Estimating Phylogenetic Multilevel Models with brms

Paul Bürkner



In the present vignette, we want to discuss how to specify phylogenetic multilevel models using brms. These models are relevant in evolutionary biology when data of many species are analyzed at the same time. The usual approach would be to model species as a grouping factor in a multilevel model and estimate varying intercepts (and possibly also varying slopes) over species. However, species are not independent as they come from the same phylogenetic tree and we thus have to adjust our model to incorporate this dependency. The examples discussed here are from the book Modern Phylogenetic Comparative Methods and the application in Evolutionary Biology (Garamszegi, 2014) – more specifically from the online practice material of Chapter 11. The necessary data can be downloaded from Some of these models may take a few minutes to fit.

Throughout the vignette, R code output is not shown to reduce installation time and size of the package.

A Simple Phylogenetic Model

Assume we have measurements of a phenotype, phen (say the body size), and a cofactor variable (say the temperature of the environment). We prepare the data using the following code.

setwd("<insert path here>")
phylo <-"phylo.nex")
data_simple <- read.table("data_simple.txt", header = TRUE)

The packages ape and MCMCglmm are required only for the data preparation and not for actually fitting the models. The phylo object contains information on the relationship between species. Using this information, we can construct a covariance matrix of species.

inv.phylo <- MCMCglmm::inverseA(phylo, nodes = "TIPS", scale = TRUE)
A <- solve(inv.phylo$Ainv)
rownames(A) <- rownames(inv.phylo$Ainv)

In contrast to MCMCglmm, brms requires the covariance matrix not its inverse. Now, we are ready to fit our first phylogenetic multilevel model:

model_simple <- brm(phen ~ cofactor + (1|phylo), data = data_simple, 
                    family = gaussian(), cov_ranef = list(phylo = A),
                    prior = c(prior(normal(0, 10), "b"),
                              prior(normal(0, 50), "Intercept"),
                              prior(student_t(3, 0, 20), "sd"),
                              prior(student_t(3, 0, 20), "sigma")))

With the exception of cov_ranef = list(phylo = A) this is a basic multilevel model with a varying intercept over species (phylo is an indicator of species in this data set). However, by using the cov_ranef argument, we make sure that species are correlated as specified by the covariance matrix A. Setting priors is not required for achieving good convergence for this model, but it improves sampling speed a bit. After fitting, the results can be investigated in detail.

plot(marginal_effects(model_simple), points = TRUE) 

The so called phylogenetic signal (often symbolize by \(\lambda\)) can be computed with the hypothesis method and is roughly \(\lambda = 0.7\) for this example.

hyp <- "sd_phylo__Intercept^2 / (sd_phylo__Intercept^2 + sigma^2) = 0"
(hyp <- hypothesis(model_simple, hyp, class = NULL))

Note that the phylogenetic signal is just a synonym of the intra-class correlation (ICC) used in the context phylogenetic analysis.

A Phylogenetic Model with Repeated Mesurements

Often, we have multiple observations per species and this allows to fit more complicated phylogenetic models.

data_repeat <- read.table("data_repeat.txt", header = TRUE)
data_repeat$spec_mean_cf <- 
  with(data_repeat, sapply(split(cofactor, phylo), mean)[phylo])

The variable spec_mean_cf just contains the mean of the cofactor for each species. The code for the repeated measurement phylogenetic model looks as follows:

model_repeat1 <- brm(phen ~ spec_mean_cf + (1|phylo) + (1|species), 
                     data = data_repeat, family = gaussian(), 
                     cov_ranef = list(phylo = A),
                     prior = c(prior(normal(0,10), "b"),
                               prior(normal(0,50), "Intercept"),
                               prior(student_t(3,0,20), "sd"),
                               prior(student_t(3,0,20), "sigma")),
                     sample_prior = TRUE, chains = 2, cores = 2, 
                     iter = 4000, warmup = 1000)

The variables phylo and species are identical as they are both identifiers of the species. However, we model the phylogenetic covariance only for phylo and thus the species variable accounts for any specific effect that would be independent of the phylogenetic relationship between species (e.g., environmental or niche effects). Again we can obtain model summaries as well as estimates of the phylogenetic signal.

plot(marginal_effects(model_repeat1), points = TRUE)
hyp <- paste(
  "sd_phylo__Intercept^2 /", 
  "(sd_phylo__Intercept^2 + sd_species__Intercept^2 + sigma^2) = 0"
(hyp <- hypothesis(model_repeat1, hyp, class = NULL))
plot(hyp, chars = NULL)

So far, we have completely ignored the variability of the cofactor within species. To incorporate this into the model, we define

data_repeat$within_spec_cf <- data_repeat$cofactor - data_repeat$spec_mean_cf

and then fit it again using within_spec_cf as an additional predictor.

model_repeat2 <- update(model_repeat1, formula = ~ . + within_spec_cf,
                        newdata = data_repeat, chains = 2, cores = 2, 
                        iter = 4000, warmup = 1000)

