Taxation of foods high in fat, sugar, and sodium in India: A modelling study of health and economic impacts
Centre for Health Economics & Policy Innovation, Department of Economics & Public Policy, Imperial College Business School, South Kensington Campus, Exhibition Rd, London, SW7 2AZ, UK
Institute of Economic Growth, Delhi University Enclave, North Campus, Delhi, 110007, India
Resolve to Save Lives Services PLC, Delhi, India, Regus, 4th Floor, Rectangle No 1, Commercial Complex, D4, Saket, New Delhi, 110017, India
Resolve to Save Lives, 1520 Belle Vista Blvd, Suite 4036, Alexandria, VA 22307, USA
ꝉCorresponding author: f.sassi@imperial.ac.ukAbstract
Background
Consumption of foods high in fat, sugar, and sodium (HFSS) and obesity are rapidly increasing in India. Taxing HFSS foods has been proposed as one of the policy interventions to promote healthier diets globally. This study estimates the effect of this approach on nutrient intake, diet-related disease, and associated health and economic burdens in India.
Methods and findings
We use nationally representative household expenditure survey data, dietary requirements, and food composition tables to model individual nutrient intake. Consumer responsiveness to food price changes for three income terciles, captured in price elasticities, is estimated using an Almost Ideal Demand System model. Longer-term policy impacts are estimated through a novel dynamic microsimulation model, Health-GPS. Modelled policy outcomes include changes in risk exposures, disease incidence and burden, and total health expenditure. On average, 9.9% of total energy intake comes from HFSS items, based on the definition by the Food Safety and Standards Authority of India’s Labelling & Display Amendment Draft Regulations 2022. Applying the highest Goods and Services Tax (GST) rate of 40% on HFSS items is associated with a persistent average per capita decrease of 0.17kg/m2 (95% CI: −0.17, −0.17) in body mass index and 45.8mg (95% CI: −45.9, −45.7) in daily sodium intake. Over 30 years, this could reduce annual disease incidence by up to 1.72% (95% CI: −1.78%, −1.66%) on average and prevent 0.63 million (95% CI: −0.71, −0.55) disability-adjusted life years per year from ischemic heart disease, chronic kidney disease, stroke, diabetes, and asthma, reducing total health expenditure by US$601 million (95% CI: −624, −578) per year. Further work is needed to incorporate potential benefits or harms associated with changes in other foods and nutrients that are not currently modelled, such as fruits, whole grains, fibre, and red and processed meats. Larger absolute health gains accrue to higher-income individuals, reflecting higher baseline HFSS food intake. Given substitution patterns and a price-inelastic demand, the tax change is expected to generate a 92.0% (95% CI: 88.2%, 95.7%) increase in tax revenue from foods and beverages with only a minor effect on household spending (+1.0%, 95% CI: +0.0%, +1.9%).
Conclusions
Higher taxation of HFSS foods could help mitigate rising incidence of diet-related diseases and morbidity in India, reduce healthcare costs, and serve as an additional source of revenue for the government.
Article notes
Competing Interest Statement
The authors have declared no competing interest.
Funding Statement
This work was supported by Resolve To Save Lives (RTSL, United States, contract PR0414 to FS) as part of their Nutrition and Surveillance programme. Authors MR, MS, and LS received salary support from RTSL during the conduct of this work. No other authors received salary support from the funder. The funders contributed to the study conceptualization, study design, and writing of the report. They had no role in data collection, data analysis, or data interpretation. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.
Summary of Updates:
Introduction
Unhealthy diets and obesity are major risk factors for non-communicable diseases, such as cardiovascular disease, cancer, and diabetes. These diseases account for substantial healthcare costs, as well as wider economic costs for governments and individuals [1]. In India, adult overweight and obesity levels have increased fourfold since 1975 [2]. Of particular concern is the increasing consumption of highly industrially processed foods rich in energy, saturated fat, sugar, and sodium [3].
Governments can promote healthy diets through fiscal policies, including the taxation of foods high in fat, sugar, and sodium (HFSS), as well as subsidies on fruits and vegetables, alongside food marketing regulation and front-of-pack nutrition labelling. There are strong health and economic rationales for using fiscal policies to address the external costs of unhealthy diets and obesity (i.e. ‘externalities’, such as collectively borne health care costs linked with obesity) while “encouraging people to avoid acting against their own self-interest” (i.e., ‘internalities’) [4,5].
India applies a Goods and Services Tax (GST), introduced in July 2017, at five different rates: 0%, 5%, 12%, 18%, and 28%. Among food and beverage items, the 28% rate only applies to caffeinated and sweetened aerated beverages, which include sugar-sweetened beverages (SSBs), along with an additional 12% compensatory Cess. This cess was part of a nationwide mechanism intended to compensate states for any shortfall in their tax revenue for five years following the introduction of the GST system, relative to a baseline assuming 14% annual growth in their 2015–16 tax revenue from taxes subsumed under GST. Adjusting existing GST rates to align with nutrition and health objectives may improve diets by shifting food preferences.
Only a few countries have implemented taxes on HFSS foods, with limited scope, targeting either specific food categories (e.g., confectionery and ice cream in Finland or French Polynesia), non-essential energy-dense foods (e.g., in Mexico), or selected nutrients (e.g., sugar and salt in Hungary) [6]. Identifying HFSS foods for taxation based on nutrient content, for example, using a nutrient profile model (NPM), may capture unhealthy foods more broadly, and is less likely to apply high rates to healthier foods or incentivise negative substitutions from ‘healthy’ to ‘less healthy’ foods [7]. NPMs are increasingly used in nutrition policies, such as front-of-pack nutrition labelling and marketing restrictions. In 2023, Colombia became the first country to introduce a broad excise tax on HFSS foods based on an NPM. The government subsequently increased the excise tax rate in 2024 and 2025, offering a promising example for other countries [8].
This study estimates the potential impacts on nutrient intakes and diet-related diseases of aligning India’s GST rate differentiation on foods and beverages to nutritional quality, placing a higher tax burden on HFSS foods, in line with suggested policy intervention globally [6]. We first estimate baseline energy and nutrient intakes, as well as consumers’ responsiveness to food price changes. We then calibrate a microsimulation model to measure the effects of alternative taxation scenarios on various health and economic outcomes, as well as their distribution across income groups, over the next 30 years.
Simulation analyses in Costa Rica, the Philippines, and the UK have suggested that NPM-based taxes can significantly improve diet [9–11]. The long-term population health and economic impacts of such a comprehensive nutrient-based approach remain understudied, especially in low- and middle-income countries. A recent study investigated the potential impact of an excise tax on HFSS foods in India on consumption and tax revenue, without considering nutrition or health outcomes [12]. While others have simulated SSB taxes and their impact on non-communicable diseases in middle-income countries [13], including in India [14], our study fills this gap for broader HFSS food taxation.
We contribute to the literature by applying a novel microsimulation model, Health-GPS, to evaluate the impacts of HFSS food taxation in India. The model’s architecture involves a hierarchy of demographic and socio-economic characteristics, risk factors, and diseases [15]. We simulate 0.5% of the Indian population to estimate policy impacts on risk factors, disease incidence, and disability-adjusted life years (DALYs) as well as associated health expenditure. We provide the most recent estimates of macronutrient intakes in the Indian population using the latest National Sample Survey Office (NSSO) Household Consumption Expenditure survey 2022-23. Lastly, we present updated estimates of the price and income elasticity of demand for foods and beverages by income groups. These represent essential metrics to estimate the potential impact of alternative fiscal reforms.
Methods
Estimating energy and nutrient intake
Our primary data source is the 2022-23 Household Consumption Expenditure survey from the NSSO. This survey collected food and beverage consumption, and expenditure from a nationally representative sample of 261,746 households between August 2022 and July 2023 (Table A1 in S1 Supporting Information provides household characteristics) [16]. Food consumption is recorded over two recall periods: 30 days for cereals, pulses, sugar, and salt, and 7 days for all other items. We harmonised these data to obtain daily household consumption for each item and matched them to their respective nutrient content. We extract nutrient information (energy, sugar, sodium, saturated fat, total fat, carbohydrates, and protein) from the National Institute of Nutrition (NIN) Indian Food Composition Tables 2017 [17], complemented by the United States Department of Agriculture (USDA) Food and Nutrient Database for Dietary Studies 2017-2018 [18]. As NSSO items do not differentiate between processing levels for meat and fish, we assume that 20% of purchased quantities are processed for these groups, based on expert opinion. We individualise household-level intake based on household members’ age and sex using the daily NIN Dietary Guidelines for Indians, 2024 [19]. Lastly, we assume that 55kg of food is wasted yearly by the average Indian to better proxy dietary intake based on data from the UN Environmental Programme [20]. The final sample size is 256,464 households after dropping households with missing food consumption information and 1% lowest outliers (i.e., daily per capita energy intake lower than 933 kcal) and 1% highest outliers (i.e., daily per capita energy intake higher than 4,673 kcal).
Estimating response to food price changes
We estimate consumer responses to food price changes, accounting for food substitutions (own- and cross-price elasticities of demand) using the 2022-23 Household Consumption Expenditure survey data and an Almost Ideal Demand System (AIDS) model. In this utility-based structural model, goods are heterogeneous, and households choose the quantity and quality of a given good as a function of its price, the price of other goods, household income, and household sociodemographic characteristics. Following Deaton (1988), we adjust for quality shading and measurement error stemming from the use of unit values as a proxy for unobservable prices (i.e., expenditure divided by quantity) [21]. This model is extensively used to estimate the demand for food in low- and middle-income countries [22]. It assumes spatially varying prices, where all households within a near geographical area, here a cluster, defined by NSSO as a village or urban block, face the same price, regardless of their income group. Any within-cluster variation in unit values is due to differences in the quality of the purchased items. Spatial variations in unit values between clusters provide an identification strategy to avoid the endogeneity of prices and address quality shading.
