Average Marginal Effects in One-Step Partially Linear Instrumental Regressions

This paper proposes a novel, easy-to-implement procedure using Reproducing Kernel Hilbert Space methods and a Bayesian bootstrap for inference to estimate and test average marginal effects in partially linear instrumental regressions, demonstrating its consistency, asymptotic normality, and strong finite-sample performance through simulations and empirical applications.

The Gist

Imagine you are a detective trying to figure out the true cause of a crime, but the only witness you have is a bit of a liar. This is the core challenge in many economic studies: trying to measure the effect of something (like education or advertising) when that "something" is influenced by hidden factors that also affect the outcome.

This paper, written by Lucas Girard and Elia Lapenta, introduces a new, smarter way to solve this detective work. Here is the breakdown in simple terms:

The Problem: The "Lying Witness" and the "Rigid Map"

In economics, we often use Instrumental Variables (IV). Think of this as a "witness" who tells the truth about a specific variable (like a treatment) but doesn't lie about the outcome.

• The Goal: We want to know the Average Marginal Effect (AME). In plain English: "On average, if I change the treatment by a tiny bit, how much does the result change?"
• The Old Way: For a long time, researchers assumed the relationship was a straight line (like a ruler). If you add one student to a class, test scores drop by exactly 0.5 points. This is easy to calculate, but it's often wrong. Real life is rarely a straight line; it's more like a winding mountain road.
• The New Challenge: If you try to map that winding road without assuming it's straight, the math gets incredibly messy. Previous methods required a complex, multi-step process where you had to tune many different "dials" (parameters). If you turned one dial wrong, your whole map was useless. It was like trying to bake a cake by adjusting the oven, the mixer, and the timer separately, with no guarantee they would work together.

The Solution: The "One-Knob Radio" and the "Flexible Rubber Sheet"

The authors propose a new method that is like upgrading from a complicated, multi-dial radio to a sleek, modern device with just one knob.

1. The One-Step Approach: Instead of doing a messy, multi-step regression (estimating the witness, then the liar, then the result), they do it all in one single step. They use a mathematical framework called Reproducing Kernel Hilbert Space (RKHS).

   • The Analogy: Imagine trying to fit a piece of fabric over a bumpy rock. Old methods tried to cut the fabric into specific shapes first. This new method treats the fabric as a smart, flexible rubber sheet that naturally stretches and molds to the shape of the rock without needing to be pre-cut. It finds the perfect fit automatically.

2. The Single Knob (Regularization): The only thing you need to adjust is one "knob" (a regularization parameter). This knob controls how much the rubber sheet is allowed to wiggle.

   • If the knob is too loose, the sheet wiggles too much and follows the noise (overfitting).
   • If it's too tight, the sheet stays too flat and misses the bumps (underfitting).
   • The beauty of this method is that you only have to tune this one knob, making it much easier for researchers to use in real life.

3. The Bayesian Bootstrap (The "What-If" Simulator): Because the math behind the rubber sheet is so complex, calculating the "margin of error" (confidence intervals) is a nightmare. The authors invented a clever trick called the Bayesian Bootstrap.

   • The Analogy: Imagine you have a bag of marbles representing your data. To see how reliable your result is, you shake the bag, pull out marbles with different weights (some heavy, some light), and re-calculate the result thousands of times. This creates a "simulation" of what could have happened. If the result stays consistent across all these simulations, you know it's solid. This avoids the need for complex, scary formulas.

Why Does This Matter? (The Real-World Tests)

The authors didn't just write theory; they tested it on three real-world scenarios to prove it works, even with small amounts of data:

1. Class Size and Grades: They re-analyzed a famous study on whether smaller class sizes improve test scores. The old linear method said "Yes, smaller classes help!" But their new, flexible method said, "Actually, the data is too messy to be sure." This suggests that the old conclusion might have been an illusion caused by forcing a straight line onto a curved reality.
2. Trade and Income: They looked at whether international trade boosts a country's income. Even with a small sample of only 150 countries, their method found a clear, positive effect, proving it works well even when data is scarce.
3. Newspapers and Ads: They studied how ads affect newspaper readership. The relationship wasn't a straight line; too many ads might actually scare readers away. Their method captured this "inverted U" shape perfectly, whereas a straight-line model would have missed it entirely.

