betaFeedback

WorksMathematics

Efficient Estimation and Consistent Class Number Selection for Semiparametric Class-Specific Index Regression

Abstract

The study introduces a parsimonious semiparametric class-specific index regression that captures heterogeneous covariate–response relationships across latent subpopulations. In such a latent class model formulation, subject-specific indices are used as a complete-data representation of the realized latent labels. We propose a unified framework for estimating model parameters and recovering latent class memberships. For the class-specific index mean model, estimation and latent class recovery are conducted via a pseudo-least-squares criterion with a separation penalty, and a semiparametric information criterion is proposed for consistently estimating the number of classes. Exploiting the response distribution, we develop a Bayes classification rule that maximizes the asymptotic probability of correct assignment and a pseudo-maximum likelihood criterion that yields class index coefficient estimators with asymptotic semiparametric efficiency. This refinement step characterizes latent class membership probabilities via sufficient dimension reduction of the covariate space. Simulations and applications to marketing and biomedical datasets demonstrate superior finite-sample performance of the proposed methodology in parameter estimation, latent class recovery, and class number selection.

Introduction

Across many scientific disciplines, the characteristics distinguishing subpopulations are often missing, unmeasured, or unobservable. This lack of observability complicates statistical analysis and motivates methods that explicitly account for heterogeneity in covariate–response relationships. Such frameworks are known as finite mixtures of regressions in statistics [19], mixtures of experts in machine learning [10], switching regression in econometrics [25], and clusterwise regression in marketing [4, 33]. Compared with existing parametric models in the literature, semiparametric class-specific index regression provides a flexible alternative, reducing the risk of misspecifying the parametric response or distribution function and avoiding the need to predefine class-specific central subspaces.

Analyzing 2018 transaction-level data from a North American supermarket chain, we observed distinct subgroups in purchase amounts in the second half of the year (H2) and an association with covariates measured in the first half (H1) that is invariant across customer segments. In another application, data from the AIDS Clinical Trials Group (ACTG) Study 175 reveal heterogeneity in treatment effects across patient subgroups. In particular, CD4 cell counts measured at 20±520 \pm5 weeks exhibit substantial heterogeneity suggestive of latent subgroups, and treatment effects—zidovudine monotherapy versus alternative therapies—on CD4 cell counts vary across these subgroups. The estimated response function from the class-specific index-mean model exhibits a non-monotonic pattern. In contrast to conventional class-specific linear, generalized linear, and quasi-likelihood models, our semiparametric class-specific index regression flexibly captures complex functional relationships without imposing restrictive parametric structural or distributional assumptions, providing a robust framework for analyzing data with heterogeneous effects.

Building on the subject-specific formulation of the proposed semiparametric class-specific index mean model, we develop a unified estimation and latent class recovery framework for uncovering latent structure without prespecified class labels. The procedure employs a separation-penalty estimation strategy, implemented via a computationally efficient alternating direction method of multipliers (ADMM; Boyd, Parikh, Chu, Peleato, and Eckstein, [2]), with a well-designed initialization scheme that ensures rapid and stable convergence, to obtain estimators of subject index coefficients and latent class memberships. Furthermore, we propose a semiparametric information criterion that consistently estimates the number of classes without relying on the response distribution, complementing the estimation and latent class recovery procedure. Under a general characterization of the response distribution, we derive a Bayes classification rule, incorporating a sufficient dimension reduction (SDR; Li [15]) representation of latent class membership probabilities, to achieve asymptotic optimality by maximizing the probability of correct assignment. We also propose a pseudo-maximum likelihood criterion to improve the efficiency of the separation-penalty estimators of the model parameters. The asymptotic variance matrix of the resulting pseudo-maximum likelihood estimators attains the semiparametric efficiency bound. As noted by Teng and Chiang [31], their pseudo-maximum likelihood method, developed under a semiparametric latent class index distribution model, is more sensitive to extreme values in the estimated density and requires more restrictive bandwidth conditions for continuous responses. Our method avoids these limitations under the proposed model formulation. Finally, by modifying the distribution estimator underlying the information criteria of Teng and Chiang [31], we construct alternative criteria for selecting the number of classes.

