drmeta_bootstrap_gamma() now short-circuits when the constrained estimate
already lies at gamma = 0. In that situation the null and alternative fits
coincide, so the likelihood-ratio statistic is zero by construction and the
simulated null distribution has a large point mass at zero. Bootstrapping
ties at zero returns a value that reads like a p-value but only records the
proportion of replicates that also reached the boundary. The function now
returns statistic = 0, p.value = 1, and B_used = 0 without running any
replicates.
The returned object gains a logical boundary component, TRUE when the
short-circuit applied and FALSE otherwise. print() explains in the
boundary case why no bootstrap was run and points to constrained = FALSE
as the appropriate directional follow-up.
drmeta_bootstrap_gamma() gains a tol argument controlling the threshold
below which the observed likelihood-ratio statistic is treated as exactly
zero. Default 1e-8.Description field cites Self and Liang (1987) for the boundary problem
and Viechtbauer and Lopez-Lopez (2022) for the general location-scale parent
model, and states that a scale model reweights studies rather than adjusting
the mean for design-linked bias.The package is repositioned as a constrained specialization of the meta-analytic location-scale model rather than a standalone "variance-function framework for causal credibility". Documentation no longer describes the design index as a measure of causal credibility.
dr_heterogeneity() no longer returns the design-residual variance
decomposition or R2_DR. That decomposition was not identified by the
model and is superseded by dr_scale_attenuation(), which is explicitly a
description of the fitted scale function and not a proportion of
heterogeneity explained.
The following functions from 0.1.0 have been removed: dr_variance(),
dr_weights(), dr_forest(), dr_funnel(), dr_plot(), and
dr_pub_bias(). dr_variance() is superseded by dr_scale_predict();
study weights are available as the weights component of a fit. The
publication-bias and plotting functions were removed pending revision
against the corrected specification.
drmeta() gains mods, constrained, gamma_max, and gamma_fixed
arguments. The pooled effect is now returned as beta, a named vector of
location coefficients, rather than a scalar mu.
Location moderators and the joint design-indexed location-scale model.
Exact estimation at the boundary gamma = 0, via direct optimization over
a finite interval plus an explicit boundary comparison.
Unrestricted scale fits (constrained = FALSE) as a directional diagnostic;
a negative estimated gradient contradicts the substantive constraint rather
than merely failing to support it.
drmeta_bootstrap_gamma(), a parametric-bootstrap test of the boundary null.
dr_scale_predict() and dr_scale_attenuation() for observed-support
summaries of the fitted scale function.
dr_loo() for leave-one-out influence on both the location and scale
components.
dr_plot_vfun() for the fitted variance function, drawn over observed
support by default.
confint() and logLik() methods. logLik() reports the criterion
actually used and refuses to relabel a REML fit as ML.
Warnings when the design index has fewer than three distinct values, when its observed range is narrow, and when the estimated gradient sits at the optimization bound.
The bundled BCG example data used a DR column alongside dr_weight, so
the documented bcg$dr silently partial-matched onto dr_weight and fed
fitted weights into the model as the design index. The column is now named
dr, the stale weight columns are removed, and examples index with [[.
Roxygen blocks in the diagnostics file were separated from their functions
by a blank line, so regenerating the documentation would have dropped
dr_scale_predict(), dr_scale_attenuation(), and dr_heterogeneity()
from the namespace.
All exported functions now document a return value.
The default gamma_max is 8, matching the optimization range reported in
the accompanying manuscript. It was 20.