middle-income countries. Sensitivity checks using classifications from 2021 or continuous classifications
yield consistent results and are available upon request.

20

https://databankfiles.worldbank.org/public/ddpext_download/site-content/OGHIST.xls
https://databankfiles.worldbank.org/public/ddpext_download/site-content/OGHIST.xls


Income elasticity Price elasticity
Data variables

Subsample 0.001 0.064
(0.019) (0.056)

Model variables
Demand system -0.054 0.196

(0.079) (0.134)
Demand system × QUAIDS -0.100*** 0.008

(0.020) (0.010)
Demand system × Other 0.016 -0.106

(0.033) (0.097)
Conditional 0.249*** 0.024

(0.049) (0.048)
Exp. measure: Income -0.197**

(0.092)
Compensated demand 0.111***

(0.032)
Sociodemographic variables

log(GDP pc) -0.120** 0.212
(0.056) (0.421)

Urbanization 2.300*** -0.118
(0.369) (1.552)

Pop >64 yo -4.399*** -4.005
(1.452) (6.267)

Regions
Africa 0.355** 0.215

(0.146) (0.384)
Americas -0.247** -0.055

(0.097) (0.487)
Asia-Oceania 0.104 0.414

(0.089) (0.496)
Publication bias correction term -0.772 0.629

(0.708) (1.489)
Constant 0.935** -2.611

(0.462) (2.396)
𝑁 6572 6701
Number of studies 186 199
Number of countries 54 55
Conditional R2 0.454 0.553
Variance of error terms 0.268 0.302
Variance of random effects 0.119 0.280
Food groups Yes Yes
Food groups × SDV Yes Yes

Table 3. Meta-regression estimates. Reference categories are underlined in Table B1. The entire model
can be found in Table C1. SDV: sociodemographic variables.
Back to Section 4.1.

Most of the estimated coefficients for the data and model variables do not have a statistically significant

effect on price elasticities. This can be viewed positively, as it suggests that data characteristics or individual

modeling choices do not significantly influence the estimated elasticities. One exception is for compensated

elasticities, which are associated with lower absolute values than uncompensated elasticities. This was

21



expected since most income elasticities are positive, and uncompensated elasticities incorporate both the

substitution effect and the (negative) income effect, whereas compensated elasticities isolate the substitution

effect. Column (5) of Table D1 shows that accounting for expenditure endogeneity, which often stems

from simultaneity bias or omitted variable bias, negatively affects price elasticity estimates (increases

their absolute value), though not statistically significant. Accounting for such endogeneity removes the

confounding influence of household income changes that would otherwise “mask” the full substitution

response. As a result, the price elasticity becomes more negative. This is in line with the demand estimation

literature, which highlights the importance of correcting for expenditure endogeneity through an instrumental

variable or control function approach (Banks et al., 1997; Blundell et al., 1998). Column (5) of Table D1 also

shows a negative coefficient on controlling for demographics, though not statistically significant. Household

characteristic differences predict consumption preferences, needs, and budget allocations. Controlling for

demographic characteristics reduces the risks of omitted variable bias and better identifies substitution

behavior (Pollak and Wales, 1981).

This study also examines how sociodemographic variables influence income and price elasticities.

Figure 3 displays the estimated coefficients and confidence intervals for food group indicators and their

interactions with sociodemographic characteristics. The reference category is ‘All foods’, so all estimates

should be interpreted as differences relative to this group.

Based on Engel’s Law, we expect the proportion of income spent on food to decrease as household income

increases, even as the absolute amount spent may rise. In this study, real income at the population level is

proxied by per capita GDP. Thus, we expect income elasticities to be larger when a country’s per capita GDP

is low, and the contrary for countries with higher incomes. Our estimates confirm this, as a 1% increase

in GDP per capita is associated with a 0.0012-point decrease in income elasticity for ‘All foods’, holding

other factors constant. The relationship between per capita GDP and income elasticities is negative for all

food groups, except for ASF, condiments & sweeteners, and FAFH. This is partially in line with Bennett’s

Law, which states that rising income leads to a notable shift in dietary patterns characterized by a decreased

reliance on coarse, calorie-dense starchy staples and increased consumption of protein, nutrient-dense, and

discretionary foods. Per capita GDP has a statistically significant positive effect on the income elasticity

for FAFH. The wealthier a country becomes, the greater share of additional income consumers allocate to

FAFH, a discretionary food. However, per capita GDP does not have a stronger negative effect on the income

elasticity for grains & starchy staples than on the elasticity for ‘All foods’ and has a nil effect on the income

