Semiparametric Generalized Linear Models

• SPGLMs are semiparametric regression models that combine a finite-dimensional parametric component with an infinite-dimensional nuisance element, such as an unspecified baseline or variance.
• They employ exponential tilting, B-spline approximations, and quasi-likelihood methods to flexibly capture response behaviors while maintaining robust inference.
• SPGLMs are applied across time-series, longitudinal, and measurement-error frameworks, offering a trade-off between efficiency and robustness in complex data structures.

Semiparametric generalized linear models (SPGLMs) are regression models that retain a GLM-type systematic component—typically a link function, a linear predictor, or a parametric conditional mean—while relaxing full distributional specification by leaving a baseline/reference distribution, a nuisance regression component, or a variance/covariance structure unspecified or only partially modeled. In the literature, SPGLM denotes several closely related constructions: exponential-tilt models with unknown $f_0$ or $c(y)$, quasi-likelihood and GEE formulations that assume only correct first or second moments, and partially linear or additive generalized models with nonparametric smooth components (Alam et al., 2024, Agnoletto et al., 2023, Cheng et al., 2014).

1. Scope and defining structures

Several model architectures recur under the SPGLM label. One family writes the conditional law as an exponential tilt of an unknown baseline distribution; another specifies only the mean, or the first two moments, and treats the variance or covariance as nuisance; a third embeds unknown smooth functions in an otherwise GLM-like predictor. These formulations differ technically, but they share the same semiparametric division of labor: a finite-dimensional regression component is preserved, while at least one infinite-dimensional nuisance object is introduced (Alam et al., 2024, Young et al., 2024, Zhang et al., 9 May 2026).

Formulation	Representative model	Semiparametric element
Reference-distribution SPGLM	$f(y\mid x)=\exp\{\theta y-b(\theta)\}f_0(y)$	Unknown baseline/reference distribution $f_0$
Moment/covariance SPGLM	$g\big(E[Y_i\mid X_i]\big)=X_i\beta$	Variance, dispersion, or covariance left unrestricted
Structured semiparametric regression	$f\{Y\mid x^T\beta+m(z),\phi\}$	Unknown smooth function $m(\cdot)$

The reference-distribution formulation appears in scored ordinal models, continuous-response likelihood-based SPGLMs, Bayesian Dirichlet-process extensions, and semiparametric time-series models. The moment-based formulation appears in quasi-likelihood, generalized Bayes, sign-flip score testing, sandwich regression, and detection-limit problems. The structured-regression formulation appears in generalized partially linear additive models, semiparametric mixed models, measurement-error-corrected GAM-like estimators, single-index longitudinal models, and manifold-valued partially linear GLMs (Lee et al., 2022, Alam et al., 25 Feb 2025, Fung et al., 2016).

This suggests that SPGLM is best understood as a family of semiparametric generalizations of GLM structure rather than as a single canonical model.

2. Exponential tilting and unknown reference distributions

A central SPGLM construction replaces the fully parametric GLM response family by an unknown baseline or reference distribution and then recovers covariate effects through exponential tilting. In a finite-support scored-ordinal formulation, the model is

$$y \sim f(y \mid x) = \exp\left\{\theta y - b(\theta)\right\} f_0(y),$$

with

$$b(\theta) = \log \int_{\mathcal{Y}} \exp(\theta y) f_0(y)\, dy,$$

and

$$\eta = x^T\beta, \qquad \mu = E(y\mid x)=g^{-1}(\eta), \qquad \mu=b'(\theta).$$

In a closely related formulation,

$c(y)$0

with

$c(y)$1

Here $c(y)$2 remains parametric, but $c(y)$3 is left unspecified (Alam et al., 2024, Alam et al., 25 Feb 2025).

For scored ordinal outcomes with support $c(y)$4, the baseline $c(y)$5 is a simplex-valued pmf, and identifiability is imposed through the mean constraint

$c(y)$6

Under a Dirichlet prior,

$c(y)$7

the induced conditional pmf has weights

$c(y)$8

described as a tilted Dirichlet random vector. In the Bayesian nonparametric extension, the same exponential-tilting idea yields a varying-weights dependent Dirichlet process with common atoms $c(y)$9 and covariate-specific weights

$f(y\mid x)=\exp\{\theta y-b(\theta)\}f_0(y)$0

so dependence across covariates is induced by varying weights rather than changing atoms (Alam et al., 2024, Alam et al., 25 Feb 2025).

