Set up a model formula for use in the brms package allowing to define (potentially non-linear) additive multilevel models for all parameters of the assumed response distribution.
1 2 3 4 5 6 7 8 9 10 11 12 13
An object of class
Optional list of formulas, which are treated in the
same way as formulas passed via the
Same argument as in
Logical; Indicates whether
Logical; Only used in non-linear models.
Indicates if the computation of the non-linear formula should be
done inside (
Logical; Indicates if the population-level design
matrix should be centered, which usually increases sampling efficiency.
See the 'Details' section for more information.
Logical; Indicates whether automatic cell-mean coding
should be enabled when removing the intercept by adding
Logical; indicates whether the population-level design matrices
should be treated as sparse (defaults to
Optional name of the decomposition used for the
population-level design matrix. Defaults to
General formula structure
formula argument accepts formulas of the following syntax:
response | aterms ~ pterms + (gterms | group)
pterms part contains effects that are assumed to be the
same across observations. We call them 'population-level' effects
or (adopting frequentist vocabulary) 'fixed' effects. The optional
gterms part may contain effects that are assumed to vary
across grouping variables specified in
call them 'group-level' effects or (adopting frequentist
vocabulary) 'random' effects, although the latter name is misleading
in a Bayesian context. For more details type
Multiple grouping factors each with multiple group-level effects
are possible. (Of course we can also run models without any
| you may use
|| in grouping terms
to prevent correlations from being modeled.
Alternatively, it is possible to model different group-level terms of
the same grouping factor as correlated (even across different formulas,
e.g., in non-linear models) by using
|<ID>| instead of
All group-level terms sharing the same ID will be modeled as correlated.
If, for instance, one specifies the terms
(1+z|2|g) somewhere in the formulas passed to
correlations between the corresponding group-level effects
will be estimated.
If levels of the grouping factor belong to different sub-populations,
it may be reasonable to assume a different covariance matrix for each
of the sub-populations. For instance, the variation within the
treatment group and within the control group in a randomized control
trial might differ. Suppose that
y is the outcome, and
x is the factor indicating the treatment and control group.
Then, we could estimate different hyper-parameters of the varying
effects (in this case a varying intercept) for treatment and control
y ~ x + (1 | gr(subject, by = x)).
You can specify multi-membership terms using the
function. For instance, a multi-membership term with two members
(1 | mm(g1, g2)), where
specify the first and second member, respectively. Moreover,
if a covariate
x varies across the levels of the grouping-factors
g2, we can save the respective covariate values
in the variables
x2 and then model the varying
(1 + mmc(x1, x2) | mm(g1, g2)).
Special predictor terms
Smoothing terms can modeled using the
t2 functions in the
of the model formula. This allows to fit generalized additive mixed
models (GAMMs) with brms. The implementation is similar to that
used in the gamm4 package. For more details on this model class
Gaussian process terms can be fitted using the
function in the
pterms part of the model formula. Similar to
smooth terms, Gaussian processes can be used to model complex non-linear
relationships, for instance temporal or spatial autocorrelation.
However, they are computationally demanding and are thus not recommended
for very large datasets.
gterms parts may contain four non-standard
effect types namely monotonic, measurement error, missing value, and
category specific effects, which can be specified using terms of the
Category specific effects can only be estimated in
ordinal models and are explained in more detail in the package's
main vignette (type
The other three effect types are explained in the following.
A monotonic predictor must either be integer valued or an ordered factor,
which is the first difference to an ordinary continuous predictor.
More importantly, predictor categories (or integers) are not assumed to be
equidistant with respect to their effect on the response variable.
Instead, the distance between adjacent predictor categories (or integers)
is estimated from the data and may vary across categories.
This is realized by parameterizing as follows:
One parameter takes care of the direction and size of the effect similar
to an ordinary regression parameter, while an additional parameter vector
estimates the normalized distances between consecutive predictor categories.
A main application of monotonic effects are ordinal predictors that
can this way be modeled without (falsely) treating them as continuous
or as unordered categorical predictors. For more details and examples
Quite often, predictors are measured and as such naturally contain
measurement error. Although most researchers are well aware of this problem,
measurement error in predictors is ignored in most
regression analyses, possibly because only few packages allow
for modeling it. Notably, measurement error can be handled in
structural equation models, but many more general regression models
(such as those featured by brms) cannot be transferred
to the SEM framework. In brms, effects of noise-free predictors
can be modeled using the
me (for 'measurement error') function.
y is the response variable and
x is a measured predictor with known measurement error
sdx, we can simply include it on the right-hand side of the
model formula via
y ~ me(x, sdx).