In estimating class-specific covariate–response relationships, subject-specific formulations have emerged as a standard approach, evolving from the early combinatorial optimization technique (Späth [29]) to contemporary regularization methods. The latter imposes concave pairwise fusion penalties ((Ma and Huang, 2017; Ma, Huang, Zhang, and Liu, 2020)) or adaptive ℓ1\ell_1 penalties ([3]). Recent innovations, including multidirectional separation penalties ([30]) and center-augmented regularization ((He, Zhou, Xia, and Lin, 2023)), aim to reduce computational complexity. Nonetheless, these approaches remain computationally burdensome and susceptible to inaccurate latent class recovery. Recent work by Teng and Chiang [31] partially addresses these challenges within a general semiparametric modeling framework. Building on their computational strategy, we further improve computational efficiency by reducing the number of ADMM loops required from two to one.

Alternatively, mixture regression models explicitly parameterize latent class membership. Early approaches assumed covariate-independent class membership probabilities ([13, 25]), but were later extended to accommodate covariate-dependent structures, including hierarchical chical mixture-of-experts models ([10, 11]) and conditional mixture models ([4]). Although robust procedures for testing the presence of latent subpopulations have been developed ([23, 27]), this class of models remains highly sensitive to misspecification of latent class membership models.

Our target remains a latent-class model. Rather than estimating the model through a marginal mixture likelihood, we introduce the subject-specific indices as a complete-data representation of the realized latent labels. The separation-penalty step estimates this realized-label representation, while the SDR-based class membership model is used in the distributional refinement step. In the proposed class-specific index density model with SDR-based class membership probabilities, direct marginal mixture-likelihood estimation is difficult because the unknown functions in these semiparametric settings create severe identifiability and computational complications.

The remainder of the article is organized as follows. § 2 formalizes the semiparametric class-specific index regression and its subject-specific formulation. § 3 introduces a unified estimation and latent class recovery procedure and a novel criterion for selecting the number of classes. § 4 refines the separation-penalty parameter estimator under a class-specific index density model. § 5 assesses the method’s finite-sample performance through simulation studies. § 6 illustrates its practical utility in personalized marketing and patient stratification. Finally, § 7 concludes with key findings and directions for future research. Appendix § 8 summarizes the notation used in § 2–§ 4.

Semiparametric Class-Specific Index Regression

Let Y∈YY \in\mathcal{Y} denote a quantitative response variable, X∈XX \in\mathcal{X} a p×1p \times1 covariate vector, kk the number of latent classes, and C∈C={1,…,k}C \in\mathcal{C} = \{1,\ldots,k\} a latent class variable. We consider a semiparametric class-specific index model (CSIM) for the conditional mean:

E[Y∣X=x,C=c]=m(γcTz),(x,c)∈X×C,(1)\mathbb{E}[Y \mid X=x,C=c] = m(\gamma_c^{\mathsf{T}}z),\quad(x,c) \in\mathcal{X} \times\mathcal{C}, \tag*{(1)}

where γc=(γc0,γc1,…,γcp)T\gamma_c = (\gamma_{c0}, \gamma_{c1}, \ldots, \gamma_{cp})^{\mathsf{T}} is the index coefficient vector for Class cc, zz is a realization of the augmented covariate vector Z=(1,XT)TZ = (1, X^{\mathsf{T}})^{\mathsf{T}}, and m(⋅)m(\cdot) is an unknown response function common across classes. This latent class model is identifiable under the following conditions. First, γ11≠0\gamma_{11} \ne0, corresponding to a continuous covariate that satisfies the continuity condition of Ichimura [9], and (γ10,γ11)(\gamma_{10}, \gamma_{11}) is normalized to (0,1)(0,1). Second, the separation condition m(γcTz)≠m(γℓTz)m(\gamma_c^{\mathsf{T}}z) \ne m(\gamma_\ell^{\mathsf{T}}z) holds for all c≠ℓc \ne\ell, c,ℓ∈Cc,\ell\in\mathcal{C}, and almost every x∈Xx \in\mathcal{X}. Under these assumptions, the CSIM (1) is identifiable up to a permutation of the class labels. Notably, the separation condition rules out periodic response functions over the covariate space in the CSIM. Since our approach is based on a subject-specific formulation of the CSIM rather than the mixture representation E⁡[Y∣X=x]=∑c=1kP(C=c∣X=x)m(γcTz)\operatorname{E}[Y \mid X = x] = \sum_{c=1}^{k} P(C = c \mid X = x)m(\gamma_c^{\mathsf{T}}z), it avoids the associated identifiability issues.

