This function calculates the intraclass-correlation coefficient (ICC) -
sometimes also called variance partition coefficient (VPC) or
repeatability - for mixed effects models. The ICC can be calculated for all
models supported by insight::get_variance(). For models fitted with the
brms-package, icc() might fail due to the large variety of
models and families supported by the brms-package. In such cases, an
alternative to the ICC is the variance_decomposition(), which is based
on the posterior predictive distribution (see 'Details').
icc(
model,
by_group = FALSE,
tolerance = 1e-05,
ci = NULL,
iterations = 100,
ci_method = NULL,
null_model = NULL,
approximation = "lognormal",
model_component = NULL,
verbose = TRUE,
...
)variance_decomposition(model, re_formula = NULL, robust = TRUE, ci = 0.95, ...)
A list with two values, the adjusted ICC and the unadjusted ICC. For
variance_decomposition(), a list with two values, the decomposed
ICC as well as the credible intervals for this ICC.
A (Bayesian) mixed effects model.
Logical, if TRUE, icc() returns the variance components
for each random-effects level (if there are multiple levels). See 'Details'.
Tolerance for singularity check of random effects, to decide
whether to compute random effect variances or not. Indicates up to which
value the convergence result is accepted. The larger tolerance is, the
stricter the test will be. See performance::check_singularity().
Confidence resp. credible interval level. For icc(), r2(), and
rmse(), confidence intervals are based on bootstrapped samples from the
ICC, R2 or RMSE value. See iterations.
Number of bootstrap-replicates when computing confidence intervals for the ICC, R2, RMSE etc.
Character string, indicating the bootstrap-method. Should
be NULL (default), in which case lme4::bootMer() is used for bootstrapped
confidence intervals. However, if bootstrapped intervals cannot be calculated
this way, try ci_method = "boot", which falls back to boot::boot(). This
may successfully return bootstrapped confidence intervals, but bootstrapped
samples may not be appropriate for the multilevel structure of the model.
There is also an option ci_method = "analytical", which tries to calculate
analytical confidence assuming a chi-squared distribution. However, these
intervals are rather inaccurate and often too narrow. It is recommended to
calculate bootstrapped confidence intervals for mixed models.
Optional, a null model to compute the random effect variances,
which is passed to insight::get_variance(). Usually only required if
calculation of r-squared or ICC fails when null_model is not specified. If
calculating the null model takes longer and you already have fit the null
model, you can pass it here, too, to speed up the process.
Character string, indicating the approximation method
for the distribution-specific (observation level, or residual) variance. Only
applies to non-Gaussian models. Can be "lognormal" (default), "delta" or
"trigamma". For binomial models, the default is the theoretical
distribution specific variance, however, it can also be
"observation_level". See Nakagawa et al. 2017, in particular supplement
2, for details.
For models that can have a zero-inflation component,
specify for which component variances should be returned. If NULL or
"full" (the default), both the conditional and the zero-inflation component
are taken into account. If "conditional", only the conditional component is
considered.
Toggle warnings and messages.
Arguments passed down to lme4::bootMer() or boot::boot()
for bootstrapped ICC, R2, RMSE etc.; for variance_decomposition(),
arguments are passed down to brms::posterior_predict().
Formula containing group-level effects to be considered in
the prediction. If NULL (default), include all group-level effects.
Else, for instance for nested models, name a specific group-level effect
to calculate the variance decomposition for this group-level. See 'Details'
and ?brms::posterior_predict.
Logical, if TRUE, the median instead of mean is used to
calculate the central tendency of the variances.
The single variance components that are required to calculate the marginal
and conditional r-squared values are calculated using the insight::get_variance()
function. The results are validated against the solutions provided by
Nakagawa et al. (2017), in particular examples shown in the Supplement 2
of the paper. Other model families are validated against results from the
MuMIn package. This means that the r-squared values returned by r2_nakagawa()
should be accurate and reliable for following mixed models or model families:
Bernoulli (logistic) regression
Binomial regression (with other than binary outcomes)
Poisson and Quasi-Poisson regression
Negative binomial regression (including nbinom1, nbinom2 and nbinom12 families)
Gaussian regression (linear models)
Gamma regression
Tweedie regression
Beta regression
Ordered beta regression
Following model families are not yet validated, but should work:
Zero-inflated and hurdle models
Beta-binomial regression
Compound Poisson regression
Generalized Poisson regression
Log-normal regression
Skew-normal regression
Extracting variance components for models with zero-inflation part is not straightforward, because it is not definitely clear how the distribution-specific variance should be calculated. Therefore, it is recommended to carefully inspect the results, and probably validate against other models, e.g. Bayesian models (although results may be only roughly comparable).
Log-normal regressions (e.g. lognormal() family in glmmTMB or gaussian("log"))
often have a very low fixed effects variance (if they were calculated as
suggested by Nakagawa et al. 2017). This results in very low ICC or
r-squared values, which may not be meaningful.
