Influence Functions: A Unified Differentiation-Based Framework
summary
The gist
This paper introduces a direct differentiation-based framework that unifies the derivation of influence functions across parametric, nonparametric…
In short
A direct differentiation-based framework unifies influence function derivations across parametric, nonparametric, and semiparametric models. It establishes verifiable conditions for efficiency equivalence between joint and plug-in estimation by showing how the influence function is obtained through centering an identification function.
Key concepts
- Unified Derivation Framework
- This core method uses direct differentiation to find the influence function for any model type (parametric, nonparametric, or semiparametric). It shows that the Riesz representer of the functional derivative is found by projecting the identification function onto mean-zero functions, leading to a common algebraic form for all influence functions.
- Local Robustness
- Local robustness ensures that the asymptotic variance remains unchanged whether you use the true nuisance parameter or an estimator of it. The framework identifies an 'oracle-based' moment function that weakly dominates the original robust version, providing a reliable benchmark for multi-step inference.
- Joint vs. Plug-in Comparison
- The method uses orthogonal decomposition of moment functions to compare joint estimation with sequential plug-in estimation. It identifies a 'companion parameter' and establishes verifiable conditions under which plug-in estimation achieves the same asymptotic variance as joint estimation, even when nuisance parameters are over-identified.
Terminology used across episodes
This episode discusses
The paper
Influence Functions: A Unified Differentiation-Based Framework · Read on arXiv
Rutgers University
This paper introduces a direct differentiation-based framework that unifies the derivation of influence functions across parametric, nonparametric, and semiparametric models. We show that the Riesz representer of the functional derivative is obtained by orthogonally projecting the identification function onto the subspace of mean-zero functions, so the influence function is a linear transformation of the moment function. This approach extends seamlessly to infinite-dimensional parameters and reveals a common algebraic form across both finite- and infinite-dimensional cases. It demonstrates that the slower nonparametric convergence rate arises from approximating the infeasible Dirac delta function. The same framework further applies to semiparametric multi-step plug-in estimation, where it automatically yields locally robust moment functions. Crucially, by employing distributional derivatives, the method accommodates non-smooth moment conditions without imposing classical differentiability.
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: "Influence Functions: A Unified Differentiation-Based Framework".
Title and authors: Ruth: Let's talk about what the paper actually does in summary. Basically, they lay out a direct differentiation-based framework that uses the functional derivative to compute influence functions across all three model types—parametric, nonparametric, and semiparametric >
Warren: And the big payoff here is that for their semiparametric multi-step plug-in estimation, this method automatically gives you locally robust moment functions and an explicit closed-form expression for the adjustment term >
Iris: But they don't rely on classical differentiability here. They use distributional derivatives to handle those non-smooth moment conditions, which is a big deal because that opens up more kinds of models than just the smooth ones >
The paper's summary: Ruth: So, the core takeaway here is that they derive easily verifiable conditions for local robustness in both parametric and semiparametric models using this framework >
Warren: They also revisit Stein's question about when you can estimate a finite dimensional parameter efficiently without knowing an infinite dimensional nuisance parameter, and they show that local robustness answers that question directly >
Iris: That means the framework ensures the asymptotic variance stays the same whether you evaluate your moment function at the true value or at some consistent estimator of that nuisance parameter, which is what local robustness achieves >
The paper's improvements: Ruth: Then they tackle a major debate: joint versus plug-in estimation. They use an orthogonal decomposition of the moment function to directly compare these two approaches >
Warren: They identify a "companion parameter," which is this kind of parameter you need to identify for the joint estimation condition, and they show that an oracle moment function, called morc-LR beta, weakly dominates the locally robust version of the original moment function >
Iris: That dominance is what establishes verifiable sufficient conditions for efficiency equivalence between those two approaches in the semiparametric setting >
Conclusion: Ruth: So to wrap up, this paper with its unified differentiation-based framework provides a way to derive influence functions and establish verifiable conditions for efficiency equivalence between joint and plug-in estimation across parametric, nonparametric, and semiparametric models >
Warren: They construct an oracle-based locally robust moment function that weakly dominates the local robust version of the original moment function, which gives us a principled benchmark for multi-step inference when nuisance parameters are over-identified >
Iris: And they nail it by diagnosing whether plug-in procedures remain adaptive in cases like the average treatment effect, or if they lose efficiency due to a lack of local robustness in situations like the average treatment effect on the treated >
More episodes
- Daily Summary for 2026-11-19
- 2608.21274-Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix
- 2408.02757-A nonparametric test for diurnal variation in spot correlation processes
- 2306.16393-High-Dimensional Canonical Correlation Analysis
- 2605.17299-Geometric Brownian motion with intermittent entries and exits
- 2607.04753-Fooling Yourself: how narratives shape beliefs
- 2408.14671-Inference on Treatment Effects with Repeatedly Mismeasured High-Dimensional Controls
- 2201.06898-Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19