Semiparametric Estimator

• Semiparametric estimators are methods that combine finite-dimensional parametric modeling with flexible nonparametric approximation for nuisance components.
• They achieve robust inference with root-n consistency and asymptotic efficiency even when parts of the model are unspecified.
• Practical implementations involve sieve approximations, two-step procedures, and influence function corrections across diverse statistical applications.

A semiparametric estimator occupies an intermediate position between parametric and nonparametric estimation, targeting finite-dimensional parameters of interest in settings where a part of the data-generating process is modeled parametrically while the remainder—often high- or infinite-dimensional—is left completely unspecified. The semiparametric approach allows for statistical inference with optimal properties (such as root-n rate and efficiency bounds) under model misspecification or high nuisance complexity, and underlies a diverse spectrum of models and estimators central to modern econometrics and statistics, including generalized linear models with unknown error distributions, partially linear and single-index models, causal effect estimation, and adaptive methods for stratified or informative sampling.

1. Foundations of Semiparametric Estimation

Semiparametric estimation is formalized in models indexed by a finite-dimensional target parameter (e.g., regression coefficients, treatment effects, or distributional parameters) and an infinite-dimensional nuisance (such as unknown link or error distributions). The canonical example is the semiparametric generalized linear model with exponential-tilt form:

$$f_{Y|X}(y|x) = \frac{\exp\{y x^T \beta + c(y)\}}{\int \exp\{t x^T \beta + c(t)\} \, d\mu(t)},$$

where $\beta \in \mathbb{R}^p$ is unknown and $c(\cdot)$ is an unrestricted function ($c(0) = 0$ for identifiability) (Lee et al., 2022). Semiparametric models generalize parametric approaches, allowing robust inference when strict likelihood or distributional assumptions are untenable, while attaining superior convergence rates and inferential precision compared to fully nonparametric estimation (Kennedy, 2017).

2. Canonical Semiparametric Estimator Construction

Estimation in semiparametric models often proceeds by combining parametrization for components of scientific interest with flexible, nonparametric approximation or explicit orthogonalization for nuisance structure. Key methodologies include:

• Approximate Maximum Likelihood/Sieve Approaches: The infinite-dimensional nuisance, such as $c(y)$ above, is approximated by a sieved basis expansion (e.g., B-splines, Hermite polynomials), yielding a finite-dimensional likelihood maximization. The estimator $(\hat\beta, \hat\gamma)$ is obtained by maximizing the approximate log-likelihood:

$$\ell_n(\beta, \gamma) = \sum_{i=1}^n \Big[Y_i X_i^T \beta + \gamma^T B(Y_i) - \log \int \exp\{t X_i^T \beta + \gamma^T B(t)\} dt\Big],$$

which is concave in $(\beta, \gamma)$ (Lee et al., 2022, Yu et al., 2022, Huang et al., 2022, Hristache et al., 2016).

• Two-Step Procedures: In models such as first-price auctions, a nonparametric estimator is first used to recover latent individual-level parameters (e.g., pseudo-values via local polynomial estimation), followed by $M$-estimation (e.g., GMM or likelihood equations) targeting the parametric component (Aryal et al., 2014). Analogous logic arises in penalized spline regression, where a parametric guide is combined with a nonparametric spline fit (Yoshida et al., 2012), and in series-based regularized IV for ill-posed Fredholm equations (Takahata et al., 2019).
• Influence Function and Orthogonalization: Estimators are constructed as solutions to estimating equations or as one-step corrections that exploit the efficient influence function to ensure double robustness and insensitivity to first-stage estimation errors, especially under high-dimensional or machine-learned nuisance fits (Kennedy, 2017, Pan et al., 2024).

3. Efficiency, Consistency, and Asymptotic Properties

Semiparametric estimators are designed to attain consistency, asymptotic normality, and (under regularity) the semiparametric efficiency bound—the minimal possible asymptotic variance achievable among regular estimators in the semi/nonparametric setting.

