Estimating Nonseparable Selection Models: A Functional Contraction Approach

summary

The gist

Estimating Nonseparable Selection Models: A Functional Contraction Approach This research introduces a novel, full-information estimation strategy for nonseparable selection models…

In short

This research proposes a new method to estimate selection models where outcomes are not separable, meaning selection and outcome error terms are linked. The method shows that potential outcome distributions can be identified nonparametrically from observed data using an iterative algorithm. This allows researchers to estimate selection model parameters without assuming the outcome equation is parametric or separable, offering a full-information approach for complex empirical settings like consumer demand and auctions.

Key concepts

Nonseparable Selection Models
These are models where the process of selecting an individual and the error in their final outcome are intertwined. Unlike simpler models, they require a more complex estimation strategy because you cannot treat the selection mechanism and the outcome error as independent components.
Functional Contraction Mapping
This is a mathematical tool used to prove that a specific iterative process will reliably find the true underlying potential outcome distributions. It ensures that starting from any guess, repeatedly applying this mapping will converge to the correct distribution, provided certain 'contraction' conditions are met.

Terminology used across episodes

This episode discusses

The paper

Estimating Nonseparable Selection Models: A Functional Contraction Approach · Read on arXiv

Fan Wu Yi Xin

Peking University HSBC Business School · California Institute of Technology

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Estimating Nonseparable Selection Models".

Warren: Here's the gist of the paper. Detailed Research Summary: Estimating Nonseparable Selection Models: A Functional Contraction Approach This research introduces a novel, full-information estimation strategy for nonseparable selection models, specifically designed to estimate unobserved potential outcome distributions and selection function parameters without imposing restrictive parametric or separability assumptions on the underlying outcome equation.

Ruth: Core Theoretical Contribution (Identification): The central insight is establishing a one-to-one mapping between the realized selected outcome distributions and the underlying potential outcome distributions.

Warren: This relationship is formalized through a fixed-point problem characterized by an operator whose fixed point corresponds to the true potential outcome distributions.

Ruth: Crucially, under suitable contraction conditions (Theorem 1), this operator is shown to be a functional contraction mapping.

Paper summary: Ruth: So this paper is about estimating these nonseparable selection models without having to put any strict rules on how the outcomes themselves are structured, right?

Warren: Yeah, it's proposing a new way to figure out the potential outcome distributions from what we actually observe in the selected outcome distributions, and they claim you can do that nonparametrically using just an iterative algorithm.

Iris: I'm interested in how they establish that link between those two types of distributions. It seems like their core idea is setting up a fixed-point problem where the true potential outcome distributions are the fixed point of some mapping.

Ruth: Exactly, and they’re relying on some contraction conditions to show that this mapping actually works as a functional contraction. That means if you start with an approximation, it gets closer to the truth with each step of that algorithm.

Warren: And they put this identification result into a two-step estimation strategy. They first estimate the outcome distributions and then use those results in a second step to find the parameters for that selection function itself.

Iris: That two-step process sounds robust, but what's the catch? What assumptions are they making about the selection function or those potential outcomes that might limit how well this works in practice?

Ruth: Well, the whole point is they don't impose parametric assumptions on the outcome equation itself, which is a big thing for flexibility. They allow for nonseparable error terms and let that unobservable component show up directly in agent preferences over alternatives.

Warren: So from my side as someone looking at strategy costs, does this full-information approach mean we can estimate selection models where the selection mechanism involves things we can't easily measure?

Iris: They explicitly say their framework permits the unobservable component in the outcome equation to enter directly into agents’ preferences, which is how they capture selection on unobservables. But they do note that this requires a specific structure for how those potential outcomes are independent conditional on observable and unobservable characteristics.

Ruth: That's a key constraint, because if you can't assume that independence conditional on observables and unobservables, the identification might break down entirely. But under the right conditions, they show consistency and asymptotic normality for their estimators.

Warren: Consistency is good for theory, but I need to know about the actual computational cost. How fast does this iterative algorithm run? Is it constant time or something that scales poorly with data size?

