Advanced Bayesian Multilevel Modeling with the R Package brms

Paul-Christian Bürkner

Abstract

The brms package allows R users to easily specify a wide range of Bayesian single-level and multilevel models, which are fitted with the probabilistic programming language Stan behind the scenes. Several response distributions are supported, of which all parameters (e.g., location, scale, and shape) can be predicted at the same time thus allowing for distributional regression. Non-linear relationships may be specified using non-linear predictor terms or semi-parametric approaches such as splines or Gaussian processes. To make all of these modeling options possible in a multilevel framework, brms provides an intuitive and powerful formula syntax, which extends the well known formula syntax of lme4. The purpose of the present paper is to introduce this syntax in detail and to demonstrate its usefulness with four examples, each showing other relevant aspects of the syntax.

Introduction

Multilevel models (MLMs) offer great flexibility for researchers across sciences (Brown and Prescott, 2015; Demidenko, 2013; Gelman and Hill, 2006; Pinheiro and Bates, 2006). They allow modeling of data measured on different levels at the same time – for instance data of students nested within classes and schools – thus taking complex dependency structures into account. It is not surprising that many packages for R have been developed to fit MLMs. Usually, however, the functionality of these implementations is limited insofar as it is only possible to predict the mean of the response distribution. Other parameters of the response distribution, such as the residual standard deviation in linear models, are assumed constant across observations, which may be violated in many applications. Accordingly, it is desirable to allow for prediction of all response parameters at the same time. Models doing exactly that are often referred to as distributional models or more verbosely models for location, scale and shape (Rigby and Stasinopoulos, 2005). Another limitation of basic MLMs is that they only allow for linear predictor terms. While linear predictor terms already offer good flexibility, they are of limited use when relationships are inherently non-linear. Such non-linearity can be handled in at least two ways: (1) by fully specifying a non-linear predictor term with corresponding parameters each of which can be predicted using MLMs (Lindstrom and Bates, 1990), or (2) estimating the form of the non-linear relationship on the fly using splines (Wood, 2004) or Gaussian processes (Rasmussen and Williams, 2006). The former are often simply called non-linear models, while models applying splines are referred to as generalized additive models (GAMs; Hastie and Tibshirani, 1990).

Combining all of these modeling options into one framework is a complex task, both conceptually and with regard to model fitting. Maximum likelihood methods, which are typically applied in classical ’frequentist’ statistics, can reach their limits at some point and fully Bayesian methods become the go-to solutions to fit such complex models (Gelman et al., 2014). In addition to being more flexible, the Bayesian framework comes with other advantages, for instance, the ability to derive probability statements for every quantity of interest or explicitly incorporating prior knowledge about parameters into the model. The former is particularly relevant in non-linear models, for which classical approaches struggle more often than not in propagating all the uncertainty in the parameter estimates to non-linear functions such as out-of-sample predictions.

Possibly the most powerful program for performing full Bayesian inference available to date is Stan (Stan Development Team, 2017b; Carpenter et al., 2017). It implements Hamiltonian Monte Carlo (Duane et al., 1987; Neal, 2011; Betancourt et al., 2014) and its extension, the No-U-Turn (NUTS) Sampler (Hoffman and Gelman, 2014). These algorithms converge much more quickly than other Markov-Chain Monte-Carlo (MCMC) algorithms especially for high-dimensional models (Hoffman and Gelman, 2014; Betancourt et al., 2014; Betancourt, 2017). An excellent non-mathematical introduction to Hamiltonian Monte Carlo can be found in Betancourt (2017).

Stan comes with its own programming language, allowing for great modeling flexibility Stan Development Team (2017b); Carpenter et al. (2017)). Many researchers may still be hesitent to use Stan directly, as every model has to be written, debugged and possibly also optimized. This may be a time-consuming and error-prone process even for researchers familiar with Bayesian inference. The brms package Bürkner (2017), presented in this paper, aims to remove these hurdles for a wide range of regression models by allowing the user to benefit from the merits of Stan by using extended lme4-like (Bates et al., 2015) formula syntax, with which many R users are familiar with. It offers much more than writing efficient and human-readable Stan code: brms comes with many post-processing and visualization functions, for instance to perform posterior predictive checks, leave-one-out cross-validation, visualization of estimated effects, and prediction of new data. The overarching aim is to have one general framework for regression modeling, which offers everything required to successfully apply regression models to complex data. To date, it already replaces and extends the functionality of dozens of other R packages, each of which is restricted to specific regression models¹¹1Unfortunately, due to the implementation via Stan, it is not easily possible for users to define their own response distributions and run them via brms. If you feel that a response distribution is missing in brms, please open an issue on GitHub (https://github.com/paul-buerkner/brms)..

The purpose of the present article is to provide an introduction of the advanced multilevel formula syntax implemented in brms, which allows to fit a wide and growing range of non-linear distributional multilevel models. A general overview of the package is already given in Bürkner (2017). Accordingly, the present article focuses on more recent developments. We begin by explaining the underlying structure of distributional models. Next, the formula syntax of lme4 and its extensions implemented in brms are explained. Four examples that demonstrate the use of the new syntax are discussed in detail. We end by describing future plans for extending the package.

Model description

The core of models implemented in brms is the prediction of the response $y$ through predicting all parameters $\theta_{p}$ of the response distribution $D$ , which is also called the model family in many R packages. We write

y_{i}\sim D(\theta_{1i},\theta_{2i},...)

to stress the dependency on the $i\textsuperscript{th}$ observation. Every parameter $\theta_{p}$ may be regressed on its own predictor term $\eta_{p}$ transformed by the inverse link function $f_{p}$ that is $\theta_{pi}=f_{p}(\eta_{pi})$ ²²2A parameter can also be assumed constant across observations so that a linear predictor is not required.. Such models are typically refered to as distributional models³³3The models described in Bürkner (2017) are a sub-class of the here described models.. Details about the parameterization of each family are given in vignette("brms_families").

