library("causatr")
fit <- causat(long_data, outcome = "Y", treatment = "A",
confounders = ~ L0, confounders_tv = ~ L,
id = "id", time = "t", estimator = "gcomp")
contrast(fit, interventions = list(
delay1 = grace_period(1),
natural = NULL
), ci_method = "bootstrap")Grace-period (delayed-initiation) natural-history treatment policy
Description
Creates a natural-history modified treatment policy (G-LMTP) that delays treatment initiation by window periods. Under the policy a patient initiates treatment exactly window periods after they would have initiated under their own natural treatment process, and treatment is absorbing once started (Diaz, Williams, Morzywolek & Rudolph 2026; the policy used in their simulation study, Section 6). For binary treatment coded 1 = on treatment, the intervened value is
A^d_t = { t - },
which depends on the natural-value history of treatment, not just the contemporaneous natural value.
This is not expressible as a dynamic() rule: the standard ICE recursion conditions on the observed lagged treatment, whereas the natural value at period \(t\) under the regime differs from the observed value whenever the policy has already perturbed the treatment trajectory (treatment-state feedback). The estimate is produced by the augmented-data sequential regression and is only valid for longitudinal g-computation (estimator = “gcomp”) on a discrete treatment.
Usage
grace_period(window = 1L, budget = 1024L)
Arguments
window
|
Non-negative integer. Number of periods by which initiation is delayed. window = 0 makes initiation absorbing at the natural initiation time (a no-delay reference); window = 1 is the canonical one-period delay from the paper.
|
budget
|
Positive integer. Maximum allowed worst-step natural-history enumeration \(|\mathcal{A}|^{\tau-1}\) before the augmentation is deemed intractable and aborted. Default 1024.
|
Value
A causatr_glmtp object (also inheriting causatr_intervention) carrying the policy closure.
References
Diaz I, Williams NT, Morzywolek P, Rudolph KE (2026). Modified treatment policies that depend on the natural history of treatment. arXiv:2605.24167.
See Also
carry_forward(), contrast(), dynamic()
Other glmtp: carry_forward()