• Consistency and Root-n Rate: Under smoothness conditions, nonparametric approximation errors are controlled to be $o(n^{-1/2})$ (Lee et al., 2022, Yu et al., 2022). Index or parametric parts converge at the fast parametric rate, provided sieve complexity grows slowly relative to sample size (Hristache et al., 2016).
• Asymptotic Normality and Information Bound: The limiting law is Gaussian, with variance determined by the Schur complement of nuisance information or, more generally, the variance of the efficient influence function. For example, the AMLE for $\beta$ in a semiparametric GLM satisfies:

$$\sqrt{n}(\hat\beta - \beta_0) \overset{d}{\to} N(0, \Sigma^{*-1}),$$

where $\Sigma^*$ is the semiparametric efficient information matrix (Lee et al., 2022). Procedures for constructing efficient estimators under equality constraints or in submodels are also provided via tangent space projection and asymptotically linear corrections (Klaassen et al., 2016).

• Attainment of the Efficiency Bound: Explicit derivation and use of the efficient score, leveraging orthogonality to the infinite-dimensional nuisance tangent space, guarantees that plug-in or one-step estimators are semiparametrically efficient (Kennedy, 2017, Lee et al., 2022, Ma, 2010, Cheng et al., 2014). Extensions ensure preservation of efficiency under various contamination schemes (e.g., fixed-size case-control sampling (Ma, 2010)).

4. Major Applications and Model Classes

Semiparametric estimator methodologies span a diverse array of econometric and applied statistical models:

• Generalized Linear and Single-Index Models: Flexible nonparametric modeling of the mean function or link, with index parameter and sieve expansion for the nuisance, enables robust estimation of regression effects and marginal/quantile effects (Lee et al., 2022, Yu et al., 2022, Hristache et al., 2016, Han et al., 2023).
• Average and Heterogeneous Treatment Effects: Under unconfoundedness or extensions for non-ignorable assignment, semiparametric estimators of ATE and HTE are constructed using partially linear, influence-function based, or series-regularized methods (Yu et al., 2022, Takahata et al., 2019, Huang et al., 2022, Athey et al., 2021). Efficient estimators under thick-tailed or contaminated outcome distributions, and robustification with trimming or weighted quantile averages, are explicitly available (Athey et al., 2021).
• Partially Linear and Additive Models with Clustered/Longitudinal Data: Spline-based semiparametric GEE estimators permit modeling of both parametric and nonparametric covariate effects, with efficiency scores adapted to longitudinal correlation and arbitrary dependence structures (Cheng et al., 2014).
• Auction Models and Time Series: In auction settings, nonparametric inversion followed by GMM yields estimators that avoid curse-of-dimensionality but retain root-n properties (Aryal et al., 2014). In long-memory multivariate time series, local Whittle-type semiparametric estimators exploit spectral density approximations for robust estimation of memory parameters (Pumi et al., 2013).
• Sampling and Selection Models: Semiparametric adaptive estimators handle inference under informative sampling by treating sample weights as random variables, constructing influence-corrected estimators that achieve the optimal efficiency bound, even under model misspecification (Morikawa et al., 2022). Locally robust semiparametric estimation in sample selection models avoids exclusion restrictions, leveraging structural nonlinearity for identification and employing cross-fitted orthogonalized moments to enable root-n inference with modern ML methods (Pan et al., 2024).

5. Practical Implementation and Computational Considerations

Semiparametric estimators are typically implemented through convex optimization (e.g., AMLE, Lagrange multipliers for norm constraints), kernel or spline smoothing, series truncation, and functional projections. Moderate sieve sizes (e.g., $m \sim O(n^{1/5})$) balance bias and variance. Standard statistical software and generic optimizers suffice for typical problem sizes (e.g., $n \sim 10^3$) (Lee et al., 2022, Cheng et al., 2014). Plug-in variance estimators are derived from empirical influence function evaluations or sandwich expressions, and cross-fitting is frequently used to prevent overfitting when incorporating high-dimensional or ML-based first-stage fits (Pan et al., 2024). Choices of basis, tuning, and moment selection are model- and application-specific, with cross-validation or penalized criteria for regularization (Yoshida et al., 2012, Takahata et al., 2019, Moriyama, 2022).