22



ASF

Beans, lentils, peas, & soy

Condiments & sweeteners

F V & N

Grains & starchy Staples

Processed meals & snacks

Oils & fats

FAFH

All beverages

ASF x log(GDP pc)

Beans x log(GDP pc)

Condiments x log(GDP pc)

F V & N x log(GDP pc)

Grains x log(GDP pc)

Processed meals x log(GDP pc)

Oils & fats x log(GDP pc)

FAFH x log(GDP pc)

All beverages x log(GDP pc)

ASF x Urbanization

Beans x Urbanization

Condiments x Urbanization

F V & N x Urbanization

Grains x Urbanization

Processed meals x Urbanization

Oils & fats x Urbanization

FAFH x Urbanization

All beverages x Urbanization

ASF x Pop >64 yo

Beans x Pop >64 yo

Condiments x Pop >64 yo

F V & N x Pop >64 yo

Grains x Pop >64 yo

Processed meals x Pop >64 yo

Oils & fats x Pop >64 yo

FAFH x Pop >64 yo

All beverages x Pop >64 yo

Income elasticity
-16 -12 -8 -4 0 4 8 12 16

Price elasticity
-16 -12 -8 -4 0 4 8 12 16

Figure 3. Estimated effects and confidence intervals of the food group variables and their interactions
with sociodemographic variables on income and price elasticities. Reference category: All foods. Confi-
dence intervals at the 95% level. Confidence intervals without end caps are truncated for ease of presentation.

elasticity for ASF.

A larger share of the population living in urban areas may reflect higher access to marketed food. Our

findings indicate that urbanization rates have a positive effect on income elasticities. This is true for all

food groups except FAFH, whose consumption is more established in urban settings (Ruel et al., 2017),

23



limiting further increases as income rises. To explore the finding that the inclusion of urbanization in the

meta-regression model reverses the sign of the coefficient on GDP per capita from positive to negative

(Table C2), we incorporate an interaction term between GDP per capita and urbanization. The coefficient

on the interaction term is negative and statistically significant, indicating that the effect of urbanization on

food demand responsiveness to household income diminishes as income rises and that the effect of income

growth on income elasticities is not uniform. Real income growth reduces income elasticity estimates in

more urbanized countries but has a limited effect in low urbanization contexts (Table D3).

We find that the income elasticity of food demand declines with the share of the population 65 years old

or above. This is consistent with lifecycle theory and less responsive consumption among older households.

However, the interaction terms are positive for processed foods, FAFH, and beverages, suggesting that older

consumers may shift toward convenience-oriented and discretionary food consumption as income rises.

Finally, sensitivity analyses including the share of population below 15 years old and a Gini index to the

meta-regression show no statistically significant effects (Table D3).

For price elasticities, we do not observe any statistically significant effects for the sociodemographic

variables (Table 3 and Figure 3). The direction of the coefficient on per capita GDP suggests that higher

income is associated with lower absolute price elasticities, which aligns with the idea that wealthier consumers

are less sensitive to price changes because food represents a smaller share of their overall budget. However,

Table C4 shows that a reverse relationship occurs in HIC. Food demand in such contexts may shift toward

luxury or more substitutable items (e.g., organic, premium brands, eating away from home), which may have

higher price elasticities. Also, substitution opportunities and market integration are higher than in LMIC.

The negative association between price elasticities and urbanization is only statistically significant at the

95% level for beans, lentils, peas & soy and oils & fats. Urban diets may rely more on market purchases

rather than home production. Beans and other pulses may compete more closely with other protein sources

in urban areas, increasing substitution when prices rise. Greater diversity in retail formats and available

substitutes can also explain these results, particularly in HIC. Finally, the share of the population aged 65 and

above is negatively associated with price elasticities in LMIC. Older consumers in these settings are more

sensitive to price changes, likely due to limited income sources (e.g., lack of pension coverage). In contrast,

elderly households in HIC may smooth consumption or be shielded from price shocks through pensions or

welfare programs.

