When a treatment can cure some patients outright and merely delay the event for others, its effect has two separate parts. The first is "whether": it changes the fraction of patients who never relapse. The second is "when": it changes how soon relapse happens among patients who are not cured. Cure fraction models separate these two parts, but the causal methods currently applied to them rely either on crude matching or on outcome models that are fit and then trusted. Causal inference offers a better tool with an odd property. Fitting the treatment-assignment probability (the propensity score) from the data, rather than plugging in the known value, makes weighted estimators more precise, even in a randomized trial. The reason is that the fitted score absorbs chance imbalances in the particular sample. The size of that gain depends on which functions of the covariates the fitted score is forced to balance.
This note proposes to derive, for each of the two cure estimands, the covariate function a propensity model should balance, and then to build weights that balance exactly those functions. The central observation is simple. The "whether" contrast needs balance on the predicted cure probability. The "when" contrast needs balance on something different: each patient's deviation in timing, scaled by that patient's probability of being susceptible to the event at all. Generic balancing, whether on raw covariates or on one predicted survival probability, targets neither function and gives up precision. The payoff is weighted Kaplan–Meier curves and plateau estimates with efficiency at the semiparametric bound, weights that stay non-negative, and a fitted-versus-true propensity gain that can be predicted separately for each component.
Core Claim
The estimands are defined per treatment arm $a$, under the standard assumption of follow-up long enough for the survival curve to reach a plateau by a horizon $\tau$. The first is the marginal cure probability $\pi_a = S_a(\tau)$, where $S_a(t)$ is the marginal survival curve under treatment $a$. The second is the survival curve among the susceptible, $S_{u,a}(t) = (S_a(t) - \pi_a)/(1-\pi_a)$, which is the mixture cure decomposition written at the population level. Write $\pi_a(X)$ for the conditional cure probability and $S_{u,a}(t\mid X)$ for the conditional survival of susceptible patients. The conditional mean part of the efficient influence function for $\pi_a$ is $\pi_a(X)$. A delta-method calculation on the ratio shows that for $S_{u,a}(t)$ it is proportional to
$$h_{a,t}(X) = \big(1-\pi_a(X)\big)\big(S_{u,a}(t\mid X) - S_{u,a}(t)\big),$$
up to an additive constant. This is the susceptibility-weighted timing deviation. Following the logic of Optimal Covariate Balancing Conditions in Propensity Score Estimation, which showed that inverse-probability-weighted estimators become efficient and doubly robust when the propensity model is fit to balance the outcome's conditional mean, we estimate the propensity score $e(X)$ by solving
$$\frac{1}{n}\sum_{i=1}^n \left(\frac{W_i}{e(X_i)} - \frac{1-W_i}{1-e(X_i)}\right) g(X_i) = 0.$$
Here $W_i$ is the treatment indicator. The balancing functions $g$ are $\hat\pi_a(X)$ for the "whether" estimand and $\hat h_{a,t}(X)$ on a grid of $t$ values for the "when" estimand. Both are supplied by a fitted mixture cure model. Censoring before $\tau$ is handled by inverse probability of censoring weights.
The formula for $h_{a,t}$ yields three testable predictions.
- Balancing the whole predicted survival curve is misdirected for the "when" contrast. The outcome-adaptive approach of Outcome adaptive propensity score methods for handling censoring and high-dimensionality balances the predicted survival curve $S(t\mid X)$. That curve contains a term $\pi(X)(1-S_u(t))$ that drives balance on covariates affecting cure, which the "when" contrast does not need. The extra constraints cost effective sample size.
- The fitted-versus-true gain can be computed per component. It equals the variance of the projection of each influence function onto the balanced functions.
- Covariates that affect timing matter less when susceptibility is low. Their influence is scaled by $1-\pi(X)$, so they can be down-weighted exactly where most patients are cured.
Research Program
The theory has three parts. The first is the influence-function derivation above, extended to the arm contrasts $\Delta\pi$ and $\Delta S_u(t)$. The second is a double-robustness and efficiency result for the balancing-weight estimator, adapted from the oCBPS argument to handle censoring and the plateau functional. The third is a variance-gap formula for randomized trials. That formula extends the fitted-versus-true propensity results of Sufficient covariates and linear propensity analysis and of Propensity score matching for estimation of pairwise marginal hazard ratios from settings where everyone eventually fails to settings with a plateau.
The balancing functions come from mixture cure models of two strengths. One is the logistic-plus-accelerated-failure-time model used by Double Variable Importance Matching. The other is the Disentangled Neural Mixture Cure model of Disentangling Whether from When in a Neural Mixture Cure Model, which separates covariates into susceptibility, timing and shared representations. That paper explicitly leaves counterfactual estimation to future work. Here its outputs serve only as balancing functions, so errors in its identification of cure versus delay are partly protected by double robustness rather than passed straight into the effect estimate. The transformation cure models of Semiparametric Transformation Models for Survival Data With a Cure Fraction provide a third, misspecification-prone choice for stress tests: the true model is proportional odds and the fitted model is proportional hazards.
Two limitations are stated up front. First, the plateau horizon $\tau$ is assumed, as in the prior work. We report sensitivity to $\tau$ rather than claim to remove it. Second, $\Delta S_u(t)$ compares the susceptible populations of two arms, which differ in composition. It is therefore a descriptive "when" contrast, matching the estimand of the double-matching paper, not a principal-stratum effect. Principal stratification is out of scope.