6. Extensions, Robustness, and Empirical Performance

Semiparametric estimators prove robust under violations of parametric assumptions and demonstrate resilience to misspecification, loss of efficiency, or pathological sampling. Simulation studies confirm that when parametric models are correct, semiparametric estimators match or surpass efficiency, and when models are misspecified or tail behavior is ill-understood, they retain small bias and valid coverage (Lee et al., 2022, Yu et al., 2022, Northrop, 2015, Moriyama, 2022). Empirical applications, for instance to non-labor income models or birth weight analysis, show that semiparametric methods may identify significant effects and distributional heterogeneity missed by parametric alternatives (Lee et al., 2022, Yu et al., 2022, Huang et al., 2022).

7. Summary Table of Semiparametric Estimator Classes

Model/Context	Parametric Component	Nuisance Component	Construction/Key Feature
Semiparametric GLM (Lee et al., 2022)	Regression $\beta$	$c(y)$ (density tilt)	Spline-approximated AMLE
Single-index (Hristache et al., 2016, Han et al., 2023)	Index $\theta$	Link function $g$	Two-step PML or LS direction
Treatment effects (Yu et al., 2022, Huang et al., 2022)	ATE, index(s)	Link, propensity, outcome	Sieve, Hermite series, IPW/DR weighting
HTE under nonignorable assignment (Takahata et al., 2019)	HTE function	Propensity, density	Series-regularized 2SLS for Fredholm eq.
Penalized spline regression (Yoshida et al., 2012)	Parametric guide	Correction $r(x)$	Penalized spline fit with bias-based model
Informative sampling (Morikawa et al., 2022)	Regression/mean params	Sampling weights	Efficient score, adaptive influence corr.

• "Semiparametric Approach to Estimation of Marginal and Quantile Effects" (Lee et al., 2022)
• "A semiparametric efficient estimator in case-control studies" (Ma, 2010)
• "Semiparametric Estimation of Average Treatment Effect with Sieve Method" (Yu et al., 2022)
• "Semiparametric Penalized Spline Regression" (Yoshida et al., 2012)
• "Semiparametrically Efficient Estimation of Euclidean Parameters under Equality Constraints" (Klaassen et al., 2016)
• "Semiparametric Estimation of First-Price Auction Models" (Aryal et al., 2014)
• "Semiparametric estimation of heterogeneous treatment effects under the nonignorable assignment condition" (Takahata et al., 2019)
• "Efficient semiparametric estimation in time-varying regression models" (Truquet, 2017)
• "A semiparametric single-index estimator for a class of estimating equation models" (Hristache et al., 2016)
• "An efficient semiparametric maxima estimator of the extremal index" (Northrop, 2015)
• "Semiparametric Single-Index Estimation for Average Treatment Effects" (Huang et al., 2022)
• "Efficient semiparametric estimation in generalized partially linear additive models for longitudinal/clustered data" (Cheng et al., 2014)
• "A Semiparametric Estimator for Long-Range Dependent Multivariate Processes" (Pumi et al., 2013)
• "A semiparametric probability distribution estimator of sample maximums" (Moriyama, 2022)
• "Semiparametric Estimation of Treatment Effects in Randomized Experiments" (Athey et al., 2021)
• "Estimation of the Directions for Unknown Parameters in Semiparametric Models" (Han et al., 2023)
• "Semiparametric adaptive estimation under informative sampling" (Morikawa et al., 2022)
• "Semiparametric estimation of McKean-Vlasov SDEs" (Belomestny et al., 2021)
• "Semiparametric theory" (Kennedy, 2017)
• "Locally robust semiparametric estimation of sample selection models without exclusion restrictions" (Pan et al., 2024)