A practical econometrics toolkit fixes bias from imperfect ML-derived product proxies, producing more reliable counterfactuals and simple standard errors; validated in calibrated simulations and an application with held-out ground truth.

Main Finding

Treating ML-derived product proxies (embeddings) as the true product attributes in multi-product demand models can yield asymptotically biased counterfactuals and invalid inference. Christensen & Compiani (2026) develop a simple, post-estimation bias-correction and diagnostic toolkit that (i) removes the asymptotic bias from imperfect proxies, (ii) is asymptotically efficient in a broad class, (iii) provides closed-form standard errors, and (iv) is computationally light and compatible with fine-tuned black-box embeddings. In simulations and an experimental application with ground-truth second-choice data, the correction substantially improves counterfactual prediction (e.g., closest-substitute prediction rises from 40% to 70%).

Key Points

• Problem framed as model misspecification (latent product attributes e known to consumers but unobserved by researcher) rather than classical measurement error. Observed proxies ˜e (e.g., embeddings from images/text or reduced numeric attributes) can be imperfect.
• Many demand models (BLP/mixed logit and micro-BLP with product fixed effects) admit a reparameterization γ(θ, e) that captures how latent attributes interact with structural parameters. Counterfactuals of interest typically depend on (θ, e) only via γ.
• The naive estimator (plug ˜e into standard workflow: GMM on moments, then compute counterfactual) can be biased through (i) direct dependence of counterfactual on ˜e and (ii) indirect effects via estimated θ and implied mean utilities ˆξ.
• Bias-corrected estimator: take the naive counterfactual and subtract a simple correction term equal to a weighted average of the first-step GMM moments (market-level) and, when available, add corrections using micro moments. The correction has a closed-form weight vector and closed-form standard-error formulas.
  • Market-level correction (no microdata): ˆκbc = (1/T)∑t [kt(ˆγ) − ˆc′ Zt ˆξt(ˆγ)].
  • With microdata: adds further terms based on discrepancies between observed and model-implied micro moments.
• Diagnostics: two easy-to-compute Lagrange Multiplier (LM)-style statistics evaluated at naive estimates to assess proxy quality and guide the choice/number of proxies (e.g., how many embedding dimensions to include).
• The method is agnostic to the stochastic structure of proxy error and accommodates data-dependent proxies, including embeddings that are fine-tuned on choice data.
• Relation to existing work:
  • Shares orthogonalization ideas with double/debiased ML (DML), but differs because the nuisance objects here are latent product attributes and the proxies often come from black-box ML without validation data.
  • Distinct from measurement-error corrections that typically require validation data linking proxies to ground truth.
• Empirical performance:
  • Simulations calibrated to Nevo (2001) and other settings show the corrected estimator reduces bias and RMSE when proxy error is material; closed-form SEs give good coverage.
  • Experimental application (Compiani et al., 2025): estimate on first-choice data and predict second-choice behavior (true counterfactual). The bias correction improves closest-substitute prediction from 40% (naive) to 70% and diagnostics identify the best proxy sets.
• Implementation: integrates with existing demand estimation packages (e.g., PyBLP, xlogit). No additional optimization or bootstrapping required beyond standard estimation; terms use the same ˆξ and instruments Zt used in GMM.

Data & Methods

• Model settings:
  • Case 1: Market-level variation in prices with instruments (BLP-style). Observables: market shares st, prices pt, observed attributes ¯xt; unobserved product-level latent attributes e (constant across markets); proxies ˜e available.
  • Case 2: Individual choice data with product fixed effects (micro-BLP / marketing setups).
  • Utility examples: mixed logit with random coefficients (normal), micro-BLP with interactions between consumer demographics and product attributes.
• Key theoretical device: reparameterize θ and e into a composite γ(θ, e) capturing all ways e and θ jointly enter choice probabilities and counterfactuals of interest. This is what the correction targets.
• Estimation flow:
• Estimate structural parameters θ using standard GMM (market moments, optionally micro moments) treating ˜e as data (the naive first step).
• Form ˆγ = γ(ˆθ, ˜e) and compute implied mean utilities ˆξt(ˆγ).
• Compute naive counterfactual ˆκ = (1/T)∑t kt(ˆγ).
• Compute bias correction using closed-form weight vector ˆc (function of derivative matrices, instruments Zt, and moment Jacobians) and, if available, micro-data correction weights ˆdt.
• Obtain ˆκbc by subtracting (and adding micro correction) the weighted moment discrepancies. Closed-form asymptotic variance for ˆκbc is provided.
• Diagnostics: LM statistics based on whether moment discrepancies predict counterfactual bias; used to compare alternative proxies or dimensionality choices.
• Validation:
  • Simulations: calibrated to canonical IO datasets (Nevo) and to the empirical application. Evaluate bias, RMSE, and CI coverage across proxy-quality regimes.
  • Experimental ground-truth: CMS dataset with first and second choices permits out-of-sample counterfactual validation; bias-correction substantially improves counterfactual predictions.
• Computational requirements: minimal extra cost; uses outputs of standard packages (ˆθ, ˆξ, instruments, Jacobians); no bootstrap; supports embeddings coming from black-box or fine-tuned ML models.

Implications for AI Economics

• Practical guidance for applied researchers using ML embeddings from unstructured data (text, images, audio, surveys) in demand estimation:
  • Do not treat embeddings as ground-truth attributes by default. Use the proposed bias correction to protect counterfactuals from proxy-induced asymptotic bias.
  • Apply the LM diagnostics to choose embedding dimensions, ML models, or whether to include fine-tuned embeddings in the demand specification.
  • The method enables use of black-box/fine-tuned embeddings without requiring costly validation datasets linking embeddings to latent ground-truth attributes.
• Policy and applied analyses that rely on substitution patterns (mergers, entry, pricing, welfare) are especially sensitive to proxy quality. The correction strengthens confidence in counterfactual policy conclusions when proxies are imperfect.
• For research combining AI and economics:
  • Low computational overhead and compatibility with existing tools lowers the barrier to responsibly integrating embeddings into structural models.
  • The approach highlights a common failure mode: poor embeddings can cause mixed-logit models to effectively collapse to plain logit (understating substitution). The correction can recover richer substitution patterns and thus more accurate counterfactuals.
• Limitations and cautions:
  • The correction targets counterfactuals that depend on (θ, e) only through γ(θ, e). It does not identify or estimate the effect of changing the latent attributes e themselves (e.g., design changes measured in the latent scale).
  • Requires the usual identification elements of the demand model (instruments, valid moments) and that the model class admits the γ reparameterization.
  • Not a panacea for all misspecification: if the structural model family itself is badly misspecified (beyond proxy error), results may still be unreliable.
• Overall: this paper provides a practical, theoretically grounded bridge between ML-derived representations and credible structural counterfactual inference in demand analysis—an important step for rigorous, policy-relevant AI×economics work.