RESEARCH DOSSIER
Empirical Demand Estimation and Household Microdata
Abstract. This Research Dossier examines the empirical foundations of household demand estimation, linking consumer theory to survey instruments, observed prices, transaction data, identification, and welfare measurement. It develops expenditure-share and Slutsky relationships, evaluates complete demand systems, examines quality and quantity errors in unit values, and distinguishes price instruments, experimental effects, conditional expenditure responses, and equilibrium counterfactuals. Historical and comparative evidence includes household budget research, Côte d'Ivoire and Indonesia, China, Tanzania, the United Kingdom, the United States, Mexico, India, and harmonized European and international survey systems. An original analytical synthesis explains how coherent economic restrictions can coexist with invalid causal inputs and how contrasting empirical estimates can represent different scientific objects. The dossier is a source-based methodological monograph with independently derived illustrative calculations rather than a new estimation of restricted household records.
Keywords: household demand; household microdata; budget shares; Engel curves; demand systems; price endogeneity; unit values; survey design; causal identification; welfare measurement; international comparability
The Scientific Object and Its Estimands
Household demand estimation investigates how feasible purchases and consumption change with prices, resources, household composition, available goods, and institutional conditions. The object is a conditional behavioral relationship defined for a specified population, horizon, commodity boundary, and economic environment. Survey records are observations generated by that environment and by a measurement system. A demand coefficient is an estimate of a particular function of those observations under assumptions about the relationship between the two.
Consider household h in market m and period t. Let p be its feasible price vector, x its available expenditure, z household characteristics, and q a vector of quantities. A static conditional demand function q(p,x,z) describes a Marshallian choice at given resources. Actual reported purchases depend additionally on inventories, the timing of transactions, credit, product availability, household production, reporting conventions, and decisions by other household members. Demand for a product, expenditure on a category, consumption of a nutrient, and the probability of any purchase are distinct outcomes. They can legitimately move in opposite directions. Demand for calories can remain nearly constant when a household replaces expensive proteins with cheaper starch. A fall in observed grocery expenditure can accompany unchanged intake if home production rises.
The target population also matters. A household-weighted elasticity gives each household equal population representation; an expenditure-weighted elasticity describes the behavior of a currency unit of expenditure; a market elasticity aggregates actual quantities. Neither household size adjustment nor a change to the survey weighting scheme automatically transforms one into another. A price increase applied universally produces an equilibrium effect if wages, profits, retail supply or transfers respond. Holding expenditure and other prices fixed defines a partial demand response. A subsidy assignment can instead identify the response to an offer with imperfect redemption. Every empirical coefficient should be labeled at these three levels: the observed outcome, the causal intervention, and the population to which the result refers.
This dossier integrates demand theory, household budget measurement, experimental evidence, and structural identification. Its emphasis is household expenditures and the data-generating restrictions that allow them to inform demand. Market power and full industrial-equilibrium conduct are related research objects. It uses a comparative review of the named studies and institutional records, mathematical derivations, and transparent hypothetical examples. No household records are newly fitted, no survey uncertainty is estimated from raw observations, and no new named theory or empirical discovery is claimed.
Historical Foundations and Engel Curves
The empirical investigation of family budgets predates modern utility-based econometrics. Ernst Engel's 1857 inquiry established the durable association between household resources and the budget fraction allocated to food. Houthakker's international expenditure comparison (1957) examined roughly forty surveys from roughly thirty countries and documented similar but unequal patterns across populations. His centenary analysis provides historical evidence of regularities in conditional expenditure, not experimental identification of the causal consequences of additional income.
Let e_i denote outlay on a good and w_i=e_i/x its share of total outlay x. The identity e_i=x w_i yields the expenditure elasticity eta_i=1+(1/w_i) times the derivative of w_i with respect to ln x. A falling food share with rising x implies eta_food<1 where the share is positive. It does not imply less food spending in absolute terms: an elasticity of 0.6 means spending rises approximately 0.6 percent for a one percent increase in total spending, locally. An increase in the quality of food purchased can explain some of even that increase. The same relationship between food shares and expenditure can arise from preference differences, demographic composition, food prices, or constraints on purchases.
An Engel curve is therefore a useful statistical description of living standards, but it cannot rank household welfare mechanically. A family with young children, a rural household consuming its own crop, and a family purchasing prepared meals face different time constraints and commodity definitions. The food share that would correspond to a common standard of living need not be the same across them. A household living in an expensive urban market may have a high housing share, reducing its recorded food share without becoming better off.
In practical estimation one can fit conditional means using splines or local polynomials and inspect curvature by household composition and region before choosing a parametric demand system. Sparse observations at either tail require wider intervals and explicit support restrictions. Expenditure measurement errors appear in both the denominator of the share and the income-like regressor, so even a well-specified smoother cannot solve correlated measurement error. Historical comparisons should retain original price regimes, reporting periods, population coverage, and commodity partitions. Later standardized categories may be useful for comparison, but the original records determine what the historical results actually mean.
From Engel Curves to Complete Demand Systems
A demand system imposes coherence across purchases. For locally nonsatiated preferences under a binding budget p·q=x, adding-up requires the sum of shares to equal one. Homogeneity of degree zero in nominal prices and nominal expenditure requires q(tp,tx)=q(p,x) for t>0. The expenditure function e(p,u) represents the minimum expenditure required to achieve utility u. Its price gradient, when differentiable, gives Hicksian quantities. The compensated substitution matrix is symmetric and negative semidefinite under the conventional differentiable cost-minimization assumptions.