We perform various sensitivity analyses using alternative model specifications in Appendix D. First, we

24



show that the meta-regression analysis results are robust to the addition of a categorical data period variable

in the model (columns (3) and (6) of Table D1). However, the negative and statistically significant effect of

GDP per capita on income elasticities observed in the main specification is attenuated. This suggests that

part of the relationship may be confounded by time-related heterogeneity in study characteristics or absorbed

by the strong correlation between income levels and time. Second, the coefficient on the publication bias

correction term is not statistically significant in our main specification. We find that results are robust to

its exclusion (columns (2) and (5) of Table D2). Third, Figure 2 shows a large dispersion and the presence

of a number of implausible values among the elasticity estimates in the meta-sample. Nevertheless, our

results are robust to the exclusion of outliers, i.e., income and price elasticity estimates outside the range

[Q1 − 3 × IQR,Q3 + 3 × IQR] (columns (3) and (6) of Table D2). Finally, results are robust to the removal

of observations with price elasticity estimates above zero, i.e., inconsistent with economic theory (column

(7) of Table D2).

4.2 Predicted elasticities

Figure 4 presents the predicted income and price elasticities by food group and region.30 As expected,

income elasticities are the highest for more discretionary food groups such as ASF and FAFH, and lowest

for grains & starchy staples. The findings confirm the geographical heterogeneity shown in the descriptive

statistics (Table B5). Predicted income elasticities for grains & starchy staples are the highest in Africa,

where they form the bulk of diets due to their affordability and caloric density. In contrast, ASF emerges as

the most income-elastic category in Africa. Similarly, other food groups like FV&N and oils & fats exhibit

higher elasticities in Africa than in other regions, indicating a broader shift toward dietary diversification

as incomes grow. Conversely, income elasticities for ASF are lowest in Europe and the Americas, where

consumption levels are high.31 In both regions, rising incomes have enabled the average consumer to move

beyond subsistence, with surplus household income now directed toward more luxurious and discretionary

consumption.

Turning to price elasticities, regional and food-group variation is also evident. Demand for staple foods

such as grains & starchy staples and oils & fats tends to be less price elastic, whereas FAFH demand is

30Table C8 provides the results in a table format. The standard errors are displayed in Table C10.
31Our World in Data. Per capita meat consumption by type, 2021, 2023 (Accessed 20 January 2025).

25

https://ourworldindata.org/grapher/per-capita-meat-type


-1

0

1

2

     

 

ASF

-1

0

1

2

     

 

Beans lentils peas & soy

-1

0

1

2

     

 

Condiments & sweeteners

-1

0

1

2

     
 

F V & N

-1

0

1

2

     

 

Grains & starchy staples

-1

0

1

2

     

 

Processed meals & snacks

-1

0

1

2

Af
ric

a

Am
er

ic
as

As
ia

-O
c

Eu
ro

pe

G
lo

ba
l

 

Oils & fats

-1

0

1

2

Af
ric

a

Am
er

ic
as

As
ia

-O
c

Eu
ro

pe

G
lo

ba
l

 

FAFH

-1

0

1

2

Af
ric

a

Am
er

ic
as

As
ia

-O
c

Eu
ro

pe

G
lo

ba
l

 

All beverages

Price Income

Figure 4. Predicted income and price elasticities, by region and food groups. Oc: Oceania. ASF:
animal-sourced foods; FAFH: food away from home; FV&N: fruits, vegetables, and nuts. Results in table
format can be found in Table C8. Standard errors can be found in Table C10.

the most price-responsive, being unit-elastic at the global level (a price elasticity of -1.0). This aligns with

expectations, as grains & starchy staples and oils & fats are often necessities with fewer close substitutes,

especially in LMIC settings. In Africa and Europe, the demand for FAFH is price elastic, while it remains

inelastic in the Americas and Asia-Oceania. North America accounts for 47% of global FAFH market

revenue,32 and South America is among the fastest-growing markets in this segment (Popkin and Reardon,

2018). Meanwhile, China and India, along with the United States, dominate the global online food delivery

sector.33 Table C9 disaggregates the results for predicted uncompensated and compensated price elasticities

by food group and region.34 In line with economic theory and the meta-regression estimates, uncompensated

elasticities are consistently larger in absolute value than compensated elasticities.

32The Brainy Insights. Food Away From Home Market, 2023 (Accessed 22 January 2025).
33Statista. Marketing Insights. eCommerce. Online Food Delivery, 2024 (Accessed 22 January 2025).
34The standard errors are available upon request.

