Regularized or constrained estimators in psychnets self-certify: each
fitted network carries a residual that reports how far it sits from the
optimum of its own objective (a KKT / stationarity residual), or – for the
structural methods (TMFG, relative importance) – whether the graph
satisfies the identity that defines it. This verb returns that certificate
as a tidy one-row data.frame, so correctness is read the same way for
every method.
Arguments
- x
A psychnet object.
- tol
Tolerance below which the fit is flagged
certified = TRUE. Default1e-6.
Value
A one-row data.frame with columns method, certificate (the
residual; smaller is better), kind ("kkt" for the optimization
certificates, "structural" for TMFG/relimp, "none" for cor/pcor), and
certified (logical: residual at or below tol; NA when no applicable
certificate is available). Correlation networks use kind "none" and are
reported as TRUE because no optimization claim is made.
Details
The residual is near machine zero for a correctly solved problem. A graph
altered after fitting (post-estimation threshold, Wald pruning in
ising_sampler()) or an external-engine fit without an available residual
reports NA rather than being presented as certified. cor and pcor have
no optimization to certify and report NA with kind "none".
Examples
S <- 0.4^abs(outer(1:6, 1:6, "-"))
certificate(ebic_glasso(cor_matrix = S, n = 250))
#> method certificate kind certified
#> 1 glasso 1.10463e-10 kkt TRUE
certificate(tmfg_network(cor_matrix = S))
#> method certificate kind certified
#> 1 tmfg 0 structural TRUE