Research & Evidence

Search for a method

Search titles, questions, territories and MSC identifiers.

    ← All methods
    METHOD DOSSIERMSC-P-020Pricing scienceVerified scientific dossier

    How do you address price endogeneity?

    Price is endogenous when it covaries with unobserved demand determinants. Controls, fixed effects and instrumental variables address different bias sources.

    Scientific editorial team: Marketing Science Center

    Direct answer

    Estimate elasticity under a declared set of identification assumptions.

    Price is endogenous when it covaries with unobserved demand determinants. Controls, fixed effects and instrumental variables address different bias sources.

    Angrist, Imbens & Rubin, 1996Villas-Boas & Winer, 1999

    01

    Why price and sales are not enough

    A price cut followed by higher sales does not prove that the cut caused the increase. The price may have been reduced precisely because demand was slowing. A promotion, a competitor stockout, seasonality or an assortment change can also move price and quantity at the same time.

    1. 01

      Business problem: price affects sales, but expected sales also affect price.

    2. 02

      Statistical problem: price is correlated with unobserved demand drivers contained in the error term.

    3. 03

      IV strategy: isolate price variation coming from a source external to unobserved demand.

    Villas-Boas & Winer (1999)

    02

    A marketing example from start to finish

    Synthetic illustration: teaching values, not observed and not reusable as a benchmark.

    The synthetic dataset follows 60 products for 10 weeks, giving 600 rows. Price responds to a supplier-cost shock and also to an unobserved demand shock. Quantity responds to price, promotion, seasonality and that same demand shock. This construction makes price endogenous while retaining an instrument that is relevant and exogenous by construction.

    Adjusted OLSβOLS = −0.577 · SE 0.046
    First stageπ = 0.434 · partial R² 0.310
    Strength, clustered covarianceWald χ²(1) = 255.749
    2SLS, clustered covarianceβIV = −1.011 · 95% CI [−1.174, −0.848]

    In this synthetic realization, OLS is biased toward zero and 2SLS moves closer to the generating coefficient of −1.20. The value −1.011 is neither a benchmark nor a universal elasticity. It only describes the price variation isolated by this synthetic protocol under its stated assumptions.

    03

    Scientific question and estimand

    For the 60 products and 10 weeks in the file, what is the log-log coefficient of price on quantity for price variation induced by the supplier-cost shock, conditional on promotion and seasonality? The estimand is β in the stated structural equation, with product-clustered inference.

    Counterfactual: for the same product i and week t, compare Qᵢₜ(p₁) with Qᵢₜ(p₀) when only log price p varies through the component induced by Z, holding controls X fixed. This contrast is defined by the stated homogeneous structural model; it is not directly observed.

    04

    Why simple regression can fail

    OLS attributes to price any residual variation shared by price and demand. Here, the unobserved demand shock raises both price and quantity. The condition E[u | ln(P), X] = 0 is therefore violated. The direction and magnitude of bias depend on the pricing process; they do not follow a universal rule.

    Villas-Boas & Winer (1999)

    05

    The causal mechanism that must be credible

    Z: instrument variationP: priceQ: quantity
    U: unobserved demand
    Expected structure: Z moves P, P affects Q, and U affects both P and Q. The missing Z → Q arrow represents the exclusion restriction, which must be argued rather than assumed from the diagram.

    Angrist, Imbens & Rubin (1996)

    06

    Required data and minimum quality

    • Unit: product × week; 600 rows, 60 product clusters and 10 periods with no missing values in the example.
    • Variables: log_quantity, log_price, supplier_cost_shock, promotion, season_index, product_id and week.
    • For real data, document: cost source, pricing calendar, stockouts, quality, assortment, promotions, competition, measurement error, attrition and price support.

