Skip to contents

Computes Cook's distance for a fitted brma object. Cook's distance measures the aggregate influence of each observation on the model coefficients.

Usage

# S3 method for class 'brma'
cooks.distance(model, ...)

Arguments

model

a fitted normal-outcome brma object without a weightfunction component.

...

additional arguments (currently ignored).

Value

A numeric vector of Cook's distance values, one for each observation.

Details

Cook's distance is computed as a PSIS leave-one-out deletion diagnostic. For each observation \(i\), normalized PSIS weights estimate the fitted values under the leave-one-out posterior. The distance is the squared posterior Mahalanobis distance between the full-data and leave-one-out fitted-value vectors: $$D_i = \Delta_i' V_\mu^+ \Delta_i$$

where \(\Delta_i = \hat{\mu} - \hat{\mu}_{(-i)}\), \(V_\mu^+\) is the generalized inverse of the full-posterior fitted-value covariance. This uses the unscaled meta-analytic convention of metafor and Viechtbauer and Cheung (2010) , without division by the fixed-effect model rank. Values can differ from metafor because RoBMA uses posterior moments and PSIS deletion rather than classical refits. For brma.mv() known-V models, Cook's distance 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. Selection brma.mv() influence is currently unavailable. The PSIS fixed-location moments do not require the inverse marginal covariance used by hatvalues().

References

(Viechtbauer and Cheung 2010)

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)

  cooks.distance(fit)
}
} # }