The Takeaway

This paper gives economists and policymakers a powerful, easy-to-use tool.

• Before: You had to assume the world was a straight line, or use a complex, error-prone multi-step process to see the curves.
• Now: You can use a "one-knob" method that lets the data tell you the shape of the relationship, whether it's a straight line, a curve, or a rollercoaster.

It's like switching from a rigid ruler to a flexible measuring tape that adapts to the object you are measuring, giving you a much truer picture of reality.

Technical Summary

1. Problem Statement

The paper addresses the challenge of estimating and conducting inference on the Average Marginal Effect (AME) in a Partially Linear Instrumental Variables (IV) model. The model is specified as:
$$Y = h_0(Z) + X^T\beta_0 + \varepsilon, \quad \text{with } E\{\varepsilon|W, X\} = 0$$
Where:

• $Y$ is the outcome variable.
• $Z$ is an endogenous continuous treatment.
• $X$ is a vector of exogenous covariates.
• $W$ is a continuous instrument.
• $h_0(\cdot)$ is an unknown nonparametric function.
• $\beta_0$ is a vector of linear coefficients.

The target parameter is the AME, defined as $\theta_0 := E\{h'_0(Z)\}$.

The Core Difficulty:
While the AME is a crucial metric for policy analysis (e.g., returns to education), estimating it in nonparametric IV settings is notoriously difficult. Existing methods typically rely on multi-step regression approaches (e.g., estimating the conditional expectation $E[Y|W,X]$ first, then $h_0$, then the derivative). These approaches suffer from:

1. Complex Tuning: They require selecting multiple regularization parameters (one for each step), which is computationally burdensome and prone to suboptimal selection.
2. Error Accumulation: Estimation errors from the first step propagate to the final estimator, complicating finite-sample performance and inference.
3. Intractable Variance: The asymptotic variance of the resulting estimators often has a complex analytical form, making standard hypothesis testing difficult.

2. Methodology

The authors propose a novel one-step estimation procedure based on Reproducing Kernel Hilbert Spaces (RKHS) and a Bayesian bootstrap for inference.

A. Estimation Framework (RKHS)

Instead of a multi-step approach, the authors reformulate the moment condition $E\{Y - h_0(Z) - X^T\beta_0 | W, X\} = 0$ using a characteristic function approach (inspired by Bierens, 2016). They define a population objective function $M(\beta, h)$ based on the $L_2$ norm of the conditional moment restriction transformed by a characteristic function.

The estimator $(\hat{\beta}, \hat{h})$ is obtained by minimizing a penalized empirical objective function:
$$(\hat{\beta}, \hat{h}) = \arg \min_{\beta, h} \left( M_n(\beta, h) + \lambda \|h\|_H^2 \right)$$
Where:

• $M_n$ is the empirical counterpart of the objective function involving a symmetric characteristic function $F_\mu$.
• $\|h\|_H$ is the norm in a Reproducing Kernel Hilbert Space (RKHS).
• Key Feature: The procedure relies on a single regularization parameter $\lambda$.

Closed-Form Solution:
By leveraging the Representer Theorem, the authors show that the solution for $\hat{h}$ lies in the span of the kernel functions evaluated at the data points: $\hat{h}(\cdot) = \sum_{i=1}^n \hat{\alpha}_i K(\cdot, Z_i)$. This reduces the infinite-dimensional optimization problem to a finite-dimensional linear system involving the Gram matrix of the kernel and the characteristic function matrix. The AME is then estimated as $\hat{\theta} = \frac{1}{n}\sum \hat{h}'(Z_i)$.