This can easily be extended to more general formulas.
x2 is another measured predictor with corresponding error
z is a predictor without error
(e.g., an experimental setting), we can model all main effects
and interactions of the three predictors in the well known manner:
y ~ me(x, sdx) * me(x2, sdx2) * z. In future version of brms,
a vignette will be added to explain more details about these
so called 'error-in-variables' models and provide real world examples.
When a variable contains missing values, the corresponding rows will
be excluded from the data by default (row-wise exclusion). However,
quite often we want to keep these rows and instead estimate the missing values.
There are two approaches for this: (a) Impute missing values before
the model fitting for instance via multiple imputation (see
brm_multiple for a way to handle multiple imputed datasets).
(b) Impute missing values on the fly during model fitting. The latter
approach is explained in the following. Using a variable with missing
values as predictors requires two things, First, we need to specify that
the predictor contains missings that should to be imputed.
y is the primary response,
x is a
predictor with missings and
z is a predictor without missings,
we go for
y ~ mi(x) + z. Second, we need to model
as an additional response with corresponding predictors and the
mi(). In our example, we could write
x | mi() ~ z. See
mi for examples with real data.
Autocorrelation terms can be directly specified inside the formula
as well. Details can be found in
Additional response information
Another special of the brms formula syntax is the optional
aterms part, which may contain multiple terms of the form
fun(<variable>) separated by
+ each providing special
information on the response variable.
fun can be replaced with
vint. Their meanings are explained below.
skew_normal, it is
possible to specify standard errors of the observations, thus allowing
to perform meta-analysis. Suppose that the variable
the effect sizes from the studies and
sei the corresponding
standard errors. Then, fixed and random effects meta-analyses can
be conducted using the formulas
yi | se(sei) ~ 1 and
yi | se(sei) ~ 1 + (1|study), respectively, where
study is a variable uniquely identifying every study.
If desired, meta-regression can be performed via
yi | se(sei) ~ 1 + mod1 + mod2 + (1|study)
yi | se(sei) ~ 1 + mod1 + mod2 + (1 + mod1 + mod2|study),
mod2 represent moderator variables.
By default, the standard errors replace the parameter
sigma in addition to the known standard errors,
sigma in function
yi | se(sei, sigma = TRUE) ~ 1.
For all families, weighted regression may be performed using
weights in the
aterms part. Internally, this is
implemented by multiplying the log-posterior values of each
observation by their corresponding weights.
Suppose that variable
wei contains the weights
yi is the response variable.
yi | weights(wei) ~ predictors
implements a weighted regression.
For multivariate models,
subset may be used in the
part, to use different subsets of the data in different univariate
models. For instance, if
sub is a logical variable and
y is the response of one of the univariate models, we may
y | subset(sub) ~ predictors so that
predicted only for those observations for which
For log-linear models such as poisson models,
rate may be used
aterms part to specify the denomintor of a response that
is expressed as a rate. The numerator is given by the actual response
variable and has a distribution according to the family as usual. Using
rate(denom) is equivalent to adding
the linear predictor of the main parameter but the former is arguably
more convenient and explicit.
With the exception of categorical, ordinal, and mixture families,
left, right, and interval censoring can be modeled through
y | cens(censored) ~ predictors. The censoring variable
censored in this example) should contain the values
2) to indicate that
the corresponding observation is left censored, not censored, right censored,
or interval censored. For interval censored data, a second variable
(let's call it
y2) has to be passed to
cens. In this case,
the formula has the structure
y | cens(censored, y2) ~ predictors.
While the lower bounds are given in
y, the upper bounds are given
y2 for interval censored data. Intervals are assumed to be open
on the left and closed on the right:
With the exception of categorical, ordinal, and mixture families,
the response distribution can be truncated using the
function in the addition part. If the response variable is truncated
between, say, 0 and 100, we can specify this via
yi | trunc(lb = 0, ub = 100) ~ predictors.
Instead of numbers, variables in the data set can also be passed allowing
for varying truncation points across observations. Defining only one of
the two arguments in
trunc leads to one-sided truncation.