This paper uses uncompensated price elasticities, which capture the effect of a change in the price of a good on demand, holding income and other prices constant. These elasticities reflect both the substitution effect and the income effect of the price change. Details of the microeconomic foundations of Deaton’s (1988) AIDS model, as well as its derivation and the steps involved in its estimation, have been described elsewhere [23]. More information is provided in Appendix C in S3 Supporting Information.
We estimate the model parameters for eleven groups: cereals, dairy, pulses, edible oils and spices (including salt and raw sugar), fruits and vegetables (including nuts, hereafter referred to as F&V), animal meats (fresh), packaged processed foods, sweets, SSBs, non-SSBs, and a numeraire group containing food-away-from-home and non-food expenditure. Adding the latter allows reallocation between food and non-food when prices or real expenditure change. We normalise the numeraire group’s unit value to one. After dropping clusters that do not have at least two households consuming at least one item for each group [24], we are left with a total of 5,432 clusters and 212,797 households. We fit this model for three income groups, defined by terciles of monthly per capita consumption expenditure.
Fiscal policy scenarios
The estimated income-group-specific own- and cross-price elasticities are used to simulate the impacts of two main scenarios consisting of increasing GST to the highest rate of 28% on foods and beverages classified as HFSS based on their nutrient composition, namely the proposed HFSS food definition from the Food Safety and Standards Authority of India (FSSAI) Labelling & Display Amendment Draft Regulations 2022 [25] (scenario 1) as well as an additional 12% (equivalent in effect to a 40% tax rate) (scenario 2). This latter scenario aligns with the “sin” goods GST rate (40%), which is applied to caffeinated and sweetened aerated beverages (including SSBs) and some tobacco products.
The proposed FSSAI definition applies uniform sodium, sugar, and saturated fat thresholds across all processed food and beverage products. We assume processed products to be items not included in Schedule IV, Category III Solid Foods/Liquid Foods, which are exempt from front-of-pack nutrition labelling under the Indian Nutrition Rating, as per the FSSAI draft regulation [25]. The impact of using an alternative NPM to define HFSS foods, namely the one developed for the South East Asian Region by the World Health Organisation (WHO SEARO) [26], based on category-specific nutrient thresholds for energy, sodium, sugar, saturated fat, and total fat content (Table A2 in S1 Supporting Information), is tested in further simulations using the same tax rates as in scenarios 1 and 2 (scenarios 3 and 4, respectively). Figure A1 and Figure A2 in S1 Supporting Information present the baseline distribution of total energy intake by GST rate and food group and by GST rate and HFSS status (based on the FSSAI definition), respectively. They show that most HFSS items are currently taxed at 12% or 18%.
The immediate impact of each policy scenario on individual energy and nutrient intake, total household expenditure on foods and beverages, and government tax revenue from foods and beverages is calculated based on the estimated price elasticities. Key assumptions, which are subsequently relaxed in sensitivity analyses (Table B1 in S2 Supporting Information), include a 100% tax passthrough to retail prices and no supply-side tax-induced product reformulation. Finally, we also test the impact of a 0% GST on F&V and pulses (equivalent to a subsidy). Results sensitivity analyses are presented in Appendix B in S2 Supporting Information.
Health-GPS microsimulation model
Long-term nutrient intake projections and policy impacts are estimated using Health-GPS, a dynamic policy microsimulation model developed by the Centre for Health Economics & Policy Innovation [27]. Health-GPS creates a synthetic population of individuals that broadly reproduces demographic and socioeconomic characteristics of the population of a given country, or sub-national jurisdiction. It simulates individual life histories from birth to death, including age, gender, socioeconomic characteristics, risk factors, and disease profiles. Health-GPS updates annually with statistical and probabilistic models, which are calibrated with real-world data to capture relationships between variables. Changes in disease risks are assumed to result solely from changes in demographic composition and risk factor distributions, with no underlying time trend in diseases.
A baseline scenario is simulated reflecting expected demographic and epidemiological changes over a time horizon in a given population. Intervention scenarios are then simulated, reflecting the introduction of policy interventions applied to targeted risk factors. Changes in risk factor exposures will, in turn, impact other risk factors, risk of diseases, and disease burden. Disease outcomes typically include incidence, prevalence, and mortality of each disease in the population. Disease burden is measured by disability-adjusted life years (DALYs), a sum of years of life lost (YLLs) and years lived with disability (YLDs). To calculate YLLs and YLDs, the model iterates over each individual in the simulated population, summing the YLLs due to premature mortality and the YLDs, weighted by the corresponding disability weights. Policy impacts are estimated as the difference in these outcomes between the baseline and intervention scenarios.
In this study, 0.5% of the Indian population is simulated at the individual level, repeated 20 times for each scenario, between 2022 and 2053, with policy introduced in 2024. The synthetic population is modelled to reflect demographic estimates and projections from the United Nations World Population Prospects [28], estimated distribution of individual energy and nutrient intakes, weight and height distributions by age and gender from NCD-RisC [29], as well as prevalence, incidence and excess mortality of diseases from the Institute for Health Metrics and Evaluation Global Burden of Disease [30]. Specifically, for each individual, we simulate two demographic characteristics: age, gender; two socio-economic characteristics: income tertile, sector (urban/rural); six risk factors including four nutrients (carbohydrate, protein, fat, and sodium), physical activity, and body mass index (BMI); disease status of five key diseases: ischemic heart disease (IHD), chronic kidney disease (CKD), stroke (including ischemic stroke, intracerebral haemorrhage and subarachnoid haemorrhage), diabetes, and asthma. A complete list of model parameters and variables definition and sources is presented in Table D1 in S4 Supporting Information.
We assume two main diet-related risk-factor-to-disease pathways in this study: the first reflects the impact of diet (and physical activity) on energy balance and thereafter BMI; the second accounts for the impact of dietary sodium. Diet composition-related benefits (e.g., increased consumption of whole grains, F&V, pulses) are not modelled. With data on two risk factors and relative risk ratios linking risk factors to diseases from the literature, changes in incidence and disease burden of five key diseases are estimated. In addition, changes in hypertension prevalence are estimated independently with projected risk factor data from Health-GPS.
We apply two alternative assumptions on the time trend of nutrient intake. First, we assume no time trends in age-, sex- and income-group-specific nutrient intakes, which means that socio-demographic shifts solely drive population-wide changes. Microsimulation outcomes under this assumption thus represent a lower bound of estimates and policy impacts (hereafter referred to as the ‘lower bound’). Our alternative assumption is an income trend which accounts for a constant 6% income growth based on IMF projections for real GDP growth [31]. It is scaled by the estimated income group- and food group-specific income elasticities (Table C1 and Figure C1 in S3 Supporting Information) and an exponential decay over time, based on Engel’s Law, which posits that, as household income increases, the proportion spent on food decreases [32]. The decay rate is nutrient-specific and estimated by regressing the logarithm of the predicted growth rate in individual nutrient intake on a time trend. The former is estimated based on a regression of individual nutrient intake on monthly per capita total expenditure and a set of household- and individual-level controls among the NSS Household Consumption Expenditure 2022-23 sample. Figure E1 in S5 Supporting Information cross-validates this approach by comparing the estimated yearly growth rate in the consumption of packaged processed foods and sweets - two food groups with a high proportion of HFSS foods (Table A6 in S1 Supporting Information) - with historical trends based on Euromonitor International Passport data and Tak et al (2022) for selected highly processed food items [3,33]. This alternative assumption gives an upper-bound estimate of policy impacts (the ‘upper bound’). More details on how we derive the parameters for the income trend assumption can be found in Appendix E in S5 Supporting Information.
Policy scenarios are applied at the beginning of 2024 (two years after the initial year) and their impact on health is estimated over 30 years until 2053. We report microsimulations outcomes including the reduction in major risk factors: sodium intake and BMI, cumulative reduction in disease incidence and DALYs, in the general population and by income group. Economic benefits are calculated as averted total health expenditures in US dollar 2024, including both government and household expenditure. Household health expenditures per disease per year are derived from the literature and converted to total health expenditure based on data from the literature and National Health Accounts Estimates for India 2021-22 [34]. Confidence intervals are constructed with results from 20 repeated simulation runs to capture stochastic variation.
A simplified model diagram describing the Health-GPS model can also be found in Figure D1 in S4 Supporting Information. More details on the methods related to its use are listed in Appendix D in S4 Supporting Information. Lastly, Figure 1 offers a simplified flow chart of the data sources and methodological steps taken in this analysis.
Results
Energy and nutrient intakes
Table 1 presents the estimated baseline average daily nutrient intake per capita. Tables A3-A5 in S1 Supporting Information show the results for low-, middle-, and high-income groups, respectively. The average daily energy intake is lower than previous NSS-based estimates, which did not account for waste [35]. Cereals represent approximately half of the total calorie intake. Total daily energy intake increases with income. While our estimated average daily sodium intake is slightly more conservative than subnational dietary recall estimates [36], we find that 80.7% of adult males, 73.0% of adult females, and 80.8% of adolescents have daily sodium intakes above the WHO-recommended level of 2,000mg/day. Three in ten adults also consume more saturated fat daily than the WHO-recommended levels (no more than 10% of total daily energy intake). For total sugars, 10.0% of adult males and 6.5% of adult females have a daily total sugar intake above the National Health Service of the United Kingdom (NHS) recommended levels (no more than 90g per day) (Figure A3 in S1 Supporting Information).