Suppressing the index $p$ for simplicity, a predictor term $\eta$ can generally be written as

\eta=\mathbf{X}\beta+\mathbf{Z}u+\sum_{k=1}^{K}s_{k}(x_{k})

In this equation, $\beta$ and $u$ are the coefficients at population-level and group-level respectively and $\mathbf{X},\mathbf{Z}$ are the corresponding design matrices. The terms $s_{k}(x_{k})$ symbolize optional smooth functions of unspecified form based on covariates $x_{k}$ fitted via splines (see Wood (2011) for the underlying implementation in the mgcv package) or Gaussian processes (Williams and Rasmussen, 1996). The response $y$ as well as $\mathbf{X}$ , $\mathbf{Z}$ , and $x_{k}$ make up the data, whereas $\beta$ , $u$ , and the smooth functions $s_{k}$ are the model parameters being estimated. The coefficients $\beta$ and $u$ may be more commonly known as fixed and random effects, but I avoid theses terms following the recommendations of Gelman and Hill (2006). Details about prior distributions of $\beta$ and $u$ can be found in Bürkner (2017) and under help("set_prior").

As an alternative to the strictly additive formulation described above, predictor terms may also have any form specifiable in Stan. We call it a non-linear predictor and write

\eta=f(c_{1},c_{2},...,\phi_{1},\phi_{2},...)

The structure of the function $f$ is given by the user, $c_{r}$ are known or observed covariates, and $\phi_{s}$ are non-linear parameters each having its own linear predictor term $\eta_{\phi_{s}}$ of the form specified above. In fact, we should think of non-linear parameters as placeholders for linear predictor terms rather than as parameters themselves. A frequentist implementation of such models, which inspired the non-linear syntax in brms, can be found in the nlme package (Pinheiro et al., 2016).

Extended multilevel formula syntax

The formula syntax applied in brms builds upon the syntax of the R package lme4 (Bates et al., 2015). First, we will briefly explain the lme4 syntax used to specify multilevel models and then introduce certain extensions that allow to specify much more complicated models in brms. An lme4 formula has the general form

response ˜ pterms + (gterms | group)The pterms part contains the population-level effects that are assumed to be the same across obervations. The gterms part contains so called group-level effects that are assumed to vary accross grouping variables specified in group. Multiple grouping factors each with multiple group-level effects are possible. Usually, group contains only a single variable name pointing to a factor, but you may also use g1:g2 or g1/g2, if both g1 and g2 are suitable grouping factors. The : operator creates a new grouping factor that consists of the combined levels of g1 and g2 (you could think of this as pasting the levels of both factors together). The / operator indicates nested grouping structures and expands one grouping factor into two or more when using multiple / within one term. If, for instance, you write (1 | g1/g2), it will be expanded to (1 | g1) + (1 | g1:g2). Instead of | you may use || in grouping terms to prevent group-level correlations from being modeled. This may be useful in particular when modeling so many group-level effects that convergence of the fitting algorithms becomes an issue due to model complexity. One limitation of the || operator in lme4 is that it only splits up terms so that columns of the design matrix originating from the same term are still modeled as correlated (e.g., when coding a categorical predictor; see the mixed function of the afex package by Singmann et al. (2015) for a way to avoid this behavior).

While intuitive and visually appealing, the classic lme4 syntax is not flexible enough to allow for specifying the more complex models supported by brms. In non-linear or distributional models, for instance, multiple parameters are predicted, each having their own population and group-level effects. Hence, multiple formulas are necessary to specify such models⁴⁴4Actually, it is possible to specify multiple model parts within one formula using interactions terms for instance as implemented in MCMCglmm (Hadfield, 2010). However, this syntax is limited in flexibility and requires a rather deep understanding of the way R parses formulas, thus often being confusing to users.. Then, however, specifying group-level effects of the same grouping factor to be correlated across formulas becomes complicated. The solution implemented in brms (and currently unique to it) is to expand the | operator into |<ID>|, where <ID> can be any value. Group-level terms with the same ID will then be modeled as correlated if they share same grouping factor(s)⁵⁵5It might even be further extended to |fun(<ID>)|, where fun defines the type of correlation structure, defaulting to unstructured that is estimating the full correlation matrix. The fun argument is not yet supported by brms but could be supported in the future if other correlation structures, such as compound symmetry or Toeplitz, turn out to have reasonable practical applications and effective implementations in Stan.. For instance, if the terms (x1|ID|g1) and (x2|ID|g1) appear somewhere in the same or different formulas passed to brms, they will be modeled as correlated.

Further extensions of the classical lme4 syntax refer to the group part. It is rather limited in its flexibility since only variable names combined by : or / are supported. We propose two extensions of this syntax: Firstly, group can generally be split up in its terms so that, say, (1 | g1 + g2) is expanded to (1 | g1) + (1 | g2). This is fully consistent with the way / is handled so it provides a natural generalization to the existing syntax. Secondly, there are some special grouping structures that cannot be expressed by simply combining grouping variables. For instance, multi-membership models cannot be expressed this way. To overcome this limitation, we propose wrapping terms in group within special functions that allow specifying alternative grouping structures: (gterms | fun(group)). In brms, there are currently two such functions implemented, namely gr for the default behavior and mm for multi-membership terms. To be compatible with the original syntax and to keep formulas short, gr is automatically added internally if none of these functions is specified.

