rollout 0.2.0
- New function
add_binomial_outcome() generalizes
add_binary_outcome() to support a binomial outcome with
more than one trial. size (the number of trials, i.e. the
denominator) can be a fixed value or the name of an existing column, so
a per-row size that varies — e.g. a Poisson-distributed eligible count
generated upstream with dplyr::mutate() — can be piped
straight in: add_binomial_outcome(size = n_eligible, ...).
add_binary_outcome() is now a size = 1 special
case of this function.
add_binary_outcome() and
add_binomial_outcome() both gain an
include_error argument (default FALSE). By
default, a .error column (if present) is no longer included
in the summed linear predictor for these two outcomes, since a
Bernoulli/binomial outcome’s variance is already implied by
p (and size). Set
include_error = TRUE to restore the previous behavior of
summing every .-prefixed column.- Breaking change:
add_binary_outcome()’s default output column is now named
"y_binary" (previously "y_bin").
- The
term argument of the eval_*() helpers
(eval_bias(), eval_greater_than(),
eval_less_than(), eval_between(), and
eval_quantile()) can now reference grouping variables, so
true values, thresholds, bounds, and probabilities may vary across
simulated parameter conditions — e.g.,
eval_bias(estimate, term = c(conditionimpl = beta)) when
results are grouped by beta. Elements of term
that do not resolve to a single value per group (for example, ones
referencing a column that was not included in group_by())
now abort with guidance instead of silently returning
NA.
eval_less_than() now correctly computes the proportion
of values below zero when term = NULL; previously
it computed the proportion above zero. Error messages in
eval_greater_than() and eval_less_than() also
now reference the correct function names.- New function
eval_coverage() computes confidence
interval coverage across simulation replicates, i.e. the proportion of
replicates whose interval [lower, upper] contains the true
value for each term. It is designed to be called inside
evaluate_model_results() alongside the other
eval_*() helpers,
e.g. eval_coverage(term = c(conditionintv = b_intv)) when
results are grouped by b_intv and term.
lower and upper default to the
conf.low and conf.high columns created by
extract_model_results(tidy_fun = \(m) broom.mixed::tidy(m, conf.int = TRUE)).
When term = NULL, coverage is computed for a true value of
zero.
rollout 0.1.0