    Download CSV · MSC-P-020

    07

    Formal model and symbols

    ln(Pᵢₜ) = α₁ + πZᵢₜ + δXᵢₜ + vᵢₜ

    ln(Qᵢₜ) = α₂ + βln(Pᵢₜ) + γXᵢₜ + uᵢₜ

    i, tproduct and week indices
    Qquantity sold; log_quantity is observed
    Pprice; log_price is observed but endogenous
    Zsupplier_cost_shock, observed candidate instrument
    Xpromotion and season_index, observed controls
    βestimated 2SLS coefficient on ln(P) in the quantity equation
    πfirst-stage coefficient of Z on ln(P)
    α₁, α₂intercepts of the price and quantity equations
    δ, γcoefficient vectors for controls X in the two equations
    ln(P)̂fitted log price projected by Z and X in the first stage
    Qᵢₜ(p)counterfactual quantity for product i in week t under log price p
    u, vunobserved errors in the structural and price equations

    08

    What the two stages actually do

    01

    Stage 1: explain price

    ln(P) = α₁ + πZ + δX + v

    Estimate the share of log price explained by instrument Z and controls X. This stage produces the fitted value of ln(P), written ln(P)̂, used by the IV model.

    02

    Stage 2: explain quantity

    ln(Q) = α₂ + β·ln(P)̂ + γX + u

    β is estimated from price variation projected by the instrument, not from all observed variation in P. For inference, use a complete 2SLS estimator: manually running a second OLS on the fitted value of ln(P) would give incorrect standard errors.

    Angrist, Imbens & Rubin (1996)

    09

    Reproducible numerical result

    The standard-library script reproduces: βOLS = −0.577216 (SE 0.046332), π = 0.434136 (SE 0.027147), partial R² = 0.310020, clustered Wald χ²(1) = 255.749005 and βIV = −1.011208 (SE 0.083104; asymptotic 95% CI, z = 1.96, 60 clusters: [−1.174092, −0.848325]). The exploratory first-stage residual coefficient is 0.628993 (SE 0.091948; descriptive t = 6.840731). Because it uses a generated regressor, this last calculation is not presented as a formal cluster-robust test and does not validate exclusion.

    Download Python · MIT

    10

    Six conditions and checks translated into business questions

    Questions and evidence to document for each condition
    Condition or checkOperational questionEvidence to document
    RelevanceDoes the instrument move price enough after controls?First-stage coefficient, partial R-squared and a design-appropriate robust statistic.
    IndependenceIs Z independent of unobserved demand conditional on X?Timing, assignment process, covariate balance and market knowledge.
    ExclusionCan Z affect sales through any path other than price?DAG, operational audit and a search for direct paths such as availability, quality or assortment.
    Instrument strengthIs the isolated variation informative enough for stable estimation?Weak-instrument diagnostics and weak-identification-robust intervals.
    HeterogeneityIs a homogeneous structural coefficient defensible?The example imposes a homogeneous β. With real data, document heterogeneity and do not automatically turn 2SLS into a universal elasticity.
    SensitivityDoes the result survive other plausible specifications?Declared alternative windows, controls, fixed effects, clusters, instruments and samples.

    Relevance and strength are measurable. Independence and exclusion cannot be tested from the file alone; they require external causal justification.

    11

    Diagnostics required before interpreting β

    1. 01

      First stage: Z coefficient, partial R-squared and robust statistic. F > 10 is not a universal validation rule.

    2. 02

      Weak instruments: complement the 2SLS interval with Anderson–Rubin or CLR-type robust inference when identification is uncertain.

    3. 03

      Uncertainty: use robust or clustered standard errors consistent with assignment unit and time dependence.

    4. 04

      Overidentification: available only with more excluded instruments than endogenous regressors. A non-rejection does not prove validity.

    5. 05

      Endogeneity: a non-significant Durbin–Wu–Hausman test does not prove price exogeneity, especially with a weak instrument.

    6. 06

      Sensitivity: compare OLS, 2SLS, alternative specifications and support domains without selecting the preferred result afterward.

    Bound, Jaeger & Baker (1995)

    12

    Interpret the result without overextending it

    In the stated homogeneous log-log model, βIV = −1.011 means that a 1% price increase induced by Z corresponds, under the IV and structural assumptions, to about 1.011% lower quantity on the studied support. For +5%, the finite calculation gives 100 × [(1.05)^−1.011208 − 1] = −4.81%. This conversion does not transport the effect beyond the studied population, period or cost mechanism.