The ICC can be interpreted as "the proportion of the variance explained by the grouping structure in the population". The grouping structure entails that measurements are organized into groups (e.g., test scores in a school can be grouped by classroom if there are multiple classrooms and each classroom was administered the same test) and ICC indexes how strongly measurements in the same group resemble each other. This index goes from 0, if the grouping conveys no information, to 1, if all observations in a group are identical (Gelman and Hill, 2007, p. 258). In other word, the ICC - sometimes conceptualized as the measurement repeatability - "can also be interpreted as the expected correlation between two randomly drawn units that are in the same group" (Hox 2010: 15), although this definition might not apply to mixed models with more complex random effects structures. The ICC can help determine whether a mixed model is even necessary: an ICC of zero (or very close to zero) means the observations within clusters are no more similar than observations from different clusters, and setting it as a random factor might not be necessary.
The coefficient of determination R2 (that can be computed with r2())
quantifies the proportion of variance explained by a statistical model, but
its definition in mixed model is complex (hence, different methods to compute
a proxy exist). ICC is related to R2 because they are both ratios of
variance components. More precisely, R2 is the proportion of the explained
variance (of the full model), while the ICC is the proportion of explained
variance that can be attributed to the random effects. In simple cases, the
ICC corresponds to the difference between the conditional R2 and the
marginal R2 (see r2_nakagawa()).
The ICC is calculated by dividing the random effect variance, σ2i, by the total variance, i.e. the sum of the random effect variance and the residual variance, σ2ε.
icc() calculates an adjusted and an unadjusted ICC, which both take all
sources of uncertainty (i.e. of all random effects) into account. While
the adjusted ICC only relates to the random effects, the unadjusted ICC
also takes the fixed effects variances into account, more precisely, the
fixed effects variance is added to the denominator of the formula to
calculate the ICC (see Nakagawa et al. 2017). Typically, the adjusted
ICC is of interest when the analysis of random effects is of interest.
icc() returns a meaningful ICC also for more complex random effects
structures, like models with random slopes or nested design (more than two
levels) and is applicable for models with other distributions than Gaussian.
For more details on the computation of the variances, see
?insight::get_variance.
Usually, the ICC is calculated for the null model ("unconditional model"). However, according to Raudenbush and Bryk (2002) or Rabe-Hesketh and Skrondal (2012) it is also feasible to compute the ICC for full models with covariates ("conditional models") and compare how much, e.g., a level-2 variable explains the portion of variation in the grouping structure (random intercept).
The proportion of variance for specific levels related to the overall model
can be computed by setting by_group = TRUE. The reported ICC is
the variance for each (random effect) group compared to the total
variance of the model. For mixed models with a simple random intercept,
this is identical to the classical (adjusted) ICC.
If model is of class brmsfit, icc() might fail due to the large
variety of models and families supported by the brms package. In such
cases, variance_decomposition() is an alternative ICC measure. The function
calculates a variance decomposition based on the posterior predictive
distribution. In this case, first, the draws from the posterior predictive
distribution not conditioned on group-level terms
(posterior_predict(..., re_formula = NA)) are calculated as well as draws
from this distribution conditioned on all random effects (by default,
unless specified else in re_formula) are taken. Then, second, the variances
for each of these draws are calculated. The "ICC" is then the ratio between
these two variances. This is the recommended way to analyse
random-effect-variances for non-Gaussian models. It is then possible to
compare variances across models, also by specifying different group-level
terms via the re_formula-argument.
Sometimes, when the variance of the posterior predictive distribution is
very large, the variance ratio in the output makes no sense, e.g. because
it is negative. In such cases, it might help to use robust = TRUE.
Hox, J. J. (2010). Multilevel analysis: techniques and applications (2nd ed). New York: Routledge.
Nakagawa, S., Johnson, P. C. D., and Schielzeth, H. (2017). The coefficient of determination R2 and intra-class correlation coefficient from generalized linear mixed-effects models revisited and expanded. Journal of The Royal Society Interface, 14(134), 20170213.
Rabe-Hesketh, S., and Skrondal, A. (2012). Multilevel and longitudinal modeling using Stata (3rd ed). College Station, Tex: Stata Press Publication.
Raudenbush, S. W., and Bryk, A. S. (2002). Hierarchical linear models: applications and data analysis methods (2nd ed). Thousand Oaks: Sage Publications.
model <- lme4::lmer(Sepal.Length ~ Petal.Length + (1 | Species), data = iris)
icc(model)
# ICC for specific group-levels
data(sleepstudy, package = "lme4")
set.seed(12345)
sleepstudy$grp <- sample(1:5, size = 180, replace = TRUE)
sleepstudy$subgrp <- NA
for (i in 1:5) {
filter_group <- sleepstudy$grp == i
sleepstudy$subgrp[filter_group] <-
sample(1:30, size = sum(filter_group), replace = TRUE)
}
model <- lme4::lmer(
Reaction ~ Days + (1 | grp / subgrp) + (1 | Subject),
data = sleepstudy
)
icc(model, by_group = TRUE)
Run the code above in your browser using DataLab