Price elasticities of demand
Table 2 shows the estimated average own- and cross-price elasticities of demand. Demand for most foods is inelastic (proportionate change in purchases smaller than proportionate change in price), except for dairy, SSBs, and non-SSBs. The demand for oils & spices is the least price-sensitive, with consumption decreasing by 3.7% for each 10% rise in oils & spices price (95% CI: 3.5%, 3.9% decrease). The demand for packaged processed foods, which mainly contain HFSS items (Table A6 in S1 Supporting Information), is close to unit-elastic (−0.92, 95% CI: −0.95, −0.89). Results indicate significant substitution and complementarity patterns between groups. F&V and fresh animal meats are complements (negative cross-price elasticities, p<0.01). Packaged processed foods and fresh animal meats are substitutes, as well as SSBs and non-SSBs. Figure 2 provides the estimated average own-price elasticity results by income group. Tables A7-A9 in S1 Supporting Information additionally display the estimated cross-price elasticities. In line with the international literature and previous studies in India [22,37], the demand from low-income households is more price-sensitive, except for dairy.
We provide sensitivity analyses estimating the model using a broader beverage category, i.e., combining SSBs and non-SSBs, to reduce the number of clusters dropped due to non-purchases (Table C2 in S3 Supporting Information). We also perform a sensitivity analysis in which we relax our assumption of unitary unit values for the numeraire group and instead assign the non-food consumer price index by state, quarter, and sector (i.e., rural or urban) (Table C3 in S3 Supporting Information) [38]. Elasticity estimates from these alternative specifications are consistent.
Immediate policy impacts
Depending on the NPM used, HFSS foods account for 9.9% (FSSAI) or 11.4% (WHO SEARO) of total baseline energy intakes by Indians (Table A6 in S1 Supporting Information). Most of the items in the sweets and packaged processed food groups are HFSS. The difference between the two approaches is primarily driven by the extensive list of items excluded from the proposed FOPNL by FSSAI [25]. Immediate policy impacts on nutrient intake are presented in Figure A4 in S1 Supporting Information. Figures B1, B4, and B7 in S2 Supporting Information show the results for the sensitivity scenarios.
All scenarios would lead to a minor increase in expenditure for all income groups, from 0.6% (95% CI: 0.1%,1.1%) for low-income households for scenario 1 to 1.3% (95% CI: 0.1%, 2.5%) for high-income households for scenario 2 (Figure A5 in S1 Supporting Information). These results are mainly driven by a significant decrease in HFSS food purchases – particularly, packaged processed foods (high absolute price elasticity of −0.92) – and substitutions from HFSS foods to dairy and fresh animal meats. Results are similar for sensitivity scenarios (passthrough, reformulation, and subsidy) (Figures B2, B5, and B8 in S2 Supporting Information).
All scenarios would significantly increase tax revenue from foods and beverages, from +51.3% for scenario 1 (95% CI: 49.8%, 52.8%) to +110.0% for scenario 4 (95% CI: 105.3%, 114.7%) (Figure A6 in S1 Supporting Information). Results are similar for sensitivity passthrough and reformulation scenarios (Figures B3 and B6 in S2 Supporting Information). Tax revenue is predictably lower for the sensitivity scenario with GST 0% on F&V and pulses (Figure B9 in S2 Supporting Information). Behaviourally, these results are driven by a quantity (substitution) effect, with consumers shifting away from HFSS items toward dairy and fresh animal meats, and a tax increase effect, as most HFSS items belong to the packaged processed food and sweets groups, whose demand is price inelastic. The latter implies that the percentage fall in quantity demanded is smaller than the tax-induced percentage increase in price, thus increasing tax revenue. Regarding the magnitude of these results, most HFSS items are taxed at 12% or 18% at baseline; therefore, our scenarios imply a 10- to 28-percentage-point increase, which represents a significant 43% to 122% relative rise in rates for these items. The tax base is also concentrated in these tax brackets: HFSS items account for 85.3%, 93.1%, and 100% of baseline energy intake taxed at 12%, 18%, and 28%, respectively, but only 1.6% and 1.1% at the 0% and 5% rates, respectively. Most non-HFSS items are zero-rated or taxed at 5%. (Figure A2 in S1 Supporting Information).
Long-term policy impacts
Figure 3 presents the upper bounds of reduction in sodium intake and BMI under scenarios 1 and 2 compared to no policy change over time, by income group (lower bound results available in Figure A7 in S1 Supporting Information). The reduction in daily sodium intake is persistent and stable over the 30 years, while the decrease in BMI tends to be stable after 4 years of policy implementation. Scenario 2 has greater impact on sodium intake and BMI than scenario 1. Under scenario 2, the population average reduction in daily sodium intake is approximately 45.8mg (95% CI: −45.9, −45.7) (−1.62% of baseline sodium intake; 95% CI: −1.623%, −1.615%), ranging from 31.8mg (95% CI: −32.1, −31.6) in the low-income group (−1.24%; 95% CI: −1.25%, −1.23%) to 66.3mg (95% CI: −66.7, −66.0) in the high-income group (−2.07%; 95% CI: −2.08%, −2.06%). After 30 years of policy implementation, the average reduction in BMI of the entire population is 0.1705kg/m2 (95% CI: −0.1709, −0.1700) (−0.796% of baseline BMI; 95% CI: −0.798%, −0.793%) under scenario 2, ranging from 0.137kg/m2 (95% CI: −0.138, −0.136) in the low-income group (−0.663%; 95% CI: −0.668%, −0.659%) to 0.219 kg/m2 (95% CI: −0.220, −0.218) in the high-income group (−0.972%; 95% CI: −0.978%, −0.967%). Sodium and BMI reductions are consistently larger for the high-income group compared to the lower-income group. Results under scenarios 3 and 4 are reported in Figures A8 and A9 in S1 Supporting Information.
A reduction in sodium intake and a decrease in BMI reduce the risk of five major diseases: IHD, CKD, stroke, diabetes, and asthma. Over 30 years, the cumulative reduction in the incidence of the five diseases in the entire population is 5.44 million (95% CI: −5.71, −5.17) cases and 9.23 million (95% CI: −9.54, −8.91) cases (equivalent to 0.18 million (95% CI: −0.17, −0.19) and 0.31 million (95% CI: −0.32, −0.30) cases per year on average) under scenarios 1 and 2, respectively (Table A10 in S1 Supporting Information, corresponding lower bound estimates in Table A11 in S1 Supporting Information). This corresponds to an average yearly reduction of 1.01% (95% CI: −1.06%, −0.96%) and 1.72% (95% CI: −1.78%, −1.66%) in the baseline incidence number of the five diseases in India (2019). As reported in Table 3, the high-income group has the largest cumulative reduction in incidence rate of the five diseases, with 915 cases per 100,000 population (95% CI: −972, −858) over 30 years under scenario 2. Among diseases, the largest reductions are observed in IHD and diabetes, which made up 7.8% and 2.5% of DALYs in India in 2019, respectively [30]. Corresponding lower bound estimates and results under scenarios 3 and 4 are reported in Tables A12-A14 in S1 Supporting Information. In addition, a lower prevalence of hypertension is observed under all scenarios. In the 30th year after policy implementation, the number of hypertension cases in India decrease by 7.07 million (95% CI: −7.66, −6.48) and 11.87 million (95% CI: −12.46, −11.28) under scenarios 1 and 2, respectively (Figure A10 in S1 Supporting Information, −3.10% (95% CI: −3.36%, −2.84%) and −5.20% (95% CI: −5.46%, −4.94%) compared to the national prevalence number in India from National Family Health Survey, 2019-21; lower bound estimates in Figure A11 in S1 Supporting Information).
As a result, associated DALYs of IHD, CKD, stroke, diabetes, and asthma are reduced. Over 30 years, in the entire population, the cumulative reduction in DALYs is 10.36 million years (95% CI: −12.47, −8.25) and 18.98 million years (95% CI: −21.37, −16.58) under scenarios 1 and 2, respectively (Table A10 in S1 Supporting Information, equivalent to 0.35 million years (95% CI: −0.42, −0.27) and 0.63 million years (95% CI: −0.71, −0.55) per year on average). The high-income group has the largest cumulative reduction in DALYs rate, up to 2,059 DALYs per 100,000 population (95% CI: −2553, −1564) under scenario 2 (Figure 4). Meanwhile, the policy scenarios are associated with a decrease in total health expenditure of US$10.7 billion (95% CI: −11.3, −10.1) and US$18.0 billion (95% CI: −18.7, −17.3) over 30 years under scenarios 1 and 2, respectively (Table 4, equivalent to US$357 million (95% CI: −376, −337) and US$601 million (95% CI: −624, −578) per year on average). Lower bound estimates and results under scenarios 3 and 4 are reported in Tables A15-A17 in S1 Supporting Information, as well as in Figures A12-A14 in S1 Supporting Information.
Discussion
Applying a GST rate of 40% to HFSS items, based on the draft FSSAI HFSS definition, could reduce annual average disease incidence by up to 1.72% (95% CI: −1.78%, −1.66%). Over 30 years, it could prevent as many as 18.98 million (95% CI: −21.37, −16.58) DALYs and avert up to US$18.0 billion (95% CI: −18.7, −17.3) in total health expenditure on IHD, CKD, stroke, diabetes and asthma (upper bound estimates). This represents an average reduction of approximately 0.6% of total health expenditure, equivalent to US$0.44 per capita and 0.02% of GDP in India per year. A 40% GST rate on HFSS items is also associated with an immediate increase in tax revenue from foods and beverages of 92.0% (95% CI: 88.2%, 95.7%) and a minor increase in total household expenditure on foods and beverages for all income groups (ranging from +0.9% (95% CI: 0.0%, 1.8%) for low-income to +1.3% for high-income (95% CI: 0.1%, 2.5%)).
The largest absolute health gains accrue to higher-income individuals because their baseline intake of HFSS calories and sodium is greater (Tables A3, A4, and A5 in S1 Supporting Information). Nevertheless, health benefits remain significant for low-income individuals, who face higher odds of incidence and intensity of catastrophic out-of-pocket health expenditure [39]. The modelled reductions in morbidity and service use lead to a decline in total health expenditure and thus out-of-pocket expenditure - representing over 40% of total health spending in the country [34]-, which could offset the rise in food and beverage expenditure, particularly for the poorest, eliminating or mitigating the net financial burden.
