Combining stated and revealed preferences: Matched and unmatched
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: I'm Ruth.
Warren: And I'm Warren, here with Iris.
Ruth: We're looking at the paper: "Combining stated and revealed preferences".
Warren: Here's the gist of the paper. Combining stated and revealed preferences allows researchers to correct for endogeneity arising from unobserved heterogeneity by using stated choices to identify this distribution, which is crucial when actual and hypothetical choices are observed in unmatched data sets.
Ruth: The key idea is to use stated choices to identify the distribution of individual unobserved heterogeneity, if this unobserved heterogeneity is the source of endogeneity, the researcher can correct for its influence in a demand function estimation using actual choices and recover causal effects.
Warren: Bounds on causal effects are derived in the case where stated choice and actual choices are observed in unmatched data sets, these data combination bounds are of independent interest.
Ruth: Identification Framework The framework posits that agents make a binary decision D based on an endogenous decision relevant attribute X.
Title and authors: Ruth: So we're looking at this paper called "Combining stated and revealed preferences: Matched and unmatched". It tackles how we can use hypothetical choices—what people *say* they would do—to fix a big problem when we only have real choices.
Warren: That big problem is endogeneity, right? When the things people actually choose are all mixed up by some hidden factors, and those hidden factors are what makes them say what they say in the hypothetical scenarios.
Iris: Exactly. The paper proposes using those stated choices to identify the distribution of that unobserved heterogeneity, which is crucial when we have actual and hypothetical choices observed in separate datasets.
Ruth: So if we can figure out that distribution, then we can correct for its influence when we're trying to estimate a demand function using the actual choices and actually recover those causal effects.
Warren: And they go further by deriving bounds on those causal effects even when the stated choice data and the actual choice data are completely unmatched. Those combination bounds are of independent interest to me, because that means we don't need them to be perfectly aligned.
Iris: The core idea hinges on using stated choices to identify the distribution of individual unobserved characteristics, which is a necessary step for correcting endogeneity in demand function estimation using actual choices >
Ruth: That sounds like it’s about using optimal transport theory to find these sharp bounds on the structural function, which is a really clever way to do it.
Warren: It's not just identifying the distribution; they are using those stated choices to recover counterfactual choice parameters, which basically means figuring out what people would have chosen under different conditions >
Iris: And they link this back to average structural function by defining it as the expectation of the choice function over all possible unobserved characteristics, which is a big conceptual step in tying everything together >
Ruth: So what does this mean for us practically? It means we can get more accurate estimates of how much things actually matter in a demand model even when our real data is messy.
Warren: The inference procedure they propose uses a bootstrap method to create one-sided confidence bounds for the upper bound, which helps us quantify how reliable those causal effects are when we only have unmatched data >
Iris: They use this procedure involving quantiles of certain distributions related to the parameter estimate, specifically involving the quantity b prime n and its quantile q over one minus alpha over two >
Ruth: It’s a lot of technical detail there, but essentially it's about getting a statistically sound way to make an inference when you're combining these two different types of data >
Warren: Simulation results show that the coverage rates for the true value of the average structural function are close to one, which is good because we expect that true value to be an interior point in the identified set >
Title and authors: Iris: They also found that those confidence bands are narrow and they get narrower with a larger sample size, which makes sense, but they also showed they shrink a bit with the tuning parameter xi over n >
Ruth: So it’s not just that it works on paper; the simulation experiments give us an idea of how much precision we can actually expect in real-world scenarios >
Warren: The authors suggest this approach is useful beyond just combining stated and revealed preferences, implying these combination bounds have broader applicability to other data integration problems >
Iris: They also point out that the inference procedure is particularly suited for empirical settings where you only have scalar or at most bivariate stated probability reports, which sets some clear boundaries on when to use it >
Ruth: So what's the big picture here? It seems like this framework offers a way to bypass some of the structural assumptions we usually have to make about how stated choices relate to actual choices >
Warren: It lets us correct for endogeneity without having those strong structural assumptions on that relation, which is pretty powerful when you’re dealing with real-world data where things are rarely perfectly clean >
Iris: The key contribution seems to be using the stated choice distribution to identify heterogeneity, which then allows us to recover causal effects in demand function estimation >
Ruth: It connects the world of hypothetical scenarios directly back to the world of what people actually do, making it a more robust way to understand human decisions >
Warren: And they are setting up future work by suggesting they might eventually identify a finite number of latent types from elicited preferences using methods from large factor models and group fixed effects literature >
Iris: That sounds like the next logical step for them, moving from continuous heterogeneity to discrete types in their modeling approach >
Ruth: So, to wrap up on this paper "Combining stated and revealed preferences: Matched and unmatched", it gives us a method to correct for unobserved factors using stated choices, even when our data is mismatched.
