Functional Effects Models Overview

• Functional Effects Models are modeling paradigms where scalar parameters are replaced by functions, curves, or surfaces, allowing dynamic modeling of effects across time, space, or covariates.
• They span diverse formulations including panel choice, survival regression, mixed-effects, and causal models, employing techniques like Gaussian processes, spline smoothing, and deep learning.
• These models enhance forecasting and heterogeneity capture while addressing challenges such as high-dimensional regularization and complex covariance estimation.

Functional Effects Model denotes a family of models in which the primary regression effect is itself functional: a coefficient function, a coefficient surface, a subject-specific latent curve, a function-valued response, or an individual-specific parameter learned as a function of observed covariates. In recent arXiv usage, the term covers panel choice models with functional intercepts and functional slopes, scalar-on-function survival models with varying functional coefficients, functional mixed-effects models with random slope functions, Gaussian-process models for function-valued responses, and causal models whose estimands are themselves curves or Fréchet-mean objects (Salvadé et al., 22 Sep 2025, Yang et al., 13 Jun 2026, Liu et al., 2016, Wiklendt et al., 2019, Raykov et al., 6 Mar 2025). This suggests that the expression is better understood as a modeling paradigm centered on functional parameterization than as a single universally standardized specification.

1. Conceptual scope and representative formulations

Across the literature, a functional effects model typically replaces a finite-dimensional coefficient vector with an object indexed by a continuum, by covariates, or by both. In some papers, the functional object is a coefficient multiplying a functional predictor. In others, it is a latent random effect over a domain, a response surface over frequency or time-frequency, or a deterministic mapping from socio-demographics to individual-specific utility parameters. The common element is that “effect” no longer means a single scalar parameter.

Usage	Representative formulation	Paper
Panel choice heterogeneity	$$V_{int} = g_{i0}(\mathbf{s}_n) + \sum_{m=1}^{M_i} g_{im}(\mathbf{s}_n)\,x_{inmt}$$	(Salvadé et al., 22 Sep 2025)
Functional AFT varying effect	$$\log(S_i) = \beta_0 + {Z}_{1i}^\top \beta_1 + \int_0^1 \beta_2({Z}_{2i}, t)\, X_i(t)\, dt + \sigma_\varepsilon \varepsilon_i$$	(Yang et al., 13 Jun 2026)
Functional linear mixed-effects regression	$$Y_{ij} = \alpha_0 + a_i + \int_{\mathcal S} [\beta(t) + b_i(t)]X_{ij}(t)\,dt + \epsilon_{ij}$$	(Liu et al., 2016)
Time-varying functional Cox model	$$\log\lambda_i(t)=\log\lambda_0(t)+X_i\beta+\int_\mathcal U Z_i(u)\gamma(u,t)\,du$$	(Du et al., 2024)
Function-valued Gaussian-process response	$$y_i(x) \sim \mathcal{GP}\!\left(\eta_i(x), \sigma_i(x,x')\right)$$	(Wiklendt et al., 2019)

This heterogeneity of usage has methodological consequences. Some formulations are primarily predictive, some are inferential, some are causal, and some are designed to preserve geometric constraints in non-Euclidean spaces. A plausible implication is that any encyclopedia treatment must distinguish the object being modeled—coefficient, response, latent effect, or causal estimand—before comparing estimation strategies.

2. Covariate-dependent coefficients and varying functional effects

One important line of work treats individual-specific coefficients as deterministic functions of observed covariates. In panel choice, the deterministic utility is written as

$$V_{int} = g_{i0}(\mathbf{s}_n) + \sum_{m=1}^{M_i} g_{im}(\mathbf{s}_n)\,x_{inmt},$$

where $g_{i0}(\mathbf{s}_n)$ is a functional intercept and $g_{im}(\mathbf{s}_n)$ is a functional slope for variable $m$ in alternative $i$. The same framework distinguishes FI, FS, and FIS specifications, and learns the effect functions with gradient boosting decision trees or deep neural networks. The paper explicitly identifies three advantages over traditional fixed and random/mixed effects models: forecasting for previously unobserved individuals, avoiding the incidental parameters problem, and not relying on strong distributional assumptions (Salvadé et al., 22 Sep 2025).