The local Slutsky identity is S_ij = ∂q_i/∂p_j + q_j ∂q_i/∂x. In elasticity form, epsilon^H_ij = epsilon^M_ij + w_j eta_i. Differentiating the budget identity at fixed prices gives sum_i w_i eta_i=1. At fixed expenditure, differentiating with respect to p_j gives sum_i w_i epsilon^M_ij=-w_j. Compensated elasticities satisfy sum_i w_i epsilon^H_ij=0. These identities allow the analyst to test internal accounting and restrictions without confusing them with empirical causality.
Deaton and Muellbauer's Almost Ideal Demand System (1980) operationalizes flexible budget shares. In standard notation, w_i = alpha_i + sum_j gamma_ij ln p_j + beta_i ln(x/P), with ln P = alpha_0 + sum_k alpha_k ln p_k + (1/2) sum_k sum_l gamma_kl ln p_k ln p_l, under symmetry. Adding-up imposes sum_i alpha_i=1, sum_i beta_i=0 and sum_i gamma_ij=0. Homogeneity imposes sum_j gamma_ij=0. Symmetry imposes gamma_ij=gamma_ji. These linear conditions are distinct from curvature; the substitution matrix must be evaluated on the relevant support. A fixed-share price index approximation is computationally convenient, but substituting contemporaneous endogenous household shares can change the statistical problem.
The exact specification provides a direct expenditure elasticity 1+beta_i/w_i. At a positive share, the uncompensated price elasticity equals -delta_ij+(gamma_ij-beta_i times ∂ln P/∂ln p_j)/w_i, where delta_ij is one if i=j and zero otherwise. This follows by differentiating q_i=xw_i/p_i. It should be evaluated using the estimated exact index if that is the model fitted. Treating an approximate linear model as the exact model can produce inconsistent welfare calculations.
Banks, Blundell, and Lewbel (1997) derive a class of integrable systems with quadratic log-expenditure terms in Quadratic Engel Curves and Consumer Demand. Their evidence from pooled British household surveys motivates a richer Engel shape and demonstrates consequences for modeled tax welfare. A quadratic term allows marginal expenditure responses to vary nonlinearly with resources, but an arbitrary quadratic share regression does not inherit their integrability conditions. The distinction is critical: adding flexibility to a statistical curve is easier than adding flexibility to an expenditure function.
A practical comparison fits a transparent baseline and flexible alternatives on common data. The researcher examines adding-up, homogeneity, symmetry, predicted shares, and compensated curvature across observed price-expenditure cells. When restrictions fail, the plausible explanations include functional-form failure, household heterogeneity, data errors, product aggregation, and endogenous prices. Imposing restrictions before diagnosing failure may obscure the scientific mechanism. Relaxing all restrictions may improve prediction yet eliminate the basis for welfare integration.
Household Survey Design and the Observation Process
Demand estimation begins with a dictionary of observation. The statistical household may be a residential unit, a group that pools resources, a consumer unit, or an individual reporter. The numerator may be cash purchases, all acquisition, consumption of home production, gifts, in-kind transfers, or a service flow from durables. The denominator should be constructed from categories defined on the same convention. A purchased refrigerator is a transaction; its annual service flow is an imputed consumption quantity. A single frequency conversion cannot reconcile the two.
Reporting windows change the distribution of values. In a seven-day food diary, a zero may indicate that a household purchased no rice during the week despite consuming from stocks. A monthly recall can capture the purchase but miss its timing and quality. A twelve-month expenditure question may be suitable for irregular purchases yet vulnerable to recall decay and telescoping. Scaling a weekly purchase by fifty-two assumes a representative week; the assumption fails when interview timing is seasonal or when infrequent spending is concentrated.
Beegle, De Weerdt, Friedman, and Gibson's Tanzanian survey experiment (2012) compared eight consumption measurement designs, documenting differences across diary and recall instruments that also vary with household characteristics. The personal diary is a comparison benchmark, not proof of perfectly observed consumption. The experiment identifies an effect of survey design on measured consumption, distinct from an effect of income on actual demand. A questionnaire reform can produce a discontinuity in apparent poverty or elasticity even without an underlying behavioral change.
The joint FAO and World Bank food data collection guidelines (2019) establish the institutional importance of designing consumption modules consistently for welfare, nutrition, and poverty measurement. An empirical data audit should specify recall period, item granularity, reporters, quantification, purchase and nonpurchase acquisition, imputed values, and how meals away from home enter the aggregate. Nonstandard units such as bundles, bowls, and locally defined containers need region-product conversion factors, with uncertain conversions propagated through quantity measures.
Survey participation is another economic selection process. Busy high-income households may decline interviews; migrants and unstable households are harder to track. Inverse-probability adjustments can account for observed patterns of participation where the required response model and overlap hold. They cannot guarantee recovery of unobserved consumption among missing households. Distinguish nonresponse, deliberate skips, genuine zeros, imputed values, and suppressed data. Treating all as zero creates errors in both extensive-margin behavior and total expenditure.
