broom tidy() for lme4 in R: Tidy Mixed-Effects Models

The broom.mixed::tidy() function turns lme4 mixed-effects model objects into one-row-per-term data frames you can pipe straight into dplyr, ggplot2, or a report. It works on lmer, glmer, and nlmer fits and returns fixed-effect estimates, random-effect standard deviations, and confidence intervals in a single call.

tidy(fit)                                          # fixed-effect estimates
tidy(fit, conf.int = TRUE)                         # add 95% Wald CIs
tidy(fit, conf.int = TRUE, conf.method = "profile") # likelihood profile CIs
tidy(fit, effects = "ran_pars")                    # random-effect SDs/variances
tidy(fit, effects = "ran_vals", conf.int = TRUE)   # group-level deviations
tidy(fit, effects = c("fixed", "ran_pars"))        # both blocks in one tibble
tidy(glmer_fit, exponentiate = TRUE)               # odds ratios for logistic glmer

What tidy() does for lme4 models in one sentence

tidy() reshapes mixed-effects model objects into rectangular data frames. A fitted lmer object is an S4 object of class lmerMod (or glmerMod for glmer) that stores fixed-effect coefficients, random-effect covariance components, the model frame, and optimisation diagnostics. broom.mixed::tidy() extracts the parts you usually need for reporting (term, estimate, standard error, statistic, and optional confidence bounds) and returns a tibble with one row per term.

For a hierarchical fit like lmer(y ~ x + (1 | group)), tidy() can also return random-effect standard deviations or the group-level conditional modes depending on the effects argument. This is much faster than digging into summary(fit)$coefficients and VarCorr(fit) by hand.

Syntax

tidy() is an S3 generic dispatched on the merMod superclass. The method lives in broom.mixed, not in broom itself, so make sure to load broom.mixed for lme4 fits even if broom is already attached.

The most useful arguments are:

• conf.int: TRUE adds conf.low and conf.high columns; default FALSE
• conf.level: nominal coverage of the interval; default 0.95
• conf.method: "Wald" (fast, default), "profile" (likelihood profile, slower but better for variance components), or "boot" (parametric bootstrap)
• effects: "fixed" (default), "ran_pars" (random-effect SDs and correlations), "ran_vals" (conditional modes per group), or "ran_modes" (synonym)
• exponentiate: TRUE exponentiates estimates and intervals, useful for logistic or Poisson glmer fits
• ddf: denominator degrees of freedom method for p-values when lmerTest is attached

These cover almost every reporting use case.

Common patterns

1. Fixed-effect table with Wald intervals

Each row is one fixed-effect term. estimate is the REML coefficient, std.error is its standard error from the Hessian, and statistic is the t-ratio. The Days slope of 10.5 ms per day, with a Wald interval well clear of zero, says reaction time grows reliably with sleep deprivation in the sleepstudy cohort.

2. Random-effect variances and correlations

effects = "ran_pars" returns one row per variance or correlation component. Here the between-subject SD on the intercept (24.7 ms) is similar in magnitude to the residual SD (25.6 ms), so subjects differ on baseline reaction time about as much as observations differ within a subject. The intercept-slope correlation near zero says subjects with slower baselines do not systematically deteriorate faster.

3. Group-level conditional modes with ran_vals

Each pair of rows is one subject's deviation from the population intercept and slope. Subject 308 has a slope 9.2 ms/day above the average, Subject 309 has a slope 8.6 ms/day below. The confidence interval is conditional on the estimated variance components, so treat it as descriptive rather than fully calibrated.

4. Combining fixed and random effects

Passing a vector to effects puts both blocks in one frame, ready for a gt::gt() table or a single forest plot grouped by effect. The ran_pars confidence bounds here are Wald, which broom.mixed warns about for variances; switch conf.method = "profile" for a publication-grade interval.

