Standardised treatment contrasts with estimate() • survalis

estimate() turns any survalis learner into a plug-in G-computation (outcome-standardisation) estimator of a covariate-adjusted contrast between the two levels of a binary treatment.

What it computes

Given formula = Surv(time, status) ~ x1 + x2 (the right-hand side lists the confounders) and treatment = "a", estimate():

fits the working model Surv(time, status) ~ a * (x1 + x2) with the chosen learner,
sets a to each level for every subject and predicts $\hat S(t \mid X_i, a)$ ,
averages over subjects to get counterfactual mean curves $\bar S_a(t) = n^{-1} \sum_i \hat S(t \mid X_i, a)$ ,

contrasts them into one of

`estimand`	quantity
`"RMST_diff"`	$\int_0^\tau \bar S_1 - \int_0^\tau \bar S_0$
`"RMST_ratio"`	$\int_0^\tau \bar S_1 \big/ \int_0^\tau \bar S_0$
`"survival_diff"`	$\bar S_1(\tau) - \bar S_0(\tau)$
`"survival_ratio"`	$\bar S_1(\tau) \big/ \bar S_0(\tau)$

Example

library(survalis)

est <- estimate(
  Surv(time, status) ~ age + karno + celltype, data = veteran,
  model = "coxph", estimand = "RMST_diff",
  treatment = "trt", tau = 300,
  R = 100, seed = 1
)
est
#> <survalis_estimate> RMST difference at tau = 300  (model: coxph)
#>   treatment: trt  (2 vs 1)
#>   estimate: -14.6732   95% CI [-38.4702, 7.7335]   (100 bootstrap, refit)
#>   plug-in G-formula; causal reading needs exchangeability, positivity,
#>   consistency, no interference and a correct outcome model.

plot(est)

Inference

The confidence interval is a nonparametric subject bootstrap. By default (refit = TRUE) the working model is refitted inside every replicate, so the interval reflects model-estimation error as well as sampling variability. For an expensive learner you can set refit = FALSE to get an interval that is conditional on the fitted model – faster, but narrower than it should be.

summary(est)
#>    estimand tau  estimate  conf_low conf_high model
#> 1 RMST_diff 300 -14.67316 -38.47022  7.733488 coxph

Assumptions

estimate() is a plug-in G-formula estimator. It is not doubly robust and carries no protection against outcome-model misspecification beyond the treatment-by-confounder interactions it adds automatically. The estimate has a causal interpretation only under:

consistency – the observed outcome equals the potential outcome under the treatment actually received;
no unmeasured confounding given the covariates in formula;
positivity – both treatment levels are possible for every covariate pattern;
no interference between subjects;
a correctly specified outcome model.

For a non-manipulable attribute (sex, a biomarker), read the result as a covariate-adjusted marginal contrast, not an effect.