Design weights approximate population representation, whereas a structural likelihood may address conditional choice among respondents. Stratification, primary sampling units, replicate weights, repeated households, and the level at which prices vary determine the appropriate uncertainty calculation. A thousand household observations in the same price-setting village do not provide a thousand independent price shocks. A defensible survey-weighted analysis estimates totals or moments using survey design and reports the effective independent variation for identification separately.
Prices, Unit Values, and Measurement Error
Market price is the price of a specified commodity at a specified place, time, quality, package and purchasing condition. The unit value v_i=e_i/q_i is the ratio of reported outlay to reported quantity. For a heterogeneous category it incorporates the household's product composition, discounts and choice of seller. If a household moves from premium rice to a cheaper variety when market prices increase, its measured unit value may increase less than the market price, remain unchanged, or fall. A zero quantity gives no observed unit value; it is not a zero market price.
Deaton's quality and spatial price analysis (1988) introduces corrections for quality variation and measurement error in geographically clustered surveys and illustrates the approach with 1979 data from Côte d'Ivoire. The cluster-level common price assumption creates an opportunity to distinguish variation in choice within locations from variation in local prices between locations. Theory restricts how within-category substitution affects the unit value. Geographic differences in unobserved tastes or seller availability can nevertheless contaminate the between-cluster price signal. Deaton's Indonesian survey extensions (1990) generalize the approach to an eleven-commodity food-demand application using 1981 survey data. An individual unit value alone cannot replace the market price without additional assumptions.
Consider a homogeneous product, with two households facing a market price of $2. The first purchases 10 units for $20, the second 5 units for $10. Their unit values correctly coincide. Suppose the first mistakenly reports 8 units, giving a unit value of $2.50, while the second reports accurately. The measured unit value and the measured quantity now differ negatively even though both households faced the same price. In logs, ln v=ln e-ln q. If quantity is reported as ln q+u and expenditure accurately, the error in ln v is -u while the error in ln q is +u. Their error covariance is -Var(u) when u is otherwise independent. This can induce a negative fitted slope without price variation at all.
A second problem arises when a spatial index mixes goods of differing quality, changing the basket. A paired price sample with fixed barcode, package and store type can reduce composition bias, but excludes products absent from one market. Seller panel data may omit informal markets or household-produced food. A community price questionnaire can broaden representation, while reporting prices retrospectively or for an unspecified grade undermines its comparability.
For each commodity the researcher should preserve raw transaction values and quantity conversions, construct a local market price where possible, and then compare alternative data paths: independently observed price; leave-household-out market-level index; corrected survey unit values; or a documented instrument. A leave-one-out average removes the household's direct contribution to the price statistic but does not remove common local demand shocks or common measurement practices. The appropriate price variable must be defended in economic terms before the estimator is selected.
Identification and Price Instruments
Demand and supply jointly determine transaction prices. In a stylized log-linear equation ln q_hmt = a + b ln p_mt + c ln x_ht + z_ht'd + u_hmt, ordinary least squares identifies a conditional linear association. A structural price response requires price variation uncorrelated with the demand disturbance after the stated controls. Retailers may raise prices when local demand strengthens. Consumers may select stores with favorable price offers. Supply bottlenecks can change both prices and the availability or quality of goods. Time and region effects absorb some confounding without guaranteeing exogeneity.
In a simple single-instrument setting with residualized variables, the instrumental variables estimand is Cov(Z,ln q)/Cov(Z,ln p). Relevance requires a nonzero denominator; the exclusion and exogeneity assumptions require that instrument variation affect the demand equation through price and be independent of the remaining demand disturbance. An instrument is not justified by a high first-stage statistic alone. A cost shifter can alter the product assortment, delivery times, seller entry, or household income. A transport-cost change can alter access to labor markets as well as food prices. Such channels require specific controls, different interventions, or a narrowed interpretation.
Angrist and Pischke's credibility revolution account (2010) stresses the centrality of research design in applied causal work. Staiger and Stock's weak instruments analysis (1997; linked earlier version) shows why conventional approximations can perform poorly when instruments are weak. Report the first-stage estimates and design-appropriate diagnostics, use weak-identification-robust tests or intervals when justified, and show reduced-form effects on the outcome. An overidentification test may reveal inconsistency under joint restrictions but passing it cannot establish the validity of every instrument.
Cross-market price instruments are plausible only if prices in other markets share cost shocks while independent local demand shocks do not contaminate the cross-market association. Common advertising campaigns, synchronized sales, national demand shifts, common demographic trends and correlated pricing algorithms make this exclusion fragile. The appropriate geographic grouping follows the pricing and distribution network, not necessarily survey administrative boundaries. A national fixed effect may absorb the common variation that made the instrument strong, revealing its practical identifying limitation.
Expenditure x is often endogenous: preferences and shocks that influence commodity purchases also influence total expenditure, and the expenditure measure includes the outcome itself. Income can be an instrument only when income changes have no direct effect on tastes, time use, rationing or household bargaining conditional on total expenditure. Benefit thresholds, randomly assigned transfers and plausibly exogenous earnings changes may identify narrower resource responses. Current income, permanent income and conditional expenditure are different regressors. Controlling for a variable affected by a price treatment may block some of its total effect; whether to condition on it depends on the estimand.
