Verified calibration methods in WFC 2.0

All public calibration methods in WFC 2.0 require the wf_design_data and wf_verified_target objects created in the verified workflow. The examples below assume design, target, and dims have already passed import and review as shown in vignette("safe-weighting-workflow").

Fixed-tolerance soft calibration

Soft calibration is for a tolerance declared before outcomes are inspected. It does not search for a target, widen tolerance automatically, or optimize a study result. The declared tolerance and any relaxed external margin appear in the result and statistical report.

soft <- wf_calibrate(
  design,
  target,
  method = "soft",
  tolerance = 0.02
)

wf_report(soft, audience = "statistician")$sections$soft_relaxation

If the declared tolerance cannot resolve infeasibility, WFC stops. Select a different external target only through a new, independently reviewed workflow.

Categorical entropy balancing

Entropy balancing may use only the categorical margins contained in the verified target.

ebal <- wf_calibrate(
  design,
  target,
  method = "ebal",
  tol = 1e-10
)

wf_report(ebal, audience = "statistician")

WFC 2.0 has no inline desired-mean interface. Continuous outcome goals, pass-rate goals, and desired intervals are not weighting inputs.

Bounded logit calibration

Bounds are reviewable settings, not values that WFC expands automatically:

bounded <- wf_calibrate(
  design,
  target,
  method = "logit",
  bounds = c(0.3, 3)
)

For a consequential analysis, record why the method and settings are suitable and include them in the human plan review.