For all continuous families, missing values in the responses can be imputed
within Stan by using the addition term
mi. This is mostly
useful in combination with
mi predictor terms as explained
above under 'Special predictor terms'.
addition should contain a variable indicating the number of trials
underlying each observation. In
lme4 syntax, we may write for instance
cbind(success, n - success), which is equivalent
success | trials(n) in brms syntax. If the number of trials
is constant across all observations, say
we may also write
success | trials(10).
Please note that the
cbind() syntax will not work
in brms in the expected way because this syntax is reserved
for other purposes.
For all ordinal families,
aterms may contain a term
thres(number) to specify the number thresholds (e.g,
thres(6)), which should be equal to the total number of response
categories - 1. If not given, the number of thresholds is calculated from
the data. If different threshold vectors should be used for different
subsets of the data, the
gr argument can be used to provide the
grouping variable (e.g,
thres(6, gr = item), if
item is the
grouping variable). In this case, the number of thresholds can also be a
variable in the data with different values per group.
A deprecated quasi alias of
cat() with which the
total number of response categories (i.e., number of thresholds + 1) can be
In Wiener diffusion models (family
wiener) the addition term
dec is mandatory to specify the (vector of) binary decisions
corresponding to the reaction times. Non-zero values will be treated
as a response on the upper boundary of the diffusion process and zeros
will be treated as a response on the lower boundary. Alternatively,
the variable passed to
dec might also be a character vector
For custom families, it is possible to pass an abitrary number of real and
integer vectors via the addition terms
respectively. An example is provided in
Multiple addition terms may be specified at the same time using the
+ operator. For example, the formula
formula = yi | se(sei) + cens(censored) ~ 1 implies a censored
The addition argument
disp (short for dispersion)
has been removed in version 2.0. You may instead use the
distributional regression approach by specifying
sigma ~ 1 + offset(log(xdisp)) or
shape ~ 1 + offset(log(xdisp)), where
the variable being previously passed to
Parameterization of the population-level intercept
By default, the population-level intercept (if incorporated) is estimated
separately and not as part of population-level parameter vector
a result, priors on the intercept also have to be specified separately.
Furthermore, to increase sampling efficiency, the population-level design
X is centered around its column means
X_means if the
intercept is incorporated. This leads to a temporary bias in the intercept
<X_means, b>, where
<,> is the scalar product. The
bias is corrected after fitting the model, but be aware that you are
effectively defining a prior on the intercept of the centered design matrix
not on the real intercept. You can turn off this special handling of the
intercept by setting argument
FALSE. For more
details on setting priors on population-level intercepts, see
This behavior can be avoided by using the reserved
(and internally generated) variable
y ~ x, you may write
y ~ 0 + Intercept + x. This way, priors can be
defined on the real intercept, directly. In addition,
the intercept is just treated as an ordinary population-level effect
and thus priors defined on
b will also apply to it.
Note that this parameterization may be less efficient
than the default parameterization discussed above.
Formula syntax for non-linear models
In brms, it is possible to specify non-linear models
of arbitrary complexity.
The non-linear model can just be specified within the
argument. Suppose, that we want to predict the response
through the predictor
x is linked to
y = alpha - beta * lambda^x, with parameters
lambda. This is certainly a
non-linear model being defined via
formula = y ~ alpha - beta * lambda^x (addition arguments
can be added in the same way as for ordinary formulas).
To tell brms that this is a non-linear model,
we set argument
Now we have to specify a model for each of the non-linear parameters.
Let's say we just want to estimate those three parameters
with no further covariates or random effects. Then we can pass
alpha + beta + lambda ~ 1 or equivalently
(and more flexible)
alpha ~ 1, beta ~ 1, lambda ~ 1
This can, of course, be extended. If we have another predictor
observations nested within the grouping factor
g, we may write for
alpha ~ 1, beta ~ 1 + z + (1|g), lambda ~ 1.
The formula syntax described above applies here as well.
In this example, we are using
g only for the
beta, but we might also use them for the other
non-linear parameters (provided that the resulting model is still
By default, non-linear covariates are treated as real vectors in Stan. However, if the data of the covariates is of type 'integer' in R (which can be enforced by the 'as.integer' function), the Stan type will be changed to an integer array. That way, covariates can also be used for indexing purposes in Stan.