Warren: It provides bounds on causal effects derived from the combination of stated and actual choices that are useful in many situations >
Iris: The inference procedure offers a one-sided confidence bound for the average structural function by incorporating exclusion restrictions, which is useful for inferring when you can exclude certain observable characteristics >
Ruth: So, it’s a method that lets us use hypothetical reports to get real causal insights without needing perfect alignment between our two data sources.
Warren: It's a tool for when the data isn't perfectly matched but we still want to know what the actual effects are >
The paper's summary: Ruth: So, to recap, this paper is about using what people say they would do in hypothetical situations—stated preferences—to fix endogeneity problems when you only have real data that doesn't match up perfectly with those hypotheticals >
Warren: Right. The main point is that you can use those stated reports to identify the underlying distribution of unobserved characteristics, and that lets you correct the demand function estimation using your actual choices, which is crucial for getting the right causal effects >
Iris: And what they do really interesting is they derive bounds on those causal effects even when your actual data and stated choice data are totally unmatched. Those combination bounds are a sharp thing >
Ruth: It means we don't need perfect alignment between our two datasets to get some reliable estimates for the average structural function >
Warren: Exactly. That’s the practical implication—you can use these methods in real-world scenarios where your data isn't perfectly clean, and you still get a sense of what actually matters when you look at how people choose >
Iris: The way they do it involves using a bootstrap procedure to get a one-sided confidence bound for the upper bound, which helps quantify how certain you are about those causal effects >
Ruth: It’s not just about getting an estimate; it’s about knowing the range of possibilities for that estimate based on what you observe >
Warren: And they show in their simulations that these confidence bands don't get too wide even when the sample size is small, which suggests this method has some stability >
Iris: But they also pointed out that the procedure works best when you only have scalar or maybe two-way stated probability reports, so it’s not a magic solution for every kind of data collection >
Ruth: True. It’s a tool tailored for specific survey types, but it shows that using stated choice information is a valid way to address endogeneity without needing those big structural assumptions about how the choices relate to each other >
Warren: That makes sense. It gives researchers an extra lever when they're stuck between having real data and not being able to run the model correctly >
Iris: So, this shifts the focus from just matching data perfectly to using stated preferences as a powerful tool for identification, even when that match isn't there >
Ruth: Exactly. And I think it opens up new avenues for how we can use these preference reports in economic modeling across different fields >
The paper's improvements: Tom: So, we’re looking at how they suggest making this whole approach better about combining stated and revealed preferences >
Iris: They propose incorporating exclusion restrictions into their inference procedure to get a one-sided confidence bound for the average structural function >
Ruth: That means they’re saying if you can rule out certain observable characteristics from affecting the outcome on average, you can make a stronger claim about the distribution of those unobserved factors >
Warren: It’s like adding a filter to your data, cutting out stuff that we know doesn't matter for the big picture when we're trying to estimate how much something actually drives demand >
Iris: That filter helps with identification because it tightens up the constraints on what the structural function can be >
Ruth: And from an econometric standpoint, that’s solid. It moves us closer to a more robust inference than just relying on the raw combination of both datasets >
Warren: The numbers suggest this tightening actually makes sense, because if you exclude something irrelevant, your uncertainty should drop too >
Iris: But they also flag that this specific procedure is really best when you’re only dealing with scalar or maybe two-way stated probability reports, which is a limitation on the scope of the improvement >
Ruth: Right. So while it makes the method stronger for those specific data types, it still keeps those same boundaries on how much general applicability it has across all kinds of choice experiments >
Warren: It’s not a universal fix; it’s an optimization for specific data setups, which is always true in applied work >
Iris: So the improvement is less about a new way to handle *any* data and more about making the existing method more precise when you know what kind of reports you have >
Ruth: Exactly. It shows that even within these limitations, there’s room to get sharper, more reliable results from combining these two kinds of information >