26

https://www.thebrainyinsights.com/report/food-away-from-home-market-13399
https://www.statista.com/outlook/emo/online-food-delivery/worldwide


4.3 Projections

Table C12 presents population-weighted averages of sociodemographic variables in 2021 by region, alongside

SSP projections for 2050 across all scenarios. Fossil-fuel development (SSP5) leads to the highest projected

GDP per capita, while regional rivalry (SSP3) yields the most limited economic growth. SSP3 is also

associated with the highest projected total population and the lowest projected rates of urbanization and

population aging. Total population is projected to rise least in the sustainability (SSP1) and SSP5 scenarios,

while urbanization and population aging rise most. Table C13 provides a time series of projected changes

in sociodemographic variables by region under SSP2 (“Middle of the Road”). This scenario is often used

as a baseline in the literature, reflecting a continuation of historical trends and moderate progress toward

sustainable development goals and greenhouse gas emission reductions.

Table 4 presents estimates of predicted income elasticities for 202135 and projections for each decade

from 2030 to 2050 for SSP2 by food group and region. Results reveal divergent projected trends in income

elasticities across regions, reflecting both differences in economic and sociodemographic trajectories and

underlying consumption saturation. In Africa, income elasticities remain high or even increase slightly

across most food groups between 2021 and 2050. This suggests that rising income and urbanization will

continue to fuel growth in demand, particularly for nutrient-rich items such as beans, lentils, peas & soy

(1.17 to 1.32), FV&N (1.05 to 1.19), and oils & fats (1.00 to 1.19). These consistently high elasticities signal

an ongoing dietary transition driven by unmet consumption needs and aspirations for more diverse diets.

In contrast, more economically advanced regions —especially Europe and the Americas —exhibit a

declining trend in income elasticities for most food groups, consistent with Engel’s Law, saturation effects,

and population aging. For example, Europe’s elasticity for ASF falls from 0.69 in 2021 to 0.38 in 2050,

while condiments & sweeteners drop from 0.73 to 0.59. Exceptions to this pattern are discretionary foods.

Income elasticities for processed meals & snacks, FAFH, and beverages are projected to rise in all regions,

highlighting a growing demand for convenience associated with higher real incomes and urban lifestyles.

35They are the same as highlighted in Figure 4 and Table C8.

27



ASF Beans Condi-
ments

F V & N Grains Processed
meals

Oils & fats FAFH All bever-
ages

Africa

2021 1.35 1.17 0.89 1.05 0.81 0.99 1.00 0.95 1.04
2030 1.38 1.32 0.93 1.12 0.89 1.14 1.09 0.89 1.09
2040 1.39 1.33 1.03 1.16 0.92 1.16 1.15 0.90 1.11
2050 1.38 1.32 1.09 1.19 0.94 1.16 1.19 1.00 1.14

Americas

2021 0.84 1.06 0.80 0.82 0.59 0.94 0.92 0.42 0.70
2030 0.74 1.10 0.74 0.83 0.63 1.03 0.92 0.75 0.75
2040 0.64 1.07 0.69 0.81 0.64 1.03 0.89 1.10 0.78
2050 0.53 1.01 0.63 0.78 0.64 1.02 0.85 1.48 0.80

Asia-
Oc

2021 0.94 0.59 0.77 0.79 0.51 0.49 0.73 1.47 0.82
2030 0.86 0.52 0.76 0.77 0.51 0.45 0.71 1.76 0.84
2040 0.75 0.53 0.72 0.78 0.55 0.52 0.71 2.17 0.89
2050 0.66 0.56 0.68 0.78 0.59 0.59 0.71 2.49 0.94

Europe

2021 0.69 0.87 0.73 0.86 0.69 0.88 0.86 1.97 0.94
2030 0.59 0.87 0.71 0.87 0.72 0.92 0.87 2.32 0.98
2040 0.49 0.86 0.66 0.85 0.74 0.95 0.85 2.69 1.02
2050 0.38 0.84 0.59 0.84 0.76 0.97 0.82 3.04 1.05

Table 4. Baseline and projected income elasticities, under scenario SSP2, by food group and region. SSP: Shared Socioeconomic Pathways.
SSP2: “Middle of the road” scenario. ASF: animal-sourced food, FAFH: food away from home, FV&N: fruits,vegetables & nuts. Oc: Oceania.

28



An exception to the general decline in income elasticities across non-discretionary food groups is grains

& starchy staples, for which elasticities are projected to rise modestly in all regions. This counterintuitive

pattern likely reflects compositional shifts within the category, as rising incomes are increasingly directed

toward higher-value, processed, or convenience-oriented grain products rather than coarse or bulk staple

consumption (Bellemare et al., 2024). In lower-income regions such as Africa, the increase may also capture

persistent caloric shortfalls and continued reliance on staple-based diets as incomes grow. In higher-income

regions, the upward trend may be driven by urbanization dynamics, with urban households relying more

heavily on packaged breads, breakfast cereals, and ready-to-eat grain product forms, which tend to be more

income-responsive than traditional staples (Harding and Lovenheim, 2017).