Slightly higher health and economic benefits are obtained under a scenario using the WHO SEARO NPM to define HFSS foods (scenarios 3 and 4, Appendix A in S1 Supporting Information), as fewer items are exempted (Table A2 and Table A6 in S1 Supporting Information). Under both the draft FSSAI and the WHO SEARO NPM HFSS food definition, significantly stronger health and economic benefits are associated with a 40% GST rate (scenarios 2 and 4, i.e., including an additional 12% tax on HFSS food) than with a 28% rate. The difference between lower- and upper-bound estimates is small, suggesting the results are robust to a plausible income trend (Appendix A in S1 Supporting Information). Our results are robust to potential strategic industry tax avoidance, namely tax under-shifting to prices and reformulation (Appendix B in S2 Supporting Information).
This study is based on the latest available nationally representative household consumption survey data, with information for more than 250,000 households. By linking this data with nutritional information, it provides the most up-to-date energy and nutrient intake estimates across the Indian population. Additionally, it estimates the income and price elasticity of demand for ten food groups and three income strata, updating previous estimates based on earlier waves of this microdata from more than a decade ago. Up-to-date intake levels and elasticities are foundational inputs for projecting diet and nutrition outcomes and for simulating policy or market shocks (e.g., taxes, subsidies, income growth, inflation). This analysis utilises a novel microsimulation model, Health-GPS, to project health impacts. It simulates a large synthetic population of over 7 million individuals, representing 0.5% of the Indian population. The individual-level simulation allows us to capture heterogeneity in risk factor exposures and disease status. Health-GPS accounts for comorbidity, reflecting real-world disease patterns, and encompasses a wide range of diseases. Mortality from other causes is captured through residual mortality in Health-GPS.
Our analysis is not exempt from limitations. Our approach is a first-order attempt to individualise energy and nutrient intake from household-level data based on members’ age and sex and dietary guidelines. However, intra-household allocation of food and nutrients in India may be influenced by age and gender biases [40]. As our demand parameters are estimated at the household level, our analysis also assumes uniform relative consumption responses to the fiscal scenarios across household members, regardless of age or sex. Inherent to demand estimations based on cross-sectional data and rather aggregated food groups [9,37], we do not account for cross-price effects within such groups. Further disaggregation comes with a trade-off. It reduces the number of non-zero consumption observations per group, which undermines the reliability and identification of demand elasticities. Future research using product-level consumer panel or scanner data in India could overcome this issue [11]. Consistent with standard practice, we assume elasticity estimates are constant over time [13,14]. However, responsiveness can drift with income growth, urbanisation, demographics, and the food environment, such that holding elasticities fixed may bias long-run impacts [41].
Our microsimulation assumes no underlying time trend of diseases, which may be inconsistent with the observed increasing risks of non-communicable diseases associated with, for example, air pollution and climate change [42]. As a result, our absolute health and cost impacts estimates should be interpreted as conservative, because any unmodelled upward drift in baseline risk would scale up the number of cases and DALYs averted by a given proportional risk reduction. The analysis accounts for two pathways from diet to health, or two dietary risks: high sodium intake and high BMI (determined by energy imbalance). It does not consider the positive health impact of improving the source of energy (e.g., from F&V or pulses rather than HFSS foods) or increasing whole grain, F&V or legume intake [43]. Because these composition effects are not captured, our health impact estimates are likely conservative, particularly for policies that promote these foods (e.g., subsidy scenario zero-rating F&V and pulses) (Figure B7 in S2 Supporting Information). A fuller treatment would require multi-risk modelling with joint intake distributions and covariance to avoid double-counting, which is not currently feasible with available data.
For ease of computability, we only account for the five diseases most attributable to excess sodium intake and high BMI in terms of DALYs in India. Our health benefit estimates therefore represent an underestimate of the potential effect. The same epidemiological data on disease incidence, prevalence, mortality, and remission are applied across income groups due to the dearth of granular data. Thus, differences in estimated absolute health benefits across income groups need to be interpreted with caution, as non-communicable diseases are not necessarily diseases of affluence in India [44]. In addition, the exposure-response functions (relative risks) used in this study are derived from global meta-analyses (e.g., Institute for Health Metrics and Evaluation Global Burden of Disease), whose underlying cohorts are disproportionately from high-income countries and typically assume generalizability across different settings. Restricted by data availability, total health expenditure is extrapolated from household expenditure on health, assuming an equal distribution of health expenditure between households and the government across diseases, following the methods adopted by WHO [45]. We did not fully propagate parameter and structural uncertainty through the model. Uncertainty was addressed by bootstrapping standard errors in our demand estimation, providing confidence intervals from 20 repeated simulation runs to capture stochastic variation, and reporting sensitivity scenario analyses. This is due to data limitations regarding the variance of risk ratio estimates and the underlying distribution of dietary risk factors. As such, our uncertainty intervals are likely conservative.
Policy implications
The 2024 Economic Survey of India and the WHO take cognisance of the increasing consumption of HFSS foods in India and outline the need to promote healthy diets [46,47]. The introduction of the 28% GST and 12% Cess on caffeinated and sweetened aerated beverages (including SSBs) in 2017, which represent only a fraction of HFSS food intake, has shown limited impacts in reducing consumption [48]. However, the 2016-2017 Kerala state-wise indirect tax of 14.5% on burgers, pizzas, tacos, doughnuts, sandwiches, pasta and bread fillings sold in restaurants (repealed following the launch of the nationwide GST by the federal government to replace all state-level indirect taxes) was associated with reduced sales [49].
Adjusting taxes within the existing GST rate structure helps avoid multiple taxation schemes. This is particularly relevant given the latest policy debates around simplifying the Indian GST rate structure to reduce financial and administrative burdens and increase revenue collection efficiency. The most recent government proposal, enacted in September 2025, includes reducing the number of rates from five to three, removing the 12% and 28% GST rates. The 12% compensatory Cess is discontinued and a 40% GST rate on “sin” goods is introduced (equivalent to the previous 28% GST + 12% Cess), still covering only caffeinated and sweetened aerated beverages and some tobacco products [50]. Curbing HFSS food consumption may require a further expansion of the “sin” goods rate to all HFSS foods.
The tax administration challenges and companies’ compliance costs associated with taxing foods based on their nutritional content should be assessed, particularly in a setting with a significant informal market. While there is no robust independent evidence for negative macroeconomic impacts of diet-related fiscal policies [51], concerns may arise regarding the economic impact of the measure on the food industry and inflation. These potential challenges should be further studied.
Increasing GST rates on HFSS foods based on nutrient content may lead to significant dietary improvements and positive health effects, reducing health expenditure and increasing government tax revenue, with only a minor impact on household spending. The latter is important since it is estimated that at least two-thirds of the rural poor could not afford a recommended healthy diet in 2011 [52]. Our nutrient intake estimates based on the latest available microdata, robust demand elasticity estimates, and simulations of various GST scenarios contribute to informing future GST reforms to improve nutrition among the Indian population.
As an increasing number of countries contemplate taxing HFSS foods [6], the results of this study may also contribute to the broader global discussion on the use of this policy. Indeed, many countries share a similar differentiated value-added or sales tax system applied to foods and beverages, as seen in India, with an opportunity to align existing rates with nutritional objectives without adding new tiers. This approach has, for example, been supported by the European Parliament in its Farm to Fork strategy [53].
Conclusions
India faces a rapid shift in diets towards HFSS foods. Tackling unhealthy diets is urgent. Other countries have highlighted the feasibility of regulating the marketing of unhealthy foods and beverages, the introduction of FOPNL, and shifts in relative prices to encourage significant changes towards healthier food purchases. Clear quantitative nutrient thresholds defining HFSS foods are crucial for the consistent development and implementation of these policies.
India introduced a higher 28% GST rate and a 12% Cess on caffeinated and sweetened aerated beverages, including SSBs, in July 2017 (equivalent to the “sin” goods 40% GST rate which will be applied from September 2025). Broader alignment of GST rates with nutrition and health objectives, targeting HFSS foods with higher rates, may represent the next critical step. This study has shown that it could result in meaningful reductions in the intake of nutrients of concern and significant long-run health benefits, yielding both lower health expenditure and higher tax revenue from foods and beverages while having a limited impact on household expenditure.
Data Availability
Input data and statistical code as well as further information on the Health-GPS microsimulation model is available at https://imperialchepi.github.io/healthgps/. Data sources used in this analysis are available publicly: NSS Household Consumption Expenditure survey (2022-23) microdata at https://microdata.gov.in/NADA/index.php/catalog/224; IHME relative risks, disease prevalence, incidence and mortality data at https://ghdx.healthdata.org/ihme_data; UN World Population Prospects database at https://population.un.org/wpp/; and NCD-RisC data at https://www.ncdrisc.org/data-downloads.html.
Acknowledgments
We thank Laura Cobb and the participants at two stakeholder meetings organized by the Institute for Economic Growth (IEG) in New Delhi, India, at the ‘Delivering for Nutrition in South Asia: Equity and Inclusion’ conference organized by the International Food Policy Research Institute (IFPRI) in Kathmandu, Nepal, and at the Indian Health Economics and Policy Association annual conference for their comments. We thank Belabess Alijadallah, Mahima Ghosh, and James Turner as well as his team members for their support with the calibration of the Health-GPS model. We thank Preeti Khanna and Abhishek Kumar for their assistance in matching food items with nutritional information and with GST rates, respectively. Finally, we thank Heather Lodge for administrative support.
Competing interests statement
The authors have declared that no competing interests exist.
