Expand description
Mixed-effects models in Rust.
This crate provides types and algorithms for fitting linear and generalized linear mixed-effects models. It is developed against Julia’s MixedModels.jl as the reference and parity target — the same PLS/PIRLS formulation, cross-checked numerically — but is an independent implementation with deliberate divergences (notably optimizer selection and the fallback strategy), not a line-for-line port.
§Tutorial
New to the crate? Read the guide module — five short, doctested
pages:
- Getting started — build a frame, parse a formula, fit an LMM.
- Reading results — coefficients, variance components, summaries, Wald vs. profile vs. bootstrap CIs.
- GLMMs — families, links, and the GLMM estimation semantics.
- When the crate refuses — typed errors and typed inference refusals (the no-fake-statistics contract from a caller’s POV).
- What is supported — a concrete inventory: model classes, families/links, formula syntax, inference paths, and what is deliberately out of scope.
§Quick start
Fit a linear mixed model y ~ 1 + x + (1 | g) on a small balanced dataset.
The prelude pulls in the common types and the
LinearMixedModelBuilder
collapses construction and the ML/REML choice into one chain:
use mixeff_rs::prelude::*;
use mixeff_rs::model::{FitOptions, LinearMixedModelBuilder};
let group_offsets = [-3.0, -1.5, 0.5, 2.0, -2.0, 1.0, 3.0, -0.5];
let jitter = [0.12, -0.20, 0.05, 0.17, -0.09, 0.22];
let mut y = Vec::new();
let mut x = Vec::new();
let mut g = Vec::new();
for (gi, off) in group_offsets.iter().enumerate() {
for (k, j) in jitter.iter().enumerate() {
let xv = k as f64;
x.push(xv);
y.push(2.0 + 1.5 * xv + off + j);
g.push(format!("g{gi}"));
}
}
let mut df = DataFrame::new();
df.add_numeric("y", y)?;
df.add_numeric("x", x)?;
df.add_categorical("g", g)?;
let model = LinearMixedModelBuilder::new(parse_formula("y ~ 1 + x + (1 | g)")?, &df)
.fit(FitOptions::reml())?; // or FitOptions::ml()
let coef = model.coef(); // fixed effects, ~[2.0, 1.5]
assert_eq!(coef.len(), 2);
assert!(coef.iter().all(|v| v.is_finite()));The lower-level form (LinearMixedModel::new(formula, &df, None)? then
model.fit(false)) remains available; the builder is purely additive.
Re-exports§
pub use nalgebra;
Modules§
- error
- Error types for the mixeff-rs crate.
- formula
- Formula parsing for mixed-effects models.
- guide
- Narrative tutorial pages (rendered on docs.rs, doctested in CI). Narrative tutorial for newcomers, in reading order.
- model
- Mixed model types and fitting algorithms.
- prelude
- Common imports for fitting and inspecting mixed models.
- stats
- Post-fit statistical summaries and inference.
- types
- Typed support values for advanced callers and tests.