Communication-Efficient Varying Coefficient Mixed-Effects Models

Scalable inference for Varying Coefficient Mixed-Effects Models (VCMMs) with large, correlated random effects. Implements sufficient-statistics, one-step communication-efficient surrogate likelihood, and SVD-stabilized (Singular Value Decomposition) estimators with Kronecker and separable covariance structures. Implements the methodology of Jalili and Lin (2025) .

cevcmm

cevcmm fits Varying Coefficient Mixed-Effects Models (VCMMs) at scale, including settings where data is split across multiple machines or doesn’t fit in memory. It implements the sufficient-statistics (SS) and one-step communication-efficient surrogate-likelihood (CSL) estimators of Jalili and Lin (2025), with diagonal, Kronecker, and separable random-effect covariance structures.

The model is

$$y_i ;=; \beta_0(t_i) ;+; \sum_{k=1}^{K} x_{ik},\beta_k(t_i) ;+; z_i^\top \alpha ;+; \varepsilon_i,$$

where each $\beta_k(t)$ is a smooth function of $t$ (cubic B-splines with a second-order-difference penalty), $\alpha \sim N(0, \Sigma_\alpha)$, and $\varepsilon_i \sim N(0, \sigma_\varepsilon^2)$.

Installation

You can install the released version of cevcmm from CRAN with:

install.packages("cevcmm")

Or the development version from GitHub:

# install.packages("remotes")
remotes::install_github("lidajalili/cevcmm")

Quick example

A 500-observation fit with one varying coefficient and three diagonal random effects:

library(cevcmm)

set.seed(1)
N <- 500L
q <- 3L

t <- runif(N)
x <- runif(N)
Z <- matrix(rnorm(N * q), N, q)
y <- 2 + sin(2 * pi * t) * x +
     as.vector(Z %*% rnorm(q, sd = 0.4)) +
     rnorm(N, sd = 0.5)

fit <- vcmm(y, X = x, Z = Z, t = t,
            control = vcmm_control(sigma_eps       = 0.5,
                                   sigma_alpha     = 0.4,
                                   update_variance = TRUE))
fit
#> <vcmm_fit>  Varying Coefficient Mixed-Effects Model fit
#>   method      : SS
#>   n_obs       : 500
#>   p (fixed)   : 12
#>   q (random)  : 3
#>   RE cov      : diag
#>   iterations  : 6 (converged)
#>   sigma_eps   : 0.5076
#>   sigma_alpha : 0.6327
#>   elapsed     : 0.0010 sec

All standard S3 methods work on the result: coef(), fixef(), ranef(), vcov(), nobs(), logLik(), AIC(), BIC(), predict(), summary(), and plot().

Choosing the random-effects structure

The package supports three re_cov modes. They specify how the random-effects vector $\alpha$ is assumed to covary — not what $Z$’s entries look like. The distribution of $Z$ (binary indicators, continuous values, or a mix) doesn’t enter the choice; only the structure of $\alpha$ does.

Each mode is demonstrated end-to-end in a vignette (links below).

Three vignettes

Three vignettes ship with the package, covering the main use cases end-to-end:

• vignette("getting-started", package = "cevcmm") — single-machine fit on a simple diagonal-random-effects example. Walks through every public S3 method.
• vignette("distributed-fitting", package = "cevcmm") — split data across $K$ nodes, compute per-node summaries with node_summary(), and recover a fit bit-equivalent to pooled vcmm() via fit_from_summaries(). Also covers the streaming-accumulator pattern for memory-constrained settings.
• vignette("od-migration", package = "cevcmm") — Kronecker covariance for origin-destination flow data, using a bundled simulated migration dataset.

Citation

If you use cevcmm in published work, please cite:

Jalili, L. and Lin, L.-H. (2025). Scalable and Communication-Efficient Varying Coefficient Mixed Effect Models: Methodology, Theory, and Applications. arXiv:2511.12732; under review at Journal of the American Statistical Association.

BibTeX:

@article{jalili2025cevcmm,
  title   = {Scalable and Communication-Efficient Varying Coefficient
             Mixed Effect Models: Methodology, Theory, and Applications},
  author  = {Jalili, Lida and Lin, Li-Hsiang},
  journal = {arXiv preprint arXiv:2511.12732},
  year    = {2025},
  note    = {Under review at Journal of the American Statistical Association}
}

License

Released under the GPL (>= 3) license.