Resolve the observation a posterior conditions on

resolve_x() and resolve_x_iid() are both thin wrappers around this: an NPE fit maps one observation to one posterior, so resolve_x() calls this with first_row = TRUE and keeps only row 1; an NLE fit's log-likelihood sums over rows as independent observations of the same parameter, so resolve_x_iid() calls this with first_row = FALSE and keeps every row. The same x_obs therefore means "the first observation" to npe() and "200 observations" to nle(), and a user moving a working call from one to the other would otherwise get a posterior conditioned on a single data point with nothing said about it. Truncating to row 1 warns rather than fails: taking row 1 of a simulation matrix is a reasonable thing to ask for.

Usage

resolve_obs(post, x, first_row, arg = if (first_row) "x" else "obs")

Arguments

post

An nsbi_posterior object.

x

Observation to condition on, or NULL to use the posterior's x_obs.

first_row

TRUE to keep row 1 only, with a warning when there was more than one row (the resolve_x() behavior); FALSE to keep every row (the resolve_x_iid() behavior).

arg

Name the caller's argument goes by, for check_numeric()'s and check_finite()'s error messages. Defaults to "x" when first_row and "obs" otherwise, matching sample()'s and log_prob()'s parameter names; the "no observation supplied" message is fixed to the same names and does not vary with arg.

Value

A matrix with dim_x columns: one row if first_row, every row otherwise.

Details

A non-finite entry is a different matter and stops here regardless of first_row. An NA passes through apply_standardizer() into de_sample() (NPE) or into mcmc_init() (NLE) and comes back as unusable output with nothing to say about the observation – for NPE the complaint lands in stats::quantile() inside summary(), for NLE in mcmc_init() complaining about initialization. There is nothing sensible to condition on either way, so this errors rather than warns.