Iris: They do detail a finite-iteration approximation scheme, showing that if you take enough iterations, say m of n tending to infinity, the approximate criterion converges uniformly to the true maximizer. That's what gives them consistency.

Ruth: So for the listener driving or cooking, think of it this way: they’ve built a mathematical engine that lets you map what you see—like prices or bids—back to what actually happened before selection occurred, without needing to guess the exact shape of those underlying potential outcomes.

Warren: It sounds like this method is really powerful because it tackles the problem from both sides simultaneously: figuring out what agents *would* do and figuring out *why* they chose what they did. That's a lot of information to process at once.

Iris: And they show it works in practice, with Monte Carlo simulations demonstrating that the approach performs well even in finite samples, which is always good news when you're dealing with real-world data from things like consumer demand models or auctions.

Ruth: It definitely opens things up for settings where traditional methods would fail, like when you only have transaction prices in a consumer demand study. It’s applicable to a wide range of empirical settings, from those to Roy-type models in labor economics with wage data.

Warren: The implication for us is that if we're building models around selection—say, predicting consumer behavior or auction outcomes—we don't have to restrict our modeling assumptions just because the underlying error terms are messy. That flexibility is where a lot of real-world modeling gets stuck.

Iris: And they mention that this method works even when the selection function incorporates alternative specific unobserved heterogeneity, which means it handles more complex choice behaviors than some simpler models do.

Ruth: So to sum up, the paper "Estimating Nonseparable Selection Models: A Functional Contraction Approach" proposes a way to nonparametrically identify potential outcome distributions from selected ones using an iterative algorithm based on contraction mapping, all while keeping the underlying outcome equation flexible and without parametric assumptions.

Warren: It’s a full-information strategy that establishes consistency and asymptotic normality for the estimators, making it a solid tool for empirical work in complex selection environments.

Iris: The main thing to keep in mind is that their identification relies on those suitable contraction conditions holding true, which is a technical hurdle you have to satisfy before you can use this method reliably.

Ruth: And the way they do the two-step estimation strategy—first getting the distributions and then optimizing the selection function parameters—is what makes it a complete picture for modeling selection.

Warren: For anyone interested in how these models relate to real decisions, this work suggests we can get a much richer view of agent behavior when we only have partial data, like just observing transaction prices.

Iris: It's also worth noting that they demonstrate the method works well in finite samples through simulations, which gives researchers a good sense of its practical utility right away.

Conclusion: Ruth: So we're wrapping up this look at "Estimating Nonseparable Selection Models: A Functional Contraction Approach." This paper is about figuring out how to estimate those unobserved potential outcome distributions without making big assumptions about the shape of the outcome equation itself.

Warren: Right. It’s essentially a new way to link what we actually observe—the selected outcomes—back to what *would* have happened if no selection happened at all.

Iris: The core idea is that you can do this nonparametrically, meaning you don't need to guess the exact functional form of those potential outcomes; you just use an iterative process based on a contraction mapping.

Ruth: That iterative algorithm, it’s the engine here for identification. It turns observed data into a distribution of what might have been.

Warren: And they show that this process is consistent and asymptotically normal, which means the estimates they get from this method are reliable when you run it over a lot of data.

Iris: But there's that technical bit about the contraction conditions being suitable; if those math rules don't hold up, the identification breaks down. That’s a key caveat.

Ruth: So what does this mean for us in the real world? It means we can tackle selection models where the error terms are messy and nonseparable, which is a huge headache for traditional methods.

Warren: It gives us a full-information approach, letting us estimate selection models without those restrictive parametric assumptions on the outcome equation itself. That’s pretty flexible.

Iris: The paper shows it works in consumer demand models and auctions with stochastic winning rules, so the applicability across different empirical settings is quite broad.

Ruth: Exactly. It opens the door to modeling agent behavior where they might be making choices based on things we can't easily measure, and it does that while keeping the underlying model flexible.

More episodes

← Home