The proposed model encompasses, as special cases, parametric mean structures arising in generalized linear models (Nelder and Wedderburn [20]) and quasi-likelihood models (Wedderburn [32]). We decompose γc\gamma_c into a class-specific index coefficient vector γ(1)c\gamma_{(1)c} and a class-invariant index coefficient vector γ(2)\gamma_{(2)}, aligned with the covariate subsets Z(1)Z_{(1)} and Z(2)Z_{(2)}, so that the class-specific index can be written as γcTZ=γ(1)cTZ(1)+γ(2)TZ(2)\gamma_c^{\mathsf{T}} Z = \gamma_{(1)c}^{\mathsf{T}} Z_{(1)} + \gamma_{(2)}^{\mathsf{T}} Z_{(2)} for each c∈Cc \in\mathcal{C}. Under this parameterization, the CSIM (1) is given by

E⁡[Y∣X=x,C=c]=m(γ(1)cTz(1)+γ(2)Tz(2)),(x,c)∈X×C.(2)\operatorname{E}[Y \mid X = x, C = c] = m\left(\gamma_{(1)c}^{\mathsf{T}} z_{(1)} + \gamma_{(2)}^{\mathsf{T}} z_{(2)}\right), \quad(x,c) \in\mathcal{X} \times\mathcal{C}. \tag*{(2)}

Given observations {(Xi,Yi)}i=1n\{(X_i,Y_i)\}_{i=1}^{n}, we adopt a subject-specific formulation of the CSIM (1) within a pseudo-estimation framework:

E⁡[Yi∣Xi,βi]=m(βiTZi),i=1,…,n,(3)\operatorname{E}[Y_i \mid X_i,\beta_i] = m\left(\beta_i^{\mathsf{T}}Z_i\right), \quad i = 1,\ldots,n, \tag*{(3)}

where βi∈{γ1,…,γk}\beta_i \in\{\gamma_1,\ldots,\gamma_k\} and βi=γc\beta_i = \gamma_c if and only if Ci=cC_i = c. Although the population model is a latent class model, the subject index coefficients βi=γCi\beta_i = \gamma_{C_i}, i=1,…,ni = 1,\ldots,n, are used as a complete-data representation of the realized latent class labels. This subject-specific index model (SSIM) enables direct estimation of the response function m(⋅)m(\cdot) via the subject index coefficient vectors {βi}i=1n\{\beta_i\}_{i=1}^{n}, without requiring the latent class memberships to be known in advance. For each observation ii, the subject index coefficient vector is partitioned as βi=(β(1)i⊤,β(2)⊤)⊤\beta_i = \left(\beta_{(1)i}^{\top}, \beta_{(2)}^{\top}\right)^{\top}, where β(1)i∈{γ(1)1,…,γ(1)k}\beta_{(1)i} \in\{\gamma_{(1)1}, \ldots, \gamma_{(1)k}\} and β(2)=γ(2)\beta_{(2)} = \gamma_{(2)}.

Specifying the distribution for YY—for example, within an exponential or location family—enables the estimation procedure to fully exploit this information, thereby improving the efficiency of parameter estimators. In particular, consider the SSIM

Yi=m(βi⊤Zi)+Ui,i=1,…,n,(4)Y_i = m\left(\beta_i^{\top} Z_i\right) + U_i,\quad i = 1,\ldots,n, \tag*{(4)}

where UiU_i is independent of (Xi,βi)(X_i,\beta_i) and has an unspecified probability density function f(⋅)f(\cdot). Its corresponding CSIM representation is

Y=m(γC⊤Z)+U,C∈C.(5)Y = m\left(\gamma_C^{\top} Z\right) + U,\quad C \in\mathcal{C}. \tag*{(5)}

Therefore, the conditional probability density function f(y∣x,c)f(y \mid x,c) of YY given (X,C)(X,C) is

f(y∣x,c)=f(y−m(γc⊤z)),(y,x,c)∈Y×X×C.(6)f(y \mid x,c) = f\left(y - m\left(\gamma_c^{\top}z\right)\right),\quad(y,x,c) \in\mathcal{Y} \times\mathcal{X} \times\mathcal{C}. \tag*{(6)}

In many applications, including bioinformatics and marketing, responses are nonnegative (e.g., biomarker concentrations or purchase amounts) and are often analyzed on the log scale. The CSIM (5) accommodates such log-linear specifications as special cases.

