Relative efficiency of a covariate-adjusted vs unadjusted analysis

In a randomized trial, covariate adjustment does not change the target estimand but it can sharpen it. getRelativeEfficiency() quantifies that precision gain by comparing the influence-function standard errors of an adjusted analysis to those of an unadjusted analysis of the same estimand, which is the quantity the FDA's 2023 covariate-adjustment guidance is about.

Supply two "ConcreteOut" tables (from getOutput() or getRMST()): one from a covariate-adjusted fit and one from an unadjusted fit. The unadjusted fit is the same workflow with treatment-only nuisance models, e.g. a marginal propensity ("SL.mean") and hazard formulas of the form Surv(time, event == j) ~ arm.

For each matched estimand the function reports the relative efficiency \(\mathrm{RE} = \mathrm{Var}_{\text{unadj}} / \mathrm{Var}_{\text{adj}}\) (values above 1 favor adjustment), the implied percentage variance reduction, and the effective sample-size multiplier: an adjusted analysis on \(n\) subjects has the precision of an unadjusted analysis on \(\mathrm{RE}\,n\).

Usage

getRelativeEfficiency(Adjusted, Unadjusted)

Arguments

Adjusted

a "ConcreteOut" table from a covariate-adjusted fit.

Unadjusted

a "ConcreteOut" table from an unadjusted (treatment-only) fit of the same estimands, interventions, events, and times.

Value

a data.table keyed by Intervention, Estimand, Event, and Time with columns seAdjusted, seUnadjusted, RelEfficiency, VarReductionPct, and EffSampleSizeMult.

See also

getOutput(), getRMST()