A second usage appears in scalar-on-function survival regression. The Bayesian functional accelerated failure-time model writes

$$\log(S_i) = \beta_0 + {Z}_{1i}^\top \beta_1 + \int_0^1 \beta_2({Z}_{2i}, t)\, X_i(t)\, dt + \sigma_\varepsilon \varepsilon_i$$0

with measurement error $$\log(S_i) = \beta_0 + {Z}_{1i}^\top \beta_1 + \int_0^1 \beta_2({Z}_{2i}, t)\, X_i(t)\, dt + \sigma_\varepsilon \varepsilon_i$$1. Here $$\log(S_i) = \beta_0 + {Z}_{1i}^\top \beta_1 + \int_0^1 \beta_2({Z}_{2i}, t)\, X_i(t)\, dt + \sigma_\varepsilon \varepsilon_i$$2 is a varying functional coefficient that changes with both time $$\log(S_i) = \beta_0 + {Z}_{1i}^\top \beta_1 + \int_0^1 \beta_2({Z}_{2i}, t)\, X_i(t)\, dt + \sigma_\varepsilon \varepsilon_i$$3 and subject-specific subgroup covariates $$\log(S_i) = \beta_0 + {Z}_{1i}^\top \beta_1 + \int_0^1 \beta_2({Z}_{2i}, t)\, X_i(t)\, dt + \sigma_\varepsilon \varepsilon_i$$4. The paper represents

$$\log(S_i) = \beta_0 + {Z}_{1i}^\top \beta_1 + \int_0^1 \beta_2({Z}_{2i}, t)\, X_i(t)\, dt + \sigma_\varepsilon \varepsilon_i$$5

assigns Gaussian process priors to the components of $$\log(S_i) = \beta_0 + {Z}_{1i}^\top \beta_1 + \int_0^1 \beta_2({Z}_{2i}, t)\, X_i(t)\, dt + \sigma_\varepsilon \varepsilon_i$$6, and corrects measurement error by combining a proxy $$\log(S_i) = \beta_0 + {Z}_{1i}^\top \beta_1 + \int_0^1 \beta_2({Z}_{2i}, t)\, X_i(t)\, dt + \sigma_\varepsilon \varepsilon_i$$7 with an instrumental variable $$\log(S_i) = \beta_0 + {Z}_{1i}^\top \beta_1 + \int_0^1 \beta_2({Z}_{2i}, t)\, X_i(t)\, dt + \sigma_\varepsilon \varepsilon_i$$8 (Yang et al., 13 Jun 2026).

A related but distinct formulation is the partially functional interaction model

$$\log(S_i) = \beta_0 + {Z}_{1i}^\top \beta_1 + \int_0^1 \beta_2({Z}_{2i}, t)\, X_i(t)\, dt + \sigma_\varepsilon \varepsilon_i$$9

This model targets conditional quantiles through $$Y_{ij} = \alpha_0 + a_i + \int_{\mathcal S} [\beta(t) + b_i(t)]X_{ij}(t)\,dt + \epsilon_{ij}$$0, allows locally sparse coefficient functions, and enforces a “main effect, interaction” hierarchy: if $$Y_{ij} = \alpha_0 + a_i + \int_{\mathcal S} [\beta(t) + b_i(t)]X_{ij}(t)\,dt + \epsilon_{ij}$$1 on a subregion, then $$Y_{ij} = \alpha_0 + a_i + \int_{\mathcal S} [\beta(t) + b_i(t)]X_{ij}(t)\,dt + \epsilon_{ij}$$2 on that subregion for each $$Y_{ij} = \alpha_0 + a_i + \int_{\mathcal S} [\beta(t) + b_i(t)]X_{ij}(t)\,dt + \epsilon_{ij}$$3. Estimation uses B-spline approximation, MCP penalties, MM for the quantile loss, and LQA for the penalties (Liang et al., 2023).