B. Inference (Bayesian Bootstrap)

The asymptotic distribution of $\sqrt{n}(\hat{\theta} - \theta_0)$ is shown to be normal, but its variance has a complex analytical form involving unknown operators. To circumvent the need to estimate this variance directly, the authors propose a Bayesian Bootstrap procedure:

1. Generate random weights $\xi_i$ (e.g., from an Exponential distribution) with mean 1.
2. Re-estimate the model using these weights in the objective function.
3. Compute the bootstrap statistic $\sqrt{n}(\hat{\theta}^b - \hat{\theta})$.
4. Construct confidence intervals and p-values using the empirical distribution of the bootstrap statistics.

3. Key Contributions

1. One-Step Estimation: The paper introduces a single-step estimator for the AME in partially linear IV models, eliminating the need for multi-step regressions and the associated accumulation of estimation errors.
2. Single Regularization Parameter: Unlike existing series-based or multi-step kernel methods that require tuning multiple parameters, this method requires tuning only one parameter ($\lambda$), significantly simplifying practical implementation.
3. RKHS Framework: It successfully adapts RKHS methods (common in machine learning) to the econometric IV setting, providing a computationally tractable closed-form solution for the nonparametric component and its derivative.
4. Valid Bootstrap Inference: The authors establish the theoretical validity of the Bayesian bootstrap for this specific estimator, proving that it correctly approximates the limiting distribution of the AME estimator even when the analytical variance is intractable.
5. Finite-Sample Performance: The method demonstrates superior performance in small to moderate samples compared to traditional two-step series estimators.

4. Theoretical Results

• Consistency and Asymptotic Normality: Under standard identification assumptions (completeness of the instrument) and source conditions (smoothness of the true function $h_0$), the estimator $\hat{\theta}$ is consistent and asymptotically normal.
• Rate Conditions: The results hold provided the regularization parameter $\lambda$ satisfies $n\lambda \to \infty$ and $n\lambda^2 = o(1)$.
• Bootstrap Validity: The paper proves that the Bayesian bootstrap distribution converges to the true limiting distribution of the statistic, ensuring correct size (Type I error control) and consistency (power) for hypothesis testing.

5. Empirical and Simulation Evidence

• Simulations:
  • The authors compare their "one-step" RKHS method against a standard "two-step" series regression method (Ai & Chen, 2007).
  • Size Control: The proposed method maintains rejection rates close to nominal levels (5%, 10%) even in small samples ($n=100$), whereas the two-step method often under-rejects.
  • Power: The one-step method exhibits significantly higher power in small samples, particularly for non-polynomial functional forms and high endogeneity.
• Empirical Applications:
  1. Class Size Effects (Angrist & Lavy, 1999): Re-analyzing the effect of class size on test scores ($n=2,024$). The RKHS method finds no statistically significant effect, contrasting with the significant negative effects found by 2SLS. This highlights the risk of misspecification in linear models.
  2. Trade and Income (Frankel & Romer, 1999): Analyzing trade shares on GDP per capita ($n=150$). The method detects a positive, significant effect, confirming the robustness of the finding even with a small sample and relaxed linearity assumptions.
  3. Newspaper Demand (Sokullu, 2016): Estimating network effects in two-sided markets ($n=117$). The method provides a formal inference procedure for the AME of advertising on reader demand, yielding results consistent with but more flexible than the parametric specifications in the original study.

6. Significance

This paper bridges the gap between advanced machine learning techniques (RKHS) and rigorous econometric inference in instrumental variable settings. Its primary significance lies in practical usability:

• It offers a "plug-and-play" solution for researchers dealing with endogenous treatments where linearity assumptions are suspect.
• By requiring only one tuning parameter and providing a simple bootstrap inference routine, it lowers the barrier to entry for conducting nonparametric IV analysis.
• It provides a robust alternative to 2SLS when the treatment effect is non-linear, preventing misleading policy conclusions derived from misspecified linear models.

The authors have made the implementation available via an R package (rkhsiv), further facilitating the adoption of this methodology in applied economics.