Quasi-Maximum Likelihood Estimator (QMLE)

• QMLE is a parametric estimator that maximizes a surrogate likelihood function to achieve robust inference under model misspecification.
• It ensures consistency and asymptotic normality for pseudo-true parameters across diverse models like GARCH, logistic regression, and factor models.
• Algorithmic implementations using Newton–Raphson or EM methods provide computational tractability while addressing issues such as incidental parameter bias.

A quasi-maximum likelihood estimator (QMLE) is a parametric estimator that maximizes a surrogate (quasi-) log-likelihood function rather than the exact likelihood, with the goal of achieving large-sample inference when the true data-generating process is unknown, misspecified, or only partially characterized. QMLE methods are foundational in modern econometrics, statistics, time series analysis, and machine learning, providing a unifying approach for estimation and inference under heterogeneity, non-Gaussianity, or structural model uncertainty.

1. General Definition and Motivation

Quasi-maximum likelihood estimation replaces the intractable or misspecified likelihood function with a tractable quasi-likelihood surrogate, typically derived from a working parametric family such as the Gaussian, logistic, Laplace, or a finite mixture (Yoshida et al., 2022). Formally, if θ is a parameter of interest and $\ell_n(\theta)$ is the quasi-log-likelihood based on sample $X=(X_1,...,X_n)$, the QMLE is defined as:

$$\hat{\theta}_n = \arg\max_{\theta \in \Theta} \ell_n(\theta).$$

The motivation is to provide an estimator that is consistent and asymptotically normal for the pseudo-true parameter (i.e., the minimizer of the Kullback–Leibler divergence between the quasi-model and the true generative process), even when the imposed likelihood is not the true one. This approach provides tractability and robustness in a wide variety of misspecified models, including but not limited to GARCH, GLM, binary choice, factor models, and high-dimensional network settings (Yoshida et al., 2022, Chang et al., 5 May 2025, Chen et al., 29 May 2025, Bai, 2023, Barigozzi, 2023, Wang et al., 31 Dec 2025).

2. Classical and Modern QMLE Theory

Under standard regularity conditions—unique maximizer, differentiability, identification, compactness, and appropriate mixing/moment assumptions—the QMLE $\hat{\theta}_n$ enjoys the following large-sample properties (Yoshida et al., 2022):

• Consistency: $\hat{\theta}_n \to_p \theta_*$, where $\theta_*$ is the minimizer of the population quasi-Kullback–Leibler criterion.
• Asymptotic normality: $$\sqrt{n}\left(\hat{\theta}_n - \theta_*\right) \xrightarrow{d} N(0, \Sigma)$$, with $\Sigma$ the “sandwich” covariance given by

$\Sigma = H^{-1} I H^{-1},$

where $H$ is the negative expected Hessian of the quasi-log-likelihood, and $I$ is the variance of the quasi-score.

Non-standard settings—including boundaries, non-identifiability, non-ergodic information, and random Fisher information—are now covered by stable-local-field arguments, anisotropic convergence rates, and local tangent-cone theory (Yoshida et al., 2022). For penalized variants (PQMLE), the same local-argmax theory applies, enabling formal analysis of selection consistency and the so-called “oracle property.”

3. QMLE in Important Model Classes

3.1 Binary Choice and Slope Consistency

In binary choice models with $Y \in \{0,1\}$ and $P(Y=1|x)=F(x'\beta)$ for a known, strictly monotonic CDF $F$, the QMLE maximizes the sum

$$L_n(\beta) = \sum_{i=1}^n \left[ y_i \log F(x_i'\beta) + (1-y_i)\log(1-F(x_i'\beta)) \right].$$

Slope consistency holds when a set of structural and regularity assumptions (index dependence, “linearity-in-expectation,” strict concavity, etc.) are satisfied: there exists a unique $(c_*, r_*)$ with $c_*>0$ such that the limiting QMLE recovers $\beta_* = c_* \beta_0$ (Chang et al., 5 May 2025). In the special case where $F$ is the logistic CDF, $c_*=1$ and logistic regression is exactly slope-consistent.

3.2 Robust and Semi/Non-Gaussian QMLE

QMLE is widely employed in volatility models (GARCH, stochastic volatility), time series models (ARMA, ARCH(∞), APARCH), double autoregressive (DAR), dynamic panels, and hidden Markov models. Innovations to the QMLE framework include:

• Laplacian QMLE: Uses an $L^1$ contrast; delivers consistency and asymptotic normality under lower moment assumptions and robust performance under heavy-tailed noise (Bardet et al., 2016).
• Exponential, Mixture, and Logistic QMLEs: Provide robustness to skewness, heavy tails, and heteroscedasticity; Normal-mixture QMLE for DAR outperforms Gaussian QMLE when tails or skewness are pronounced (Chen et al., 29 May 2025, Liu et al., 2020, Wang et al., 11 Mar 2025).
• Adaptive and two-step QMLEs: Used when the innovation distribution is unknown—estimate auxiliary nuisance parameters (scales, mixture weights, pseudo-variances) in a preliminary step, or adaptively select the optimal quasi-likelihood (Qi et al., 2010, Jiang et al., 2019, Armillotta et al., 2023).