An illustrative sign reversal: suppose the true price coefficient is -0.6 and the omitted-demand component contributes +0.9 to the observational regression coefficient. The fitted association is +0.3. This arithmetic is a hypothetical decomposition, not evidence that upward slopes are always spurious. Likewise an exogenous income shock can induce positive demand for a normal good without establishing the shape of compensated demand. Credibility requires matching estimands, signs and assumptions rather than preferring a coefficient because its sign appears economically familiar.
Experimental and Quasi-Experimental Designs
A randomized price subsidy can generate exogenous price variation for eligible households. The intention-to-treat effect compares outcomes across randomized assignment groups. If redemption varies, the assigned subsidy and effective price differ. An instrumental variables interpretation of assignment as an instrument for received treatment requires a defensible exclusion restriction, a first stage and assumptions about heterogeneous compliance. Incentives may also change product salience or perceived quality independently of the paid price. A pilot must measure these pathways.
The methods discussed by Duflo, Glennerster, and Kremer in the randomization research toolkit (2007) place experimental design and interpretation in an economic context. Randomization balances expected pre-treatment covariates for the assigned population, but does not repair attrition, spillovers, treatment contamination, or measurement failures after randomization. Report assignment probabilities, clusters, baseline outcomes, missingness, and prespecified estimands. If sellers change list prices in response to vouchers, an individual assignment comparison can differ from a market-wide policy effect.
Jensen and Miller's Giffen behavior study (2008) provides an instructive example: staple price subsidies for very poor households in China yield stronger evidence of an upward uncompensated demand response for rice in Hunan and weaker evidence for wheat in Gansu. The relevant mechanism is a large negative income effect for an inferior staple in a subsistence-constrained budget. It is context-dependent. The result is consistent with nonpositive compensated own-price substitution because the income component can outweigh it. The evidence concerns the experimental population and commodity, not all staple foods or households.
Other policy designs exploit tax reforms, transport openings, cost shocks, or benefit eligibility changes. In differences-in-differences, the identifying restriction concerns counterfactual trends; the relevant treatment timing and pre-exposure effects must be specified. For a tax change, the price pass-through may be incomplete, and purchases may move across borders. An event study can reveal pre-treatment movements, but absence of a statistically significant pretrend is not proof of parallel counterfactuals. Regression discontinuity identifies a local effect near an eligibility cutoff if the continuity and manipulation restrictions are credible. Anticipated threshold changes, bunching, and concurrent rule changes complicate interpretation.
A price experiment typically covers a bounded range. Extrapolation to removal of a subsidy, a large commodity shock, or a different population requires behavioral structure or credible evidence on nonlinearities. A price intervention that lasts one month may estimate reallocation of shopping dates while a permanent price change estimates a different response. Panel outcomes collected before, during, and after interventions can help distinguish inventory movements, substitution and adaptation. Research design and economic modeling are complementary when each addresses the part of the counterfactual that its evidence actually supports.
Zero Purchases, Heterogeneity, and Dynamic Adjustment
Household budget data contain many zeros even for goods consumed by nearly everyone over a sufficiently long period. Some reflect nonparticipation, some infrequency, some censoring and some reporting omission. A two-part representation takes expected quantity as Pr(q>0) times E(q|q>0). Where both quantities are positive and differentiable, the elasticity of expected quantity is the sum of the participation elasticity and the elasticity of the conditional mean. Estimating only positive observations generally omits the first channel and changes the population.
A Tobit likelihood treats zeros as censored realizations of a particular latent process; corner solutions, discrete transactions and inventories can generate different relationships. A hurdle model separates the purchase decision from positive quantities but does not by itself ensure coherent system-wide budget shares. A selection correction depends on a credible specification of participation and its exclusion variables. Transformations such as ln(1+q) are unit-dependent and do not recover a conventional elasticity automatically. Report effects in physical units or use a model with clearly defined mean derivatives.
Heterogeneity in baseline expenditure, demographics and elasticities is substantively important. For example, an increase in child-related expenditure after a birth can reflect changed needs and changed household income or time allocation. Equivalence scales encode a convention or estimated welfare relationship; they do not mechanically convert observed household expenditure into each member's consumption. Intrahousehold resource allocation can cause a transfer paid to one member to have distinct effects from an otherwise identical household-level transfer.
Dynamic choice creates further distinctions. A price promotion can induce brand switching, movement of purchases from next month, or extra consumption. Durable goods are purchased intermittently and deliver services for years. A short panel may be excellent for measuring transactions and poor for identifying long-run service demand. Demand models with inventory states, durable replacement or habit persistence need longer histories and assumptions on expectations. A fixed household effect controls time-constant unobserved heterogeneity but does not control time-varying shocks, and can remove variation needed to estimate demographic effects.
When the scientific objective is a conditional food-demand system, weak separability of food from other goods and two-stage budgeting may simplify the problem. That restriction implies that conditional food purchases depend on the food budget and within-group prices rather than independently on all nonfood prices. A food-budget elasticity therefore need not equal a total-expenditure elasticity. A change in rent can reduce the food budget even with unchanged within-food relative prices. A full welfare analysis must include the first-stage allocation or explicitly restrict itself to within-food substitution.
Differentiated Products and Transaction Microdata
Barcode transactions and household purchase panels measure detailed product switching and the price actually paid by participating shoppers. They add empirical variation within categories but define their own sampling frames. Cash purchases outside participating stores, unlabeled foods, informal markets and households without reliable scanner participation can be absent. A store-level promotion may be targeted to expected demand. The analyst needs both the price opportunities of nonpurchased alternatives and a record of purchased products, rather than prices conditional on choice alone.
