hypothesis.brmsfit: Non-Linear Hypothesis Testing

Description

Perform non-linear hypothesis testing for all model parameters.

Usage

# S3 method for brmsfit
hypothesis(x, hypothesis, class = "b", group = "",
  scope = c("standard", "ranef", "coef"), alpha = 0.05, seed = NULL,
  ...)
hypothesis(x, ...)
# S3 method for default
hypothesis(x, hypothesis, alpha = 0.05, ...)

Arguments

An R object. If it is no brmsfit object, it must be coercible to a data.frame.

hypothesis

A character vector specifying one or more non-linear hypothesis concerning parameters of the model.

class

A string specifying the class of parameters being tested. Default is "b" for population-level effects. Other typical options are "sd" or "cor". If class = NULL, all parameters can be tested against each other, but have to be specified with their full name (see also parnames)

group

Name of a grouping factor to evaluate only group-level effects parameters related to this grouping factor.

scope

Indicates where to look for the variables specified in hypothesis. If "standard", use the full parameter names (subject to the restriction given by class and group). If "coef" or "ranef", compute the hypothesis for all levels of the grouping factor given in "group", based on the output of coef.brmsfit and ranef.brmsfit, respectively.

alpha

The alpha-level of the tests (default is 0.05; see 'Details' for more information).

seed

A single numeric value passed to set.seed to make results reproducible.

...

Currently ignored.

Value

A brmshypothesis object.

Details

Among others, hypothesis computes an evidence ratio (Evid.Ratio) for each hypothesis. For a one-sided hypothesis, this is just the posterior probability (Post.Prob) under the hypothesis against its alternative. That is, when the hypothesis is of the form a > b, the evidence ratio is the ratio of the posterior probability of a > b and the posterior probability of a < b. In this example, values greater than one indicate that the evidence in favor of a > b is larger than evidence in favor of a < b. For an two-sided (point) hypothesis, the evidence ratio is a Bayes factor between the hypothesis and its alternative computed via the Savage-Dickey density ratio method. That is the posterior density at the point of interest divided by the prior density at that point. Values greater than one indicate that evidence in favor of the point hypothesis has increased after seeing the data. In order to calculate this Bayes factor, all parameters related to the hypothesis must have proper priors and argument sample_prior of function brm must be set to "yes". When interpreting Bayes factors, make sure that your priors are reasonable and carefully chosen, as the result will depend heavily on the priors. In particular, avoid using default priors.

The Evid.Ratio may sometimes be 0 or Inf implying very small or large evidence, respectively, in favor of the tested hypothesis. For one-sided hypotheses pairs, this basically means that all posterior samples are on the same side of the value dividing the two hypotheses. In that sense, instead of 0 or Inf, you may rather read it as Evid.Ratio smaller 1 / S or greater S, respectively, where S denotes the number of posterior samples used in the computations.

The argument alpha specifies the size of the credible interval (i.e., Bayesian confidence interval). For instance, if we tested a two-sided hypothesis and set alpha = 0.05 (5%) an, the credible interval will contain 1 - alpha = 0.95 (95%) of the posterior values. Hence, alpha * 100% of the posterior values will lie outside of the credible interval. Although this allows testing of hypotheses in a similar manner as in the frequentist null-hypothesis testing framework, we strongly argue against using arbitrary cutoffs (e.g., p < .05) to determine the 'existence' of an effect.

Examples

Run this code

# NOT RUN {
## define priors
prior <- c(set_prior("normal(0,2)", class = "b"),
           set_prior("student_t(10,0,1)", class = "sigma"),
           set_prior("student_t(10,0,1)", class = "sd"))

## fit a linear mixed effects models
fit <- brm(time ~ age + sex + disease + (1 + age|patient),
           data = kidney, family = lognormal(),
           prior = prior, sample_prior = "yes", 
           control = list(adapt_delta = 0.95))

## perform two-sided hypothesis testing
(hyp1 <- hypothesis(fit, "sexfemale = age + diseasePKD"))
plot(hyp1)
hypothesis(fit, "exp(age) - 3 = 0", alpha = 0.01)

## perform one-sided hypothesis testing
hypothesis(fit, "diseasePKD + diseaseGN - 3 < 0")

hypothesis(fit, "age < Intercept", 
           class = "sd", group  = "patient")

## test the amount of random intercept variance on all variance
h <- paste("sd_patient__Intercept^2 / (sd_patient__Intercept^2 +",
           "sd_patient__age^2 + sigma^2) = 0")
(hyp2 <- hypothesis(fit, h, class = NULL))
plot(hyp2)

## test more than one hypothesis at once
h <- c("diseaseGN = diseaseAN", "2 * diseaseGN - diseasePKD = 0")
(hyp3 <- hypothesis(fit, h))
plot(hyp3, ignore_prior = TRUE)

## compute hypotheses for all levels of a grouping factor
hypothesis(fit, "age = 0", scope = "coef", group = "patient")

## use the default method
dat <- as.data.frame(fit)
hypothesis(dat, "b_age > 0")
# }
# NOT RUN {
# }