    13

    What you may conclude

    Supported wording

    Under the identification assumptions and the homogeneous structural assumption explicitly imposed in this example, β describes the causal response to instrument-induced price variation within the analyzed population, period and support. A 1% increase then corresponds approximately to a β% quantity change; the exact finite-change calculation is 100 × [(1.01)^β − 1].

    Overstated wording

    This elasticity necessarily applies to every product, customer, period, price level and future pricing policy.

    14

    Possible marketing decision

    Possible decision: compare a price-increase scenario bounded by the observed support with a no-increase scenario, reporting volume, margin and uncertainty separately. Unsupported decision: automate an optimal price or extrapolate to uncovered products, customers or price levels. If exclusion or independence remain fragile, postpone the decision or use a safe experiment.

    15

    When not to use this method

    • No credible instrument is available.
    • The candidate instrument directly affects demand, availability, quality or assortment.
    • The first stage is too weak to yield an informative estimate.
    • There are too few observations or clusters for the required uncertainty assessment.
    • An ethical and operationally safe pricing experiment is feasible.
    • A known pricing rule enables a more direct identification strategy.

    16

    Reference implementations

    Python 3.13 · standard library · reference

    python msc-p020-reference.py --csv msc-p020-price-endogeneity.csv
    Download

    R 4.5 · ivreg / sandwich / lmtest · companion syntax not executed in this build

    library(ivreg); library(sandwich); library(lmtest)
    d <- read.csv('msc-p020-price-endogeneity.csv')
    fit <- ivreg(log_quantity ~ log_price + promotion + season_index |
      supplier_cost_shock + promotion + season_index, data=d)
    summary(fit, diagnostics=TRUE)
    coeftest(fit, vcov.=vcovCL(fit, cluster=d$product_id, type='HC1'))
    Official documentation

    IBM SPSS Statistics 31 · secondary syntax

    2SLS log_quantity WITH log_price promotion season_index
     /INSTRUMENTS=promotion season_index supplier_cost_shock
     /ENDOGENOUS=log_price
     /PRINT=COV.
    Official documentation

    SAS 9.4 · secondary syntax

    proc syslin data=pricing 2sls;
     endogenous log_price;
     instruments supplier_cost_shock promotion season_index;
     model log_quantity = log_price promotion season_index;
    run;
    Official documentation

    Python is the only reference executed in this build. The R block is companion syntax not executed here; SPSS and SAS are secondary and do not reproduce clustered inference. No syntax makes the instrument valid: verify versions, covariance and weak-instrument diagnostics in your environment.

    17

    Expected deliverable for a pricing decision

    1. 01

      Pricing question, population, unit of analysis, period and log-log estimand.

    2. 02

      DAG and written justification of relevance, independence and exclusion.

    3. 03

      Complete first stage, instrument strength and identified variation domain.

    4. 04

      2SLS coefficient, suitable interval, weak-instrument-robust inference and clustering plan.

    5. 05

      OLS/IV comparison, sensitivity checks, transport limits and explicitly unsupported decisions.

    18

    Scientific anchors

    1. Shows why endogeneity in marketing variables, including price, can bias parameters and decision uses of choice models.

    2. Clarifies that causal IV interpretation depends on explicit assumptions about the instrument, assignment and structural model.

    3. Establishes the bias and instability risk when instruments explain little of the endogenous variable.

    19

    Evidence level and transport limits

    Dossier level: method illustrated and replicated on synthetic data, with three full texts verified at claim level. This is neither empirical validation of a real instrument nor evidence of transportability. With real data, relevance is measurable; independence and exclusion remain causal assumptions to defend through the business process, causal graph and sensitivity analyses.

    Dataset · Tool

    Dataset · MSC-P-020-PRICE-ENDOGENEITYSynthetic endogenous price and cost shock→

    Method connections

    Parent territoryPricing science: connecting price, demand and contributionLimitsHow do you estimate price elasticity and its uncertainty?Compare withFixed or random effects: which panel model should you choose?

    Read next

    MSC-H-003Pricing science: connecting price, demand and contribution→MSC-P-022How do you estimate price elasticity and its uncertainty?→MSC-P-018Predictive or causal regression: what are you trying to estimate?→MSC-P-021Fixed or random effects: which panel model should you choose?→