While some non-linear relationships, such as quadratic relationships, can be expressed within the basic R formula syntax, other more complicated ones cannot. For this reason, it is possible in brms to fully specify non-linear predictor terms similar to how it is done in nlme, but fully compatible with the extended multilevel syntax described above. Suppose, for instance, we want to model the non-linear growth curve

y=b_{1}(1-\exp(-(x/b_{2})^{b_{3}})

between $y$ and $x$ with parameters $b_{1}$ , $b_{2}$ , and $b_{3}$ (see Example 3 in this paper for an implementation of this model with real data). Furthermore, we want all three parameters to vary by a grouping variable $g$ and model those group-level effects as correlated. Additionally $b_{1}$ should be predicted by a covariate $z$ . We can express this in brms using multiple formulas, one for the non-linear model itself and one per non-linear parameter:

y ˜ b1 * (1 - exp(-(x / b2) ˆ b3)b1 ˜ z + (1|ID|g)b2 ˜ (1|ID|g)b3 ˜ (1|ID|g)The first formula will not be evaluated using standard R formula parsing, but instead taken literally. In contrast, the formulas for the non-linear parameters will be evaluated in the usual way and are compatible with all terms supported by brms. Note that we have used the above described ID-syntax to model group-level effects as correlated across formulas.

There are other syntax extensions implemented in brms that do not directly target grouping terms. Firstly, there are terms formally included in the pterms part that are handled separately. The most prominent examples are smooth terms specified through the s and t2 functions of the mgcv package (Wood, 2011). Other examples are category specific effects cs, monotonic effects mo, noise-free effects me, or Gaussian process terms gp. The former is explained in Bürkner (2017), while the latter three are documented in help(brmsformula). Internally, these terms are extracted from pterms and not included in the construction of the population-level design matrix. Secondly, making use of the fact that | is unused on the left-hand side of $\sim$ in formula, additional information on the response variable may be specified via

response | aterms ˜ <predictor terms>The aterms part may contain multiple terms of the form fun(<variable>) separated by + each providing special information on the response variable. This allows among others to weight observations, provide known standard errors for meta-analysis, or model censored or truncated data. As it is not the main topic of the present paper, we refer to help("brmsformula") and help("addition-terms") for more details.

To set up the model formulas and combine them into one object, brms defines the brmsformula (or short bf) function. Its output can then be passed to the parse_bf function, which splits up the formulas in separate parts and prepares them for the generation of design matrices and related data. Other packages may re-use these functions in their own routines making it easier to offer support for the above described multilevel syntax.

Examples

The idea of brms is to provide one unified framework for multilevel regression models in R. As such, the above described formula syntax in all of its variations can be applied in combination with all response distributions supported by brms (currently about 35 response distributions are supported; see help("brmsfamily") and vignette("brms_families") for an overview).

In this section, we will discuss four examples in detail, each focusing on certain aspects of the syntax. They are chosen to provide a broad overview of the modeling options. The first is about the number of fish caught be different groups of people. It does not actually contain any multilevel structure, but helps in understanding how to set up formulas for different model parts. The second example is about housing rents in Munich. We model the data using splines and a distributional regression approach. The third example is about cumulative insurance loss payments across several years, which is fitted using a rather complex non-linear multilevel model. Finally, the fourth example is about the performance of school children, who change school during the year, thus requiring a multi-membership model.

Despite not being covered in the four examples, there are a few more modeling options that we want to briefly describe. First, brms allows fitting so called phylogenetic models. These models are relevant in evolutionary biology when data of many species are analyzed at the same time. Species are not independent as they come from the same phylogenetic tree, implying that different levels of the same grouping-factor (i.e., species) are likely correlated. There is a whole vignette dedicated to this topic, which can be found via vignette("brms_phylogenetics"). Second, there is a canonical way to handle ordinal predictors, without falsely assuming they are either categorical or continuous. We call them monotonic effects and discuss them in vignette("brms_monotonic"). Last but not least, it is possible to account for measurement error in both response and predictor variables. This is often ignored in applied regression modeling (Westfall and Yarkoni, 2016), although measurement error is very common in all scientific fields making use of observational data. There is no vignette yet covering this topic, but one will be added in the future. In the meantime, help("brmsformula") is the best place to start learning about such models as well as about other details of the brms formula syntax.

Example 1: Catching fish

An important application of the distributional regression framework of brms are so called zero-inflated and hurdle models. These models are helpful whenever there are more zeros in the response variable than one would naturally expect. Here, we consider an example dealing with the number of fish caught by various groups of people. On the UCLA website (https://stats.idre.ucla.edu/stata/dae/zero-inflated-poisson-regression), the data are described as follows: “The state wildlife biologists want to model how many fish are being caught by fishermen at a state park. Visitors are asked how long they stayed, how many people were in the group, were there children in the group and how many fish were caught. Some visitors do not fish, but there is no data on whether a person fished or not. Some visitors who did fish did not catch any fish so there are excess zeros in the data because of the people that did not fish.”

zinb <- read.csv("http://stats.idre.ucla.edu/stat/data/fish.csv")zinb$camper <- factor(zinb$camper, labels = c("no", "yes"))head(zinb)

nofish livebait camper persons child xb zg count1 1 0 no 1 0 -0.8963146 3.0504048 02 0 1 yes 1 0 -0.5583450 1.7461489 03 0 1 no 1 0 -0.4017310 0.2799389 04 0 1 yes 2 1 -0.9562981 -0.6015257 05 0 1 no 1 0 0.4368910 0.5277091 16 0 1 yes 4 2 1.3944855 -0.7075348 0As predictors we choose the number of people per group, the number of children, as well as whether or not the group consists of campers. Many groups may not catch any fish just because they do not try and so we fit a zero-inflated Poisson model. For now, we assume a constant zero-inflation probability across observations.

