Conjoint or Discrete Choice Model¶
Choice model — instantiates Demand Curve Calibration and Response Design
Reconstructs demand from the ground up by making people choose among attribute bundles, recovering how much each feature — including price — is worth.
A Conjoint or Discrete Choice Model builds the demand picture from the bottom up. Instead of watching what people already buy and reading a curve off the trace, it hands respondents a sequence of constructed alternatives — each one a bundle of attributes set at specific levels, always including price — and asks which they would pick. From the pattern of those forced choices it estimates a part-worth: an implicit weight for every attribute level, with the marginal pain of a dollar carried on the same scale. Its defining move is that demand is assembled out of hypothetical tradeoffs among whole bundles, which is exactly what lets it price a feature — or an entire product — that does not yet exist to be observed.
Example¶
A streaming service is redesigning its subscription tiers and cannot simply watch the market, because the packages it is considering have never been sold. It runs a choice study: each respondent sees a dozen tasks, and every task offers three fictional plans that vary on monthly price ($8 / $12 / $16), ad load (none / light / heavy), simultaneous streams (1 / 2 / 4), and maximum resolution (HD / 4K), plus a "none of these" option. Nobody is asked "how much is 4K worth?" — they just keep choosing plans.
The fitted model turns those choices into part-worths: it might report that dropping from heavy to no ads is worth about as much as a $5 price cut, while going HD-to-4K barely moves anyone outside one segment. Because price is on the same scale, the team can read a willingness-to-pay for each feature and simulate the share each candidate bundle would capture against the others. The output is not "people like 4K" but "a $14 ad-free / 2-stream / HD plan out-pulls the current lineup and cannibalizes the top tier less than feared" — a demand estimate for a product that was, until that moment, only a proposal.
How it works¶
- Design the attribute space. Choose the handful of attributes that actually drive choice and the levels each can take, spanning a realistic range (including a credible price ceiling and floor).
- Present forced tradeoffs. Respondents repeatedly pick from small sets of full bundles; realism comes from making them give something up to get something.
- Fit a utility model. A logit-family model converts choices into part-worths — the estimated contribution of each level to the odds of being chosen.
- Simulate the market. Combine part-worths across a defined choice set to predict shares for bundles that were never actually offered, then vary price to trace sensitivity.
Tuning parameters¶
- Attribute and level set — more attributes capture more of the real decision but overload respondents and thin the data per level; keep to the few that move choice.
- Model form — a single average logit is cheap but hides heterogeneity; a mixed / hierarchical-Bayes model recovers per-segment part-worths at the cost of data and complexity.
- No-choice option — including "buy none" anchors the model to real walk-away behavior; excluding it forces artificial purchase and inflates demand.
- Price range and increments — the span over which price sensitivity can be read; too narrow and you cannot see the threshold, too wide and it strays past believable levels.
When it helps, and when it misleads¶
Its strength is pricing the counterfactual: it estimates demand for configurations that do not exist yet, decomposes a product into the value of its parts, and puts features and dollars in one currency so tradeoffs become explicit rather than argued. For launch and packaging decisions that is often the only evidence available.
Its central weakness is that stated choices are not revealed choices — respondents in a survey face no budget constraint and no consequence, so they routinely over-value features and under-weight price (hypothetical bias). The model is also vulnerable to the independence of irrelevant alternatives trap: a naïve logit predicts that adding a near-duplicate option steals share proportionally from everything, when in reality it mostly cannibalizes its twin.[n1] The guarding discipline is to keep a credible no-buy option, validate predicted shares against a holdout task or a small real-behavior check, and treat the part-worths as a structured hypothesis to be confirmed once actual purchases exist.
How it implements the components¶
generalized_cost_vector— price is entered as just one attribute among nonmoney costs (ad exposure, effort, restriction), so the estimate spans the whole cost vector rather than money alone.choice_set_and_substitution_map— every task is an explicit choice set; how respondents switch among bundles reveals the substitution structure directly.elasticity_and_threshold_profile— the price part-worth traces own-price sensitivity and locates the level at which a tier stops being chosen (its threshold).consumer_surplus_or_value_proxy— expressing each part-worth in dollars yields a willingness-to-pay proxy for every feature.
It does not measure how real, existing goods trade off when a rival's price moves (cross_elasticity_matrix, demand_shift_classifier) — that is Cross-Elasticity Matrix — nor does it record actual purchases (observed_response_evidence_base), which come from the Price Sensitivity Experiment and Revealed Preference Choice Log; this model works entirely from stated, hypothetical choices.
Related¶
- Instantiates: Demand Curve Calibration and Response Design — supplies a demand estimate for configurations that have never been sold.
- Sibling mechanisms: Cross-Elasticity Matrix · Demand Curve Estimation Workbook · Demand Segmentation Dashboard · Equity Access Impact Review · Price Sensitivity Experiment · Revealed Preference Choice Log · Scenario Demand Stress Test · Shadow Price Probe · Waitlist and Stockout Analysis
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Reconstructs demand from the ground up by making people choose among attribute bundles, recovering how much each feature — including price — is worth, making its operative form a computation, comparison, model, or analytic representation used to infer, estimate, or choose.
Independent corroboration: The frozen evidence defines Conjoint or Discrete Choice Model as 'Reconstructs demand from the ground up by making people choose among attribute bundles, recovering how much each feature — including price — is worth', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Economics & Finance
Origin pattern: Convergent development
Present-day reach: Multi-domain
Rationale: Random-utility and discrete-choice economics cohered inference of demand and willingness to pay from repeated choices among attribute bundles.
Related originating lineages:
- Psychology — Mathematical psychology developed conjoint measurement of preferences over multi-attribute alternatives.
- Statistics & Experimental Design — Designed choice experiments and estimation methods make part-worth recovery from constructed alternatives possible.
Review resolution: The reviewers agree on economics_finance with psychology and statistics_experimental_design as formative lineages. The source joins random-utility discrete choice with conjoint measurement and statistical estimation, a genuine convergence used across multiple substantive domains.
Attribution caveat: Modern conjoint choice models join discrete-choice economics to an earlier mathematical-psychology lineage.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] Independence of irrelevant alternatives — the property (of a plain multinomial logit) that the relative odds between two options do not depend on what else is available. The classic "red bus / blue bus" counterexample shows why: introducing a near-identical alternative should mostly split its twin's share, not draw evenly from every option. Mixed-logit and nested specifications relax it. ↩