Conclusion: Tom: So, to wrap up this discussion on "Combining stated and revealed preferences: Matched and unmatched," the main idea is that using stated choices helps us identify hidden factors, which lets us correct for endogeneity when estimating demand >
Ruth: It's about using what people *say* they would do to fix the problems we run into when we only have real choices, and it lets us recover causal effects even when our data isn't perfectly matched >
Warren: The combination bounds are a sharp thing, so you can get a reliable range for those effects even with messy data >
Iris: And they showed that these bounds are quite tight, which is good because it means the uncertainty in the estimate is actually quite controlled by the sampling interval >
Ruth: It’s not just about getting an estimate; it’s about knowing exactly how much wiggle room you have in your demand function parameters >
Warren: And I think that stability in the bounds is what makes this useful for a quantitative trader, because we can't afford huge uncertainties when making decisions >
Iris: From my econometric side, the identification method relies heavily on those exclusion restrictions to ensure we’re looking at the right variation doing the work >
Ruth: So we've seen how they use stated preferences as an identifier and how they set up a solid framework for inference with this paper >
Warren: It gives us a way to get causal insights without needing those huge structural assumptions about how people actually behave in the market >
Iris: It’s definitely useful for empirical work, though we have to remember that it’s best suited for scalar or maybe two-way stated probability reports >
Ruth: True. So this paper shows how we can use hypothetical choices to get real causal insights without needing perfect alignment between our two data sources >
Warren: It’s a tool for when the data isn't perfectly matched but you still want to know what the actual effects are, and it's got some good numerical grounding >
Iris: So we’re left with this method offering a way to use stated choice information to correct for endogeneity without needing those strong structural assumptions on the relationship between stated and actual choices >
R. Méango, M. Henry, Í MourifiƎ
University of Oxford · Penn State · Washington University in St Louis
econ.EM
Submitted: 2025-07-17
Updated: 2026-11-07
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 87/100
The gist: Combining stated and revealed preferences allows researchers to correct for endogeneity arising from unobserved heterogeneity by using stated choices to identify this distribution…
Key concepts
- Endogeneity Correction
- This method uses stated preferences, which reflect underlying unobserved traits (heterogeneity), to identify those traits. This allows researchers to adjust their analysis of actual choices so that the resulting estimates reflect true causal effects rather than biased correlations caused by hidden individual differences.
- Identification Framework
- The framework assumes a binary decision D depends on an endogenous attribute X, influenced by unobserved characteristics η. By modeling how stated preferences relate to this underlying characteristic vector η, the researchers aim to uniquely identify the structural choice function and subsequently recover counterfactual choice parameters.
- Data Combination Bounds
- These bounds provide rigorous limits on the causal effects when data sets are incomplete or unmatched. They show how much uncertainty remains in estimating demand functions even when combining stated and actual preferences, offering a measure of reliability for the derived causal estimates.
Terminology
Summary
Combining stated and revealed preferences allows researchers to correct for endogeneity arising from unobserved heterogeneity by using stated choices to identify this distribution, which is crucial when actual and hypothetical choices are observed in unmatched data sets. p. 0 The key idea is to use stated choices to identify the distribution of individual unobserved heterogeneity, if this unobserved heterogeneity is the source of endogeneity, the researcher can correct for its influence in a demand function estimation using actual choices and recover causal effects. p. 0 Bounds on causal effects are derived in the case where stated choice and actual choices are observed in unmatched data sets, these data combination bounds are of independent interest. p. 0
Identification Framework
The framework posits that agents make a binary decision D based on an endogenous decision relevant attribute X. Assumption 1 states that the structural choice function E[D(x)η] satisfies m(x, η):= E[D(x)X = x, η] = E[DX = x, η], so that recovering η would allow identification of m(x, η) and hence of the following counterfactual choice parameters. p. 5 The average structural function is defined as µ(x):= E[D(x)] = Z m(x, η) dFη. p. 5
Stated Preference Modeling
Stated preferences are modeled by assuming the stated choice function m˜ (xt, η) is driven by the same vector η of unobserved characteristics as the structural choice function m(xt, η). In general, the agent’s report Pt is driven by the same vector η of unobserved characteristics. p. 6 Assumption 2 guarantees that a unique η can be recovered from the system P:= (P1,. p. 7.., Pd) = (m˜ (x1, η),..., m˜ (xd, η)):= m˜ (x, η).
