Poisson Lognormal with Dimension Reduction for Count Matrices with Missing Data

Impute missing data in count tables using a Poisson lognormal model with dimension reduction, with and without zero inflation, as described in Bricout et al. (2026) . Parameters are estimated with a variational expectation-maximization algorithm. The package also provides the asymptotic variance of the estimators, prediction intervals for the imputed counts, and model selection criteria (Bayesian information criterion, BIC, and integrated completed likelihood, ICL) to choose the dimension of the latent space.

Documentation: https://BBricout.github.io/colvR/

colvR

colvR provides modeling tools for count data — especially overdispersed data with many zeros — and includes missing-data imputation. It also exposes fast Rcpp/Armadillo utilities for matrix–vector conversions and low-rank ZIP/PLN optimization.

Installation

From GitHub:

# install.packages("devtools")
devtools::install_github("BBricout/colvR")

library(colvR)

When on CRAN, you’ll be able to do: install.packages(“colvR”)

Example

This is a basic example which shows you how to solve a common problem:

library(colvR)
#> 
#> Attachement du package : 'colvR'
#> L'objet suivant est masqué depuis 'package:stats':
#> 
#>     BIC

n <- 1500 ; p <- 30 ; q <- 4 ; d <- 4
dim <- list(n = n, p = p, d = d, q = q)
X <- cbind(rep(1, n*p), rnorm(n*p), rep(rnorm(n), p))
trend <- rep(1:p, each = n)
X <- cbind(X, scale(trend))
B <- c(2, 0.7, 0.5, -0.6) ; D <- c(-0.5, 0.5, 0.4, -0.4) ; C <- matrix(rnorm(p*q)/2, nrow = p, ncol = q)
theta <- list(B = B, D = D, C = C)

sim <- Simul(X, theta, dim) ; Y <- sim$Y

fit <- Miss.ZIPLNPCA(Y, X, q)