fit_zinb1 <- brm(count ˜ persons + child + camper, data = zinb, family = zero_inflated_poisson("log"))The model is readily summarized via

summary(fit_zinb1)

Family: zero_inflated_poisson (log)Formula: count ˜ persons + child + camper Data: zinb (Number of observations: 250)Samples: 4 chains, each with iter = 2000; warmup = 1000; thin = 1; total post-warmup samples = 4000 WAIC: Not computedPopulation-Level Effects: Estimate Est.Error l-95% CI u-95% CI Eff.Sample RhatIntercept -1.01 0.17 -1.34 -0.67 2171 1persons 0.87 0.04 0.79 0.96 2188 1child -1.36 0.09 -1.55 -1.18 1790 1camper 0.80 0.09 0.62 0.98 2950 1Family Specific Parameters: Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhatzi 0.41 0.04 0.32 0.49 2409 1Samples were drawn using sampling(NUTS). For each parameter, Eff.Sampleis a crude measure of effective sample size, and Rhat is the potentialscale reduction factor on split chains (at convergence, Rhat = 1).A graphical summary is available through

marginal_effects(fit_zinb1)

Refer to caption — Figure 1: Marginal effects plots of the fit_zinb1 model.

(see Figure 1). In fact, the marginal_effects method turned out to be so powerful in visualizing effects of predictors that I am using it almost as frequently as summary. According to the parameter estimates, larger groups catch more fish, campers catch more fish than non-campers, and groups with more children catch less fish. The zero-inflation probability zi is pretty large with a mean of 41%. Please note that the probability of catching no fish is actually higher than 41%, but parts of this probability are already modeled by the Poisson distribution itself (hence the name zero-inflation). If you want to treat all zeros as originating from a separate process, you can use hurdle models instead (not shown here).

Now, we try to additionally predict the zero-inflation probability by the number of children. The underlying reasoning is that we expect groups with more children to not even try catching fish, since children often lack the patience required for fishing. From a purely statistical perspective, zero-inflated (and hurdle) distributions are a mixture of two processes and predicting both parts of the model is natural and often very reasonable to make full use of the data.

fit_zinb2 <- brm(bf(count ˜ persons + child + camper, zi ˜ child), data = zinb, family = zero_inflated_poisson())To transform the linear predictor of zi into a probability, brms applies the logit-link, which takes values within $[0,1]$ and returns values on the real line. Thus, it allows the transition between probabilities and linear predictors.

summary(fit_zinb2)

Family: zero_inflated_poisson (log)Formula: count ˜ persons + child + camper zi ˜ child Data: zinb (Number of observations: 250)Samples: 4 chains, each with iter = 2000; warmup = 1000; thin = 1; total post-warmup samples = 4000 WAIC: Not computedPopulation-Level Effects: Estimate Est.Error l-95% CI u-95% CI Eff.Sample RhatIntercept -1.07 0.18 -1.43 -0.73 2322 1persons 0.89 0.05 0.80 0.98 2481 1child -1.17 0.10 -1.37 -1.00 2615 1camper 0.78 0.10 0.60 0.96 3270 1zi_Intercept -0.95 0.27 -1.52 -0.48 2341 1zi_child 1.21 0.28 0.69 1.79 2492 1Samples were drawn using sampling(NUTS). For each parameter, Eff.Sampleis a crude measure of effective sample size, and Rhat is the potentialscale reduction factor on split chains (at convergence, Rhat = 1).

According to the model, trying to fish with children not only decreases the overall number fish caught (as implied by the Poisson part of the model) but also drastically increases your chance of catching no fish at all (as implied by the zero-inflation part), possibly because groups with more children spend less time or no time at all fishing. Comparing model fit via leave-one-out cross validation as implemented in the loo package (Vehtari et al., 2016a, b).

LOO(fit_zinb1, fit_zinb2)

LOOIC SEfit_zinb1 1639.52 363.30fit_zinb2 1621.35 362.39fit_zinb1 - fit_zinb2 18.16 15.71reveals that the second model using the number of children to predict both model parts has better fit. However, when considering the standard error of the LOOIC difference, improvement in model fit is apparently modest and not substantial. More examples of distributional model can be found in vignette("brms_distreg").

Example 2: Housing rents

In their book about regression modeling, Fahrmeir et al. (2013) use an example about the housing rents in Munich from 1999. The data contains information about roughly 3000 apartments including among others the absolute rent (rent), rent per square meter (rentsqm), size of the apartment (area), construction year (yearc), and the district in Munich (district), where the apartment is located. The data can be found in the gamlss.data package (Stasinopoulos and Rigby, 2016):

data("rent99", package = "gamlss.data")head(rent99)

rent rentsqm area yearc location bath kitchen cheating district1 109.9487 4.228797 26 1918 2 0 0 0 9162 243.2820 8.688646 28 1918 2 0 0 1 8133 261.6410 8.721369 30 1918 1 0 0 1 6114 106.4103 3.547009 30 1918 2 0 0 0 20255 133.3846 4.446154 30 1918 2 0 0 1 5616 339.0256 11.300851 30 1918 2 0 0 1 541Here, we aim at predicting the rent per square meter with the size of the apartment as well as the construction year, while taking the district of Munich into account. As the effect of both predictors on the rent is of unknown non-linear form, we model these variables using a bivariate tensor spline (Wood et al., 2013). The district is accounted for via a varying intercept.

fit_rent1 <- brm(rentsqm ˜ t2(area, yearc) + (1|district), data = rent99, chains = 2, cores = 2)We fit the model using just two chains (instead of the default of four chains) on two processor cores to reduce the model fitting time for the purpose of the present paper. In general, using the default option of four chains (or more) is recommended.