Funding Disclosure statement
This work was supported by Resolve To Save Lives (RTSL, United States, contract PR0414 to FS) as part of their Nutrition and Surveillance programme. Authors MR, MS, and LS received salary support from RTSL during the conduct of this work. No other authors received salary support from the funder. The funders contributed to the study conceptualization, study design, and writing of the report. They had no role in data collection, data analysis, or data interpretation. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.
Data Sharing statement
Input data and statistical code as well as further information on the Health-GPS microsimulation model is available at https://imperialchepi.github.io/healthgps/. Data sources used in this analysis are available publicly: NSS Household Consumption Expenditure survey (2022-23) microdata at https://microdata.gov.in/NADA/index.php/catalog/224; IHME relative risks, disease prevalence, incidence and mortality data at https://ghdx.healthdata.org/ihme_data; UN World Population Prospects database at https://population.un.org/wpp/; and NCD-RisC data at https://www.ncdrisc.org/data-downloads.html.
Supplementary information files
Supporting information file 1
Appendix A. Additional tables and figures
Appendix A forms part of the revised submission.
Supplement to: Roche M, Zhu J, Olney J, Laydon DJ, Joe W, Sharma M, Steele L, Sassi F. Taxation of foods high in saturated fat, sugar, and sodium in India: A modelling study of health and economic impacts. Submitted after revision on 19 September 2025.
References
- 1Ministry of Statistics and Programme Implementation. Household Consumption Expenditure Survey 2022–23 [Internet]. New Delhi: National Statistical Office, Government of India; 2024 [cited 2025 April 20]. Available from: https://microdata.gov.in/NADA/index.php/catalog/237
- 2Food Safety and Standards Authority of India. Labelling & Display Amendment Draft Regulations, 2022 (44272/2022/REGULATION-FSSAI) [Internet]. New Delhi: FSSAI; 2022 [cited 2023 Jun 25]. Available from: https://fssai.gov.in/upload/uploadfiles/files/Draft_Notification_HFSS_20_09_2022.pdf
- 3World Health Organization, Regional Office for South-East Asia. WHO nutrient profile model for the South-East Asia Region [Internet]. New Delhi: WHO; 2017 [cited 2023 Sep 6]. Available from: https://www.who.int/publications/i/item/9789290225447
- 4Deaton A. Quality, quantity, and spatial variation of price. The American Economic Review. 1988 Jun 1:418–30.
- 5Institute for Health Metrics and Evaluation. Global Burden of Disease [Internet]. Seattle (WA): IHME, University of Washington; [cited 2023 Nov 20]. Available from: https://www.healthdata.org/research-analysis/gbd
Supporting information file 2
Appendix B. Sensitivity scenarios
Appendix B forms part of the revised submission.
Supplement to: Roche M, Zhu J, Olney J, Laydon DJ, Joe W, Sharma M, Steele L, Sassi F. Taxation of foods high in saturated fat, sugar, and sodium in India: A modelling study of health and economic impacts. Submitted after revision on 19 September 2025.
References
- 1Andreyeva T, Marple K, Marinello S, Moore TE, Powell LM. Outcomes following taxation of sugar-sweetened beverages: a systematic review and meta-analysis. JAMA Network Open. 2022 Jun 1;5(6):e2215276–.
- 2Food Safety and Standards Authority of India. Labelling & Display Amendment Draft Regulations, 2022 (44272/2022/REGULATION-FSSAI) [Internet]. New Delhi: FSSAI; 2022 [cited 2023 Jun 25]. Available from: https://fssai.gov.in/upload/uploadfiles/files/Draft_Notification_HFSS_20_09_2022.pdf
- 3World Health Organization, Regional Office for South-East Asia. WHO nutrient profile model for the South-East Asia Region [Internet]. New Delhi: WHO; 2017 [cited 2023 Sep 6]. Available from: https://www.who.int/publications/i/item/9789290225447
Supporting information file 3
Appendix C. Almost Ideal Demand System (AIDS) demand model
Appendix C forms part of the revised submission.
Supplement to: Roche M, Zhu J, Olney J, Laydon DJ, Joe W, Sharma M, Steele L, Sassi F. Taxation of foods high in saturated fat, sugar, and sodium in India: A modelling study of health and economic impacts. Submitted after revision on 19 September 2025.
C1.The model
As with most other household budget surveys, the NSSO Household Consumption Expenditure survey 2023-24 does not provide price information, thus we rely on unit values (i.e., expenditure divided by quantity). However, unit values are subject to measurement errors in both expenditure and quantity and may lead to quality shading, where consumers may not only adjust quantity but also the quality of the goods they buy as a response to price changes. We thus estimate consumer responses to food price changes (technically, own- and cross-price elasticities of demand) using Deaton (1988)’s Almost Ideal Demand System (AIDS) model adjusting for quality shading and measurement error [1]. where wjic represents the share of total expenditure on food and beverages (xic) that household i in cluster c spends on good j; pkc is the price of good k which does not vary in cluster c; fjc are unobservable cluster fixed effects; vjic is the unit value of good j; and ujic are idiosyncratic error terms. zic is a vector of household characteristics, including the logarithm of household size, the sector of the household (urban vs. rural), the sex of the head of household and a dummy for completion of primary education, and dummies capturing the religion of the household. We remove regional and seasonal effects using dummies for the six Indian administrative zones and the quarter of the year in which the household reports information.
The specification nests rich income (Engel-curve) responses and flexible cross-price substitution patterns while remaining consistent with standard consumer theory. Adding-up, homogeneity, and Slutsky symmetry are imposed so the estimates are integrable and interpretable as arising from well-behaved preferences in the price-independent generalized logarithmic preferences class (PIGLOG) [2].
The model assumes spatially varying prices, where all households within a near geographical area face the same price. As defined by NSSO, a cluster is equivalent to a village, or an urban block, surveyed during the same period. Any within-cluster variation in unit values is due to differences in the quality of the purchased items.
Spatial variations in unit values between clusters provide an identification strategy to avoid the endogeneity of prices and address quality shading [1]. We argue this assumption is justified in a middle-income country like India where transport is more difficult and relatively costlier, and markets are not always well-integrated [2]. To abide by our main assumption of spatially varying prices, we force households to face the same price if they are drawn from the same cluster, regardless of their income group. This takes care of a potential endogeneity in average prices that might arise if households from different income groups face different average prices despite being in the same cluster [3]. We drop clusters not having at least two households consuming at least an item for each group. Following the demand system literature [4], we normalize the numeraire group’s unit value to one.
We recover uncompensated price elasticity of demand estimates at the means through a three-step method including demeaned within-cluster regressions, between-cluster regressions of adjusted budget share on adjusted unit values estimated in the first step using an error-in-variable estimator, and finally separating quality and price effects. where δjk is the Kronecker delta equal to one if j = k (i.e., for own-price elasticities) or zero otherwise.
This approach has been used extensively to model the demand for food in low- and middle-income countries [5]. A key strength is its rigorous treatment of measurement noise. By residualizing within clusters, averaging to the cluster, and then applying measurement-error corrections that scale with buyer and household counts, it strips out spurious variation that would otherwise attenuate price effects while also tempering quality shading.
Nevertheless, the use of this model with NSSO Household Consumption Expenditure survey data has some limitations. First, censoring is only mitigated, not solved. Unit-value regressions exclude non-buyers while budget share regressions include everyone. Clusters with very few buyers are dropped to avoid unstable means and the measurement-error corrections down-weight clusters where buyer counts are small. These steps reduce the distortion created by many zeros, but they cannot remove selection bias if the decision not to buy is correlated with unobserved preferences, access, or local conditions that also relate to prices. Second, using household budget survey data, the underlying demand model cannot account for cross-price effects within food groups. Future research using detailed consumer panel data (e.g., home-scan) could estimate demand systems using more disaggregated groups, for example, by bundling items based on nutritional impact. Third, the NSSO also lacks information on household disposable income, which is commonly used as an instrumental variable to address the endogeneity in total expenditure arising from simultaneity bias [6].
Details regarding the microeconomic foundations of Deaton (1988)’s AIDS model, as well as its derivation and the steps involved in its estimation, have been described elsewhere [2]. The paper uses uncompensated price elasticities as opposed to compensated price elasticities, as uncompensated elasticities measure the effect of a price change for one good while keeping income and the prices of other goods constant, thereby reflecting both income and substitution effects.
C.2.Food groups
We estimate the model parameters for eleven groups: cereals, dairy, pulses, edible oils and spices (including salt and raw sugar), fruits and vegetables (including nuts, hereafter referred to as F&V), animal meats (fresh), packaged processed foods, sweets, SSBs, non-SSBs, and a numeraire group containing food-away-from-home and non-food expenditure. While we account for served processed foods (on-trade sector, i.e., food-away-from-home, like bars, restaurants, etc.) in the estimation of total nutrient intake, we do not include this aggregate as a standalone food group in our demand system. Instead, we group it with non-food expenditure within the numeraire.
The GST is applied differently to the on-trade sector in India, based on industry categorisation or types of premises rather than the specific type of food or beverage items served. Therefore, the GST scenarios simulated in this study only apply to the off-trade sector. In addition, food-away-from-home is not well captured in the NSS Household Consumption Expenditure survey, including only a limited number of broadly defined items (cooked meals purchased, cooked meals received free in the workplace, cooked meals received as assistance, cooked snacks purchased, other served processed food) [7]. While the simulated GST scenarios may impact the price of raw materials, it is not possible, given the available data, to estimate the extent to which this may affect the price of served processed foods.
References
- 1Deaton A. Quality, quantity, and spatial variation of price. The American Economic Review. 1988 Jun 1:418–30.
- 2Deaton A. The Analysis of Household Surveys: A Microeconometric Approach to Development Policy [Internet]. Washington, D.C.: World Bank; 2019 [cited 2023 Dec 19]. Available from: http://hdl.handle.net/10986/30394
- 3John RM, Tullu FT, Gupta R. Price elasticity and affordability of aerated or sugar-sweetened beverages in India: implications for taxation. BMC Public Health. 2022 Jul 17;22(1):1372.