High-dimensional varying-coefficient work extends the same theme to random-effects settings. A sparse high-dimensional varying coefficients model with random effects allows covariates and coefficients to have a functional dependence with time, assumes sampling times that are either fixed and common or random and independent amongst individuals, proposes a projection procedure for empirical estimation of all varying coefficients, and extends the estimator to construct confidence bands for a fixed number of varying coefficients (Law et al., 2021).

3. Functional random effects and mixed-effects structure

Another major meaning of the term centers on functional random effects. In the functional linear mixed-effects model,

$$Y_{ij} = \alpha_0 + a_i + \int_{\mathcal S} [\beta(t) + b_i(t)]X_{ij}(t)\,dt + \epsilon_{ij}$$4

$$Y_{ij} = \alpha_0 + a_i + \int_{\mathcal S} [\beta(t) + b_i(t)]X_{ij}(t)\,dt + \epsilon_{ij}$$5 is the population slope function and $$Y_{ij} = \alpha_0 + a_i + \int_{\mathcal S} [\beta(t) + b_i(t)]X_{ij}(t)\,dt + \epsilon_{ij}$$6 is a subject-specific random slope function. The model assumes $$Y_{ij} = \alpha_0 + a_i + \int_{\mathcal S} [\beta(t) + b_i(t)]X_{ij}(t)\,dt + \epsilon_{ij}$$7, $$Y_{ij} = \alpha_0 + a_i + \int_{\mathcal S} [\beta(t) + b_i(t)]X_{ij}(t)\,dt + \epsilon_{ij}$$8, and $$Y_{ij} = \alpha_0 + a_i + \int_{\mathcal S} [\beta(t) + b_i(t)]X_{ij}(t)\,dt + \epsilon_{ij}$$9. Penalized spline smoothing estimates both $$\log\lambda_i(t)=\log\lambda_0(t)+X_i\beta+\int_\mathcal U Z_i(u)\gamma(u,t)\,du$$0 and $$\log\lambda_i(t)=\log\lambda_0(t)+X_i\beta+\int_\mathcal U Z_i(u)\gamma(u,t)\,du$$1, while a REML-based EM algorithm estimates $$\log\lambda_i(t)=\log\lambda_0(t)+X_i\beta+\int_\mathcal U Z_i(u)\gamma(u,t)\,du$$2, $$\log\lambda_i(t)=\log\lambda_0(t)+X_i\beta+\int_\mathcal U Z_i(u)\gamma(u,t)\,du$$3, and the covariance of the random slope coefficients (Liu et al., 2016).

Inference for fixed effects in complex functional mixed-effects data can be decoupled from full covariance estimation. One approach estimates the fixed effects under the independence of functional residuals assumption and then bootstraps independent units, typically subjects, to estimate variability and conduct hypothesis testing. The method delivers pointwise confidence intervals, simultaneous confidence bands, and a bootstrap test of

$$\log\lambda_i(t)=\log\lambda_0(t)+X_i\beta+\int_\mathcal U Z_i(u)\gamma(u,t)\,du$$4

with simulations showing excellent coverage probability of the confidence intervals and size of tests (Park et al., 2016).

Prediction-style inference for functional random effects extends the mixed-effects perspective to generalized responses. The Functional Random-effect Inference Method combines local GLMMs, structured FPCA, and a global Bayesian multilevel model conditional on estimated eigenfunctions. Its target is the subject- and visit-specific latent deviation on the linear predictor scale,

$$\log\lambda_i(t)=\log\lambda_0(t)+X_i\beta+\int_\mathcal U Z_i(u)\gamma(u,t)\,du$$5

with posterior samples used to construct pointwise credible intervals for latent random-effect trajectories (Zhou et al., 14 Jan 2025).

