Survival treatment effects get confidence intervals for change over time: one shared spline replaces per-time fits
Synopsis
For right-censored survival data, a treatment effect varies not only across patients but also over time; earlier methods mostly estimated conditional average treatment effects at single or fixed evaluation times and could not give valid inference for how much the effect changes between nearby times. This work models covariate effects with a neural network and the temporal trajectory with a shared B-spline basis, then builds inference from doubly robust pseudo-outcomes and a subsampled ensemble. In simulations under the Bernstein truth, as the grid refined from 5 to 97 points the joint fit's time-normalized contrast RMSE stayed at 0.31–0.34, while separately fitted DNNs rose from 0.46 to above 7; at one grid both reached a level RMSE of about 0.11.
Interpretation
Estimation error for temporal contrasts no longer has to grow as the evaluation grid refines, because the whole treatment-effect trajectory is fitted as one object with shared temporal structure. Earlier theory for flexible survival CATEs characterized accuracy at individual times or on a fixed evaluation grid, and separately fitted estimators amplify temporal-contrast error by the inverse of the gap between adjacent times as the grid densifies. In simulations with the Bernstein truth and true nuisance functions, as the grid refined from 5 to 97 points the joint fit's time-normalized contrast RMSE stayed at 0.31–0.34 while separately fitted DNNs rose from 0.46 to above 7; at one grid both had a level RMSE of about 0.11.
Beyond point estimates, pointwise CATEs, temporal contrasts and covariate contrasts all receive asymptotically valid confidence intervals, including contrasts over shrinking time intervals. Inference for nonlinear survival CATE estimators was underdeveloped, and existing subsampling and ensemble results did not directly cover jointly modeled survival CATEs with estimated nuisance functions. Across 200 simulation replications, coverage was 0.947–0.948 for pointwise CATEs, 0.939–0.947 for temporal contrasts and 0.957–0.958 for covariate contrasts, staying near nominal when either the propensity or the survival model was misspecified.
Covariate heterogeneity is handled by a neural network while time is expanded in a fixed-order B-spline basis, so the network outputs spline coefficients rather than per-time effect values. Separate fitting trains one model per evaluation time, whereas joint fitting lets all times share one mapping from covariates to coefficients. The joint estimator used cubic clamped B-splines and networks with two hidden layers of 128 and 64 units, with 200 Monte Carlo replications per configuration.
Perspective
The demonstrated scope is right-censored survival data, conditional average treatment effects on a prespecified evaluation grid and their temporal and covariate contrasts; simulations use three CATE surfaces generated from known truth, and the application is an adjuvant chemotherapy analysis of stage II and III surgically resected patients in the Boston Lung Cancer Study.
The theory relies on unconfoundedness, positivity, conditionally independent censoring and several regularity and bias-control conditions; the Boston Lung Cancer Study analysis is observational and may carry residual confounding and treatment-timing bias, so the authors caution against causal interpretation. Simulations also report that coverage deteriorates in one setting even with a smaller bias-to-standard-error ratio, indicating that bias reduction alone does not guarantee valid inference.