summary(fit_rent1)

Family: gaussian(identity)Formula: rentsqm ˜ t2(area, yearc) + (1 | district) Data: rent99 (Number of observations: 3082)Samples: 2 chains, each with iter = 2000; warmup = 1000; thin = 1; total post-warmup samples = 2000 ICs: LOO = NA; WAIC = NA; R2 = NASmooth Terms: Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhatsds(t2areayearc_1) 4.93 2.32 1.61 10.77 1546 1.00sds(t2areayearc_2) 5.78 2.87 1.58 13.15 1175 1.00sds(t2areayearc_3) 8.09 3.19 3.66 16.22 1418 1.00Group-Level Effects:˜district (Number of levels: 336) Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhatsd(Intercept) 0.60 0.06 0.48 0.73 494 1.01Population-Level Effects: Estimate Est.Error l-95% CI u-95% CI Eff.Sample RhatIntercept 7.80 0.11 7.59 8.02 2000 1.00t2areayearc_1 -1.00 0.09 -1.15 -0.83 2000 1.00t2areayearc_2 0.75 0.17 0.43 1.09 2000 1.00t2areayearc_3 -0.07 0.16 -0.40 0.24 1579 1.00Family Specific Parameters: Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhatsigma 1.95 0.03 1.90 2.01 2000 1.00Samples were drawn using sampling(NUTS). For each parameter, Eff.Sampleis a crude measure of effective sample size, and Rhat is the potentialscale reduction factor on split chains (at convergence, Rhat = 1).For models including splines, the output of summary is not tremendously helpful, but we get at least some information. Firstly, the credible intervals of the standard deviations of the coefficients forming the splines (under ’Smooth Terms’) are sufficiently far away from zero to indicate non-linearity in the (combined) effect of area and yearc. Secondly, even after controlling for these predictors, districts still vary with respect to rent per square meter by a sizable amount as visible under ’Group-Level Effects’ in the output. To further understand the effect of the predictor, we apply graphical methods:

marginal_effects(fit_rent1, surface = TRUE)In Figure 2, the marginal effects of both predictors are displayed, while the respective other predictor is fixed at its mean. In Figure 3, the combined effect is shown, clearly demonstrating an interaction between the variables. In particular, housing rents appear to be highest for small and relatively new apartments.

In the above example, we only considered the mean of the response distribution to vary by area and yearc, but this my not necessarily reasonable assumption, as the variation of the response might vary with these variables as well. Accordingly, we fit splines and effects of district for both the location and the scale parameter, which is called sigma in Gaussian models.

bform <- bf(rentsqm ˜ t2(area, yearc) + (1|ID1|district), sigma ˜ t2(area, yearc) + (1|ID1|district))fit_rent2 <- brm(bform, data = rent99, chains = 2, cores = 2)If not otherwise specified, sigma is predicted on the log-scale to ensure it is positive no matter how the predictor term looks like. Instead of (1|district) as in the previous model, we now use (1|ID1|district) in both formulas. This results in modeling the varying intercepts of both model parts as correlated (see the description of the ID-syntax above). The group-level part of the summary output looks as follows:

Group-Level Effects:˜district (Number of levels: 336) Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhatsd(Intercept) 0.60 0.06 0.49 0.73 744 1.00sd(sigma_Intercept) 0.11 0.02 0.06 0.15 751 1.00cor(Intercept,sigma_Intercept) 0.72 0.17 0.35 0.98 648 1.00As visible from the positive correlation of the intercepts, districts with overall higher rent per square meter have higher variation at the same time. Lastly, we want to turn our attention to the splines. While marginal_effects is used to visualize effects of predictors on the expected response, marginal_smooths is used to show just the spline parts of the model:

marginal_smooths(fit_rent2)The plot on the left-hand side of Figure 4 resembles the one in Figure 3, but the scale is different since only the spline is plotted. The right-hand side of 4 shows the spline for sigma. Since we apply the log-link on sigma by default the spline is on the log-scale as well. As visible in the plot, the variation in the rent per square meter is highest for relatively small and old apartments, while the variation is smallest for medium to large apartments build around the 1960s.

Example 3: Insurance loss payments

On his blog, Markus Gesmann predicts the growth of cumulative insurance loss payments over time, originated from different origin years (see http://www.magesblog.com/2015/11/loss-developments-via-growth-curves-and.html). We will use a slightly simplified version of his model for demonstration purposes here. It looks as follows:

cum_{AY,dev}\sim N(\mu_{AY,dev},\sigma)

\mu_{AY,dev}=ult_{AY}\left(1-\exp\left(-\left(\frac{dev}{\theta}\right)^{\omega}\right)\right)

The cumulative insurance payments $cum$ will grow over time, and we model this dependency using the variable $dev$ . Further, $ult_{AY}$ is the (to be estimated) ultimate loss of accident each year. It constitutes a non-linear parameter in our framework along with the parameters $\theta$ and $\omega$ , which are responsible for the growth of the cumulative loss and are for now assumed to be the same across years. We load the data

url <- paste0("https://raw.githubusercontent.com/mages/", "diesunddas/master/Data/ClarkTriangle.csv")loss <- read.csv(url)head(loss)

AY dev cum1 1991 6 357.8482 1991 18 1124.7883 1991 30 1735.3304 1991 42 2182.7085 1991 54 2745.5966 1991 66 3319.994and translate the proposed model into a non-linear brms model.

nlform <- bf(cum ˜ ult * (1 - exp(-(dev / theta)ˆomega)), ult ˜ 1 + (1|AY), omega ˜ 1, theta ˜ 1, nl = TRUE)nlprior <- c(prior(normal(5000, 1000), nlpar = "ult"), prior(normal(1, 2), nlpar = "omega"), prior(normal(45, 10), nlpar = "theta"))fit_loss1 <- brm(formula = nlform, data = loss, family = gaussian(), prior = nlprior, control = list(adapt_delta = 0.9))

In the above functions calls, quite a few things are going on. The formulas are wrapped in bf to combine them into one object. The first formula specifies the non-linear model. We set argument nl = TRUE so that brms takes this formula literally and instead of using standard R formula parsing. We specify one additional formula per non-linear parameter (a) to clarify what variables are covariates and what are parameters and (b) to specify the predictor term for the parameters. We estimate a group-level effect of accident year (variable AY) for the ultimate loss ult. This also shows nicely how a non-linear parameter is actually a placeholder for a linear predictor, which in the case of ult, contains only a varying intercept for year. Both omega and theta are assumed to be constant across observations so we just fit a population-level intercept.