Multilevel two-way functional data require a richer random-effects hierarchy. In the Bayesian mixed-effects model for EEG,

$$\log\lambda_i(t)=\log\lambda_0(t)+X_i\beta+\int_\mathcal U Z_i(u)\gamma(u,t)\,du$$6

$$\log\lambda_i(t)=\log\lambda_0(t)+X_i\beta+\int_\mathcal U Z_i(u)\gamma(u,t)\,du$$7 is a subject-level random effect shared across conditions, $$\log\lambda_i(t)=\log\lambda_0(t)+X_i\beta+\int_\mathcal U Z_i(u)\gamma(u,t)\,du$$8 is a subject-by-condition random effect, and $$\log\lambda_i(t)=\log\lambda_0(t)+X_i\beta+\int_\mathcal U Z_i(u)\gamma(u,t)\,du$$9 is a covariate-dependent condition-level mean surface. The fixed effects are represented by a covariate-dependent CP decomposition with Stiefel-manifold constraints, and WAIC strongly favors the specification with both $$y_i(x) \sim \mathcal{GP}\!\left(\eta_i(x), \sigma_i(x,x')\right)$$0 and $$y_i(x) \sim \mathcal{GP}\!\left(\eta_i(x), \sigma_i(x,x')\right)$$1 (Ju et al., 27 Jul 2025).

A further extension combines real longitudinal and pseudo longitudinal growth data. After clustering cross-sectional observations into pseudo subjects, the resulting functional mixed effects model allows the covariance functions of the random effects and the variance functions of the measurement errors for pseudo and real longitudinal data to be different. This is designed for settings in which pseudo longitudinal data have larger variability than genuine longitudinal data (Chen et al., 23 Sep 2025).

4. Function-valued responses, spatial maps, and non-Euclidean domains

In many formulations, the effect is functional because the response itself is a curve or surface. A Bayesian functional mixed-effects model for wavelet spectra treats observed functional responses $$y_i(x) \sim \mathcal{GP}\!\left(\eta_i(x), \sigma_i(x,x')\right)$$2 as Gaussian processes,

$$y_i(x) \sim \mathcal{GP}\!\left(\eta_i(x), \sigma_i(x,x')\right)$$3

with mean

$$y_i(x) \sim \mathcal{GP}\!\left(\eta_i(x), \sigma_i(x,x')\right)$$4

and heteroscedastic residual scale

$$y_i(x) \sim \mathcal{GP}\!\left(\eta_i(x), \sigma_i(x,x')\right)$$5

The responses may be one-dimensional over frequency or two-dimensional over frequency and phase, with separable product kernels and Kronecker structure used to accelerate computation (Wiklendt et al., 2019).

Spatial neuroimaging introduces a related but distinct effect notion. In hierarchical covariate ICA,

$$y_i(x) \sim \mathcal{GP}\!\left(\eta_i(x), \sigma_i(x,x')\right)$$6

subject-specific source signals at voxel $$y_i(x) \sim \mathcal{GP}\!\left(\eta_i(x), \sigma_i(x,x')\right)$$7 are decomposed into a population map $$y_i(x) \sim \mathcal{GP}\!\left(\eta_i(x), \sigma_i(x,x')\right)$$8, a voxel-specific covariate effect matrix $$y_i(x) \sim \mathcal{GP}\!\left(\eta_i(x), \sigma_i(x,x')\right)$$9, and subject random effects $$V_{int} = g_{i0}(\mathbf{s}_n) + \sum_{m=1}^{M_i} g_{im}(\mathbf{s}_n)\,x_{inmt},$$0. Exact and subspace-based approximate EM algorithms estimate the model, and voxel-wise approximate inference avoids computationally expensive covariance estimation when testing covariate effects on spatial brain networks (Shi et al., 2014).