5. Odds ratios from a logistic glmer

With exponentiate = TRUE, estimate is the odds ratio for each period relative to baseline. Every period sits below 1 with a confidence interval clear of 1, so the odds of contagious bovine pleuropneumonia drop after the first observation period in every herd.

tidy() vs summary(fit) and parameters::model_parameters

Three tools cover the same job from different angles. Pick by what you do next with the output.

Use tidy() whenever the next step is code: filtering terms, joining several models, drawing a forest plot, or writing to CSV. Use parameters::model_parameters() when you need Satterthwaite or Kenward-Roger denominator df without attaching lmerTest and re-fitting.

Common pitfalls

Pitfall 1: loading broom instead of broom.mixed. The tidy.merMod method lives in broom.mixed. If only broom is attached, tidy(fit) may dispatch to a stale fallback that returns just the fixed effects without the effect column, or warn about missing methods. Always library(broom.mixed) for lme4 fits.

Pitfall 2: expecting p-values in the default output. Base lme4 deliberately omits p-values for fixed effects because the reference distribution is unclear for unbalanced designs. tidy() mirrors this: you get statistic (the t-ratio) but no p.value column. Attach lmerTest before fitting to get Satterthwaite-approximated p-values via summary(fit), or compute them with parameters::model_parameters(fit).

Pitfall 3: forgetting effects = "ran_pars" for hierarchical fits. The default effects = "fixed" only returns population-level coefficients. For lmer and glmer fits, variance components and group-level deviations require explicit effects arguments. Calling tidy(fit) alone silently omits the random-effect SDs that motivated the hierarchical model in the first place.

Try it yourself

Related broom.mixed functions for lme4

After mastering tidy(), look at:

• glance(): one-row model summary with sigma, logLik, AIC, BIC, deviance, and df.residual
• augment(): per-observation fitted values, residuals, and conditional modes
• lme4::VarCorr(): the raw variance-covariance matrix of the random effects
• parameters::model_parameters(): similar coefficient table with degrees-of-freedom-aware p-values
• performance::check_model(): diagnostic plots for merMod fits

For a coefficient forest plot, pipe tidy(fit, conf.int = TRUE) into ggplot2::geom_pointrange() with x = estimate, xmin = conf.low, xmax = conf.high, y = term and add a geom_vline(xintercept = 0) reference line.

See the official broom.mixed reference for lme4 methods for the full argument list and additional effects options.

FAQ

How do I get p-values from broom tidy on an lme4 model?

Attach lmerTest before fitting, then call tidy(fit, effects = "fixed"). lmerTest re-classes the fit so the Satterthwaite degrees-of-freedom and corresponding p-values flow through to the tibble's p.value column. For a glmer fit, p-values come from the Wald z-test by default and appear without any extra package.

What is the difference between effects = "ran_pars" and effects = "ran_vals"?

ran_pars returns one row per random-effect parameter (standard deviations and correlations of the covariance matrix), so the size depends on the random-effect structure, not the data. ran_vals returns one row per group per random term (the conditional modes, also called BLUPs), so it grows with the number of groups. Use ran_pars to report variance components, ran_vals to plot per-group deviations.

Can broom.mixed tidy return profile confidence intervals?

Yes. Set conf.method = "profile" and tidy() calls confint(fit, method = "profile") under the hood, returning likelihood profile bounds for both fixed effects and variance components. The call is several times slower than the Wald default but produces intervals that respect the natural boundary at zero for SDs.

Why does tidy(fit) show different standard errors than summary(fit)?

They should match exactly for fixed effects under conf.method = "Wald". If the numbers differ, you have probably loaded lmerTest after fitting, which can change the summary() standard errors when Kenward-Roger adjustments kick in. Refit the model after attaching lmerTest, or call tidy() on the lmerTest-classed object to bring the two into agreement.

Does tidy() work with nlme::lme objects?

No. tidy.merMod only dispatches on lme4's S4 classes. For nlme::lme fits, broom.mixed provides a separate tidy.lme method that takes similar arguments but lives in the same package. Both share the column conventions (effect, group, term, estimate, std.error, conf.low, conf.high) so downstream code can stay model-agnostic.