Data Combination Bounds
The fundamental idea is that stated preferences are useful insofar as they provide valuable information on individual heterogeneity. The sharp identified set for the structural function m(x, p) contains all the conditional expectations E[DX = x, P = p] compatible with marginal distributions FDX and FP. p. 9 Corollary 1 provides bounds on the average structural function µ(x) based on knowledge of the distributions of DX from revealed preferences and of PX from stated preferences. p. 3
Inference Procedure
Inference with unmatched data involves deriving a one-sided confidence bound for the upper bound UBx using a bootstrap procedure proposed by Fang and Santos (2019). This procedure yields the (1 − α)-level confidence region hϕ(ˆθx) − r−1n q 1−α/2, ϕ¯(ˆθx) − r−1n ϕ¯(⌊B(1−α/2)⌋) x, where q 1-α/2 and q¯α/2 are quantiles of the distribution of bϕ′n (rn(θ∗ − ˆθ)), and bϕ¯′n (rn(θ∗ − ˆθ)) respectively. p. 4
Simulation Results
Simulation experiments assess the informativeness of the data combination bounds and the performance of the proposed inference procedure. Table 1 reports the rate of coverage of the true value of µ(0) and the excess length of the confidence region, relative to the identified set. p. 8 Coverage rates close to 1 are expected, since the true value is an interior point in the identified set. p. 8 The confidence bands are narrow and decline both with the sample size, as expected, and somewhat also with the value of the tuning parameter ξn.
Discussion
The paper proposes to use stated choice probabilities to correct for endogeneity without structural assumptions on the relation between stated choice probabilities and actual choice. The bounds are useful beyond the combination of stated and revealed preferences considered in this work. p. 2 The inference procedure is best suited to empirical environments with scalar or at most bivariate stated probability reports. p. 8 The authors propose, in ongoing research, to identify a finite number of latent types from elicited preferences using methods from the large factor model and group fixed effects literature.
References
Adams-Prassl, A. and Andrew, A. p. 2 (2019). Preferences and beliefs in the marriage market for young brides. p. 1 CEPR Discussion Paper No. p. 0 DP13567.
Alm˚as, I. p. 1, Attanasio, O. p. 1 p. 1, and Jervis, P. (2024). Presidential address: Economics and measurement: New measures to model decision making. p. 1
Ameriks, J. p. 1 p. 1, Briggs, J. p. 2 p. 2, Caplin, A., Lee, M., Shapiro, M. p. 2 p. 2 D., and Tonetti, C. (2020a). Older americans would work longer if jobs were flexible. p. 3
Ameriks, J., Briggs, J., Caplin, A., Shapiro, M. D., and Tonetti, C. (2020b). Long-term-care utility and late-in-life saving. p. 1
Arcidiacono, P. p. 1, Hotz, V. J., Maurel, A., and Romano, T. (2020). Ex ante returns and occupational choice. p. 2
Attanasio, O., Cunha, F., and Jervis, P. (2019). Subjective parental beliefs. p. 1 their measurement and role. p. 1
Batista, C. p. 1, Costa, D. M., Freitas, P., Lima, G., and Reis, A. B. (2025). What matters for the decision to study abroad? p. 3 a lab-in-the-field experiment in cape verde. p. 0
Bernheim, B. p. 2 D., Bj¨orkegren, D., Naecker, J., and Pollmann, M. (2021). Causal inference from hypothetical evaluations. p. 1
Blass, A. A., Lach, S., and Manski, C. p. 1 p. 1 p. 1 F. (2010). Using elicited choice probabilities to estimate random utility models: Preferences for electricity reliability. p. 2
Bonnans, J. F. and Shapiro, A. p. 2 (2013). Perturbation analysis of optimization problems. p. 9
Bontemps, C. p. 4, Florens, J.-P., and Meddahi, N. (2025). Functional ecological inference. p. 3
Boyer, M. M., De Donder, P., Fluet, C., Leroux, M.-L., and Michaud, P.-C. (2020). Long-term care insurance: Information frictions and selection. p. 1
Briggs, J., Caplin, A., Leth-Petersen, S., Tonetti, C., and Violante, G. (2020). Estimating marginal treatment effects with survey instruments. p. 2
Cross, P. J. and Manski, C. F. (2002). Regressions, short and long. p. 1
de Bresser, J. p. 2 and van Soest, A. p. 2 (2019). The predictive power of subjective probabilities: probabilistic and deterministic polling in the dutch 2017 election. p. 2
Delavande, A. p. 1 and Zafar, B. p. 0 p. 0 p. 0 p. 0 (2019). University choice: The role of expected earnings, nonpecuniary outcomes, and financial constraints. p. 0
D’Haultfoeuille, X. p. 4 p. 4, Gaillac, C. p. 0 p. 0, and Maurel, A. (2024). Linear regressions with combined data. p. 4
D’Haultfoeuille, X., Gaillac, C., and Maurel, A. (2025). Partially linear models under data combination. p. 4
Fan, Y., Pass, B., and Shi, X. (2025). Partial identification in moment models with incomplete data via optimal transport. p. 4
Fan, Y., Sherman, R., and Shum, M. (2014). Identifying treatment effects under data combination. p. 2
Fan, Y., Sherman, R., and Shum, M. (2016). Estimation and inference in an ecological inference model. p. 0
Fang, Z. and Santos, A. p. 4 (2019). Inference on directionally differentiable functions. p. 4
Giustinelli, P. p. 0 p. 0 (2023). Expectations in education. p. 1
Giustinelli, P. and Shapiro, M. p. 2 D. (2024). Seate: Subjective ex ante treatment effect of health on retirement. p. 2
Gong, Y., Stinebrickner, T., Stinebrickner, R., and Yao, Y. (2022). The role of non-pecuniary considerations: Locations decisions of college graduates from low income backgrounds. p. 2
Haghani, M. p. 2 p. 2, Bliemer, M. C., Rose, J. M., Oppewal, H., and Lancsar, E. (2021a). Hypothetical bias in stated choice experiments: Part i. p. 1 macro-scale analysis of literature and integrative synthesis of empirical evidence from applied economics.
Haghani, M., Bliemer, M. C., Rose, J. M., Oppewal, H., and Lancsar, E. (2021b). Hypothetical bias in stated choice experiments: Part ii. p. 1 conceptualisation of external validity, sources and explanations of bias and effectiveness of mitigation methods. p. 1
Hainmueller, J. p. 2, Hangartner, D., and Yamamoto, T. (2015). Validating vignette and conjoint survey experiments against real-world behavior.
Hausman, J. p. 2 (2012). Contingent valuation: from dubious to hopeless.
Heckman, J. p. 4 p. 4 J. and Robb, R. p. 5 (1985). Alternative methods for evaluating the impact of interventions: An overview. p. 2
Heckman, J. J., Smith, J., and Clements, N. (1997). Making the most out of programme evaluations and social experiments: Accounting for heterogeneity in programme impacts. p. 7
Hong, H. and Li, J. (2018). The numerical delta method. p. 0
Horowitz, J. p. 3 L. and Manski, C. F. (1995). Identification and robustness with contaminated and corrupted data. p. 2
Hurd, M. D. (2009). Subjective probabilities in household surveys. p. 1
Ichimura, H. p. 4 and Martinez-Sanchis, E. p. 4 (2005). Identification and estimation of gmm models by combining two data sets. p. 4
Juster, F. T. and Shay, R. p. 1 P. (1964). Consumer sensitivity to finance rates.
Kesternich, I. p. 1, Heiss, F., McFadden, D. p. 1 p. 1, and Winter, J. (2013). Suit the action to the word, the word to the action: Hypothetical choices and real decisions in medicare part d. p. 2
Ko¸sar, G. p. 1 p. 1 and O’Dea, C. p. 2 (2023). Expectations data in structural microeconomic models. p. 4
Ko¸sar, G., Ransom, T., and Van der Klaauw, W. p. 2 p. 2 (2022). Understanding migration aversion using elicited counterfactual choice probabilities. p. 5
Lee, R. (2019). Identification and estimation of treatment effects with instrumental variables under data combination. p. 2
Low, C. (2024). Pricing the biological clock: The marriage market costs of aging to women. p. 1
Maestas, N. p. 2, Mullen, K. J., Powell, D., Von Wachter, T., and Wenger, J. B. (2023). The value of working conditions in the united states and implications for the structure of wages. p. 1
Manski, C. F. (2004). Measuring expectations. p. 1
Mas, A. and Pallais, A. p. 2 (2017). Valuing alternative work arrangements. p. 2
Mas-Colell, A. p. 7 (1979). Homeomorphisms of compact, convex sets and the jacobian matrix. p. 3
McFadden, D. (2017). Stated preference methods and their applicability to environmental use and non-use valuations. p. 1
M´eango, R. p. 0 p. 0 and Girsberger, E. p. 2 M. (2024). Identification of ex ante returns using elicited choice probabilities: an application to preferences for public-sector jobs. p. 6
M´eango, R. and Poinas, F. p. 0 (2023). The option value of overstaying. p. 7
Morikawa, T. p. 2, Ben-Akiva, M., and McFadden, D. p. 1 (2002). Discrete choice models incorporating revealed preferences and psychometric data. p. 2
Murphy, J. J., Allen, P. G., Stevens, T. H., and Weatherhead, D. (2005). A meta-analysis of hypothetical bias in stated preference valuation. p. 1
Pacini, D. (2019). Two-sample least squares projection. p. 8
Pantano, J. p. 2 and Zheng, Y. p. 2 (2013). Using subjective expectations data to allow for unobserved heterogeneity in hotz-miller estimation strategies. p. 0
R¨uschendorf, L. p. 4 (1991). Bounds for distributions with multivariate marginals. p. 4
Train, K. E. (2009). Discrete choice methods with simulation. p. 3
Van der Klaauw, W. (2012). On the use of expectations data in estimating structural dynamic choice models. p. 4
Villani, C. p. 9 (2003). Topics in optimal transportation, volume 58. p. 0
Wiswall, M. p. 2 p. 2 p. 2 and Zafar, B. (2015). Determinants of college major choice: Identification using an information experiment. p. 3
Wiswall, M. and Zafar, B. (2018). Preference for the workplace, investment in human capital, and gender. p. 1
Wiswall, M. and Zafar, B. (2021). Human capital investments and expectations about career and family. p. 1
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Counterfactual analysis correction: The improved system can correct for endogeneity arising from unobserved heterogeneity by using stated choices to identify this heterogeneity, allowing researchers to
recover causal effects
in demand function estimation using actual choices. -
Data combination bounds derivation: The system can derive bounds on causal effects when stated choice and actual choices are observed in unmatched data sets by utilizing the dual of optimal transport problems to establish sharp bounds on the structural function E[D|X = x, P = p].
-
Inference under exclusion restrictions: The improved system can provide a one-sided confidence bound for the average structural function by incorporating exclusion restrictions, allowing inference when observable characteristics Z are excluded from affecting the potential outcome D on average.
Sources
- Using Experiments to Correct for Selection in Observational Studies
- Linear Regressions with Combined Data
- Partial Identification in Moment Models with Incomplete Data--A Conditional Optimal Transport Approach
- Using Probabilistic Stated Preference Analyses to Understand Actual Choices
Related papers
- Stationarity with Occasionally Binding Constraints
- Backtesting High-Dimensional Expected Shortfall
- Higher-Order Neyman Orthogonality in Moment-Condition Models
- Online Bayesian Inference for a Dirichlet Process Mixture of Binary Logit Models
- Debiased Conditional Independence Tests for Causal Effects
- Sharp Testable Implications of Encouragement Designs