Non-Euclidean data force the concept of functional effect onto geometric manifolds. A joint model of brain shape and functional connectivity represents cortical surfaces by diffeomorphic deformation fields and connectivity by log-Euclidean tangent coordinates of covariance matrices, then combines the resulting coefficients in a matrix-normal mixed model

$$V_{int} = g_{i0}(\mathbf{s}_n) + \sum_{m=1}^{M_i} g_{im}(\mathbf{s}_n)\,x_{inmt},$$1

with $$V_{int} = g_{i0}(\mathbf{s}_n) + \sum_{m=1}^{M_i} g_{im}(\mathbf{s}_n)\,x_{inmt},$$2 structured by a known relatedness matrix and $$V_{int} = g_{i0}(\mathbf{s}_n) + \sum_{m=1}^{M_i} g_{im}(\mathbf{s}_n)\,x_{inmt},$$3, and $$V_{int} = g_{i0}(\mathbf{s}_n) + \sum_{m=1}^{M_i} g_{im}(\mathbf{s}_n)\,x_{inmt},$$4 structured by $$V_{int} = g_{i0}(\mathbf{s}_n) + \sum_{m=1}^{M_i} g_{im}(\mathbf{s}_n)\,x_{inmt},$$5 and $$V_{int} = g_{i0}(\mathbf{s}_n) + \sum_{m=1}^{M_i} g_{im}(\mathbf{s}_n)\,x_{inmt},$$6. The off-diagonal blocks of $$V_{int} = g_{i0}(\mathbf{s}_n) + \sum_{m=1}^{M_i} g_{im}(\mathbf{s}_n)\,x_{inmt},$$7 and $$V_{int} = g_{i0}(\mathbf{s}_n) + \sum_{m=1}^{M_i} g_{im}(\mathbf{s}_n)\,x_{inmt},$$8 encode shape–connectivity co-variation due to genetics and environment. On the HCP data, the paper reports overall $$V_{int} = g_{i0}(\mathbf{s}_n) + \sum_{m=1}^{M_i} g_{im}(\mathbf{s}_n)\,x_{inmt},$$9, with size $g_{i0}(\mathbf{s}_n)$0, shape $g_{i0}(\mathbf{s}_n)$1, and connectivity $g_{i0}(\mathbf{s}_n)$2 (Lila et al., 2020).

These formulations show that a functional effects model need not involve a functional predictor. It may instead involve a function-valued response, a spatial latent map, or a geometry-preserving latent representation.

5. Survival and causal versions

Survival analysis provides a direct route from functional regression to functional effects modeling. The time-varying functional linear Cox model is

$g_{i0}(\mathbf{s}_n)$3

where the coefficient surface $g_{i0}(\mathbf{s}_n)$4 varies over both the functional domain $g_{i0}(\mathbf{s}_n)$5 and follow-up time $g_{i0}(\mathbf{s}_n)$6. The paper estimates $g_{i0}(\mathbf{s}_n)$7 with penalized tensor product splines, using cyclic cubic regression splines in $g_{i0}(\mathbf{s}_n)$8 and cubic regression splines in $g_{i0}(\mathbf{s}_n)$9. It proposes a Cox-Poisson approach for small-to-medium datasets and a landmark approach for large datasets and high-dimensional functional predictors, both implemented in mgcv, and reports attenuation of diurnal effects on mortality over an 8-year follow-up in NHANES (Du et al., 2024).

The functional AFT model described earlier belongs to the same survival family but emphasizes measurement error and subgroup modification rather than proportional hazards violations. Its empirical application uses the REGARDS study, where weekday activity is treated as a proxy, weekend activity as an instrumental variable, and the full measurement-error-corrected Bayesian model slightly improves C-index and integrated Brier score relative to a naive model (Yang et al., 13 Jun 2026).