A second likelihood-based SPGLM parameterizes the nuisance part directly as an unknown function $f(y\mid x)=\exp\{\theta y-b(\theta)\}f_0(y)$1. For continuous $f(y\mid x)=\exp\{\theta y-b(\theta)\}f_0(y)$2,

$f(y\mid x)=\exp\{\theta y-b(\theta)\}f_0(y)$3

with $f(y\mid x)=\exp\{\theta y-b(\theta)\}f_0(y)$4, $f(y\mid x)=\exp\{\theta y-b(\theta)\}f_0(y)$5 for identifiability, and centered covariates $f(y\mid x)=\exp\{\theta y-b(\theta)\}f_0(y)$6. The nuisance $f(y\mid x)=\exp\{\theta y-b(\theta)\}f_0(y)$7 is approximated by a B-spline sieve,

$f(y\mid x)=\exp\{\theta y-b(\theta)\}f_0(y)$8

and the resulting approximate log-likelihood is concave, so the optimizer is unique and can be found by standard convex optimization routines (Lee et al., 2022).

The same exponential-tilt logic extends to dependent data. In semiparametric time-series GLMs,

$f(y\mid x)=\exp\{\theta y-b(\theta)\}f_0(y)$9

with

$f_0$0

and the conditional mean constraint

$f_0$1

The underlying response distribution $f_0$2 is treated as an infinite-dimensional parameter and estimated jointly with $f_0$3 by maximum empirical likelihood (Fung et al., 2016).

3. Mean-model semiparametrics, quasi-likelihood, and robust testing

A different SPGLM tradition treats the conditional mean as the only fully trusted parametric object. In quasi-likelihood formulations,

$f_0$4

and the log-quasi-likelihood is

$f_0$5

Exponentiating $f_0$6 with a prior gives the quasi-posterior

$f_0$7

This yields a generalized Bayes posterior that only requires correct specification of the first two moments, recovers ordinary Bayes when the model is a correctly specified exponential family, converges in total variation to a normal limit, and is asymptotically calibrated when $f_0$8 is correctly set or consistently estimated (Agnoletto et al., 2023).

Variance robustness can also be handled by score randomization. In GLMs with mean model

$f_0$9

and potentially misspecified variances, the effective score for $g\big(E[Y_i\mid X_i]\big)=X_i\beta$0 is

$g\big(E[Y_i\mid X_i]\big)=X_i\beta$1

If $g\big(E[Y_i\mid X_i]\big)=X_i\beta$2 is a random diagonal matrix with independent $g\big(E[Y_i\mid X_i]\big)=X_i\beta$3 entries, the sign-flipped score is

$g\big(E[Y_i\mid X_i]\big)=X_i\beta$4

and the standardized statistic

$g\big(E[Y_i\mid X_i]\big)=X_i\beta$5

is asymptotically valid under arbitrary variance misspecification, while improving small-sample level control relative to the unstandardized effective score test (Santis et al., 2022).

In clustered and longitudinal settings, the same mean-only semiparametric stance leads to covariance-robust estimator selection. The model is

$g\big(E[Y_i\mid X_i]\big)=X_i\beta$6

with no model imposed on $g\big(E[Y_i\mid X_i]\big)=X_i\beta$7. A parametric class of QML estimators is indexed by a working covariance parameter $g\big(E[Y_i\mid X_i]\big)=X_i\beta$8, and sandwich regression chooses $g\big(E[Y_i\mid X_i]\big)=X_i\beta$9 by minimizing an estimate of the variance of the target $f\{Y\mid x^T\beta+m(z),\phi\}$0: $f\{Y\mid x^T\beta+m(z),\phi\}$1 Its finite-sample empirical loss is based on leave-one-cluster-out refits,

$f\{Y\mid x^T\beta+m(z),\phi\}$2

The method is explicitly targeted at fixed-effect precision rather than covariance fit (Young et al., 2024).