Priors on population-level effects are required and, for the present model, are actually mandatory to ensure identifiability. Otherwise, we may observe that different Markov chains converge to different parameter regions as multiple posterior distribution are equally plausible. Setting prior distributions is a difficult task especially in non-linear models. It requires some experience and knowledge both about the model that is being fitted and about the data at hand. Additionally, there is more to keep in mind to optimize the sampler’s performance: Firstly, using non- or weakly informative priors in non-linear models often leads to problems even if the model is generally identified. For instance, if a zero-centered and reasonably wide prior such as normal(0, 10000) it set on ult, there is little information about theta and omega for samples of ult being close to zero, which may lead to convergence problems. Secondly, Stan works best when parameters are roughly on the same order of magnitude (Stan Development Team, 2017a). In the present example, ult is of three orders larger than omega. Still, the sampler seems to work quite well, but this may not be true for other models. One solution is to rescale parameters before model fitting. For instance, for the present example, one could have downscaled ult by replacing it with ult * 1000 and correspondingly the normal(5000, 1000) prior with normal(5, 1).

In the control argument we increase adapt_delta to get rid of a few divergent transitions (cf. Stan Development Team, 2017a; Bürkner, 2017). Again the model is summarized via

summary(fit_loss1)

Family: gaussian (identity)Formula: cum ˜ ult * (1 - exp(-(dev / theta)ˆomega)) ult ˜ 1 + (1 | AY) omega ˜ 1 theta ˜ 1 Data: loss (Number of observations: 55)Samples: 4 chains, each with iter = 2000; warmup = 1000; thin = 1; total post-warmup samples = 4000 WAIC: Not computedGroup-Level Effects:˜AY (Number of levels: 10) Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhatsd(ult_Intercept) 745.74 231.31 421.05 1306.04 916 1Population-Level Effects: Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhatult_Intercept 5273.70 292.34 4707.11 5852.28 798 1omega_Intercept 1.34 0.05 1.24 1.43 2167 1theta_Intercept 46.07 2.09 42.38 50.57 1896 1Family Specific Parameters: Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhatsigma 139.93 15.52 113.6 175.33 2358 1Samples were drawn using sampling(NUTS). For each parameter, Eff.Sampleis a crude measure of effective sample size, and Rhat is the potentialscale reduction factor on split chains (at convergence, Rhat = 1).as well as

marginal_effects(fit_loss1)

(see Figure 5). We can also visualize the cumulative insurance loss over time separately for each year.

conditions <- data.frame(AY = unique(loss$AY))rownames(conditions) <- unique(loss$AY)me_year <- marginal_effects(fit_loss1, conditions = conditions, re_formula = NULL, method = "predict")plot(me_year, ncol = 5, points = TRUE)

(see Figure 6). It is evident that there is some variation in cumulative loss across accident years, for instance due to natural disasters happening only in certain years. Further, we see that the uncertainty in the predicted cumulative loss is larger for later years with fewer available data points.

In the above model, we considered omega and delta to be constant across years, which may not necessarily be true. We can easily investigate this by fitting varying intercepts for all three non-linear parameters also estimating group-level correlation using the above introduced ID syntax.

nlform2 <- bf(cum ˜ ult * (1 - exp(-(dev / theta)ˆomega)), ult ˜ 1 + (1|ID1|AY), omega ˜ 1 + (1|ID1|AY), theta ˜ 1 + (1|ID1|AY), nl = TRUE)fit_loss2 <- update(fit_loss1, formula = nlform2, control = list(adapt_delta = 0.90))We could have also specified all predictor terms more conveniently within one formula as

ult + omega + theta ˜ 1 + (1|ID1|AY)because the structure of the predictor terms is identical. To compare model fit, we perform leave-one-out cross-validation.

LOO(fit_loss1, fit_loss2)

LOOIC SEfit_loss1 715.44 19.24fit_loss2 720.60 19.85fit_loss1 - fit_loss2 -5.15 5.34

Since smaller values indicate better expected out-of-sample predictions and thus better model fit, the simpler model that only has a varying intercept over parameter ult is preferred. This may not be overly surprising, given that three varying intercepts as well as three group-level correlations are probably overkill for data containing only 55 observations. Nevertheless, it nicely demonstrates how to apply the ID syntax in practice. More examples of non-linear models can be found in vignette("brms_nonlinear").

Example 4: Performance of school children

Suppose that we want to predict the performance of students in the final exams at the end of the year. There are many variables to consider, but one important factor will clearly be school membership. Schools might differ in the ratio of teachers and students, the general quality of teaching, in the cognitive ability of the students they draw, or other factors we are not aware of that induce dependency among students of the same school. Thus, it is advised to apply a multilevel modeling techniques including school membership as a group-level term. Of course, we should account for class membership and other levels of the educational hierarchy as well, but for the purpose of the present example, we will focus on schools only. Usually, accounting for school membership is pretty-straight forward by simply adding a varying intercept to the formula: (1 | school). However, a non-negligible number of students might change schools during the year. This would result in a situation where one student is a member of multiple schools and so we need a multi-membership model. Setting up such a model not only requires information on the different schools students attend during the year, but also the amount of time spend at each school. The latter can be used to weight the influence each school has on its students, since more time attending a school will likely result in greater influence. For now, let us assume that students change schools maximally once a year and spend equal time at each school. We will later see how to relax these assumptions.