3.3 QMLE in Dependent and High-Dimensional Settings

• Dynamic Panel Models with Interactive Effects: The QMLE matches or attains the efficiency bound under the normality assumption, even without assuming true normality; it dominates the traditional fixed-effects estimator in bias and efficiency (Bai, 2023, Wang et al., 31 Dec 2025).
• Approximate Factor Models: QMLE and Principal Components estimators are asymptotically equivalent as $n \rightarrow \infty$, and both achieve the “blessing of dimensionality” effect (Barigozzi, 2023).
• Network and Spatial Models: Recent work establishes bias correction for QMLE in spatial autoregression with many covariates or fixed effects, substantially improving finite-sample properties (Martellosio et al., 2019, Wang et al., 31 Dec 2025).

4. Model Misspecification, Robustness, and Adaptivity

QMLE is specifically designed to address misspecification, delivering estimators that are consistent for the “pseudo-true” parameter minimizing Kullback–Leibler divergence, irrespective of the exact functional form of the data-generating density. The essential ingredients for this robustness are:

• The quasi-score is unbiased under the pseudo-true parameter, and higher (e.g., sandwich) variance formulas account for loss of efficiency under misspecification.
• In semiparametric and nonstationary regimes (e.g., time-varying volatility, nonstationary mean), QMLE can remain adaptive, achieving the parametric convergence rate even when nonparametric trends are present in nuisance components (Jiang et al., 2019).
• For penalized QMLE involving boundary or non-regular models (e.g., variance-component models, selection via $L_q$ penalties), theory establishes limit distributions and variable selection properties precisely (Yoshida et al., 2022).

5. Algorithmic Implementation and Inference

QMLE estimation operates by maximizing the quasi-log-likelihood, typically via Newton–Raphson, EM-type, or quasi-Newton algorithms. Key features include:

• Uniqueness and strict concavity of the quasi-likelihood under shape restrictions on the link (e.g., logistic or strictly log-concave) (Chang et al., 5 May 2025, Wang et al., 11 Mar 2025).
• Plug-in estimators for sandwich variance matrices (score variance and Hessian), with Wald and Lagrange Multiplier tests derived via standard M-estimation approaches (Wang et al., 11 Mar 2025, Jiang et al., 2019).
• Model selection (e.g., number of mixture components in NM-QMLE) is handled using BIC, ICL, or entropy-corrected information criteria, with provable consistency and robust finite-sample performance (Chen et al., 29 May 2025).

6. Limitations and Open Questions

Despite its versatility, QMLE requires careful attention to identification, regularity, and information conditions—particularly in boundary or non-ergodic models (Yoshida et al., 2022). Key challenges include:

• Relaxing “index-dependence” and “linearity-in-expectation” in binary choice for unconditional slope recovery remains open (Chang et al., 5 May 2025).
• Handling incidental parameter bias in high-dimensional panels or with many fixed effects demands bias corrections or fully profiled likelihoods (Wang et al., 31 Dec 2025, Martellosio et al., 2019).
• Further analysis of finite-sample bias, regularization in high-dimensional parameter spaces, and explicit construction of confidence regions under nonregular asymptotics (Yoshida et al., 2022).

7. Applications and Practical Impact

QMLE is central to modern empirical research in economics, finance, machine learning, and signal processing:

• Gaussian and non-Gaussian QMLEs are default tools for GARCH, stochastic volatility, and heavy-tailed time-series modeling (Qi et al., 2010, Liu et al., 2020, Jiang et al., 2019, Chen et al., 29 May 2025).
• Logistic QMLE underpins the theory and practice of logistic regression in large-scale binary classification tasks in econometrics and machine learning (Chang et al., 5 May 2025).
• QMLE frameworks for factor, dynamic network, and spatial autoregressive models with many fixed effects underpin leading-edge work on high-dimensional panel data and networks (Barigozzi, 2023, Wang et al., 31 Dec 2025, Bai, 2023).
• Robust QMLE (Laplacian, logistic) and mixture QMLEs enable valid inference in the presence of outliers, heavy-tails, and adversarial data contamination (Wang et al., 11 Mar 2025, Bardet et al., 2016, Chen et al., 29 May 2025).

In sum, the QMLE provides a foundational framework for inference in parametric and semiparametric models under a wide variety of misspecification or non-regular behavior, combining computational tractability, broad applicability, and rigorous theoretical guarantees (Yoshida et al., 2022, Chang et al., 5 May 2025, Chen et al., 29 May 2025, Bai, 2023, Barigozzi, 2023, Wang et al., 11 Mar 2025, Wang et al., 31 Dec 2025).