Remark 1. When YY follows a distribution in the exponential family, the conditional probability density function of YY given (X,C)(X,C) can be written as

f(y∣x,c)=exp⁡(yθ(m(γc⊤z))−b(θ(m(γc⊤z)))ϕ+c(y,ϕ)),(y,x,c)∈Y×X×C,(7)f(y \mid x,c) = \exp\left(\frac{y\theta\left(m\left(\gamma_c^{\top}z\right)\right)-b\left(\theta\left(m\left(\gamma_c^{\top}z\right)\right)\right)}{\phi}+c(y,\phi)\right),\quad(y,x,c) \in\mathcal{Y} \times\mathcal{X} \times\mathcal{C}, \tag*{(7)}

where θ(⋅)\theta(\cdot), b(⋅)b(\cdot), and c(⋅,⋅)c(\cdot,\cdot) are known functions determined with the chosen exponential family, and ϕ\phi is a scale parameter. An estimation procedure can be constructed in close analogy to that under the class-specific index probability density model (6), while the corresponding theoretical analysis proceeds similarly but is simplified by the absence of any need to estimate the error density f(⋅)f(\cdot).

Estimation and Class Number Selection

Let θ∗\theta_* denote the unknown components of a generic parameter vector θ\theta. Corresponding to the alternative expression (2) of the CSIM (1) and the SSIM (3), let γ=(γ(1)⊤,γ(2)⊤)\gamma= \left(\gamma_{(1)}^{\top},\gamma_{(2)}^{\top}\right) and β=\beta= (β(1)T,β(2)T)T(\beta_{(1)}^{\mathsf T},\beta_{(2)}^{\mathsf T})^{\mathsf T}, where γ(1)=(γ(1)1T,…,γ(1)kT)T\gamma_{(1)}=(\gamma_{(1)1}^{\mathsf T},\ldots,\gamma_{(1)k}^{\mathsf T})^{\mathsf T} and β(1)=(β(1)1T,…,β(1)nT)T\beta_{(1)}=(\beta_{(1)1}^{\mathsf T},\ldots,\beta_{(1)n}^{\mathsf T})^{\mathsf T}. The true parameters are denoted by γo\gamma^{\mathrm{o}} and βo\beta^{\mathrm{o}}. For the estimators defined below, we adopt the convention that a ratio of the form 0/00/0 is assigned the value zero.

Pseudo-Least-Squares Estimation with a Separation Penalty

Given a coefficient vector β\beta, we estimate the response function m(v)m(v) using the kernel estimator

m^h(v;β)=∑i=1nYiKh(βiTZi−v)∑i=1nKh(βiTZi−v)(8)\widehat{m}_h(v;\beta)=\frac{\sum_{i=1}^{n}Y_iK_h(\beta_i^{\mathsf T}Z_i-v)}{\sum_{i=1}^{n}K_h(\beta_i^{\mathsf T}Z_i-v)} \tag*{(8)}