- 4Richards TJ, Padilla L. Promotion and fast food demand. American Journal of Agricultural Economics. 2009 Feb;91(1):168–83.
- 5Green R, Cornelsen L, Dangour AD, Turner R, Shankar B, Mazzocchi M, Smith RD. The effect of rising food prices on food consumption: systematic review with meta-regression. BMJ. 2013 Jun 17;346.
- 6Banks J, Blundell R, Lewbel A. Quadratic Engel curves and consumer demand. Review of Economics and statistics. 1997 Nov 1;79(4):527–39.
- 7Ministry of Statistics and Programme Implementation. Household Consumption Expenditure Survey 2022–23 [Internet]. New Delhi: National Statistical Office, Government of India; 2024 [cited 2025 April 20]. Available from: https://microdata.gov.in/NADA/index.php/catalog/237
Supporting information file 4
Appendix D. Health-GPS microsimulation model
Appendix D forms part of the revised submission.
Supplement to: Roche M, Zhu J, Olney J, Laydon DJ, Joe W, Sharma M, Steele L, Sassi F. Taxation of foods high in saturated fat, sugar, and sodium in India: A modelling study of health and economic impacts. Submitted after revision on 19 September 2025.
D1.Overview
Health-GPS is a microsimulation model that simulates individuals over time, estimating their risks of developing a variety of NCDs (e.g., stroke) given their exposure to particular risk factors (e.g., excessive sodium intake). Health-GPS then models possible health policies (e.g., changes in taxation on foods high in sodium), that might affect these risk factors, and therefore may impact the incidence and outcomes of NCDs at the population level.
The model creates a synthetic population of individuals that broadly reproduces demographic and socioeconomic characteristics of the population of a given country, or sub-national jurisdiction, and simulates individual life histories from birth to death. Because it simulates a population at the individual level, Health-GPS can capture heterogeneity in risk exposures, diseases, and multi-morbidity patterns. Correlations between risk factor exposures and between diseases are reflected in simulated populations. Health-GPS is calibrated to real-world data to capture relationships between variables while matching to estimated and projected demographic and epidemiological metrics from international databases.
In Health-GPS, a baseline scenario is simulated reflecting expected demographic and epidemiological changes over a time horizon in a given population, to match current demographic projections over that same time horizon. Intervention scenarios are then simulated, reflecting the introduction of health policies applied to targeted individuals and risk factors. Changes in risk exposures will, in turn, impact other risk factors, diseases, and mortality through the model equations. Outcomes simulated through the model typically include, but are not limited to, the prevalence of risk factors, the incidence of, and mortality from, NCDs, years of life lost (YLLs), years lived with disability (YLDs), disability-adjusted life years (DALYs), and life expectancy. Policy impacts are estimated as the difference in these outcomes between the baseline and intervention scenarios. Given the random variation in model outputs due to Health-GPS’ stochastic nature, the model repeats these steps multiple times and averages the differences to obtain the effect of a given health policy. Another way of limiting the impact of stochastic variation on simulation results is to model large population samples.
This appendix is divided into the following sections describing four key Health-GPS component modules, illustrated in Figure D1:
- 1. Demographic module: births, deaths, immigration, population and socioeconomic status.
- 2. Risk factor modules: including energy balance and high dietary sodium intake in the version used for the study described in this paper.
- 3. Disease modules: we explicitly model the following diseases: asthma; chronic kidney disease; diabetes; ischemic heart disease; and stroke (including intracerebral haemorrhage, ischemic stroke, and subarachnoid haemorrhage). Other diseases are accounted for in a residual mortality aggregate.
- 4. Disease burden module: including metrics used to estimate population-level outcomes of interventions, specifically YLLs, YLDs, and DALYs.
In the last section, we additionally include sources and methods used to estimate the health expenditure impact of interventions.
D2.Demographic module
The demographic module manages the yearly numbers of births, deaths, and migrations, as well as socioeconomic status. This module aims to produce a baseline scenario consistent with projected demographic data.
D2.1.Data
Health-GPS uses the United Nations (UN)’s World Population Prospects projections for yearly births, deaths, total population, and net migration, for all countries, by age and sex [1]. In the next sections, we explain how these data are calibrated in the demographic module.
D2.2.Births and deaths
The demographic module reproduces historical and projected births and deaths in a baseline scenario. The UN estimates and projects birth and death rates for all countries from 1950 until 2100. These estimates and projections are available in five-year intervals, and the module interpolates the data linearly to account for missing years.
These data determine the number of births at the beginning of each year of the simulation, assigning sex based on a sex ratio. Newborns are disease-free at the beginning of the simulation, and their initial risk factor exposures are generated using the initialisation equations of the risk factor modules. For all individuals in Health-GPS, risk and disease profiles are updated yearly until they either die or migrate.
Similarly, UN data provide estimates of the number of deaths in each year of the simulation. Health-GPS calculates an individual’s probability of death each year based on their sex, age and disease profile. The model simulates multiple diseases but allows for death from other non-modelled causes, i.e. residual mortality.
D2.3.Residual mortality
Let the ith person (who has neither died nor emigrated) have age ai and sex gi. Let mdag(t) be the annual ‘excess mortality’ associated with disease d, for age a, and sex g, during calendar year t. We use the Institute for Health Metrics and Evaluation (IHME) definition of excess mortality [2], that is disease-specific mortality divided by disease prevalence. Then let denote
where
. Then the probability si(t) that the ith person survives diseases 1, …, ND in year t is given by
Then the number of people Nsag(t) of age a, and sex g who survive diseases 1, …, ND in year t is given by and so the proportion of people Sag(t) of age a, and sex g who survive diseases 1, …, ND in year t is given by where N is the number of people in the population, and Nag(t) denotes the number of people of age a and sex g at time t.
If Mag (t) and RMag(t) respectively denote the population-level mortality rate and residual mortality rate (of diseases not modelled in Health-GPS), for people of age a and sex g, then we have and so rearranging gives
This equation is used throughout the model. It is calculated in the baseline scenario, and the corresponding values are then used in the simulation of intervention scenarios. This assumes that mortality from non-modelled causes of death is constant across scenarios. However, diseases are not always assumed to be independent in Health-GPS, which means that variations in morbidity from modelled diseases that are produced by an intervention may impact non-modelled causes of death, and therefore residual mortality. Therefore, the assumption that residual mortality is constant across scenarios is an approximation and may lead to a (small) underestimation of the effects of an intervention when such an intervention produces a reduction in morbidity from modelled diseases.
At any given time t, each individual i is exposed to a specific risk of death in connection with the diseases they have, plus a residual risk in connection with causes of death that are not explicitly modelled in the simulation. Individuals who reach age 110 (i.e., the maximum attainable age in Health-GPS) are assumed to die at the next time step.
In the simulation, each year individuals die with probability .
D2.4.Population
The demographic module uses historical and projected population changes by year, sex, and age. UN population projections are available in 5-year age groups, and we linearly interpolate the population for each age group between two years. Similarly to deaths, population numbers are smoothed. This conserves the total annual population by sex.
D2.5.Net migration
While data on births and deaths are publicly available in several databases, finding accurate data on migration is more challenging. We model migration using the following formula: where TP(a, t), B(t), D(a, t) and I(a, t) respectively denote the total population, number of births, number of deaths, and net migration for age a in year t.Therefore, migration is simulated indirectly to replicate expected numbers from the UN World Population Prospects database [1]. Consequently, depending on the sign of net migration, individuals will be added or removed from the population. In particular, the attributes of migrants are bootstrapped (i.e., sampled with replacement) from the distributions of the existing individuals of the same sex and age.
D3.Risk factor modules
Health-GPS models risk factor exposures accounting for hierarchical dependencies. For instance, characteristics of the living environment (e.g., neighbourhood deprivation) may influence individual behaviours (e.g., dietary intake), and this in turn, may influence the likelihood of conditions (e.g., obesity, hypertension) typically linked with the incidence of NCDs. Diseases themselves can be risk factors for other diseases. Diabetes, for example, is associated with a higher incidence of cardiovascular disease or chronic kidney disease [3,4]. The associations between risk factors are complex and dynamic. They can change due to exogenous shocks such as changes in food prices, or the implementation of new interventions and policies. As individuals age in the simulation, their risk exposures are updated in line with the latest observed age- and sex-specific prevalences (those used to initialise the synthetic population at the start of the simulation). Such age- and sex-specific prevalences are typically held constant throughout the projection period unless specific underlying trends are assumed (in some of the analyses in this project, for instance, a future trend in ultra-processed food consumption has been assumed). Future trends over time in risk factor exposures may apply uniformly for the entire population, or differently for different individuals or population sub-groups.
In the version of the Health-GPS model used for the evaluation of food tax policy scenarios in India, the most upstream risk exposure is food prices, which are assumed to influence food purchases and eventually nutrient intakes. All of the above risk exposures are modelled at the household level outside Health-GPS, using a separate demand system model, described in the paper. Individual nutrient intakes are calculated based on household food purchases and are used as inputs into Health-GPS microsimulations. Nutrient intakes are assumed to generate health impacts through two main pathways: energy and sodium. These two pathways are described in the following sections, 3.1 and 3.2.
D3.1.Energy balance model
Health-GPS models body weight dynamics through an energy balance model (EBM) following the mathematical framework developed by Hall et al (2022) [5]. The purpose of this model is to understand and predict how various factors, such as diet and physical activity, influence an individual’s body weight over time. In this section, we provide an overview of how Health-GPS incorporates the Hall et al (2022) model [5]. Full details can be found in Hall et al (2022) [5] and all Health-GPS model codes (including the implementation of the Hall et al (2022) model) can be found at https://github.com/imperialCHEPI/healthgps.