Global average projected elasticities results for the other SSP scenarios are presented in Table C14. The

findings reveal variation across scenarios, driven by differences in projected sociodemographic trajectories.

Notably, SSP3 yields the lowest projected income elasticities, reflecting its limited rise in urbanization —a

factor strongly and positively associated with income elasticities, as shown in Table 3.

Table 5 presents the results of our global food demand projection exercise under SSP2. We compare re-

sults using constant predicted income elasticities estimated as of 2021 and time-varying projected elasticities

estimated using population-weighted average projected sociodemographic characteristics. The projections

based on constant income elasticities are lower than those based on time-varying projected elasticities, except

for ASF and condiments & sweeteners. The gap between demand projections based on time-varying income

elasticities and those based on constant elasticities widens over time, reflecting the growing influence of

changing sociodemographic conditions on food demand. By 2050, the projections based on time-varying

elasticities are higher by 1.6% for beans, lentils, peas & soy, 1.8% for FV&N, 5.7% for grains & starchy

staples, and 1.9% for oils & fats. They are lower than the projections based on constant elasticities for ASF

by 10.7% and for condiments & sweeteners by 1.9%.

Table C16 presents the projection results for each region under SSP2. As expected from the projected

elasticity estimates in Table 4, projected demand based on time-varying elasticities is higher for all food

groups in Africa under this scenario. However, it is similar to or lower than projections based on constant

elasticities in other regions, except for grains & starchy staples. Table C17 presents the global projections

for all other scenarios. Projected global ASF demand based on time-varying elasticities is lower for all SSP

scenarios. The contrary is true for other food groups, except for SSP3 for which projected global demand

based on time-varying elasticities is lower for all food groups.

29



Baseline Constant 2021 elasticities Time-varying elasticities

2021 2030 2040 2050 2030 2040 2050

ASF 1259.0 1628.8 2087.8 2538.1 1608.1 1976.8 2265.9
Beans 76.3 94.6 116.9 137.8 94.5 117.3 140.0
Condiments 260.2 323.4 400.0 472.2 323.3 397.0 463.1
F V & N 1900.9 2390.5 2987.9 3558.0 2394.2 3012.5 3623.1
Grains 1962.3 2336.5 2782.0 3183.7 2347.7 2852.6 3364.2
Oils & fats 128.6 160.6 199.6 236.5 160.9 201.3 240.9

Table 5. Projected global food consumption levels, under scenario SSP2, by food group, in million
tons. ASF: animal-sourced food, FV&N: fruits, vegetables & nuts. SSP: Shared Socioeconomic Pathways.
SSP2: “Middle of the road” scenario.

5 Discussion

In line with the literature, we find that most data and model choices do not have a statistically significant effect

on demand elasticity estimates (Bouyssou et al., 2024; Colen et al., 2018; Cornelsen et al., 2016). There are

some notable exceptions for income elasticities. More flexible demand systems are associated with lower

income elasticity estimates. As expected from economic theory and in line with Bouyssou et al. (2024)’s

findings for ASF, income elasticities tend to be higher when estimated conditionally on food demand or on

broader, more aggregated food categories. Using household income rather than total expenditure to measure

income elasticities is associated with lower estimates, which is consistent with results found by Colen et al.

(2018) in their meta-analysis in Africa. For price elasticities, compensated elasticities are associated with

lower absolute values than uncompensated elasticities. This result aligns with theory for normal goods, since

the (negative) income effect reinforces the substitution effect in the uncompensated elasticity.

Our findings provide empirical support for Engel’s Law, particularly in more urbanized settings. Support

for Bennett’s Law is comparatively weaker. While discretionary foods like FAFH become more income-

elastic as GDP per capita increases, real income growth has no statistically significant impact on ASF

elasticities. Moreover, the responsiveness of grain & starchy staples demand to household income does

not diminish more steeply with real income growth than it does for total food demand. A limitation of

our analysis is the inability to examine substitution patterns within aggregated food groups — for example,

between coarse and fine cereals or across quality tiers within a category. This reflects a common constraint

in meta-analyses, where food group definitions vary across primary studies and must be harmonized to a

30



common set of aggregated categories. As a result, our analysis may obscure within-group heterogeneity

in income responsiveness. For example, recent micro-level evidence by Bellemare et al. (2024) in LMIC

finds that as income rises, households substitute coarse staples with fine staples, and coarse staples with

protein-rich foods, but not fine staples with protein-rich foods, suggesting a more nuanced hierarchy in food

transitions than captured by broad staple or protein categories alone. Our aggregated grains & starchy staples

grouping likely masks these distinctions.