References

  1. [1]Bowman, A. W. (1984). An alternative method of cross-validation for the smoothing of density estimates. Biometrika 71, 353–360.DOI
  2. [2]Boyd, S., Parikh, N., Chu, E., Peleato, B., and Eckstein, J. (2011). Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends in Machine Learning 3, 1–122.
  3. [3]Chen, J., Tran-Dinh, Q., Kosorok, M. R., and Liu, Y. (2021). Identifying heterogeneous effect using latent supervised clustering with adaptive fusion. Journal of Computational and Graphical Statistics 30, 43–54.
  4. [4]DeSarbo, W. S. and Cron, W. L. (1988). A maximum likelihood methodology for clusterwise linear regression. Journal of Classification 5, 249–282.DOI
  5. [5]Giné, E. and Guillou, A. (2002). Rates of strong uniform consistency for multivariate kernel density estimators. Annales de l’Institut Henri Poincaré Probability and Statistics 38, 907–921.
  6. [6]Hammer, S. M., Katzenstein, D. A., Hughes, M. D., Gundacker, H., Schooley, R. T., Haubrich, R. H., Henry, W. K., Lederman, M. M., Phair, J. P., Niu, M., Hirsch, M. S., and Merigan, T. C. (1996). A trial comparing nucleoside monotherapy with combination therapy in HIV-infected adults with CD4 cell counts from 200 to 500 per cubic millimeter. New England Journal of Medicine 335, 1081–1090.pubmed.ncbi.nlm.nih.gov/8813038
  7. [8]Hoeffding, W. (1948). A class of statistics with asymptotically normal distribution. Annals of Mathematical Statistics 19, 293–325.
  8. [9]Ichimura, H. (1993). Semiparametric least squares (SLS) and weighted SLS estimation of single-index models. Journal of Econometrics 58, 71–120.
  9. [10]Jacobs, R. A., Jordan, M. I., Nowlan, S. J., and Hinton, G. E. (1991). Adaptive mixtures of local experts. Neural Computation 3, 79–87.DOI
  10. [11]Jordan, M. I. and Jacobs, R. A. (1994). Hierarchical mixtures of experts and the EM algorithm. Neural Computation 6, 181–214.
  11. [12]Khalili, A. (2010). New estimation and feature selection methods in mixture-of-experts models. Canadian Journal of Statistics 38, 519–539.DOI
  12. [13]Kiefer, N. M. (1978). Discrete parameter variation: efficient estimation of a switching regression model. Econometrica 46, 427–434.DOI
  13. [14]Kosorok, M. R. (2008). Introduction to Empirical Processes and Semiparametric Inference. New York: Springer.DOI
  14. [15]Li, K. C. (1991). Sliced inverse regression for dimension reduction. Journal of the American Statistical Association 86, 316–327.
  15. [19]McLachlan, G. J. and Peel, D. (2000). Finite Mixture Models. New York: Wiley.DOI
  16. [20]Nelder, J. A. and Wedderburn, R. W. (1972). Generalized linear models. Journal of the Royal Statistical Society A135, 370–384.
  17. [21]Nolan, D. and Pollard, D. (1987). U-processes: rates of convergence. Annals of Statistics 15, 780–799.DOI
  18. [22]Pakes, A. and Pollard, D. (1989). Simulation and the asymptotics of optimization estimators. Econometrica 57, 1027–1057.DOI
  19. [23]Peng, F., Jacobs, R. A., and Tanner, M. A. (1996). Bayesian inference in mixtures-of-experts and hierarchical mixtures-of-experts models with an application to speech recognition. Journal of the American Statistical Association 91, 953–960.DOI
  20. [24]Pollard, D. (1984). Convergence of Stochastic Processes. New York: Springer.
  21. [25]Quandt, R. E. (1972). A new approach to estimating switching regressions. Journal of the American Statistical Association 67, 306–310.DOI
  22. [26]Rand, W. M. (1971). Objective criteria for the evaluation of clustering methods. Journal of the American Statistical Association 66, 846–850.DOI
  23. [27]Shen, J. and He, X. (2015). Inference for subgroup analysis with a structured logistic-normal mixture model. Journal of the American Statistical Association 110, 303–312.DOI
  24. [28]Sherman, R. P. (1994). Maximal inequalities for degenerate U-processes with applications to optimization estimators. Annals of Statistics 22, 439–459.DOI
  25. [29]Späth, H. (1979). Algorithm 39: clusterwise linear regression. Computing 22, 367–373.DOI
  26. [30]Tang, X., Xue, F., and Qu, A. (2021). Individualized multidirectional variable selection. Journal of the American Statistical Association 116, 1280–1296.DOI
  27. [31]Teng, J.-C. and Chiang, C.-T. (2025). Unsupervised learning in a general semiparametric clusterwise index distribution model. arXiv preprint arXiv:2509.24987.arxiv.org/abs/2509.24987
  28. [32]Wedderburn, R. W. (1974). Quasi-likelihood functions, generalized linear models, and the Gauss-Newton method. Biometrika 61, 439–447.DOI
  29. [33]Wedel, M. and Kamakura, W. A. (2000). Market Segmentation: Conceptual and Methodological Foundations. Boston: Springer.

Paper details

vlnc:W_ztrm2uh9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents