Abstract
Multinomial choice models are fundamental for empirical modeling of economic choices among discrete alternatives. This paper analyzes the identification of binary and multinomial choice models when choice utilities are nonseparable in observed attributes and multidimensional unobserved heterogeneity. Using both cross-section and panel data, we demonstrate that derivatives of choice probabilities with respect to continuous attributes are weighted averages of utility derivatives. We extend these results to models with endogenous heterogeneity using control functions and panel data. In time-stationary panel models, we find that differences in derivatives over time identify utility derivatives "on the diagonal"—where observed attributes are identical across periods. However, we also prove that time stationarity does not identify structural derivatives "off the diagonal" in either continuous or multinomial choice panel models.
Introduction
In empirical economics, understanding how individuals choose between discrete alternatives—such as brands of consumer goods or modes of transportation—is a cornerstone of policy and market analysis. Traditional models often rely on "separability," assuming that unobserved individual preferences simply add a layer of noise to the utility function. However, real-world behavior is rarely so simple. A consumer's hidden risk aversion or brand loyalty often interacts with observable factors like price or quality in complex, non-additive ways. This paper, authored by Victor Chernozhukov, Iván Fernández-Val, and Whitney K. Newey, provides a rigorous theoretical framework for identifying these "nonseparable" effects in both snapshot (cross-section) and over-time (panel) data.
The Challenge of Nonseparability
The core problem addressed is "identification": can we determine the true structural effect of a variable (like a price increase) on utility if we cannot observe the complex, multidimensional hidden traits of the individuals? In a nonseparable model, the unobserved heterogeneity is not just an error term; it is a vector of traits that can change how an individual responds to any given attribute. Previous research often required restrictive assumptions to solve this. This paper relaxes those constraints, focusing on the mathematical relationship between observed choice probabilities and the underlying utility functions.
Methodological Framework
The authors utilize advanced econometric theory to derive identification results, primarily through the analysis of derivatives:
1. Cross-Sectional Analysis
The paper proves that the derivative of a choice probability—the rate at which the likelihood of choosing an option changes as an attribute changes—is a weighted average of the underlying utility derivatives. In the specific case of random coefficient models, the authors show that at the point of indifference (zero threshold), the probability derivative is directly proportional to the population mean of the coefficients.
2. Handling Endogeneity
To account for endogeneity (where observed variables are correlated with unobserved ones, such as high-end products being marketed specifically to high-income individuals), the authors employ control functions. This technique "cleans" the data, allowing researchers to recover structural effects even when the independence between observed and unobserved factors is violated.
3. Panel Data and Time Stationarity
In panel data settings where the same individuals are observed over two periods, the authors assume time stationarity—the idea that the distribution of hidden traits remains stable over time. By comparing the derivatives of choice probabilities across these two periods, they can cancel out the bias introduced by time-invariant hidden factors.
Key Insights and Results
The "Indifference" Filter
A vital insight of the paper is that probability derivatives act as a filter. They primarily capture the effects for "marginal" individuals—those who are currently on the fence between two choices. The resulting math shows that while the population is diverse, the observable change in probability is driven by these specific cases.
The Power and Limits of the Diagonal
The research introduces a distinction between "on the diagonal" and "off the diagonal" identification:
- On the Diagonal: When an individual's attributes are the same in both time periods, the utility derivatives are perfectly identified.
- Off the Diagonal: Conversely, the authors provide a formal non-identification proof showing that time stationarity cannot identify utility effects when attributes change between periods. This means that two different, equally plausible models could produce the same observed data, making it impossible to distinguish between them without further assumptions.
Formal Theorems
- Theorems 1 & 2: Establish that in binary choice, the derivative of choice probability is a scalar multiple of the average utility derivative at the indifference point.
- Theorem 4: Extends this to multinomial choices (choosing between $J$ alternatives), showing the derivative is a sum of weighted utility effects across all options.
- Theorem 9: Provides identification for multinomial panel data, showing that coefficients for continuous regressors are identified "up to scale" even with bounded data.
Conclusion and Future Directions
This paper advances the field by providing the "minimal conditions" required to understand complex discrete choices. While it highlights the power of derivative-based identification, it also sounds a cautionary note regarding the limitations of two-period panel data for identifying non-marginal effects. Future work may explore how to overcome the "off-diagonal" identification gap by utilizing more than two periods of data or by finding additional conditions that allow for broader population-level inferences.
Summary updated