Non-linear models may not be uniquely identified and / or show bad convergence.
For this reason it is mandatory to specify priors on the non-linear parameters.
For instructions on how to do that, see
For some examples of non-linear models, see
Formula syntax for predicting distributional parameters
It is also possible to predict parameters of the response distribution such
as the residual standard deviation
sigma in gaussian models or the
hu in hurdle models. The syntax closely resembles
that of a non-linear parameter, for instance
sigma ~ x + s(z) +
(1+x|g). For some examples of distributional models, see
mu exists for every family and can be used as an
alternative to specifying terms in
formula. If both
formula are given, the right-hand side of
formula is ignored.
Accordingly, specifying terms on the right-hand side of both
mu at the same time is deprecated. In future versions,
formula might be updated by
The following are
distributional parameters of specific families (all other parameters are
treated as non-linear parameters):
sigma (residual standard
deviation or scale of the
shape (shape parameter of the
negbinomial, and related zero-inflated
/ hurdle families);
nu (degrees of freedom parameter of the
parameter of the
kappa (precision parameter of the
beta (mean parameter of the exponential component of the
quantile (quantile parameter of the
zi (zero-inflation probability);
hu (hurdle probability);
coi (conditional one-inflation probability);
disc (discrimination) for ordinal models;
bias (boundary separation, non-decision time, and initial bias of
wiener diffusion model). By default, distributional parameters
are modeled on the log scale if they can be positive only or on the logit
scale if the can only be within the unit interval.
Alternatively, one may fix distributional parameters to certain values.
However, this is mainly useful when models become too
complicated and otherwise have convergence issues.
We thus suggest to be generally careful when making use of this option.
quantile parameter of the
is a good example where it is useful. By fixing
one can perform quantile regression for the specified quantile.
quantile = 0.25 allows predicting the 25%-quantile.
bias parameter in drift-diffusion models,
is assumed to be
0.5 (i.e. no bias) in many applications.
To achieve this, simply write
bias = 0.5.
Other possible applications are the Cauchy distribution as a
special case of the Student-t distribution with
nu = 1, or the geometric distribution as a special case of
the negative binomial distribution with
shape = 1.
Furthermore, the parameter
disc ('discrimination') in ordinal
models is fixed to
1 by default and not estimated,
but may be modeled as any other distributional parameter if desired
(see examples). For reasons of identification,
can only be positive, which is achieved by applying the log-link.
In categorical models, distributional parameters do not have
fixed names. Instead, they are named after the response categories
(excluding the first one, which serves as the reference category),
with the prefix
'mu'. If, for instance, categories are named
cat3, the distributional parameters
will be named
Some distributional parameters currently supported by
have to be positive (a negative standard deviation or precision parameter
does not make any sense) or are bounded between 0 and 1 (for zero-inflated /
hurdle probabilities, quantiles, or the initial bias parameter of
However, linear predictors can be positive or negative, and thus the log link
(for positive parameters) or logit link (for probability parameters) are used
by default to ensure that distributional parameters are within their valid intervals.
This implies that, by default, effects for such distributional parameters are
estimated on the log / logit scale and one has to apply the inverse link
function to get to the effects on the original scale.
Alternatively, it is possible to use the identity link to predict parameters
on their original scale, directly. However, this is much more likely to lead
to problems in the model fitting, if the parameter actually has a restricted range.
brmsfamily for an overview of valid link functions.
Formula syntax for mixture models
The specification of mixture models closely resembles that
of non-mixture models. If not specified otherwise (see below),
all mean parameters of the mixture components are predicted
using the right-hand side of
formula. All types of predictor
terms allowed in non-mixture models are allowed in mixture models
Distributional parameters of mixture distributions have the same
name as those of the corresponding ordinary distributions, but with
a number at the end to indicate the mixture component. For instance, if
you use family
mixture(gaussian, gaussian), the distributional
Distributional parameters of the same class can be fixed to the same value.
For the above example, we could write
sigma2 = "sigma1" to make
sure that both components have the same residual standard deviation,
which is in turn estimated from the data.
In addition, there are two types of special distributional parameters.
The first are named
mu<ID>, that allow for modeling different
predictors for the mean parameters of different mixture components.
For instance, if you want to predict the mean of the first component
x and the mean of the second component using
z, you can write
mu1 ~ x as well as
mu2 ~ z.
