stan_polr: Bayesian ordinal regression models via Stan

Description

Bayesian inference for ordinal (or binary) regression models under a proportional odds assumption.

Usage

stan_polr(formula, data, weights, ..., subset,
  na.action = getOption("na.action", "na.omit"), contrasts = NULL,
  model = TRUE, method = c("logistic", "probit", "loglog", "cloglog",
  "cauchit"), prior = R2(stop("'location' must be specified")),
  prior_counts = dirichlet(1), shape = NULL, rate = NULL,
  prior_PD = FALSE, algorithm = c("sampling", "meanfield", "fullrank"),
  adapt_delta = NULL, do_residuals = NULL)
stan_polr.fit(x, y, wt = NULL, offset = NULL, method = c("logistic",
  "probit", "loglog", "cloglog", "cauchit"), ...,
  prior = R2(stop("'location' must be specified")),
  prior_counts = dirichlet(1), shape = NULL, rate = NULL,
  prior_PD = FALSE, algorithm = c("sampling", "meanfield", "fullrank"),
  adapt_delta = NULL, do_residuals = algorithm == "sampling")

Arguments

formula, data, subset

Same as polr, 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.

weights, na.action, contrasts, model

Same as polr, but rarely specified.

...

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 possibly to specify iter, chains, cores, refresh, etc.

method

One of 'logistic', 'probit', 'loglog', 'cloglog' or 'cauchit', but can be abbreviated. See polr for more details.

prior

Prior for coefficients. Should be a call to R2 to specify the prior location of the \(R^2\) but can be NULL to indicate a standard uniform prior. See priors.

prior_counts

A call to dirichlet to specify the prior counts of the outcome when the predictors are at their sample means.

shape

Either NULL or a positive scalar that is interpreted as the shape parameter for a GammaDistribution on the exponent applied to the probability of success when there are only two outcome categories. If NULL, which is the default, then the exponent is taken to be fixed at \(1\).

rate

Either NULL or a positive scalar that is interpreted as the rate parameter for a GammaDistribution on the exponent applied to the probability of success when there are only two outcome categories. If NULL, which is the default, then the exponent is taken to be fixed at \(1\).

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.

do_residuals

A logical scalar indicating whether or not to automatically calculate fit residuals after sampling completes. Defaults to TRUE if and only if algorithm="sampling". Setting do_residuals=FALSE is only useful in the somewhat rare case that stan_polr appears to finish sampling but hangs instead of returning the fitted model object.

A design matrix.

A response variable, which must be a (preferably ordered) factor.

A numeric vector (possibly NULL) of observation weights.

offset

A numeric vector (possibly NULL) of offsets.

Value

A stanreg object is returned for stan_polr.

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

Details

The stan_polr function is similar in syntax to polr but rather than performing maximum likelihood estimation of a proportional odds model, Bayesian estimation is performed (if algorithm = "sampling") via MCMC. The stan_polr function calls the workhorse stan_polr.fit function, but it is possible to call the latter directly.

As for stan_lm, it is necessary to specify the prior location of \(R^2\). In this case, the \(R^2\) pertains to the proportion of variance in the latent variable (which is discretized by the cutpoints) attributable to the predictors in the model.

Prior beliefs about the cutpoints are governed by prior beliefs about the outcome when the predictors are at their sample means. Both of these are explained in the help page on priors and in the rstanarm vignettes.

Unlike polr, stan_polr also allows the "ordinal" outcome to contain only two levels, in which case the likelihood is the same by default as for stan_glm with family = binomial but the prior on the coefficients is different. However, stan_polr allows the user to specify the shape and rate hyperparameters, in which case the probability of success is defined as the logistic CDF of the linear predictor, raised to the power of alpha where alpha has a gamma prior with the specified shape and rate. This likelihood is called “scobit” by Nagler (1994) because if alpha is not equal to \(1\), then the relationship between the linear predictor and the probability of success is skewed. If shape or rate is NULL, then alpha is assumed to be fixed to \(1\).

Otherwise, it is usually advisible to set shape and rate to the same number so that the expected value of alpha is \(1\) while leaving open the possibility that alpha may depart from \(1\) a little bit. It is often necessary to have a lot of data in order to estimate alpha with much precision and always necessary to inspect the Pareto shape parameters calculated by loo to see if the results are particularly sensitive to individual observations.

Users should think carefully about how the outcome is coded when using a scobit-type model. When alpha is not \(1\), the asymmetry implies that the probability of success is most sensitive to the predictors when the probability of success is less than \(0.63\). Reversing the coding of the successes and failures allows the predictors to have the greatest impact when the probability of failure is less than \(0.63\). Also, the gamma prior on alpha is positively skewed, but you can reverse the coding of the successes and failures to circumvent this property.

References

Nagler, J., (1994). Scobit: An Alternative Estimator to Logit and Probit. American Journal of Political Science. 230 -- 255.

Examples

Run this code

# NOT RUN {
if (!grepl("^sparc",  R.version$platform)) {
 fit <- stan_polr(tobgp ~ agegp, data = esoph, method = "probit",
          prior = R2(0.2, "mean"), init_r = 0.1, seed = 12345,
          algorithm = "fullrank") # for speed only
 print(fit)
 plot(fit)
}

# }

Run the code above in your browser using DataLab