In a discrete-choice model individual utility from alternative j can be represented by delta_j plus an individual taste component and idiosyncratic error. The simplest multinomial logit with utility delta_j-alpha p_j imposes a strong pattern of substitution: for constant alpha and positive shares s, ∂s_j/∂p_j=-alpha s_j(1-s_j), whereas ∂s_j/∂p_k=alpha s_j s_k for k≠j. The outside alternative must be defined to match the market boundary. A random-coefficient specification allows taste and price sensitivity to vary; integration over the distribution produces aggregate shares and more flexible diversion. Additional household moments can help constrain heterogeneity but the endogeneity of prices and assortment remains.
Berry, Levinsohn, and Pakes' automobile market equilibrium study (1995; linked working-paper record) illustrates joint differentiated-product demand and supply modeling. It addresses automobile market equilibrium, not a complete household budget. Transferring its logic to grocery data requires a choice occasion and an outside good consistent with multiple-item baskets, quantities, and availability. A choice model over mutually exclusive brands cannot simply describe a household that simultaneously buys several brands without a suitable extension.
Handbury and Weinstein's goods prices and availability study (2015; linked working-paper summary) demonstrates that controlling product heterogeneity and availability can alter the measured relationship between urban size and food prices. Local price indexes constructed from observed expenditure mix are therefore vulnerable to composition, even in rich transaction datasets. Constructing a common-product comparison solves part of the problem at the cost of coverage. Evaluating the cost of living with genuinely different available goods requires a model or explicit index convention.
Allcott and colleagues (2019), in Food Deserts and the Causes of Nutritional Inequality, use store entry, household moves, and structural grocery-demand modeling. Their model-based counterfactual assigns a relatively small share of measured U.S. nutritional inequality to unequal product and price access. The result depends on their market definition, outcomes, population and model; it does not establish that physical access is irrelevant everywhere. Their geographically varying chain-product instrument illustrates an attempt to separate price and product access from local demand. A chain entering poor neighborhoods can affect travel costs, local employment, and competition at other stores, so the exclusion argument must be examined for the specified design.
For household demand research, granular transaction data and budget surveys should be treated as complementary. The former measures price-product switching among included channels; the latter can capture nonmarket acquisition, housing, transfers, and spending outside the sampled retail universe. Record-linkage consent and safe access matter when joining sources. An unrepresentative transaction panel can improve internal substitution estimates yet reduce population coverage. External validation against survey totals and market shares should use aligned units and time periods rather than assume one source is a universal benchmark.
International Survey Systems and Comparability
International household research requires distinguishing harmonized labels from harmonized observations. Identical classification codes can represent different collection instruments, service flows and coverage. A reproducible cross-country release states the national survey vintage, geographic exclusions, household definition, reference period, sampling design, currency and price index, commodity crosswalk, nonmarket valuation, and access conditions. Repeated cross sections should not be labeled individual panels. Aggregate expenditure tables should not be labeled household records.
The United States has distinct Interview and Diary components. The Bureau of Labor Statistics Consumer Expenditure methods describes quarterly interviews and diaries maintained for two consecutive weeks. The instruments emphasize different purchase frequencies and use separate samples. Analysts cannot merge them as if they were two interchangeable records of the same households. For total expenditure, preserve the official integration conventions and distinguish a modeled imputation from a directly observed purchase history.
Eurostat's Household Budget Survey microdata documentation lists scientific-use files for 2010, 2015 and 2020 and identifies national methodological variation. These documented rounds provide a common institutional framework, but timing, questionnaire content and survey methods remain relevant to causal comparability. The existence of a harmonized output code does not imply common retail price measurement or variance design. Scientific-use access is not identical to open download. The analysis does not infer subsequent wave-level access from this documentation.
LIS's harmonized household database provides documented mappings for income, expenditure flows and imputed rent. It is particularly useful for comparing resource distributions with national metadata; the detail and coverage of consumption variables must be checked at each country-wave. Treating standardized missing categories as recorded zeros would distort comparison.
Tanzania's National Panel Survey Wave 5 (2020–21, catalog 5639) covers consumption alongside agriculture, household enterprises and other living-standard measures. The catalog distinguishes panel households from a booster sample and specifies household tracking. This makes it possible in principle to distinguish market purchase from production and to examine repeated outcomes with appropriate linkage and weights. This dossier uses catalog-level information; it has neither downloaded the survey records nor verified every variable, code and conversion factor.
India's official Household Consumption Expenditure Survey release describes the 2023–24 survey conducted from August 2023 to July 2024 across rural and urban areas with specified geographic exclusions. Its reported expenditure aggregates belong to that survey design and time frame. Before treating a historical comparison as growth or estimating a price coefficient, one must examine questionnaire redesign, free-item valuation, reference periods and the spatial price deflator. Conversion into a common currency, including PPP conversion, cannot create the local relative price vector required for a demand system.
These examples suggest a defensible comparison sequence: first reproduce the national observation definitions; second map to a common set of commodity and household concepts; third identify the country-specific price variation and estimate population-specific relationships; fourth compare responses at matched expenditure levels, commodity definitions and intervention ranges. A pooled regression with country dummies does not repair different measurement errors or income-price support. Reweighting across countries can align observed covariate distributions, but transport of causal elasticities also requires stability of conditional mechanisms and overlap. If the relevant covariate or price support has no common region, separate national estimates are scientifically preferable to a misleading common estimate.