The Hall et al (2022) model takes into account the principles of energy balance, which is the relationship between the calories consumed through food and beverages and the calories expended through metabolism and physical activity [5]. The model incorporates complex physiological processes and feedback mechanisms to provide a more realistic representation of weight dynamics than simpler, linear models.
In this framework, energy intake and expenditure are the result of diet, physical activity and metabolic patterns. Body weight changes in the microsimulation are associated with dynamic imbalances between energy intake (EI, the calories consumed from food and beverages), and energy expenditure (EE, the energy expended throughout the day). When an individual is in a steady state energy balance (i.e., “energy in” equals “energy out”), body weight is assumed to remain stable. A surplus of energy is stored as fat, leading to weight gain. Conversely, an energy deficit leads to weight loss.
Energy intake and expenditure are determined by nutrient intakes (fat, carbohydrates, protein), physical activity, and prior body mass index (BMI) as a proxy for metabolic energy requirements, which in turn are determined by individual characteristics including age, sex and socioeconomic status (Figure D2). The output of the energy balance model is a (time-dependent) change in BMI.
Each layer in Figure D2 influences subsequent layers. Risk factors are simulated in two steps: 1) initialisation, using static risk factor equations; 2) projection, which updates values for each risk factor at every time cycle during the simulation using dynamic risk factor equations.
D3.2.Sodium intake
Sodium intake is modelled alongside nutrient intake that contribute to the EBM. Sodium intake affects disease incidence directly, through relative risks derived from existing studies. Most of the sodium-attributable risk of disease is generated through an increase in blood pressure caused by excessive sodium intakes, however, the individual steps of this causal chain are not modelled in Health-GPS. The relative risks used to model increases in disease risk linked with sodium intake account for the increased blood pressure that mediates the effect of sodium intake on diseases. On the other hand, changes in hypertension prevalence are calculated in parallel as a function of sodium intake.
D3.3.Risk factor initialization and propagation
We use a two-stage approach to initialize risk factors within each person in Health-GPS that combines regression modelling, data transformation, and stochastic variation.
Nutrient initialisation
First, we use a regression model to determine average values of nutrient intake by age, sex, income (i.e., high, middle and low) and sector (i.e., rural or urban). Second, for each person and each nutrient, we add a stochastic variation term to these average values to generate the required population heterogeneity. This stochastic variation term is sampled from the residuals of the regression performed in the first step.
To ensure normality of the nutrient intake variables, a Box-Cox transformation is applied: where ýk is the transformed variable, yk is the dependent variable (nutrient intake) and λk is the transformation variable for nutrient k = 1, …, 4.
After transforming the variables, average daily nutrient intake ýk is initialized using the equation below for each of the following four nutrients: fat, protein and carbohydrates (measure in grams per day); and sodium (measured in milligrams per day). For nutrient k, we have where βk0 is the intercept and ɛk is the error term for nutrient k. Data input for socio-economic variables and nutrient intake are based on NSS 68th round 2011-2012 and adjusted to December 2022 using the methods detailed in the Methods section of the paper.
To capture realistic variation and maintain correlations between nutrients, stochastic residuals (ɛ) are generated using Cholesky decomposition.
Energy intake and energy expenditure initialisation
After daily nutrient intake values have been assigned for each person, they are inputted into the Hall et al (2022) model,5 together with each person’s Physical Activity Level (PAL), to determine each person’s energy intake (EI), and energy expenditure (EE). Physical Activity Level (PAL) is the ratio between energy expended through physical activity and the person’s baseline metabolic rate. Individual PAL values are initialized from a normal distribution and subsequently adjusted to maintain population-level expected values. Energy expenditure (EE) is calculated as a function of body composition and several metabolic parameters, integrating multiple physiological processes, including resting metabolic rate, thermic effect of food, adaptive thermogenesis, and energy partitioning. Further details can be found in Hall et al (2022) [5].
Energy intake (EI) quantifies the total caloric content consumed and is a function of the daily nutrient intakes calculated above.
BMI calculation
Body Mass Index (BMI) is calculated using the standard definition where Wi(t) and Hi(t) respectively denote individual i’s weight and height at time t.
Wi(t) is initialized as a function of age, sex, energy intake and physical activity. Hi is calculated based on population-level expected values given a person’s age and sex, their weight, as well as randomly sampled individual variation. Height is assumed to remain constant after age 19. Individual BMI values are calibrated to the BMI distribution for India in the NCD-RisC database through an adjustment of weight and nutrient intake values [6].
Using the initialized values for weight, height, energy intake and physical activity, the Hall et al (2022) model is used to initialize values of body fat, lean tissue, glycogen, extracellular fluid and energy expenditure for each person [5]. The changes in these values (governed by changes in nutrient intake and energy intake) are then used to update each individual’s weight, and thus their BMI.
D4.Disease modules
Health-GPS is calibrated to match empirical disease distributions. Changes in the distributions of risk factors and diseases are then transferred to other diseases. Although we assume the associations of risk factors to diseases to be constant throughout the simulation, any deviation of a risk factor from its initial distribution will translate into increased or decreased disease incidence through relative risk equations (the epidemiological literature does not always distinguish clearly between relative risks and hazard ratios; when it is not possible to identify values for both, available values are used interchangeably in Health-GPS). Therefore, any change in the prevalence of a disease is solely caused by changes in risk factor distributions, alongside the ageing of the population.
The disease module uses multiple data sources, including IHME for prevalence, incidence, mortality and remission rates for non-cancerous diseases [2]. When cancers are modelled (which is not the case in this study), Health-GPS uses data from the International Agency for Research on Cancer (IARC) for prevalence at 1, 3 and 5 years (from diagnosis), incidence, and mortality rates [7]. Relative risks from BMI to diseases are derived from IHME attributable risk data, while those for sodium to ischaemic heart disease, stroke, and chronic kidney disease are derived from specific studies [8,9].
Probabilities of disease for each individual are updated throughout the simulation to reflect changes in risk exposures and disease profiles. Prevalence in the initial population is used to calibrate Health-GPS parameters. However, as time progresses, risk factor distributions and demographics are the sole drivers of prevalence.
The excess mortality rate is the proportion of excess deaths within a designated population of “cases” (people with a medical condition) over the course of the disease. These rates are kept constant over time in the simulation. Remission is the rate at which people with a disease revert to a state with the same survival prospect and disability level as those without the disease. Remission rates are kept constant over time in Health-GPS. In this study, remission rates are taken from IHME [2]. When not available, they are assumed zero.
Disease is modelled through 3 types of events: onset, remission and fatality. An individual’s probability of acquiring a disease is calculated based on their risk exposures, existing conditions, and the population-level incidence of each disease. Therefore, this incidence is adjusted for each individual to account for their sex, age, risk and disease profiles. For each individual in the population and for each disease, we compute an Individual Relative Risk (IRR), which reflects the aggregate change in disease incidence associated with the individual’s risk factors and diseases. After disease onset, the individual may either go into remission or die, over time.
Let RRdfag be the relative risk of disease d associated with risk factor f = 1, …, Nf for age a and sex g. Let DRdagj be the relative risk of disease d associated with the presence of another disease j (where j = 1, …, ND). for age a and sex g. If and then individual i’s overall relative risk IRRid (t) of disease d is given by
At the beginning of the simulation, each individual’s disease status is assigned based on: (a) their IRRid (t0); (b) the mean value of this quantity among people of the same age and sex; and (c) the prevalence of disease d in the same age and sex group. Therefore, individual i is assigned disease d with probability equal to where Pdag (t0) and |IRRd |ag respectively denote the population level prevalence, and the average relative risk of disease d for age a and sex g, at the beginning of the simulation. Note that this formulation assumes relative risks combine multiplicatively and are independent.
Similarly, if Idag(t0) denotes the population level incidence of disease d for age a and sex g, as currently observed empirically (at t0), then individual i develops disease d at time t with probability equal to and
I∗is calculated in the baseline scenario and is held constant throughout the simulation. It is also used throughout the simulation of intervention scenarios.
Finally, each person of age a and sex g with disease d has probability ρdag of going into remission each year, and dies with probability each year (see above section on Residual Mortality).
D5.Burden of disease module
If DWd represents the disability weight associated with disease d, let
Now define the product Xi(t) of the compliment of disability weights Xi(t) by and the average of this product for age a and sex g is where Nag (t) denotes the number of people of age a and sex g at time t. If YLDag (0) is the empirically observed years lived with disability for age a and sex g over all diseases at the time of initialization, and if RDWag is the residual disability weight for diseases not included in Health-GPS for age a and sex g, then at time t = 0, we have
Rearranging gives
RDWag is initialized as above at the beginning of the simulation and unchanged after. More generally, across diseases (modelled in Health-GPS or otherwise) at time t, each person i has disability weight DWi (t) given by
Disability-adjusted life years (DALYsag(t)) are calculated as the sum of the years lost due to premature mortality (YLLag (t)) and the years lived with a disability YLDag (t) over the time horizon of the simulation, for age a and sex g. YLD(t) at time t is given by
The contribution of person i to years of life lost at time t is given by where ai is their age, TDi gives their time of death, and LE(t) = max {LEMale(t), LEFenale (t)} is the life expectancy at time t. Then
The DALYs at time t are then given by
D6.Health expenditure
Total health expenditure (THE) consists of current health expenditure (CHE) and capital expenditure. According to National Health Accounts Estimates for India 2021-22, CHE is 87.32% of THE, and of the CHE, governments’ share is 38.79% and the households’ share is 50.55% (including out-of-pocket expenditure and insurance contributions). Out-of-pocket expenditure (OOPE) specifically account for 45.11% of THE [10]. Therefore, households’ expenditure and households’ OOPE on health accounts for 44.14% and 39.39% of THE, respectively.