The same logic underlies SPGLM methodology for an explanatory variable censored by a detection limit. The primary regression is modeled only through

$f\{Y\mid x^T\beta+m(z),\phi\}$3

with the surrogacy condition

$f\{Y\mid x^T\beta+m(z),\phi\}$4

When $f\{Y\mid x^T\beta+m(z),\phi\}$5 is censored, the observed-data mean becomes

$f\{Y\mid x^T\beta+m(z),\phi\}$6

and $f\{Y\mid x^T\beta+m(z),\phi\}$7 is estimated by the GEE

$f\{Y\mid x^T\beta+m(z),\phi\}$8

Here the primary model is semiparametric even when the auxiliary model for the censored covariate is parametric (Zhang et al., 13 Jul 2025).

4. Smooth components, longitudinal structure, and non-Euclidean covariates

Many SPGLMs place the semiparametric component in the predictor rather than in the response law. In generalized partially linear additive models for longitudinal or clustered data,

$f\{Y\mid x^T\beta+m(z),\phi\}$9

with identifiability constraints

$m(\cdot)$0

Each $m(\cdot)$1 is approximated by B-splines,

$m(\cdot)$2

and estimation proceeds through extended GEE with a working covariance $m(\cdot)$3. The estimator remains consistent and asymptotically normal if $m(\cdot)$4 is misspecified, and is semiparametrically efficient if $m(\cdot)$5 equals the true covariance $m(\cdot)$6 (Cheng et al., 2014).

Measurement error introduces a second semiparametric layer. In semiparametric generalized regression models with one smooth predictor,

$m(\cdot)$7

with penalized-spline decomposition

$m(\cdot)$8

Under classical error,

$m(\cdot)$9

the estimator integrates out $$y \sim f(y \mid x) = \exp\left\{\theta y - b(\theta)\right\} f_0(y),$$0, approximates the first two conditional moments of $$y \sim f(y \mid x) = \exp\left\{\theta y - b(\theta)\right\} f_0(y),$$1 by Monte Carlo, and fits the induced heteroscedastic mixed model

$$y \sim f(y \mid x) = \exp\left\{\theta y - b(\theta)\right\} f_0(y),$$2

This produces the observed semiparametric measurement error estimator (OSMEE) (Hattab et al., 2021).

Longitudinal SPGLM variants also include partially linear single-index models and semiparametric regression on manifolds. For repeated measurements,

$$y \sim f(y \mid x) = \exp\left\{\theta y - b(\theta)\right\} f_0(y),$$3

and estimation combines local linear smoothing of $$y \sim f(y \mid x) = \exp\left\{\theta y - b(\theta)\right\} f_0(y),$$4 with a semiparametric GEE. On a Riemannian manifold $$y \sim f(y \mid x) = \exp\left\{\theta y - b(\theta)\right\} f_0(y),$$5, generalized partially linear models take the form

$$y \sim f(y \mid x) = \exp\left\{\theta y - b(\theta)\right\} f_0(y),$$6

where $$y \sim f(y \mid x) = \exp\left\{\theta y - b(\theta)\right\} f_0(y),$$7 enters through an unknown smooth function estimated with kernel weights based on Riemannian distance and volume density. The manifold construction is specialized to binary logistic and ordinal proportional-odds models on Kendall’s three-dimensional shape space (Chen et al., 2015, Simó et al., 2018).

High-dimensional longitudinal SPGLM extensions add random effects and penalization. In the generalized semiparametric mixed model,

$$y \sim f(y \mid x) = \exp\left\{\theta y - b(\theta)\right\} f_0(y),$$8

the smooth $$y \sim f(y \mid x) = \exp\left\{\theta y - b(\theta)\right\} f_0(y),$$9 is approximated by regression splines, the fixed effects are penalized by SCAD, the posterior distribution of random effects is approximated by a Metropolis algorithm, and the resulting estimator enjoys oracle properties under a diverging-dimension regime (Taavoni et al., 2019).