Welfare, Distribution, and Policy Counterfactuals
Demand elasticities become policy-relevant when the welfare metric and intervention are specified. Let p0 be initial prices and p1 new prices, and let u0=v(p0,x0) be baseline utility. For a price increase at fixed nominal resources, the amount needed to restore baseline utility is CV=e(p1,u0)-e(p0,u0). The equivalent variation measured at initial prices for the utility loss is e(p0,u0)-e(p0,u1), using a positive-loss convention and u1=v(p1,x0). They coincide to first order for small changes but generally differ for large changes.
By Shephard's lemma, the first-order compensating transfer for a small change in prices is q(p0,x0)·dp. Divide by initial expenditure x0 to obtain sum_i w_i dln p_i. This is a local cost-of-living approximation at fixed baseline utility. For finite price changes, demand substitution changes the cost path. The second-order term is one half dp' S dp at the baseline, where S is the expenditure function's compensated Hessian in price units. Negative semidefiniteness makes this term nonpositive; the fixed initial basket provides a feasible upper bound for compensation as long as it remains purchasable at the new prices. Product disappearance or rationing changes the feasible set, invalidating that comparison.
A transparent hypothetical calculation uses two households each with $100 in monthly resources. Household A spends $60 on food and B spends $20. A ten percent proportional rise in the price of the food basket, with all other prices and resources fixed, costs $6 and $2 respectively for their unchanged baskets. Compensation relative to baseline spending is six versus two percent before substitution. If households can substitute across broad goods, actual expenditure needed to maintain utility may be lower. These figures are constructed examples. They do not describe any specific country or population. Their analytical implication is that average inflation weighted by aggregate expenditures is insufficient to identify the distribution of welfare losses.
Attanasio, Di Maro, Lechene, and Phillips' rural Mexico food-price study (2013) evaluates welfare consequences of food price increases and simulates alternative policy responses. Its cash-transfer versus commodity-subsidy comparison illustrates the dependence of welfare on initial budgets, substitution and targeting. A cash grant relaxes the budget constraint while a commodity subsidy changes relative prices. With complete markets and a distributional objective, relative-price distortions may be undesirable; with nutrition externalities, missing credit, rationing or incomplete targeting, the evaluation requires extra mechanisms. The cited empirical comparison is specific to its rural Mexican setting and modeled interventions.
Households may also be producers. A price increase in a crop raises consumption costs but can increase farm profits. Holding production quantities fixed, a local first-order effect on net resources relative to the cost of maintaining food consumption is (sales quantity minus purchased or consumed quantity as consistently defined) times the price change. More precisely, with a profit function encompassing production and a consumption expenditure function, the relevant differential is the price derivative of profit less the compensated consumption derivative. Use net seller or net buyer positions measured on a single consistent production-consumption account; avoid counting own production twice. Changes in wages, input prices, farm output and public transfers are additional channels for a general-equilibrium counterfactual.
A distributional assessment states whether weights represent households or persons, how household size and resources are normalized, and which welfare scale is used. A national mean effect can hide large losses at the bottom of the resource distribution or gains among producers. The empirical target determines if quantiles, subgroup effects, expenditure-weighted means, or welfare aggregates are appropriate. A structurally estimated demand system is useful for large counterfactuals precisely because it imposes coherent cross-price responses; its welfare conclusions remain conditional on identification and the validity of the modeled feasible set.
Estimation, Uncertainty, and Reproducibility
A reproducible estimation workflow starts with a declared estimand and a dated data inventory. Preserve raw identifiers under access restrictions and record public metadata, country-wave versions, item crosswalks and derived variables. Create a ledger of exclusions, missing observations and imputation. Show how each numerator and denominator was formed. Prices and quantities require explicit unit checks; negative returns or reimbursement records must follow a documented transaction convention. When merging market prices with household records, report match rates and whether missing matches vary by income or location.
Fit descriptive demand patterns before imposing structural restrictions. Report simple mean shares, zero frequencies, support by market and expenditure, and weighted versus unweighted summaries. For a price design, report the independent variation and the first stage before the structural estimates. Distinguish standard errors for parameters from uncertainty in calculated elasticities, welfare measures and simulations. The delta method or appropriate resampling can propagate nonlinear transformations; a cluster bootstrap should resample the units that generate independent variation, not merely household rows when assignment occurred at village level.
For system estimation with adding-up, one share equation is redundant; when estimating a covariance-based system it may be omitted and its parameters recovered from restrictions, with results checked for invariance under the implementation. Depending on model and error covariance, seemingly unrelated regression may improve efficiency but cannot fix endogeneity. A generalized method of moments estimator can combine demand restrictions with instruments; convergence, multiple initializations, gradient tolerances and sensitivity to weighting matrices should be documented. A predictive fit alone cannot identify structural prices or welfare.
A specification grid should vary reasonable recall conversions, category mappings, price measures, sample restrictions, treatment windows and functional forms. Distinguish planned sensitivity checks from post hoc selection of a favorable result. Report failures as part of the evidence and preserve the result set when feasible. Adjust hypotheses and intervals for multiplicity where multiple goods or heterogeneous responses are tested jointly. Where identification remains partial, show the assumptions underlying bounds rather than replace them with point estimates.