To probe further, we re-estimate the meta-regression models separately for ASF, FV&N, and grains &

starchy staples (Table C5, Table C6, and Table C7). Results confirm the negative association between GDP

per capita and income elasticities, with statistical significance for rice. However, this classification does

not clearly align with the fine and coarse staple distinctions in Bellemare et al. (2024), particularly given

that this classification may be context-dependent. In the ASF category, we observe an overall statistically

significant negative association with GDP per capita but a positive association for eggs and dairy. This is

in line with Bouyssou et al. (2024)’ findings. These same two categories show negative interactions with

urbanization, which may reflect a transition from home production to market purchases. For FV&N, we find

a divergence between fruits and vegetables as income elasticities for fruits and nuts increase with GDP per

capita, whereas those for vegetables decline. This asymmetry may stem from differences in price sensitivity,

perishability, and perceived necessity (Miller et al., 2016). As the literature almost exclusively aggregates

these two subgroups, this result points to the need for future research with more granular food classifications

to unpack these heterogeneous effects.

We find that urbanization is associated with higher income elasticity estimates but has no statistically

significant effects on price elasticity estimates, consistent with findings from other meta-analyses (Bouyssou

et al., 2024; Colen et al., 2018). Urbanization is a key factor influencing food consumption patterns. It is

associated with reduced reliance on home-grown foods (Mendez and Popkin, 2004). Urban centers are more

closely tied to world markets and, therefore, urban residents tend to have greater access to a diversity of

food types and sources (Rosegrant et al., 2001). When exploring the interaction effect between real income

growth and urbanization, we find that urbanization has a diminishing effect on the income elasticity of food

demand as countries become wealthier. In LMIC settings, increases in urbanization are associated with a

positive effect on income elasticities, likely reflecting improved access to markets, dietary diversification,

and changing consumption preferences. The marginal impact of urbanization is smaller or negligible in

HIC contexts, as diets are already diversified and further income gains exert limited influence on food

31



demand patterns. We additionally explore the urbanization-income elasticities relationship by food group.

We find that the positive association is more modest for ASF and beverages, likely explained by higher meat

consumption and sugary drinks availability in urban settings (York and Gossard, 2004; Malik et al., 2013).

The negative link between urbanization and FAFH income elasticities, and the weaker positive association

for ASF and beverages, may reflect early adoption patterns and saturation effects in urban areas.

Population aging is associated with lower income elasticities, in line with the expectation that older indi-

viduals require fewer calories and exhibit more stable dietary preferences. To further probe this mechanism,

we examined the role of youth and overall population dependency, finding no significant effects (Table D3).

This suggests that the observed demographic effects on demand elasticities are primarily driven by aging

rather than broader shifts in population structure.

Finally, we explored whether income inequality moderates the relationship between income and food

demand by including the Gini index as an additional sociodemographic covariate. While the estimated

interaction was not statistically significant, the coefficient was negative, suggesting that greater income

inequality could possibly attenuate the responsiveness of demand to income. This direction aligns with the

theoretical predictions of Cirera and Masset (2010), who argue that high income inequality between countries

leads to a flattening of Engel curves when aggregating demand across heterogeneous populations, lowering

the observed income elasticity at the macro level. In our case, the absence of a statistically significant effect

may reflect limited variation in inequality across studies or measurement challenges in harmonizing Gini

estimates across time and data sources. Nevertheless, our findings are consistent with the view that high

inequality could possibly dampen the strength of aggregate income responses, even if not precisely estimated

here.

Our predicted income elasticities are broadly consistent with those reported by Colen et al. (2018)

in Africa, especially regarding relative magnitudes across food groups. As in their analysis, we find the

highest income elasticity for ASF and the lowest for grains & starchy staples. However, our predicted

elasticities for Africa are notably higher. This may stem from several methodological differences between

the two studies. We estimate a larger and more up-to-date meta-sample. More importantly, we control

for the conditionality of estimations (e.g., whether elasticities are conditional on food demand or broader

expenditure) and explicitly model interactions with sociodemographic drivers. This feature allows us to

more accurately capture heterogeneities in food demand responsiveness. We find that beverages exhibit the

highest income elasticity in Africa —a pattern consistent with market analyses predicting rapid growth in

32



carbonated soft drink consumption across the continent.36 This is also in line with Colen et al. (2018)’s

predicted elasticity results classifying beverages as income-elastic.