5. Estimation algorithms, Bayesian computation, and bias correction

SPGLM estimation spans empirical likelihood, sieve likelihood, GEE, MCMC, Metropolis–Hastings, and online variational Bayes. In time-series SPGLMs the unknown $$b(\theta) = \log \int_{\mathcal{Y}} \exp(\theta y) f_0(y)\, dy,$$0 is replaced by discrete masses $$b(\theta) = \log \int_{\mathcal{Y}} \exp(\theta y) f_0(y)\, dy,$$1 on the observed support, and one maximizes an empirical log-likelihood under normalization and mean constraints. In continuous-response SPGLMs with nuisance $$b(\theta) = \log \int_{\mathcal{Y}} \exp(\theta y) f_0(y)\, dy,$$2, one replaces $$b(\theta) = \log \int_{\mathcal{Y}} \exp(\theta y) f_0(y)\, dy,$$3 by a B-spline expansion and maximizes a concave approximate likelihood. In additive longitudinal SPGLMs, spline approximation reduces the model to a finite-dimensional extended GEE problem (Fung et al., 2016, Lee et al., 2022, Cheng et al., 2014).

Bayesian SPGLMs place priors directly on the semiparametric component. Dir-SPGLM uses a Dirichlet prior on the finite-support baseline $$b(\theta) = \log \int_{\mathcal{Y}} \exp(\theta y) f_0(y)\, dy,$$4 and computes the posterior with Metropolis–Hastings proposals for $$b(\theta) = \log \int_{\mathcal{Y}} \exp(\theta y) f_0(y)\, dy,$$5 and $$b(\theta) = \log \int_{\mathcal{Y}} \exp(\theta y) f_0(y)\, dy,$$6. DPGLM replaces the fixed baseline by a gamma completely random measure whose normalization is a Dirichlet process, introduces auxiliary variables

$$b(\theta) = \log \int_{\mathcal{Y}} \exp(\theta y) f_0(y)\, dy,$$7

and exploits a conditional posterior decomposition

$$b(\theta) = \log \int_{\mathcal{Y}} \exp(\theta y) f_0(y)\, dy,$$8

Posterior simulation combines Ferguson–Klass sampling for the CRM part with Metropolis–Hastings updates for $$b(\theta) = \log \int_{\mathcal{Y}} \exp(\theta y) f_0(y)\, dy,$$9, $$\eta = x^T\beta, \qquad \mu = E(y\mid x)=g^{-1}(\eta), \qquad \mu=b'(\theta).$$0, $$\eta = x^T\beta, \qquad \mu = E(y\mid x)=g^{-1}(\eta), \qquad \mu=b'(\theta).$$1, and latent $$\eta = x^T\beta, \qquad \mu = E(y\mid x)=g^{-1}(\eta), \qquad \mu=b'(\theta).$$2 (Alam et al., 2024, Alam et al., 25 Feb 2025).

Streaming computation has also been developed in a broad semiparametric regression framework that explicitly includes generalized additive models, generalized linear mixed models, geostatistical models, wavelet nonparametric regression models, and their combinations. The online MFVB updates maintain sufficient statistics such as $$\eta = x^T\beta, \qquad \mu = E(y\mid x)=g^{-1}(\eta), \qquad \mu=b'(\theta).$$3, $$\eta = x^T\beta, \qquad \mu = E(y\mid x)=g^{-1}(\eta), \qquad \mu=b'(\theta).$$4, and $$\eta = x^T\beta, \qquad \mu = E(y\mid x)=g^{-1}(\eta), \qquad \mu=b'(\theta).$$5, and extend to Bernoulli models of the form

$$\eta = x^T\beta, \qquad \mu = E(y\mid x)=g^{-1}(\eta), \qquad \mu=b'(\theta).$$6

using the Jaakkola–Jordan variational approximation (Luts et al., 2012).

Finite-sample bias correction has become a separate SPGLM theme. In the broad semiparametric class

$$\eta = x^T\beta, \qquad \mu = E(y\mid x)=g^{-1}(\eta), \qquad \mu=b'(\theta).$$7

SABRE starts from a spline-based initial estimator $$\eta = x^T\beta, \qquad \mu = E(y\mid x)=g^{-1}(\eta), \qquad \mu=b'(\theta).$$8, simulates synthetic data under the approximating model, and solves

$$\eta = x^T\beta, \qquad \mu = E(y\mid x)=g^{-1}(\eta), \qquad \mu=b'(\theta).$$9

For generalized partially linear models, the paper gives

$c(y)$00

$c(y)$01

and

$c(y)$02

while preserving the asymptotic variance of $c(y)$03 (Zhang et al., 9 May 2026).

