Skip to contents

Tipping-point sensitivity analysis for informative censoring

Source: R/senseCensoring.R
senseCensoring.Rd

concrete's primary analysis assumes censoring is independent of the event given the measured covariates. senseCensoring() probes robustness to departures from that assumption with a transparent tipping-point (bounds) analysis: a fraction delta of the subjects who are censored before the target time are assumed to have actually experienced the event of interest, and the analysis is re-fit for each delta. delta = 0 is the optimistic bound (censored subjects never have the event), delta = 1 the pessimistic bound (all do), and the primary inverse-probability-of-censoring analysis sits between them. The tipping point is the smallest delta at which the conclusion changes – a sensitivity artifact recommended by ICH E9(R1).

Unlike scaling the censoring weight (which leaves a doubly-robust estimator's target unchanged), imputing the event status of the censored changes the estimand and so produces a genuine, interpretable sensitivity curve. Because it re-fits the estimator for every delta, it is computationally heavier than the primary analysis.

When the fit carries a treatment-switching (crossover) model (i.e. formatArguments() was called with Crossover), the censored subjects are a mix of two intercurrent events with two different untestable assumptions: ordinary dropout (conditionally-independent censoring) and crossover (the no-switching counterfactual). mechanism selects which pool is imputed, so the two assumptions can be probed individually or jointly:

  • "dropout" – tip only the genuinely-censored (drop-out) subjects, holding the switching handled by the crossover hazard;

  • "crossover" – tip only the subjects re-censored at their switch time, i.e. probe “what if switchers would have had the event had they not switched”, holding dropout handled by the censoring weight;

  • "all" (default) – tip both pools jointly (the original behaviour).

With no crossover model, all censored subjects are dropout and the three modes coincide.

Usage

senseCensoring(
  ConcreteArgs,
  deltas = seq(0, 1, by = 0.25),
  Estimand = c("RD", "RR", "Risk"),
  Intervention = c(1, 2),
  mechanism = c("all", "dropout", "crossover"),
  Signif = 0.05,
  Verbose = FALSE
)

Arguments

ConcreteArgs

a "ConcreteArgs" object from formatArguments().

deltas

numeric in [0, 1]: fractions of pre-target-time censored subjects imputed as having the event of interest. Should include 0.

Estimand

one of "RD" (default), "RR", "Risk".

Intervention

length-2 numeric: treatment and control indices.

mechanism

one of "all" (default), "dropout", "crossover": which pool of censored subjects to impute (see Details). "crossover" requires a fit built with Crossover.

Signif

numeric (default 0.05): two-sided alpha.

Verbose

logical.

Value

a data.table of estimate / CI / p-value by delta x event x time, with a leading mechanism column, the tipping point in attr(., "tippingPoint"), and the mechanism in attr(., "mechanism").