There is comparatively more existing literature on predicted price elasticities than income elasticities for

specific food groups at the global level. One of the most widely cited works and the latest study available

is Green et al. (2013), which synthesizes demand elasticities across food groups and income levels using

studies published up to August 2011. Our analysis extends the evidence base by incorporating a broader

set of more recent studies and accounting for important sources of heterogeneity such as urbanization and

demographic characteristics.

Our results align well with Green et al. (2013)’s estimates for comparable food aggregates. Our predicted

global elasticity for oils & fats is −0.51, which falls well within the range they report for low- to high-income

countries (–0.42 to –0.60). The same is true for ASF (−0.69), FV&N (–0.67), and grains & starchy staples

(–0.58), despite slight differences in group definitions.37 Overall, we find that Asia-Oceania is the least

price-sensitive region, in line with Bouyssou et al. (2024). In the absence of global estimates for FAFH price

elasticity, our finding that FAFH is the most price-responsive food group is consistent with evidence from the

U.S. (Andreyeva et al., 2010). Our global estimate for beverages (–0.76) also matches closely with the values

reported by Andreyeva et al. (2010) for soft drinks and juice in the U.S. (–0.79 and –0.76, respectively).

In regions like the Americas and Europe, where sugar-sweetened beverages (SSB) make up a large share

of total beverage consumption, this aggregate likely reflects the high price-responsiveness of the beverages

category (Andreyeva et al., 2022). Conversely, in Africa and Asia, where per capita SSB consumption is

lower, this category may be dominated by water, tea, or other non-SSB, typically associated with lower price

responsiveness (Muhammad et al., 2019).

We further predict income and price elasticities for disaggregated subgroups within ASF, FV&N, and

grains & starchy staples, which represent the three Level 1 food groups with the largest sample sizes in our

meta-sample (Figure 5).38 Among ASF, pork has the lowest predicted income elasticity, while chicken &

other poultry and dairy & whole milk products are the most income-elastic. Lower global income elasticities

for pork and beef likely reflect cultural and religious dietary restrictions in highly populated regions such

36Grand View Research, Inc. Middle East & Africa Carbonated Soft Drink Market Size & Outlook, 2021 (Accessed 14 July
2025).

37Green et al. (2013) report the following ranges: meat, fish, and dairy (–0.6 to –0.8), fruit and vegetables (–0.53 to –0.72), and
cereals: (–0.43 to –0.61).

38Table C11 shows the results in a table format.

33

https://www.grandviewresearch.com/horizon/outlook/carbonated-soft-drink-market/mea


-1

-.5

0

.5

1

C
hi

ck
en

 &
 o

th
er

 p
ou

ltr
y

Be
ef

 &
 la

m
b

Po
rk

Fi
sh

Eg
gs

D
ai

ry
 &

 w
ho

le
 m

ilk
 p

ro
du

ct
s

   
   

   
   

   
   

   
Al

l f
ru

its
 &

 n
ut

s

   
   

   
   

   
   

   
 A

ll 
ve

ge
ta

bl
es

St
ar

ch
y 

ro
ot

s 
& 

tu
be

rs

W
he

at
, m

ai
ze

 &
 o

th
er

 g
ra

in
s

R
ic

e

 

Price Income

Figure 5. Predicted income and price elasticities, by Level 3 food groups. Based on the full sample.
ASF: Animal source foods; FV&N: fruits, vegetables, and nuts. Results in table format can be found in
Table C11. Standard errors are available upon request.

as Northern and Western Africa, the Middle East, and South Asia. Predicted income elasticities for grains

& starchy staple subgroups are low and homogeneous, with values between 0.53 and 0.68, reflecting their

status as dietary essentials in many parts of the world. In the FV&N category, predicted income and absolute

price elasticities are notably higher for fruits & nuts than for vegetables, in line with McCullough et al.

(2024)’s findings in Sub-Saharan Africa. Our predicted price elasticities for meat types (−0.86 to −0.97)

fall within the range of Gallet (2010)’s predicted elasticity results for beef, pork, lamb, poultry, and fish

(−0.78 to −1.17). Overall, disaggregation from Level 1 food groups reveals higher predicted price elasticity

estimates, consistent with expectations that substitution possibilities are more pronounced at finer levels of

food classification.

