3 Methods

Most of the *.stanreg methods are in R/stanreg-methods.R, but as long as things are done appropriately in the .fit file and in stanreg.R all the methods here should work fine.

3.1 predict

The main thing here is to make sure predict works appropriately when the user declares new data. As a rough check, the predictions should match the predictions made by the function you’re emulating.

Also, if no new data is declared then predict(fit) and fit$fitted.values should be identical.

3.2 posterior_predict

This is a little more involved than the predict method. Essentially you need to return and \(N \times S\) dimensional matrix where \(N\) is the number of observations and \(S\) is the number of draws from the posterior distribution. There are two parts to this:

  1. Specify pp_fun
  • pp_fun will call on the posterior prediction function of the form .pp_*. So you need to specify the (stochastic) data generating process within .pp_*. We use sapply() to iterate over the number of draws and compute the fitted values.
  1. Specify pp_args
  • Include anything you might need for posterior predictions within the args list in the pp_args function. (Make sure you do any necessary link function transformations here.)

3.3 posterior_linpred

3.4 loo and log_lik

You need to check whether,

  1. loo() is using the correct log likelihood specified in log_lik.R. This is the log likelihood function that corresponds to object$family (or some other identifier that you can subset from object). If it does then you’re done.
  2. If not then you need to specify the appropriate log likelihood to be used in loo().

Getting the loo function to work on a stanreg object can be tricky. It involves creating a log likelihood function for the posterior llfun and a set of arguments to be passed through this function llargs.

3.4.1 llfun

The best way to think about this is that you want to create a \(S \times N\) matrix point-wise log likelihood, where \(S\) is the number of draws and \(N\) is the number of observations (i.e. you’re evaluating the log-likelihood of the posterior for each datum and draw from the marginal posterior).

The approach taken with using loo on a stanreg object is to declare a function that iterates over the data, rather than specifying the entire point-wise log likelihood matrix.

3.4.2 llargs

Within the llargs list data needs to be a data frame or matrix that can be iterated over \(N\) times. draws should be a list containing the draws of \(S\) dimension. One way to think about it is that data is what you need to iterate over and draws is fixed. This is useful in cases where some variables may be considered as data but you don’t actually want to iterate over them, or in cases where you only have one observation and actually need to iterate over the draws (e.g. a multinormal outcome with correlated errors.)

3.5 prior_summary

The prior_summary function is used to report the prior distributions specified on the parameters when the sampler iterates over the target distribution (which is not necessarily identical to what the user declares).

  1. Define a summarize_*_prior function at the end of the model’s .fit file to capture all the prior information. See stan_glm.fit for a comprehensive example or stan_sp.fit for a simple example.
    • If the user can call prior_aux then you need to give this parameter a name in $prior_aux$aux_name = "prior_aux_name_here". (e.g. in spatial models we have $prior_aux$aux_name = "rho" and in stan_betareg we have $prior_aux$aux_name = "phi")
  2. Call prior_info <- summarize_*_prior(...) before you do any model fitting.
  3. At end of the "optimizing" and "sampling" conditionals make sure you return(structure(stanfit, prior.info = prior_info)).

If you do this right then everything should work out swimmingly in the prior_summary.R file. If it so happens that you’ve introduced a new prior then you’ll need to update the conditional in the relevant .prior_*_prior function to pick this information up.