Data confidentiality determines which artifacts can be distributed. A replication package can include code, metadata, synthetic test data, declared software versions and results checksums without redistributing protected household records. Disclosure review is needed for small cells and locations, particularly when linking detailed scans to household characteristics. A synthetic fixture validates computation and accounting; it cannot establish an empirical result about the target population. This distinction applies equally to the present dossier, whose numerical demonstrations are hypothetical and whose empirical results are attributed to the original studies.
Independent Comparative Synthesis
The comparative evidence makes a substantive methodological conclusion possible. The scientific content of an estimated household demand relationship depends jointly on the observation definition, independent variation and counterfactual mapping. Improving one stage does not automatically repair the others. The Tanzania measurement experiment gives unusually clear evidence on the first stage; the China subsidy experiment addresses price variation for a specific population; the British quadratic Engel study addresses flexibility and welfare integration; the spatial unit-value studies connect surveys to prices under theory; U.S. transaction research adds detailed product choice and access. Each supplies evidence for a different link in the same inferential task.
A useful way to see this is to compare hypothetical studies answering what seems to be the same question: how strongly does food demand respond to price? One regresses kilograms of purchased food on household unit values across income groups. Its result is vulnerable to measurement and quality selection. A second randomizes a staple subsidy among poor households and measures total staple acquisition. Its local causal estimate can be scientifically strong while referring to a different population, commodity definition and intervention. A third estimates a full demand system with external price instruments and uses consumer theory to simulate an economy-wide increase. Its modeled counterfactual is broader but rests on the instrument and structural stability. Disagreement among their coefficients need not constitute a contradiction. First align their objects, horizons, conditioning variables and populations, then investigate residual disagreement.
For a concrete internal-consistency example, assume two exhaustive goods, food and other consumption, with budget shares 0.4 and 0.6. Let food's expenditure elasticity be 0.7. The adding-up identity requires the other elasticity to be (1-0.4×0.7)/0.6=1.2. Suppose the uncompensated food own-price elasticity is -0.5. Its compensated value is -0.5+0.4×0.7=-0.22. By homogeneity, the food cross-price Marshallian elasticity for the only other good is -(-0.5)-0.7=-0.2. Its compensated cross-price elasticity is -0.2+0.6×0.7=0.22. Symmetry implies the compensated cross-price elasticity for other consumption with respect to food price is (0.4/0.6)×0.22=0.146666..., and the compensated own-price elasticity of other consumption is its negative under compensated homogeneity. Converting back with the expenditure elasticities yields -0.333333... and -0.866666... respectively for other consumption's Marshallian price responses. The expenditure-weighted price-column sums are -0.4 and -0.6. The compensated matrix has the required symmetry after conversion to quantity-price derivatives and is negative semidefinite in the two-good case because the normalized compensated matrix has opposite equal diagonal and off-diagonal elements. These numbers are a constructed local example; existence of globally integrable positive demand on a wider domain would require additional conditions.
Replacing the hypothetical food own-price coefficient with an observed unit-value slope would invalidate the behavioral interpretation even if the arithmetic restrictions remained satisfied. An analyst could enforce adding-up and symmetry and report an apparently sophisticated welfare table whose input is not a price intervention. Conversely an experimentally well-identified staple response alone does not determine all cross-price terms. The complete welfare matrix needs additional variation or justified structure. The example demonstrates why empirical design, data quality and theory must be reconciled rather than scored as substitutes.
A second implication concerns international generalization. Poverty, retail structure and budget shares condition both the magnitude and sometimes the sign of the response. A common sign across independent country-specific studies can be meaningful evidence for a regularity if the estimands match. It cannot provide a common magnitude without support for transport. Heterogeneity should be treated as an object of explanation: compare market access, production status, relative prices, resource distributions, commodity coding and measurement design before attributing a residual to preference differences. This analysis integrates established methods and published findings; it introduces no new named model or theoretical doctrine.
Open Scientific Questions and Research Agenda
Future research should pair household consumption records with independent local market prices at known product quality. An especially informative design collects otherwise identical product prices for buyers and nonbuyers, with package, outlet, location and date recorded, together with a survey module documenting inventory and own-produced consumption. Randomizing recording protocols within the same price environment can quantify measurement effects independently of price effects. Combining those experiments with a price intervention can test whether estimated elasticities are stable under alternative observation designs.
A second agenda is estimating price responses among households facing intermittent access or nonmarket acquisition. In such environments, purchases are an incomplete proxy for consumption. Joint information about household stock, gifts, transfers, own production, commuting and product availability can separate participation from physical intake. Weather and transport interventions may provide variation, but they can directly affect household production and income. Factorial experiments or measured channels may offer stronger evidence than mechanically treating those variables as instruments.
A third agenda concerns linking transaction panels to representative budgets and testing whether calibrated samples reproduce population-level expenditure distributions. Validation requires outcomes observed on common definitions and overlapping periods. Access constraints and selective participation should be part of the intended design. Household digital receipts can improve timing while excluding cash and informal market channels. Linkage itself may change the composition of observed households.
