bayesglm {arm}R Documentation

Bayesian generalized linear models.

Description

Bayesian functions for generalized linear modeling with independent normal, t, or Cauchy prior distribution for the coefficients.

Usage

bayesglm (formula, family = gaussian, data, 
    weights, subset, na.action, 
    start = NULL, etastart, mustart, 
    offset, control = glm.control(...), 
    model = TRUE, method = "glm.fit", 
    x = FALSE, y = TRUE, contrasts = NULL, 
    prior.mean = 0, prior.scale = 2.5, 
    prior.scale.for.intercept = 10, prior.df = 1, 
    scaled = TRUE, n.iter = 100, ...)
    
bayesglm.fit (x, y, weights = rep(1, nobs), 
    start = NULL, etastart = NULL, 
    mustart = NULL, offset = rep(0, nobs), family = gaussian(), 
    control = glm.control(), intercept = TRUE,
    prior.mean=0, prior.scale=2.5, prior.scale.for.intercept.default=10,
    prior.df=1, scaled=TRUE)

Arguments

formula a symbolic description of the model to be fit. The details of model specification are given below.
family a description of the error distribution and link function to be used in the model. This can be a character string naming a family function, a family function or the result of a call to a family function. (See family for details of family functions.)
data an optional data frame, list or environment (or object coercible by as.data.frame to a data frame) containing the variables in the model. If not found in data, the variables are taken from environment(formula), typically the environment from which glm is called.
weights an optional vector of weights to be used in the fitting process. Should be NULL or a numeric vector.
subset an optional vector specifying a subset of observations to be used in the fitting process.
na.action a function which indicates what should happen when the data contain NAs. The default is set by the na.action setting of options, and is na.fail if that is unset. The “factory-fresh” default is na.omit. Another possible value is NULL, no action. Value na.exclude can be useful.
start starting values for the parameters in the linear predictor.
etastart starting values for the linear predictor.
mustart starting values for the vector of means.
offset this can be used to specify an a priori known component to be included in the linear predictor during fitting. This should be NULL or a numeric vector of length either one or equal to the number of cases. One or more offset terms can be included in the formula instead or as well, and if both are specified their sum is used. See model.offset.
control a list of parameters for controlling the fitting process. See the documentation for glm.control for details.
model a logical value indicating whether model frame should be included as a component of the returned value.
method the method to be used in fitting the model. The default method "glm.fit" uses iteratively reweighted least squares (IWLS). The only current alternative is "model.frame" which returns the model frame and does no fitting.
x, y For glm: logical values indicating whether the response vector and model matrix used in the fitting process should be returned as components of the returned value.
For glm.fit: x is a design matrix of dimension n * p, and y is a vector of observations of length n.
contrasts an optional list. See the contrasts.arg of model.matrix.default.
intercept logical. Should an intercept be included in the null model?
prior.mean prior mean for the coefficients:default is 0. Can be a vector of length equal to the number of predictors (including the intercept, if any). If it is a scalar, it is expanded to the length of this vector.
prior.scale prior scale for the coefficients: default is 2.5. Can be a vector of length equal to the number of predictors (including the intercept, if any). If it is a scalar, it is expanded to the length of this vector.
prior.scale.for.intercept prior scale for the intercept: default is 10.
prior.df for t distribution: default is 1 (Cauchy). Set to Inf to get normal prior distributions. Can be a vector of length equal to the number of predictors (including the intercept, if any). If it is a scalar, it is expanded to the length of this vector.
scaled if scaled = TRUE, then the prior distribution is rescaled: default is TRUE
n.iter default is 100.
prior.scale.for.intercept.default prior scale for the intercept: default is 10.
... further arguments passed to or from other methods.

Details

The program is a simple alteration of glm() that uses an approximate EM algorithm to update the betas at each step using an augmented regression to represent the prior information.

We use Student-t prior distributions for the coefficients. The prior distribution for the constant term is set so it applies to the value when all predictors are set to their mean values.

If scaled=TRUE, the scales for the prior distributions of the coefficients are determined as follows: For a predictor with only one value, we just use prior.scale. For a predictor with two values, we use prior.scale/range(x). For a predictor with more than two values, we use prior.scale/(2*sd(x)).

We include all the glm() arguments but we haven't tested that all the options (e.g., offests, contrasts, deviance for the null model) all work.

The new arguments here are: prior.mean, prior.scale, prior.scale.for.intercept, prior.df, and scaled.

Value

See glm for details.

Author(s)

Andrew Gelman gelman@stat.columbia.edu; Yu-Sung Su ys463@columbia.edu; Maria Grazia Pittau grazia@stat.columbia.edu; Aleks Jakulin Jakulin@stat.columbia.edu

References

Andrew Gelman, Aleks Jakulin, Maria Grazia Pittau and Yu-Sung Su, A default prior distribution for logistic and other regression models, unpublished paper available at http://www.stat.columbia.edu/~gelman/standardize/

See Also

glm, bayespolr

Examples

  n <- 100
  x1 <- rnorm (n)
  x2 <- rbinom (n, 1, .5)
  b0 <- 1
  b1 <- 1.5
  b2 <- 2
  y <- rbinom (n, 1, invlogit(b0+b1*x1+b2*x2))

  M1 <- glm (y ~ x1 + x2, family=binomial(link="logit"))
  display (M1)  # (using the display() function from regression.R)

  M2 <- bayesglm (y ~ x1 + x2, family=binomial(link="logit"), 
    prior.scale=Inf, prior.df=Inf)
  display (M2)  # just a test:  this should be identical to classical logit

  M3 <- bayesglm (y ~ x1 + x2, family=binomial(link="logit"))  
    # default Cauchy prior with scale 2.5
  display (M3)

  M4 <- bayesglm (y ~ x1 + x2, family=binomial(link="logit"), 
    prior.scale=2.5, prior.df=1)  
    # Same as M3, explicitly specifying Cauchy prior with scale 2.5
  display (M4)

  M5 <- bayesglm (y ~ x1 + x2, family=binomial(link="logit"), 
    prior.scale=2.5, prior.df=7)   # t_7 prior with scale 2.5
  display (M5)

  M6 <- bayesglm (y ~ x1 + x2, family=binomial(link="logit"), 
    prior.scale=2.5, prior.df=Inf)  # normal prior with scale 2.5
  display (M6)

# Create separation:  set y=1 whenever x2=1
# Now it should blow up without the prior!

  y <- ifelse (x2==1, 1, y)

  M1 <- glm (y ~ x1 + x2, family=binomial(link="logit"))
  display (M1)

  M2 <- bayesglm (y ~ x1 + x2, family=binomial(link="logit"), 
    prior.scale=Inf, prior.df=Inf) # Same as M1
  display (M2)

  M3 <- bayesglm (y ~ x1 + x2, family=binomial(link="logit"))
  display (M3)

  M4 <- bayesglm (y ~ x1 + x2, family=binomial(link="logit"), 
    prior.scale=2.5, prior.df=1)  # Same as M3
  display (M4)

  M5 <- bayesglm (y ~ x1 + x2, family=binomial(link="logit"), 
    prior.scale=2.5, prior.df=7)
  display (M5)

  M6 <- bayesglm (y ~ x1 + x2, family=binomial(link="logit"), 
    prior.scale=2.5, prior.df=Inf)
  display (M6)
 
  # bayesglm with gaussian family (bayes lm)
  sigma <- 5
  y2 <- rnorm (n, b0+b1*x1+b2*x2, sigma)
  M7 <- bayesglm (y2 ~ x1 + x2, prior.scale=Inf, prior.df=Inf)
  display (M7)

[Package arm version 1.0-17 Index]