For each disease in asthma, chronic kidney disease, diabetes, ischemic heart disease, and stroke (including ischemic stroke, intracerebral haemorrhage, and subarachnoid haemorrhage), either households’ expenditure, or OOPE on health, per person per year in India is derived from previous studies [11–13], depending on availability, and is then converted into THE per case following the method adopted by WHO in NCD investment cases [14]. Finally, the averted annual THE of each disease under intervention is calculated as the product of reduced cases under intervention and THE per case.
D7.Hypertension
Changes in hypertension prevalence are estimated independently with data on hypertension prevalence in India from the National Family Health Survey (NFHS-5), 2019-21 [15], estimates on changes in sodium and BMI under intervention from Health-GPS, and the association between sodium, BMI and the risk of hypertension from literature [16,17].
A linear association is assumed between BMI and the risk of hypertension, and between sodium and the risk of hypertension. One unit change in sodium (mg) and BMI (kg/m2) is associated with a change of 0.00003 and 0.186 in the risk of hypertension, respectively [16,17]. Reduction in prevalence number of hypertension by age and sex is calculated as follows.
Where a denotes age and s denotes sex. prevalencea,s is the baseline prevalence number of hypertension from NFHS-5. Δsodiuma,s and Δbmia,s are policy effect estimates from Health-GPS. slopesodium = 0.00003. slopebmi = 0.186.
References
- 1United Nations, Department of Economic and Social Affairs, Population Division. World Population Prospects 2022 [Internet]. New York: United Nations; 2022 [cited 2023 Nov 26]. Available from: https://population.un.org/wpp/
- 2Institute for Health Metrics and Evaluation. Global Burden of Disease [Internet]. Seattle (WA): IHME, University of Washington; [cited 2023 Nov 20]. Available from: https://www.healthdata.org/research-analysis/gbd
- 3Matheus AS, Tannus LR, Cobas RA, Palma CC, Negrato CA, Gomes MD. Impact of diabetes on cardiovascular disease: an update. International Journal of Hypertension. 2013;2013(1):653789.
- 4Koye DN, Magliano DJ, Nelson RG, Pavkov ME. The global epidemiology of diabetes and kidney disease. Advances in Chronic Kidney Disease. 2018 Mar 1;25(2):121–32
- 5Hall KD, Sacks G, Chandramohan D, Chow CC, Wang YC, Gortmaker SL, Swinburn BA. Quantification of the effect of energy imbalance on bodyweight. The Lancet. 2011 Aug 27;378(9793):826–37.
- 6Institute for Health Metrics and Evaluation. Global Burden of Disease [Internet]. Seattle (WA): IHME, University of Washington; [cited 2023 Nov 20]. Available from: https://www.healthdata.org/research-analysis/gbd
- 7International Agency for Research on Cancer. Global Cancer Observatory [Internet]. Lyon: IARC; 2022 [cited 2023 Nov 10]. Available from: https://gco.iarc.who.int/en
- 8Tuomilehto J, Jousilahti P, Rastenyte D, Moltchanov V, Tanskanen A, Pietinen P, Nissinen A. Urinary sodium excretion and cardiovascular mortality in Finland: a prospective study. The Lancet. 2001 Mar 17;357(9259):848–51.
- 9Liu N, Sun W, Xing Z, Ma F, Sun T, Wu H, Dong Y, Xu Z, Fu Y, Yuan H. Association between sodium intakes with the risk of chronic kidney disease: evidence from a meta-analysis. International Journal of Clinical and Experimental Medicine. 2015;8(11):20939.
- 10Ministry of Health and Family Welfare, Government of India. National Health Accounts: estimates for India 2021–22 [Internet]. New Delhi: MoHFW; 2024 [cited 2024 Sep 25]. Available from: https://nhsrcindia.org/sites/default/files/2024-09/NHA2021-22.pdf
- 11Yadav J, Menon GR, John D. Disease-specific out-of-pocket payments, catastrophic health expenditure and impoverishment effects in India: an analysis of National Health Survey data. Applied Health Economics and Health Policy. 2021 Sep;19:769–82.
- 12Kwatra G, Kaur P, Toor G, Badyal DK, Kaur R, Singh Y, Pandian JD. Cost of stroke from a tertiary center in northwest India. Neurology India. 2013 Nov 1;61(6):627–32.
- 13Kansra P, Oberoi S. Cost of diabetes and its complications: results from a STEPS survey in Punjab, India. Global Health Research and Policy. 2023 Apr 7;8(1):11.
- 14Pan American Health Organization; World Health Organization; United Nations Inter-Agency Task Force; United Nations Development Programme. The case for investment in prevention and control of noncommunicable diseases and mental disorders in Suriname: a return-on-investment analysis [Internet]. Washington (DC): PAHO; 2023 [cited 2025 Aug 9]. Available from: https://www.paho.org/en/node/93097
- 15International Institute for Population Sciences (IIPS) and ICF. 2021. National Family Health Survey (NFHS-5), 2019–21: India: Volume I. Mumbai: IIPS. https://dhsprogram.com/pubs/pdf/FR375/FR375.pdf
- 16Gelber RP, Gaziano JM, Manson JE, Buring JE, Sesso HD. A prospective study of body mass index and the risk of developing hypertension in men. Am J Hypertens. 2007 Apr;20(4):370–7. doi: 10.1016/j.amjhyper.2006.10.011. PMID: 17386342; PMCID: PMC1920107.
- 17Filippini T, Malavolti M, Whelton PK, Vinceti M. Sodium Intake and Risk of Hypertension: A Systematic Review and Dose-Response Meta-analysis of Observational Cohort Studies. Curr Hypertens Rep. 2022 May;24(5):133–144. doi: 10.1007/s11906-022-01182-9. Epub 2022 Mar 4. PMID: 35246796.
- 18Almdal T, Scharling H, Jensen JS, Vestergaard H. The independent effect of type 2 diabetes mellitus on ischemic heart disease, stroke, and death: A population-based study of 13 000 men and women with 20 years of follow-up. Arch Intern Med. 2004;164(13):1422–6.
- 19Mueller NT, Koh WP, Odegaard AO, Gross MD, Yuan JM, Pereira MA. Asthma and the risk of type 2 diabetes in the Singapore Chinese Health Study. Diabetes Res Clin Pract [Internet]. 2013 Feb [cited 2025 Sep 15];99(2):192–9. Available from: https://pubmed.ncbi.nlm.nih.gov/23260853/
Supporting information file 5
Appendix E. Income trend assumption
Appendix E forms part of the revised submission.
Supplement to: Roche M, Zhu J, Olney J, Laydon DJ, Joe W, Sharma M, Steele L, Sassi F. Taxation of foods high in saturated fat, sugar, and sodium in India: A modelling study of health and economic impacts. Submitted after revision on 19 September 2025.
Description
The income trend factor fi,j,0 for income group i and food group j for period t = 0 is estimated as the product of the income elasticity estimate (ηi,j) (Table C1) and an income growth assumption. We assume a constant 6% income growth for India based on IMF projections [2].
Food consumption for period t = 1 is then estimated at the food item level by multiplying fi,j,0 with the household level total daily consumption at t = 0. We derive household level daily intake of energy and nutrient by multiplying the estimated daily consumption for each item at t = 1 with each item’s energy and nutrient content derived from food composition tables. We sum this up at household-level to obtain the estimated household daily total intake of energy and nutrient. We then individualise daily intake based on household members’ age and sex using the daily NIN Dietary Guidelines for Indians 2024 to obtain the estimated daily individual nutrient intake (q̂i,n,1) for individual i, nutrient n, and period t = 1. The baseline individual-level nutrient-specific income trend factor fi,n,0 for individual i and nutrient n is recovered as below:
We then run the below cross-sectional OLS regression:
Where MPCEi,h represents the per capita monthly total expenditure for household h and Xi,h a vector of household and individual characteristics including the household sector, size, religion, and the individual sex, age, and education. We obtain the vector of coefficients β̂.
We recover the predicted individual level nutrient intake for future periods as:
Where MPCEi,h,t = MPCEi,h,0 × 1.06t based on the constant IMF projection assumption for real GDP per capita.
The nutrient-specific decay rate b̂n is finally recovered as the coefficient from the below OLS regression of the logarithm of the predicted growth rate in nutrient intake on a time trend:
As expected from Engel’s Law, which posits that, as household income increases, the proportion spent on food decreases, bn < 0. To align with Engel’s Law [2], the individual-level nutrient-specific income trend factor for period t is scaled by an exponential decay over time and is recovered using the below formulae:
Figure C1 shows that our income elasticities are well-behaved, i.e., higher for low-income households across food groups, in line with Engel’s Law [2]. Figure E1 cross-validates this approach by comparing the estimated yearly growth rate in the consumption of packaged processed foods and sweets - two food groups with a high proportion of HFSS foods (Table A6) - with historical trends based on Euromonitor International Passport data and Tak et al (2022) for selected highly processed food items [3,4].
References
- 1International Monetary Fund. Countries: India [Internet]. Washington (DC): IMF; [cited 2025 May 20]. Available from: https://www.imf.org/en/Countries/IND
- 2Engel E. Die Productions und Consumptionsverhältnisse des Königreichs Sachsen. Zeitschrift des Statistischen Bureaus des Königlich Sächsischen Ministerium des Inneren. 1857. Reprinted in: Engel E. Die Lebenskosten belgischer Arbeiter-Familien. Dresden: C Heinrich; 1895. German
- 3Euromonitor International. Passport global market information database [Internet]. London: Euromonitor International; [cited 2025 Aug 6]. Available from: https://www.euromonitor.com/our-expertise/passport
- 4Tak M, Law C, Green R, Shankar B, Cornelsen L. Processed foods purchase profiles in urban India in 2013 and 2016: a cluster and multivariate analysis. BMJ open. 2022 Oct 1;12(10):e062254