6. Efficiency, empirical behavior, and recurring issues

A recurrent theoretical theme is that SPGLM flexibility does not preclude strong asymptotics. In different formulations, the literature proves consistency, asymptotic normality, semiparametric efficiency, oracle properties, total-variation Bernstein–von Mises behavior, and robust sandwich-type variance estimation. Efficient influence functions and information bounds are derived for marginal and quantile effects in likelihood-based SPGLMs; Euclidean parameters in generalized partially linear additive models attain the semiparametric information bound when the correct covariance is used; quasi-posteriors converge to normal limits and yield asymptotically valid credible sets; and penalized high-dimensional longitudinal GSMM estimators have oracle properties (Lee et al., 2022, Cheng et al., 2014, Agnoletto et al., 2023, Taavoni et al., 2019).

Empirical behavior is often reported as a robustness–efficiency trade-off. In the scored-ordinal Bayesian SPGLM, small-sample simulations report about 11%–21% lower RMSE for $c(y)$04, 7%–12% shorter intervals, and coverage near nominal relative to ML-SPGLM. For severe exceedance in the AHEAD study, the reported AUC is 0.79 for Dir-SPGLM versus 0.75 for ML-SPGLM, while for moderate exceedance it is 0.71 versus 0.70 (Alam et al., 2024).

In multilevel and longitudinal mean-only semiparametric GLMs, sandwich regression is comparable to EQML and GEE when the covariance is well specified and typically better when it is misspecified. In the Dominican Republic mercado study, sandwich regression reduces the estimated variance of the treatment effect by about 18.4% relative to unweighted OLS, while the mixed-effects model gives only a small improvement and GEE is slightly worse than OLS (Young et al., 2024).

For censored explanatory variables due to detection limit, the distinction between semi-para and semi-semi is explicit: the first combines a semiparametric primary regression with a parametric auxiliary model and is described as more efficient, whereas the second combines semiparametric components in both stages and is described as more robust. The same paper reports that sample splitting and cross fitting improve computational efficiency over bootstrap by 450 times; in the ESPINA application, an earlier semiparametric-likelihood approach took about 7.5 hours, whereas the SSCF-based approach took less than one minute (Zhang et al., 13 Jul 2025).

Other applications reinforce the breadth of the SPGLM program. In RNA sequencing count data, the standardized sign-flip score test is reported to be much less sensitive to variance-model misspecification than DESeq2 (Santis et al., 2022). In semiparametric time-series GLMs, the method performs nearly as well as correctly specified parametric models under correct specification and is much more robust under misspecification, especially for dependence parameters (Fung et al., 2016). In generalized semiparametric graphical models for mixed data, nodewise conditionals are SPGLMs with unspecified base measure functions, and a symmetric pairwise score test is proposed so that inferential results are invariant with respect to different parametrizations of the same edge (Yang et al., 2014). In manifold GPLMs for children’s garment fit, the best leave-one-out accuracies reported are 68.57\% for boys and 71.23\% for girls, outperforming related Euclidean alternatives (Simó et al., 2018). In measurement-error-corrected semiparametric generalized regression, the naive estimator is heavily biased, whereas OSMEE substantially reduces that bias across Poisson, logistic, negative binomial, and gamma settings, with REML reported as markedly more stable than GCV (Hattab et al., 2021).

A common misconception is that semiparametric GLM methodology is a single inferential recipe. The published formulations instead span empirical likelihood, approximate likelihood, GEE, quasi-likelihood, generalized Bayes, BNP priors on baseline distributions, randomization tests, and simulation-based bias correction. Another misconception is that semiparametric flexibility necessarily sacrifices rigorous inference; the available results show the opposite in many settings. A plausible implication is that the main unresolved choice in practice is not whether to use an SPGLM, but which nuisance object should be left unspecified—baseline law, variance, covariance, smooth predictor, measurement-error distribution, or latent reference measure—and which efficiency loss is acceptable for the robustness gained.