Real educational data are usually relatively large and complex so that we simulate our own data for the purpose of this tutorial paper. We simulate 10 schools and 1000 students, with each school having the same expected number of 100 students. We model 10% of students as changing schools.

data_mm <- sim_multi_mem(nschools = 10, nstudents = 1000, change = 0.1)head(data_mm)

s1 s2 w1 w2 y1 8 9 0.5 0.5 16.274222 10 9 0.5 0.5 18.713873 5 3 0.5 0.5 23.653194 3 5 0.5 0.5 22.352045 5 3 0.5 0.5 16.380196 10 6 0.5 0.5 17.63494The code of function sim_multi_mem can be found in the online supplement of the present paper. For reasons of better illustration, students changing schools appear in the first rows of the data. Data of students being only at a single school looks as follows:

data_mm[101:106, ]

s1 s2 w1 w2 y101 2 2 0.5 0.5 27.247851102 9 9 0.5 0.5 24.041427103 4 4 0.5 0.5 12.575001104 2 2 0.5 0.5 21.203644105 4 4 0.5 0.5 12.856166106 4 4 0.5 0.5 9.740174Thus, school variables are identical, but we still have to specify both in order to pass the data appropriately. Incorporating no other predictors into the model for simplicity, a multi-membership model is specified as

fit_mm <- brm(y ˜ 1 + (1 | mm(s1, s2)), data = data_mm)The only new syntax element is that multiple grouping factors (s1 and s2) are wrapped in mm. Everything else remains exactly the same. Note that we did not specify the relative weights of schools for each student and thus, by default, equal weights are assumed.

summary(fit_mm)

Family: gaussian (identity)Formula: y ˜ 1 + (1 | mm(s1, s2)) Data: data_mm (Number of observations: 1000)Samples: 4 chains, each with iter = 2000; warmup = 1000; thin = 1; total post-warmup samples = 4000 WAIC: Not computedGroup-Level Effects:˜mms1s2 (Number of levels: 10) Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhatsd(Intercept) 2.76 0.82 1.69 4.74 682 1.01Population-Level Effects: Estimate Est.Error l-95% CI u-95% CI Eff.Sample RhatIntercept 19 0.93 17.06 20.8 610 1Family Specific Parameters: Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhatsigma 3.58 0.08 3.43 3.75 2117 1Samples were drawn using sampling(NUTS). For each parameter, Eff.Sampleis a crude measure of effective sample size, and Rhat is the potentialscale reduction factor on split chains (at convergence, Rhat = 1).

With regard to the assumptions made in the above example, it is unlikely that all children who change schools stay in both schools equally long. To relax this assumption, we have to specify weights. First, we amend the simulated data to contain non-equal weights for students changing schools. For all other students, weighting does of course not matter as they stay in the same school anyway.

data_mm[1:100, "w1"] <- runif(100, 0, 1)data_mm[1:100, "w2"] <- 1 - data_mm[1:100, "w1"]head(data_mm)

s1 s2 w1 w2 y1 8 9 0.3403258 0.65967423 16.274222 10 9 0.1771435 0.82285652 18.713873 5 3 0.9059811 0.09401892 23.653194 3 5 0.4432007 0.55679930 22.352045 5 3 0.8052026 0.19479738 16.380196 10 6 0.5610243 0.43897567 17.63494Incorporating these weights into the model is straight forward.

fit_mm2 <- brm(y ˜ 1 + (1 | mm(s1, s2, weights = cbind(w1, w2))), data = data_mm)The summary output is similar to the previous, so we do not show it here. The second assumption that students change schools only once a year, may also easily be relaxed by providing more than two grouping factors, say, mm(s1, s2, s3).

Conclusion

The present paper is meant to introduce R users and developers to the extended lme4 formula syntax applied in brms. Only a subset of modeling options were discussed in detail, which ensured the paper was not too broad. For some of the more basic models that brms can fit, see Bürkner (2017). Many more examples can be found in the growing number of vignettes accompanying the package (see vignette(package = "brms") for an overview).

To date, brms is already one of the most flexible R packages when it comes to regression modeling. However, for the future, there are quite a few more features that I am planning to implement (see https://github.com/paul-buerkner/brms/issues for the current list of issues). In addition to smaller, incremental updates, I have two specific features in mind: extended multivariate models and missing value imputation. I receive ideas and suggestions from users almost every day – for which I am always grateful – and so the list of features that will be implemented in the proceeding versions of brms will continue to grow.

Acknowledgments

First of all, I would like to thank the Stan Development Team for creating the probabilistic programming language Stan, which is an incredibly powerful and flexible tool for performing full Bayesian inference. Without it, brms could not fit a single model. Furthermore, I want to thank Heinz Holling, Donald Williams and Ruben Arslan for valuable comments on earlier versions of the paper. I also would like to thank the many users who reported bugs or had ideas for new features, thus helping to continuously improve brms.

