Skip to contents

Validation defaults for a model

Source: R/validation-defaults.R
validation_defaults.Rd

[Experimental]

Usage

validation_defaults(model = NULL)

Arguments

model

NULL for every entry; a model name ("mixture2p") or a bmmodel for that model's.

Value

With model = NULL, a named list of entries. Otherwise the entry, or NULL with no error for a model without one. An entry is a list: model_args (the arguments its constructor is called with when validation_plan() is given the name), pars, sds and cors (the truth on the link scale), grid (n_subjects by n_trials), reps, sampler (passed to the fitter), sbc (the calibration's design point, number of simulations, prior and the reason that prior is not bmm's default, levels and sampler), prior_check (its design point, draws and sampler), reference, owner, version and sections (model-specific card sections).

Details

The generating values, the standard grid and the budget a validation_plan() fills from when an argument is left NULL. An entry exists for a model once its validation card is made; a model without one, including a model of your own, is planned with the generic grid and needs pars.

The standard grid is 20 and 50 subjects by 30, 100 and 300 trials per subject, ten replications each, fitted with 4 chains of 2000 iterations (1000 warmup) under cmdstanr; SBC runs 100 simulations at 20 subjects by 50 trials with 2 chains of 500. Ten replications give a coverage a Monte Carlo standard error of about 0.07 to 0.1, which the card prints next to every rate.

Two entries ship:

  • mixture2p: kappa = log(8), thetat = qlogis(0.75), SDs 0.3 and 0.5. Its calibration runs under a narrower prior than bmm's. The default SD prior on log kappa of a bmm development build drew values at which bmm::rmixture2p() overflows the stack; the defaults of bmm 1.4.1 from CRAN rarely reach them, and the narrower prior keeps the calibration from moving with bmm's defaults. The entry records the reason and the card prints it.

  • ezdm (three parameters): drift = 1.5, bound = 0, ndt = log(0.3), SDs 0.3, 0.2 and 0.2, a correlation of 0.4 between drift and bound. Its calibration also runs under a narrower prior than bmm's, because bmm's cauchy(0, 1) drift intercept and exponential SD priors draw parameter values (a drift of -31, an SD of 2 on log bound) at which some fits fail the card's convergence criterion.

Each entry names its owner, who changes its values by pull request; the card prints the owner and the plan version.

See also

validation_plan(), which fills a plan from an entry.

Examples

names(validation_defaults())
#> [1] "mixture2p" "ezdm"     
validation_defaults("ezdm")$pars
#>     drift     bound       ndt 
#>  1.500000  0.000000 -1.203973 
validation_defaults("sdm")
#> NULL