Others have previously attempted to project food demand accounting for changes in real income. In

China, Zhou et al. (2020) investigate the relationship between income elasticities for meat and cereals and

income growth through a meta-analysis, and then project such elasticities to 2030, accounting for changes in

real per capita income. Yu et al. (2004) also argue that global food demand projections should rely on flexible

demand systems that reflect declining income elasticities as consumers become wealthier. Building on these

efforts, our study further incorporates changes in urbanization and population aging —two sociodemographic

34



drivers that our meta-regression results identify as significant determinants of income elasticities.

Accounting for these dynamics and time-varying income elasticities leads to notable differences in

projected global food demand. While the percentage differences with projections based on constant income

elasticities might seem relatively small, they translate into large absolute quantity differences. By 2050,

these differences are about −272.2 million tons for ASF, 2.2 million tons for beans, lentils, peas & soy, −9.1

million tons for condiments & sweeteners, 65.1 million tons for FV&N, 180.5 million tons for grains &

starchy staples, and 4.4 million tons for oils & fats. For comparison, differences for ASF and grains & starchy

staples represent approximately the total consumption in the Americas as of 2021 (Table C16). Zhou et al.

(2020) also predict lower future income elasticities for meat, and consequently lower future demand, when

accounting for time-varying elasticities in China. Their results for cereals are mixed, with higher projected

future income elasticities for wheat but lower for other cereals.

Overall, the findings from this projection exercise underscore the importance of using time-varying

income elasticities when projecting long-term consumption patterns, rather than assuming static preferences

that may obscure structural dietary shifts. A limitation of our approach, and of related work such as Zhou

et al. (2020), is the assumption of linear relationships between income elasticity estimates and sociode-

mographic drivers. While this assumption aids interpretation, it does not impose theoretical bounds on

income elasticities, such as the expectation that they decline and eventually plateau as real income rises.

Introducing non-linear interactions (e.g., quadratic terms) in the meta-regression would markedly increase

the number of parameters, consuming degrees of freedom, and leaving too little scope to meaningfully assess

between-study variability in our two-level random intercept model. For this reason, non-linear interactions

with sociodemographic variables are not used in the literature (e.g., Bouyssou et al. (2024), Colen et al.

(2018)).

Further limitations deserve consideration. By construction, meta-regression studies rely on sociodemo-

graphic variables that are often partially correlated. While we model their joint effects, this multicollinearity

may reduce precision in projections and obscure underlying mechanisms. Although we account for an

exhaustive list of methodological moderators in our main specification and sensitivity meta-regressions,

information on the econometric method (e.g., seemingly unrelated regressions, maximum likelihood, least

squares) was not explicitly reported in most primary studies. As a result, we could not systematically control

for this factor, which may influence elasticity estimates (Gallet, 2010). Nevertheless, we do include demand

model choice in our meta-regression model, which is correlated with the econometric method used. As in

35



other global meta-regression analyses of food demand, our regional aggregates may mask within-region het-

erogeneity. For instance, Asia-Oceania-wide population-weighted average predicted estimates are influenced

heavily by populous countries like China and India, and may not reflect trends in smaller economies.

6 Conclusions

This paper aimed to examine the extent to which empirical choices, such as data characteristics and estimation

methods, account for the observed heterogeneity in food demand elasticities, assess whether theoretical

predictions about the role of sociodemographic factors are supported by empirical evidence, and simulate

how food demand is expected to vary with projected sociodemographic changes. It provides the most

comprehensive global meta-analysis of food demand elasticities to date. We synthesized over four decades

of peer-reviewed empirical evidence and collected more than 13,000 elasticity estimates covering 57 countries

and five continents.

We find that while methodological and data choices generally have limited effects on demand estimates,

income elasticities are more sensitive to model specification —particularly the choice of demand system

model, conditional demand, and the choice of household income measure. Our large database additionally

allowed us to add a set of sociodemographic variables and interactions in the meta-regressions characterizing

the heterogeneity across world regions and products. Our findings indicate empirical support for Engel’s

Law, as food income elasticities decline with real income growth. This decline becomes more significant

with urbanization. In contrast, support for Bennett’s Law is mixed. While income elasticities increase

with real income for some discretionary foods, they do not decline more steeply for staples than for total

food demand, nor rise consistently for protein-rich foods like ASF. Importantly, we show that urbaniza