Functional causal inference pushes the concept from association to potential outcomes. For a functional treatment $g_{im}(\mathbf{s}_n)$0, the causal estimand is the average dose-response functional $g_{im}(\mathbf{s}_n)$1, modeled as

$g_{im}(\mathbf{s}_n)$2

The paper proposes three estimators for this functional causal effect: the functional stabilized weight estimator, the outcome regression estimator, and the doubly robust estimator. The functional stabilized weight is

$g_{im}(\mathbf{s}_n)$3

introduced because the usual generalized propensity score is not directly applicable to a functional treatment variable (Tan et al., 2022).

A broader causal framework replaces ordinary expectations with Fréchet means in metric spaces of functions. Its dynamic average treatment effect is

$g_{im}(\mathbf{s}_n)$4

and estimation proceeds through empirical Fréchet means, kernel ridge regression, and operator-valued kernels. The same framework also accommodates registered covariates and outcomes through Fisher–Rao geometry and the square-root slope function $g_{im}(\mathbf{s}_n)$5 (Raykov et al., 6 Mar 2025).

6. Regularization, computation, and recurring interpretive questions

Because functional effects are high-dimensional, regularization is central. One approach proposes adaptive shrinkage of smooth functional effects towards a predefined functional subspace rather than towards zero. With spline basis coefficients $g_{im}(\mathbf{s}_n)$6, the prior precision is

$g_{im}(\mathbf{s}_n)$7

and the shrinkage coefficient

$g_{im}(\mathbf{s}_n)$8

controls shrinkage toward the null space spanned by a user-defined matrix $g_{im}(\mathbf{s}_n)$9. The model combines an adapted horseshoe prior with a smoothing penalty similar to the random walk prior commonly applied in Bayesian P-spline priors (Wiemann et al., 2021).

Other regularization strategies target sparsity or rank. The partially functional interaction model uses MCP penalties to induce local sparsity and enforce the “main effect, interaction” hierarchy (Liang et al., 2023). The EEG model places a spike-and-slab-LASSO-inspired mixture prior on CP-rank weights $m$0, so redundant components are shrunk toward zero without imposing an ordering on component importance (Ju et al., 27 Jul 2025). The wavelet-spectrum model relies on weakly informative Gaussian-process priors, log-frequency kernels, and Kronecker algebra (Wiklendt et al., 2019).

Computation varies sharply by model class. The panel choice framework uses gradient boosting decision trees and deep neural networks, and its synthetic experiment compares FI-RUMBoost, FI-DNN, and a random-intercept mixed logit (Salvadé et al., 22 Sep 2025). The functional AFT model is fit by a Metropolis-within-Gibbs sampler and uses the Bhattacharya et al. fast Gaussian sampler for high-dimensional varying-effect coefficients (Yang et al., 13 Jun 2026). The wavelet-spectrum model uses Hamiltonian Monte Carlo in Stan (Wiklendt et al., 2019). Hierarchical covariate ICA employs exact EM and a subspace-based approximate EM, with the approximate method reducing the effective latent-state space from size $m$1 to size $m$2 under background-state concentration (Shi et al., 2014). The fixed-effects bootstrap strategy is explicitly motivated by the difficulty of estimating a huge covariance operator or matrix in complex functional data (Park et al., 2016).

A recurrent misconception is that “functional effects model” names a single model comparable to ordinary fixed effects or random effects. The literature does not support that reading. In some papers, the effect is a deterministic function learned from covariates; in others it is a random function; in others it is a response surface or a causal estimand. Another misconception is that functional modeling always concerns functional predictors. The examples above include functional predictors, function-valued responses, spatial source maps, and tangent-space coefficients on manifolds. This suggests that the most stable defining property is not the data type alone, but the decision to place the effect of interest in a functional object and then estimate, regularize, or infer that object with domain-aware structure.

Functional Effects Models therefore occupy a broad methodological territory linking functional data analysis, mixed models, Gaussian processes, causal inference, survival analysis, manifold statistics, and machine learning. Their diversity is substantive rather than terminological: each variant answers a different scientific question, but all replace a low-dimensional effect parameter with a functional one.