The second are named
theta<ID>, which constitute the mixing
proportions. If the mixing proportions are fixed to certain values,
they are internally normalized to form a probability vector.
If one seeks to predict the mixing proportions, all but
one of the them has to be predicted, while the remaining one is used
as the reference category to identify the model. The
function is applied on the linear predictor terms to form a
For more information on mixture models, see
the documentation of
Formula syntax for multivariate models
Multivariate models may be specified using
or with help of the
y2 are response variables
x is a predictor. Then
mvbind(y1, y2) ~ x
specifies a multivariate model.
The effects of all terms specified at the RHS of the formula
are assumed to vary across response variables.
For instance, two parameters will be estimated for
one for the effect on
y1 and another for the effect on
This is also true for group-level effects. When writing, for instance,
mvbind(y1, y2) ~ x + (1+x|g), group-level effects will be
estimated separately for each response. To model these effects
as correlated across responses, use the ID syntax (see above).
For the present example, this would look as follows:
mvbind(y1, y2) ~ x + (1+x|2|g). Of course, you could also use
any value other than
2 as ID.
It is also possible to specify different formulas for different responses.
If, for instance,
y1 should be predicted by
should be predicted by
z, we could write
mvbf(y1 ~ x, y2 ~ z).
brmsformula objects can be added to
specify a joint multivariate model (see 'Examples').
An object of class
is essentially a
list containing all model
formulas as well as some additional information.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81
# multilevel model with smoothing terms brmsformula(y ~ x1*x2 + s(z) + (1+x1|1) + (1|g2)) # additionally predict 'sigma' brmsformula(y ~ x1*x2 + s(z) + (1+x1|1) + (1|g2), sigma ~ x1 + (1|g2)) # use the shorter alias 'bf' (formula1 <- brmsformula(y ~ x + (x|g))) (formula2 <- bf(y ~ x + (x|g))) # will be TRUE identical(formula1, formula2) # incorporate censoring bf(y | cens(censor_variable) ~ predictors) # define a simple non-linear model bf(y ~ a1 - a2^x, a1 + a2 ~ 1, nl = TRUE) # predict a1 and a2 differently bf(y ~ a1 - a2^x, a1 ~ 1, a2 ~ x + (x|g), nl = TRUE) # correlated group-level effects across parameters bf(y ~ a1 - a2^x, a1 ~ 1 + (1|2|g), a2 ~ x + (x|2|g), nl = TRUE) # define a multivariate model bf(mvbind(y1, y2) ~ x * z + (1|g)) # define a zero-inflated model # also predicting the zero-inflation part bf(y ~ x * z + (1+x|ID1|g), zi ~ x + (1|ID1|g)) # specify a predictor as monotonic bf(y ~ mo(x) + more_predictors) # for ordinal models only # specify a predictor as category specific bf(y ~ cs(x) + more_predictors) # add a category specific group-level intercept bf(y ~ cs(x) + (cs(1)|g)) # specify parameter 'disc' bf(y ~ person + item, disc ~ item) # specify variables containing measurement error bf(y ~ me(x, sdx)) # specify predictors on all parameters of the wiener diffusion model # the main formula models the drift rate 'delta' bf(rt | dec(decision) ~ x, bs ~ x, ndt ~ x, bias ~ x) # fix the bias parameter to 0.5 bf(rt | dec(decision) ~ x, bias = 0.5) # specify different predictors for different mixture components mix <- mixture(gaussian, gaussian) bf(y ~ 1, mu1 ~ x, mu2 ~ z, family = mix) # fix both residual standard deviations to the same value bf(y ~ x, sigma2 = "sigma1", family = mix) # use the '+' operator to specify models bf(y ~ 1) + nlf(sigma ~ a * exp(b * x), a ~ x) + lf(b ~ z + (1|g), dpar = "sigma") + gaussian() # specify a multivariate model using the '+' operator bf(y1 ~ x + (1|g)) + gaussian() + cor_ar(~1|g) + bf(y2 ~ z) + poisson() # model missing values in predictors bf(bmi ~ age * mi(chl)) + bf(chl | mi() ~ age) + set_rescor(FALSE) # model sigma as a function of the mean bf(y ~ eta, nl = TRUE) + lf(eta ~ 1 + x) + nlf(sigma ~ tau * sqrt(eta)) + lf(tau ~ 1)
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.