The results are almost unchanged, with apparently no relationship between the phenotype and the within species variance of cofactor.

plot(model_repeat2, N = 6)
plot(marginal_effects(model_repeat2), points = TRUE) 

Also, the phylogenetic signal remains more or less the same.

hyp <- paste(
  "sd_phylo__Intercept^2 /", 
  "(sd_phylo__Intercept^2 + sd_species__Intercept^2 + sigma^2) = 0"
(hyp <- hypothesis(model_repeat2, hyp, class = NULL))
plot(hyp, chars = NULL)

A Phylogenetic Meta-Analysis

Let’s say we have Fisher’s z-transformated correlation coefficients \(Zr\) per species along with corresponding sample sizes (e.g., correlations between male coloration and reproductive success):

data_fisher <- read.table("data_effect.txt", header = TRUE)
data_fisher$obs <- 1:nrow(data_fisher)

We assume the sampling variance to be known and as \(V(Zr) = \frac{1}{N - 3}\) for Fisher’s values, where \(N\) is the sample size per species. Incorporating the known sampling variance into the model is straight forward. One has to keep in mind though, that brms requires the sampling standard deviation (square root of the variance) as input instead of the variance itself. The group-level effect of obs represents the residual variance, which we have to model explicitly in a meta-analytic model.

model_fisher <- brm(Zr | se(sqrt(1 / (N - 3))) ~ 1 + (1|phylo) + (1|obs), 
                    data = data_fisher, family = gaussian(), 
                    cov_ranef = list(phylo = A),
                    prior = c(prior(normal(0, 10), "Intercept"),
                              prior(student_t(3, 0, 10), "sd")),
                    control = list(adapt_delta = 0.95),
                    chains = 2, cores = 2, iter = 4000, warmup = 1000)

A summary of the fitted model is obtained via


The meta-analytic mean (i.e., the model intercept) is \(0.16\) with a credible interval of \([0.07, 0.24]\). Thus the mean correlation across species is positive according to the model.

A phylogenetic count-data model

Suppose that we analyze a phenotype that consists of counts instead of being a continuous variable. In such a case, the normality assumption will likely not be justified and it is recommended to use a distribution explicitely suited for count data, for instance the Poisson distribution. The following data set (again retrieved from provides an example.

data_pois <- read.table("data_pois.txt", header = TRUE)
data_pois$obs <- 1:nrow(data_pois)

As the poisson distribution does not have a natural overdispersion parameter, we model the residual variance via the group-level effects of obs (e.g., see Lawless, 1987).

model_pois <- brm(phen_pois ~ cofactor + (1|phylo) + (1|obs), 
                  data = data_pois, family = poisson("log"), 
                  cov_ranef = list(phylo = A),
                  chains = 2, cores = 2, iter = 4000,
                  control = list(adapt_delta = 0.95))

Again, we obtain a summary of the fitted model via

plot(marginal_effects(model_pois), points = TRUE) 

Now, assume we ignore the fact that the phenotype is count data and fit a linear normal model instead.

model_normal <- brm(phen_pois ~ cofactor + (1|phylo), 
                    data = data_pois, family = gaussian(), 
                    cov_ranef = list(phylo = A),
                    chains = 2, cores = 2, iter = 4000,
                    control = list(adapt_delta = 0.95))

We see that cofactor has a positive relationship with the phenotype in both models. One should keep in mind, though, that the estimates of the Poisson model are on the log-scale, as we applied the canonical log-link function in this example. Therefore, estimates are not comparable to a linear normal model even if applied to the same data. What we can compare, however, is the model fit, for instance graphically via posterior predictive checks.


Apparently, the distribution of the phenotype prediced by the Poisson model resembles the original distribution of the phenotype pretty closely, while the normal models fails to do so. We can also apply leave-one-out cross-validation for direct numerical comparison of model fit.

LOO(model_pois, model_normal)

Since smaller values of LOO indicate better fit, it is again evident that the Poisson model fits the data better than the normal model. Of course, the Poisson model is not the only reasonable option here. For instance, you could use a negative binomial model (via family negative_binomial), which already contains an overdispersion parameter so that modeling a varying intercept of obs becomes obsolete.

Phylogenetic models with multiple group-level effects

In the above examples, we have only used a single group-level effect (i.e., a varying intercept) for the phylogenetic grouping factors. In brms, it is also possible to estimate multiple group-level effects (e.g., a varying intercept and a varying slope) for these grouping factors. However, it requires repeatedly computing Kronecker products and Cholesky factors of covariance matrices while fitting the model. This will be very slow especially when the grouping factors have many levels and matrices are thus large. I hope to increase efficiency for such models in the future.


Garamszegi, L. Z. (2014). Modern Phylogenetic Comparative Methods and the application in Evolutionary Biology. London, UK: Springer.

Lawless, J. F. (1987). Negative binomial and mixed Poisson regression. Canadian Journal of Statistics, 15(3), 209-225.