References

Bates et al. (2015) D. Bates, M. Mächler, B. Bolker, and S. Walker. Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1):1–48, 2015.
Betancourt (2017) M. Betancourt. A conceptual introduction to hamiltonian monte carlo. arXiv preprint, 2017. URL https://arxiv.org/pdf/1701.02434.pdf.
Betancourt et al. (2014) M. Betancourt, S. Byrne, S. Livingstone, and M. Girolami. The geometric foundations of hamiltonian monte carlo. arXiv preprint arXiv:1410.5110, 2014.
Brown and Prescott (2015) H. Brown and R. Prescott. Applied Mixed Models in Medicine. John Wiley & Sons, 2015.
Bürkner (2017) P.-C. Bürkner. brms: An R package for bayesian multilevel models using stan. Journal of Statistical Software, 2017.
Carpenter et al. (2017) B. Carpenter, A. Gelman, M. Hoffman, D. Lee, B. Goodrich, M. Betancourt, M. A. Brubaker, J. Guo, P. Li, and A. Ridell. Stan: A probabilistic programming language. Journal of Statistical Software, 2017.
Demidenko (2013) E. Demidenko. Mixed Models: Theory and Applications with R. John Wiley & Sons, 2013.
Duane et al. (1987) S. Duane, A. D. Kennedy, B. J. Pendleton, and D. Roweth. Hybrid monte carlo. Physics Letters B, 195(2):216–222, 1987.
Fahrmeir et al. (2013) L. Fahrmeir, T. Kneib, S. Lang, and B. Marx. Regression: models, methods and applications. Springer Science & Business Media, 2013.
Gelman and Hill (2006) A. Gelman and J. Hill. Data Analysis Using Regression and Multilevel/Hierarchical Models. Cambridge University Press, 2006.
Gelman et al. (2014) A. Gelman, J. B. Carlin, H. S. Stern, and D. B. Rubin. Bayesian Data Analysis, volume 2. Taylor & Francis, 2014.
Hadfield (2010) J. D. Hadfield. Mcmc methods for multi-response generalized linear mixed models: the MCMCglmm R package. Journal of Statistical Software, 33(2):1–22, 2010.
Hastie and Tibshirani (1990) T. J. Hastie and R. J. Tibshirani. Generalized Additive Models, volume 43. CRC Press, 1990.
Hoffman and Gelman (2014) M. D. Hoffman and A. Gelman. The no-u-turn sampler: Adaptively setting path lengths in hamiltonian monte carlo. The Journal of Machine Learning Research, 15(1):1593–1623, 2014.
Lindstrom and Bates (1990) M. J. Lindstrom and D. M. Bates. Nonlinear mixed effects models for repeated measures data. Biometrics, pages 673–687, 1990.
Neal (2011) R. M. Neal. Handbook of Markov Chain Monte Carlo, volume 2, chapter MCMC Using Hamiltonian Dynamics. CRC Press, 2011.
Pinheiro and Bates (2006) J. Pinheiro and D. Bates. Mixed-Effects Models in S and S-PLUS. Springer-Verlage Science & Business Media, 2006.
Pinheiro et al. (2016) J. Pinheiro, D. Bates, S. DebRoy, D. Sarkar, and R Core Team. nlme: Linear and Nonlinear Mixed Effects Models, 2016. URL http://CRAN.R-project.org/package=nlme. R package version 3.1-124.
Rasmussen and Williams (2006) C. E. Rasmussen and C. K. I. Williams. Gaussian processes for machine learning. 2006.
Rigby and Stasinopoulos (2005) R. A. Rigby and D. M. Stasinopoulos. Generalized additive models for location, scale and shape. Journal of the Royal Statistical Society C (Applied Statistics), 54(3):507–554, 2005.
Singmann et al. (2015) H. Singmann, B. Bolker, and J. Westfall. afex: Analysis of Factorial Experiments, 2015. URL https://CRAN.R-project.org/package=afex. R package version 0.15-2.
Stan Development Team (2017a) Stan Development Team. Stan: A C++ Library for Probability and Sampling, Version 2.17.0, 2017a. URL http://mc-stan.org/.
Stan Development Team (2017b) Stan Development Team. Stan Modeling Language: User’s Guide and Reference Manual, 2017b. URL http://mc-stan.org/manual.html.
Stasinopoulos and Rigby (2016) M. Stasinopoulos and B. Rigby. gamlss.data: GAMLSS Data, 2016. URL https://CRAN.R-project.org/package=gamlss.data. R package version 5.0-0.
Vehtari et al. (2016a) A. Vehtari, A. Gelman, and J. Gabry. loo: Efficient Leave-One-Out Cross-Validation and WAIC for Bayesian Models., 2016a. URL https://github.com/stan-dev/loo. R package version 1.0.0.
Vehtari et al. (2016b) A. Vehtari, A. Gelman, and J. Gabry. Practical bayesian model evaluation using leave-one-out cross-validation and waic. Statistics and Computing, pages 1–20, 2016b.
Westfall and Yarkoni (2016) J. Westfall and T. Yarkoni. Statistically controlling for confounding constructs is harder than you think. PloS one, 11(3):e0152719, 2016.
Williams and Rasmussen (1996) C. K. Williams and C. E. Rasmussen. Gaussian processes for regression. In Advances in neural information processing systems, pages 514–520, 1996.
Wood (2004) S. N. Wood. Stable and efficient multiple smoothing parameter estimation for generalized additive models. Journal of the American Statistical Association, 99(467):673–686, 2004.
Wood (2011) S. N. Wood. Fast stable restricted maximum likelihood and marginal likelihood estimation of semiparametric generalized linear models. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 73(1):3–36, 2011.
Wood et al. (2013) S. N. Wood, F. Scheipl, and J. J. Faraway. Straightforward intermediate rank tensor product smoothing in mixed models. Statistics and Computing, pages 1–20, 2013.

Paul-Christian Bürkner
Faculty of Psychology, University of Münster
Fliednerstr. 21, 48151 Münster
Germany
paul.buerkner@gmail.com