flexyBayes

flexyBayes fits linear and generalised linear mixed models written in ASReml or brms syntax. It fits them with INLA, or with Stan through brms. If the chosen engine cannot fit a model, flexyBayes stops and says why.

Version 0.18.1. Not on CRAN, and the functions may still change.

Installation

install.packages(c("remotes", "knitr", "rmarkdown"))
remotes::install_github("AAGI-AUS/flexyBayes", build_vignettes = TRUE)

Keep build_vignettes = TRUE. Without it, vignette(package = "flexyBayes") lists nothing. Building the vignettes needs knitr, rmarkdown and pandoc (RStudio includes pandoc).

You need at least one engine. brms is on CRAN. INLA is not:

install.packages("brms")
install.packages("INLA",
  repos = c(getOption("repos"),
            INLA = "https://inla.r-inla-download.org/R/stable"))

Dependencies

flexyBayes needs R 4.1.0 or later. Its required packages install from CRAN with it. The two engines, INLA and brms, are optional, but at least one is needed. With backend = "auto", a model INLA can fit goes to INLA when it is installed, and to brms otherwise. Packages used only by particular functions or vignettes are listed under Suggests in DESCRIPTION.

Example

An alpha design from agridat: 24 genotypes in three replicates, with blocks nested in replicates. The call is written as it would be in ASReml.

library(flexyBayes)
data(john.alpha, package = "agridat")

fit <- flexybayes(
  fixed   = yield ~ rep,
  random  = ~ rep:block + gen,
  data    = john.alpha,
  backend = "inla",
  verbose = FALSE
)

summary(fit)$varcomp[, c("component", "estimate", "conf.low", "conf.high")]
#>      component estimate conf.low conf.high
#> 1        sigma    0.293    0.229     0.371
#> 2 sd_rep:block    0.288    0.163     0.467
#> 3       sd_gen    0.397    0.274     0.561

The estimates are posterior means of the three standard deviations, with 95% credible intervals. REML estimates from lme4::lmer() for the same model are 0.286, 0.265 and 0.378.

Heritability is computed on each posterior draw, so its interval stays between 0 and 1:

vpredict(fit, h2 ~ sd_gen^2 / (sd_gen^2 + sigma^2))
#>   functional estimate median std.error conf.low conf.high
#> 1         h2    0.636  0.639       0.1    0.429     0.821

What it fits

The table lists the models at least one engine fits. Notes on each row, and the models neither engine fits, are in system.file("KNOWN_ISSUES.md", package = "flexyBayes").

Model class	Spelling	INLA	brms
Gaussian LMM, simple random intercept	random = ~ g / (1 | g)	fits	fits
GLMM (binomial, Poisson, negative binomial, gamma, beta), simple random effect	(1 | g) with family =	fits	fits
Hurdle gamma (zero mass plus a positive gamma part)	family = "hurdle_gamma"	refuses	fits
Uncorrelated random slope	(x || g)	fits	fits
Factor-by-numeric fixed interaction	y ~ f * x with numeric x	fits	fits
Correlated random slope	(x | g)	refuses	fits
Nested / interaction random effects, multi-stratum	~ gen:env, ~ env:rep:block	fits	fits
Heterogeneous variance by factor level	~ diag(f):g, ~ idh(f):g, ~ at(f):g	fits	fits
Unstructured genotype-by-environment covariance	~ us(f):g	refuses	fits
Heterogeneous residual by factor level	residual = ~ dsum(~ units | f) / ~ at(f):units	fits	fits
Baseline heteroscedastic MET: iid genotype, iid genotype-by-environment and a sectioned residual	random = ~ gen + gen:env with the dsum residual	fits	fits
Known-covariance genomic / pedigree random effect	~ vm(g, K), ~ ped(a, A)	fits	fits
Separable AR1 spatial field	random = ~ ar1(row):ar1(col), random = ~ ar1(t)	fits	refuses
Per-trial separable AR1 field	random = ~ id(trial):ar1(row):ar1(col)	fits	refuses
Grouped temporal AR1 field	random = ~ ar1(t):g, ~ g:ar1(t), ~ ar1(t):id(g)	fits	refuses
Univariate P-spline	~ spl(x)	fits	refuses
Observation weights (Gaussian, identity link)	weights = w	fits	fits
Exact sufficient-statistic aggregation (binomial, Poisson)	aggregate = TRUE, flexybayes_stream()	fits	n/a

fits: the engine fits it, and a test checks the fit. refuses: flexyBayes stops with a message saying why and what to use instead. n/a: the request does not apply to that engine.

Documentation

The manual is in the package vignettes listed below.

Vignette	Content
flexyBayes-01-mixed-models	Mixed models in brief
flexyBayes-02-bayesian	Bayesian estimation in brief
flexyBayes-03-syntax	The ASReml syntax flexyBayes covers
flexyBayes-04-illustration	ASReml and flexyBayes on the same trials
flexyBayes-06-two-stage	A two-stage analysis of a series of trials
flexyBayes-07-from-asreml	Moving from ASReml-R

Reporting problems

Please report bugs at https://github.com/AAGI-AUS/flexyBayes/issues, with a reproducible example and the output of sessionInfo().

Contributing

Fixes, documentation and discussion of new features are welcome. See CONTRIBUTING.md for the development setup and the test tiers. The project follows the Code of Conduct.

Citation

@software{flexyBayes,
  title  = {flexyBayes: Bayesian Mixed Models with ASReml Syntax
            via INLA and brms},
  author = {Moldovan, Max},
  year   = {2026},
  url    = {https://github.com/AAGI-AUS/flexyBayes}
}

Acknowledgement

flexyBayes is developed within Analytics for the Australian Grains Industry (AAGI), funded by the Grains Research and Development Corporation (GRDC) under projects CUR2210-005OPX and UOA2301-005OPX.

Licence

MIT. See LICENSE.md.