Computes DFFITS, Cook's distance, COVRATIO, and other influence diagnostics for a fitted brma object.
Usage
# S3 method for class 'brma'
influence(model, ...)Value
An object of class "infl.brma", which corresponds to the
structure of metafor::influence objects. It is a list containing:
- inf
A data frame with columns:
rstudent,dffits,cook.d,cov.r,tau.del, andhat, where available. Thetau.delcolumn is the PSIS leave-one-out posterior mean of the root mean square of total row-level heterogeneity over the remaining scale-model rows after each deletion, the root mean square over the observed design thatsummary_heterogeneityreports. It differs from the arithmetic mean of the observation-specific tau that 4.0.0 used, which is never larger. It is not a pooled log-scale design prediction.- dfbs
A data frame with DFBETAS values for the model coefficients.
Undefined COVRATIO and coefficient DFBETAS values are reported as
NaN with an explanatory note. DFFITS returns NA when its
leave-one-out posterior fitted-value SD is zero, without a separate note.
For binomial and Poisson GLMMs the rstudent component is the
descriptive normal approximation of rstudent.brma (and the
DFFITS, Cook's distance, and hat columns are not available).
For brma.mv() objects, influence returns only components with
implemented estimate-unit targets. DFFITS and Cook's distance report
influence on fixed-location fitted values \(\mu = X\beta\); tau.del
is currently deferred.
With correlated known-V data, the returned PSIS influence components
use conditional estimate deletion rather than independent-study deletion.
Details
The hat column uses the design of each draw's member for ensembles
whose members differ in their PET and PEESE terms, as described in
hatvalues.brma.
Selection brma.mv() models are currently unavailable through this
aggregate diagnostic. Standalone DFFITS and Cook's distance are available
for normal outcomes without a weightfunction component.
For known-V models, the hat values require applying the inverse of
the marginal covariance. DFFITS and Cook's distance use PSIS fixed-location
moments and do not apply this inverse. A covariance that is too
ill-conditioned for that (a reciprocal condition number of about 2e-10
or less in any posterior draw used) is unavailable: the call stops with an
error of class RoBMA_known_v_precision_unavailable, whose parent class
is RoBMA_numerical_unavailable, instead of returning a rounded number.
The condition carries the draw positions (rows) and their
reciprocal_condition.