Finally, demand estimation should incorporate rising variation in environmental shocks, regulated prices, transfers, and product availability while preserving the difference between partial effects and economy-wide incidence. Policy simulations should show uncertainty from parameter estimates, price pass-through, product exit and alternative welfare metrics. International collaborative research could make source dictionaries and comparable estimand specifications public even where records remain confidential. Such practice would permit comparison of experimental and model-driven evidence across very different statistical systems without claiming precision beyond their information content.
Conclusion
Empirical demand estimation transforms household observations into knowledge about choices under constraints only when the outcome, feasible prices, treatment variation, budget definition and target population are explicit. Household microdata provide unusually rich evidence of heterogeneity, composition and expenditure allocation, but the measured values are products of survey institutions and market behavior. Unit values can contain quality choices and correlated quantity error; household expenditure includes timing and recording choices; survey harmonization does not erase national differences.
Demand theory gives restrictions and a coherent welfare representation. Econometric research design determines when observed price movements support causal interpretation. Experimental and quasi-experimental methods identify specified changes for specified populations, whereas structurally estimated systems can extend analysis to more complex counterfactuals under additional assumptions. Neither approach automatically confers external validity on a different country, time horizon, or distribution of resources.
The strongest practical conclusion is the need for complete alignment of measured outcomes, identifying variation and welfare interpretation. A descriptive association, a local treatment response and a policy-equilibrium prediction are distinct scientific products. Their integration requires evidence for each transition. The comparative research and explicit derivations in this dossier establish the reasons for that distinction and a concrete set of methods for auditing it. They do not purport to deliver newly estimated national elasticities or a newly instituted theoretical system. The substantive objective remains reliable knowledge of household demand and the distributional consequences of economic change.
References
- Allcott, H., Diamond, R., Dubé, J.-P., Handbury, J., Rahkovsky, I., & Schnell, M. (2019). Food deserts and the causes of nutritional inequality. Quarterly Journal of Economics, 134(4), 1793–1844. doi:10.1093/qje/qjz015. Source
- Angrist, J. D., & Pischke, J.-S. (2010). The credibility revolution in empirical economics: How better research design is taking the con out of econometrics. Journal of Economic Perspectives, 24(2), 3–30. doi:10.1257/jep.24.2.3. Source
- Attanasio, O., Di Maro, V., Lechene, V., & Phillips, D. (2013). Welfare consequences of food prices increases: Evidence from rural Mexico. Journal of Development Economics, 104, 136–151. doi:10.1016/j.jdeveco.2013.03.009. Source
- Banks, J., Blundell, R., & Lewbel, A. (1997). Quadratic Engel curves and consumer demand. Review of Economics and Statistics, 79(4), 527–539. doi:10.1162/003465397557015. Source
- Beegle, K., De Weerdt, J., Friedman, J., & Gibson, J. (2012). Methods of household consumption measurement through surveys: Experimental results from Tanzania. Journal of Development Economics, 98(1), 3–18. Source
- Berry, S., Levinsohn, J., & Pakes, A. (1995). Automobile prices in market equilibrium. Econometrica, 63(4), 841–890. Earlier version: NBER Working Paper 4264 (1993). Source
- Deaton, A. (1988). Quality, quantity, and spatial variation of price. American Economic Review, 78(3), 418–430. Source
- Deaton, A. (1990). Price elasticities from survey data: Extensions and Indonesian results. Journal of Econometrics, 44(3), 281–309. Source
- Deaton, A., & Muellbauer, J. (1980). An almost ideal demand system. American Economic Review, 70(3), 312–326. Source
- Duflo, E., Glennerster, R., & Kremer, M. (2007). Using randomization in development economics research: A toolkit. In T. P. Schultz & J. Strauss (Eds.), Handbook of Development Economics, Vol. 4, 3895–3962. Elsevier. doi:10.1016/S1573-4471(07)04061-2. Source
- Eurostat. (n.d.). Household budget survey: Microdata. Accessed November 17, 2026. Source
- Food and Agriculture Organization of the United Nations & World Bank. (2019). Food data collection in household consumption and expenditure surveys: Guidelines for low- and middle-income countries. Source
- Government of India, Ministry of Statistics and Programme Implementation. (2025). Household consumption expenditure survey 2023–24. Press Information Bureau, release 2109812. Source
- Handbury, J., & Weinstein, D. E. (2015). Goods prices and availability in cities. Review of Economic Studies, 82(1), 258–296. doi:10.1093/restud/rdu033. Earlier version: Working Paper 755, Zell/Lurie Real Estate Center. Source
- Houthakker, H. S. (1957). An international comparison of household expenditure patterns, commemorating the centenary of Engel's law. Econometrica, 25(4), 532–551. Source
- Jensen, R. T., & Miller, N. H. (2008). Giffen behavior and subsistence consumption. American Economic Review, 98(4), 1553–1577. doi:10.1257/aer.98.4.1553. Source
- LIS Cross-National Data Center. (n.d.). LIS Database. Accessed November 17, 2026. Source
- National Bureau of Statistics, Tanzania. (2022). Tanzania National Panel Survey 2020–21, Wave 5. World Bank Microdata Library, catalog 5639. Source
- Staiger, D., & Stock, J. H. (1997). Instrumental variables regression with weak instruments. Econometrica, 65(3), 557–586. Earlier version: NBER Technical Working Paper 151 (1994). Source
- U.S. Bureau of Labor Statistics. (n.d.). Consumer expenditures and income: Collections & data sources. Handbook of Methods. Accessed November 17, 2026. Source