Computes DFFITS (Difference in FITS, standardized) for a fitted brma object. DFFITS measures how much the fitted value for observation \(i\) changes if observation \(i\) is removed, standardized by the estimated standard error of the fit.
Details
DFFITS values are computed as a PSIS leave-one-out deletion diagnostic. For each observation \(i\), the leave-one-out posterior mean fitted value at that observation is estimated with normalized PSIS weights and compared to the full-posterior fitted value: $$DFFITS_i = \frac{\hat{\mu}_i - \hat{\mu}_{i(-i)}}{SD_{(-i)}(\mu_i)}$$
This targets deletion influence on fitted values directly. It does not use
LOO-PIT residuals, which are predictive outlier diagnostics rather than
fitted-value deletion diagnostics.
For brma.mv() known-V models, DFFITS uses estimate-unit PSIS
weights. With correlated known-V, deletion is conditional estimate
deletion and the reported fitted-value target is the fixed-location mean
\(\mu = X\beta\); sampled or marginalized random effects are not included
in the reported fitted value.
Estimate-unit LOO must first be computed with
model <- add_loo(model, unit = "estimate"). If the leave-one-out
posterior SD of a fitted value is zero, the corresponding DFFITS value
is returned as NA, without a separate explanatory note.
The diagnostic is available for normal outcomes without a weightfunction
component; selection brma.mv() influence is currently unavailable.
Fixed-location PSIS moments do not require the inverse marginal covariance
used by hatvalues().
Examples
if (FALSE) { # \dontrun{
if (requireNamespace("metadat", quietly = TRUE)) {
data(dat.lehmann2018, package = "metadat")
fit <- bPET(yi = yi, vi = vi, data = dat.lehmann2018, measure = "SMD")
fit <- add_loo(fit)
dffits(fit)
}
} # }