stan_lm: Bayesian regularized linear models via Stan
In rstanarm: Bayesian Applied Regression Modeling via Stan

stan_aov

R Documentation

Bayesian regularized linear models via Stan

Description

Bayesian inference for linear modeling with regularizing priors on the model parameters that are driven by prior beliefs about R^2, the proportion of variance in the outcome attributable to the predictors. See priors for an explanation of this critical point. stan_glm with family="gaussian" also estimates a linear model with normally-distributed errors and allows for various other priors on the coefficients.

Usage

stan_aov(
  formula,
  data,
  projections = FALSE,
  contrasts = NULL,
  ...,
  prior = R2(stop("'location' must be specified")),
  prior_PD = FALSE,
  algorithm = c("sampling", "meanfield", "fullrank"),
  adapt_delta = NULL
)

stan_lm(
  formula,
  data,
  subset,
  weights,
  na.action,
  model = TRUE,
  x = FALSE,
  y = FALSE,
  singular.ok = TRUE,
  contrasts = NULL,
  offset,
  ...,
  prior = R2(stop("'location' must be specified")),
  prior_intercept = NULL,
  prior_PD = FALSE,
  algorithm = c("sampling", "meanfield", "fullrank"),
  adapt_delta = NULL
)

stan_lm.wfit(
  x,
  y,
  w,
  offset = NULL,
  singular.ok = TRUE,
  ...,
  prior = R2(stop("'location' must be specified")),
  prior_intercept = NULL,
  prior_PD = FALSE,
  algorithm = c("sampling", "meanfield", "fullrank"),
  adapt_delta = NULL
)

stan_lm.fit(
  x,
  y,
  offset = NULL,
  singular.ok = TRUE,
  ...,
  prior = R2(stop("'location' must be specified")),
  prior_intercept = NULL,
  prior_PD = FALSE,
  algorithm = c("sampling", "meanfield", "fullrank"),
  adapt_delta = NULL
)

Arguments

`formula`, `data`, `subset`	Same as `lm`, but we strongly advise against omitting the `data` argument. Unless `data` is specified (and is a data frame) many post-estimation functions (including `update`, `loo`, `kfold`) are not guaranteed to work properly.
`projections`	For `stan_aov`, a logical scalar (defaulting to `FALSE`) indicating whether `proj` should be called on the fit.
`...`	Further arguments passed to the function in the rstan package (`sampling`, `vb`, or `optimizing`), corresponding to the estimation method named by `algorithm`. For example, if `algorithm` is `"sampling"` it is possible to specify `iter`, `chains`, `cores`, and other MCMC controls. Another useful argument that can be passed to rstan via `...` is `refresh`, which specifies how often to print updates when sampling (i.e., show the progress every `refresh` iterations). `refresh=0` turns off the iteration updates.
`prior`	Must be a call to `R2` with its `location` argument specified or `NULL`, which would indicate a standard uniform prior for the `R^2`.
`prior_PD`	A logical scalar (defaulting to `FALSE`) indicating whether to draw from the prior predictive distribution instead of conditioning on the outcome.
`algorithm`	A string (possibly abbreviated) indicating the estimation approach to use. Can be `"sampling"` for MCMC (the default), `"optimizing"` for optimization, `"meanfield"` for variational inference with independent normal distributions, or `"fullrank"` for variational inference with a multivariate normal distribution. See `rstanarm-package` for more details on the estimation algorithms. NOTE: not all fitting functions support all four algorithms.
`adapt_delta`	Only relevant if `algorithm="sampling"`. See the adapt_delta help page for details.
`na.action`, `singular.ok`, `contrasts`	Same as `lm`, but rarely specified.
`model`, `offset`, `weights`	Same as `lm`, but rarely specified.
`x`, `y`	In `stan_lm, stan_aov`, logical scalars indicating whether to return the design matrix and response vector. In `stan_lm.fit or stan_lm.wfit`, a design matrix and response vector.
`prior_intercept`	Either `NULL` (the default) or a call to `normal`. If a `normal` prior is specified without a `scale`, then the standard deviation is taken to be the marginal standard deviation of the outcome divided by the square root of the sample size, which is legitimate because the marginal standard deviation of the outcome is a primitive parameter being estimated. Note: If using a dense representation of the design matrix —i.e., if the `sparse` argument is left at its default value of `FALSE`— then the prior distribution for the intercept is set so it applies to the value when all predictors are centered. If you prefer to specify a prior on the intercept without the predictors being auto-centered, then you have to omit the intercept from the `formula` and include a column of ones as a predictor, in which case some element of `prior` specifies the prior on it, rather than `prior_intercept`. Regardless of how `prior_intercept` is specified, the reported estimates of the intercept always correspond to a parameterization without centered predictors (i.e., same as in `glm`).
`w`	Same as in `lm.wfit` but rarely specified.

Details

The stan_lm function is similar in syntax to the lm function but rather than choosing the parameters to minimize the sum of squared residuals, samples from the posterior distribution are drawn using MCMC (if algorithm is "sampling"). The stan_lm function has a formula-based interface and would usually be called by users but the stan_lm.fit and stan_lm.wfit functions might be called by other functions that parse the data themselves and are analogous to lm.fit and lm.wfit respectively.

In addition to estimating sigma — the standard deviation of the normally-distributed errors — this model estimates a positive parameter called log-fit_ratio. If it is positive, the marginal posterior variance of the outcome will exceed the sample variance of the outcome by a multiplicative factor equal to the square of fit_ratio. Conversely if log-fit_ratio is negative, then the model underfits. Given the regularizing nature of the priors, a slight underfit is good.

Finally, the posterior predictive distribution is generated with the predictors fixed at their sample means. This quantity is useful for checking convergence because it is reasonably normally distributed and a function of all the parameters in the model.

The stan_aov function is similar to aov, but does a Bayesian analysis of variance that is basically equivalent to stan_lm with dummy variables. stan_aov has a somewhat customized print method that prints an ANOVA-like table in addition to the output printed for stan_lm models.

Value

A stanreg object is returned for stan_lm, stan_aov.

A stanfit object (or a slightly modified stanfit object) is returned if stan_lm.fit or stan_lm.wfit is called directly.

References

Lewandowski, D., Kurowicka D., and Joe, H. (2009). Generating random correlation matrices based on vines and extended onion method. Journal of Multivariate Analysis. 100(9), 1989–2001.

Examples

if (.Platform$OS.type != "windows" || .Platform$r_arch != "i386") {

op <- options(contrasts = c("contr.helmert", "contr.poly"))
fit_aov <- stan_aov(yield ~ block + N*P*K, data = npk,
         prior = R2(0.5), seed = 12345)
options(op)
print(fit_aov)

}
if (.Platform$OS.type != "windows" || .Platform$r_arch !="i386") {
(fit <- stan_lm(mpg ~ wt + qsec + am, data = mtcars, prior = R2(0.75), 
                # the next line is only to make the example go fast enough
                chains = 1, iter = 300, seed = 12345, refresh = 0))
plot(fit, "hist", pars = c("wt", "am", "qsec", "sigma"), 
     transformations = list(sigma = "log"))
}

rstanarm documentation built on May 29, 2024, 5:51 a.m.