PDRobust: Principal-stratification treatment-effect estimation
Source:R/PDRobust-package.R
PDRobust-package.RdPDRobust estimates treatment effects for longitudinal outcomes that may be
unavailable after death. The effects apply to the always-survivor principal
stratum: subjects who would survive through the selected cutoff time under
either treatment. Mapping() identifies the variables and analysis times,
and DataStandard() prepares the data for the remaining functions.
Details
PSPred(), PrinPred(), and OutPred() fit their models each time they are
called and return numeric predictions. HTESepT() estimates effects
separately at user-selected times, whereas HTEAllT() estimates a joint
trajectory over every observed time from baseline through cutoff.
Treatment coding
The implemented estimator uses treatment 1 as the survival-favorable arm:
potential survival satisfies \(S^1 \ge S^0\) at cutoff. Its always-survivor
principal score is therefore the survival probability under treatment 0.
If the survival-favorable arm is coded as 0 in the raw data,
recode the raw treatment as 1 - A before mapping and
standardizing. To report the original contrast, negate the package estimate
and transform an interval [lower, upper] to [-upper, -lower].
Mapping() does not infer or reverse treatment coding.
Interpretation and assumptions
The target population comprises subjects who would survive through the
selected cutoff under either treatment. The effect model describes outcome
differences at earlier analysis times within that fixed population.
Continuous-outcome models describe effects on a linear scale. Binary-outcome
models transform the model's linear predictor with
2 * plogis(eta) - 1; their coefficients are not odds ratios.
Causal interpretation requires consistency, no interference, adequate treatment and survival overlap, treatment ignorability conditional on the measured covariates, the stated survival monotonicity, and principal ignorability. Checking the data and covariate balance cannot establish these assumptions. Triple robustness means that, under the method's assumptions and regularity conditions, at least two of the propensity score, principal score, and outcome mean models must be correctly specified. It does not guarantee unbiased results in every finite sample. Limiting extreme probabilities can also affect the estimates.
SA() examines sensitivity by adding random noise to the outcome. It does
not vary the principal-ignorability assumption. See the method-and-coding
vignette for treatment coding and other implementation limits.
References
Zhang, Y., Shardell, M., Falvey, J., McCoy, R., Stuart, E., and Chen, C. (2026). A Novel Tool for Evaluating Effect Modification in Older Adults with ADRD Using Medicare Claims. arXiv:2608.06654. doi:10.48550/arXiv.2608.06654 .