Baselines
Naive reference points. The first is the unadjusted Kaplan–Meier plateau difference. The second, used in randomized trials, is inverse weighting with the known propensity of 0.5. The latter is the reference point against which the fitted-versus-true gain is measured.
Standard weighting. A logistic propensity model is fit by maximum likelihood on the raw covariates, followed by weighted Kaplan–Meier. This isolates whether estimand-specific balance adds anything beyond ordinary estimation of the propensity score.
Cox model with inverse weighting. A weighted Cox model is fit, with cured patients treated as censored. It is included because practitioners use it. Wang and Liang show it underestimates hazard for susceptible patients, and it produces no separate "whether" and "when" estimates.
Outcome-adaptive propensity score. Following Outcome adaptive propensity score methods, one predicted survival probability is added to the propensity model. That paper reports root mean squared error reductions of 26.3% to 41.6% for random forests with this addition, which makes it the strongest "generic balancing" competitor.
Double variable importance matching. This is the MCM KNN method of Double Variable Importance Matching, the closest existing method for these estimands. It reports cutting cure-probability mean absolute error "by more than half" relative to propensity score matching when cure and timing covariates overlap. No absolute error value appears on Lacuna.
Augmented inverse weighting with the same mixture cure outcome model. This is the strongest competitor. It reaches the same efficiency bound when both models are correct.
All methods receive the same fitted mixture cure model, the same censoring model and the same tuning budget. Neural components get an equal number of hyperparameter trials, selected by cross-validated likelihood that never uses treatment-effect error.
Ablations. Two ablations isolate the key ingredient. The first swaps $h_{a,t}$ for the generic $S(t\mid X)$ as the balancing function. The second compares fitted and true propensity within the same trial.
What counts as winning. Against augmented weighting, lower Monte Carlo variance and correct coverage under weak overlap or single-model misspecification, with no loss when both models are correct. Against MCM KNN, lower mean absolute error on both estimands in all four of its simulation scenarios.
Why not just…?
Why not just use augmented inverse weighting with a mixture cure outcome model? This is the real threat. When both models are correct, it has the same asymptotic variance as our estimator. The argument for balancing weights concerns finite samples and misspecification. Propensity score matching for estimation of pairwise marginal hazard ratios reports augmented weighting variance of 31.3, against 22.2 for simple weighting and 17.2 for matching, under weak overlap. Optimal Covariate Balancing Conditions shows that balancing estimators converge faster under local propensity misspecification. We run augmented weighting both with and without weight trimming. If trimmed augmented weighting matches our estimator in every regime, we will report that the transferable contribution is the function $h_{a,t}$ itself. It can be used as the outcome-regression target inside augmented weighting, and we would present it that way.
Why not just apply standard oCBPS to the plateau indicator $I(T>\tau)$? For $\Delta\pi$ alone, this is nearly our method. The only differences are that the mixture model estimates $\pi(X)$ better, because it borrows information from patients censored before $\tau$. For $\Delta S_u(t)$, plateau balancing targets the wrong function. The ablation that swaps in generic balancing functions tests exactly this, and a null result on $\Delta\pi$ would be expected and reported.
Why not just match on two importance-weighted distances, as MCM KNN does? Matching gives no efficiency bound and no valid bootstrap when matching with replacement. The double-matching paper provides no closed-form variance for its estimators. Its distance weights also inherit misspecification from the fitted parametric model, whereas our weights are protected by double robustness. The four-scenario simulation replicating its design decides this comparison.
Proposed Experiments
Simulations. We reproduce the MCM KNN design: 20,000 samples, 20 covariates, and four scenarios in which treatment affects cure only, timing only, disjoint covariates, or overlapping covariates. We add sample sizes of 500 and 2,000, strong and weak overlap, and proportional-odds generation fitted with a proportional-hazards cure model. The metrics are bias, Monte Carlo variance, 95% interval coverage, and effective sample size for $\Delta\pi$ and $\Delta S_u(t)$ at three horizons. Intervals come from the influence-function variance and are checked against a resampling scheme, since the double-resampling method of the hazard-ratio matching paper raised coverage from 91.3% to 95.4% under weak overlap.
Semi-synthetic benchmarks. On NACD and SUPPORT, following the DNMC paper, treatment and cure status are simulated on real covariates so that the ground truth is known.
Randomized trial. On E1690 (427 patients, 30 confirmed cured), we test the fitted-versus-true gain directly. We compare the known propensity of 0.5 against estimand-specific fitted weights, and check the observed variance ratio against the predicted projection formula. The original analysis found interferon did not significantly change long-term relapse-free survival. No marginal cure difference with an interval is reported on Lacuna for E1690.
Observational cohort. On the 1,157-patient acute lymphoblastic leukemia cohort, we compare against the roughly 5% leukemia-free survival improvement reported by MCM KNN, with attention to interval width and to whether the "whether" and "when" components agree in sign.
Decisive checks. Three results would support the proposal. The ablation using $h_{a,t}$ should reduce $\Delta S_u(t)$ variance relative to generic $S(t\mid X)$ balancing mainly when timing effects are concentrated in patients with low cure probability. The fitted-versus-true gain on E1690 should match its theoretical prediction. The estimator should beat trimmed augmented weighting under weak overlap without losing to it under strong overlap. Failure on the first check would falsify the core claim.
References
- Optimal Covariate Balancing Conditions in Propensity Score Estimation
- Outcome adaptive propensity score methods for handling censoring and high-dimensionality: Application to insurance claims
- Sufficient covariates and linear propensity analysis
- Propensity score matching for estimation of pairwise marginal hazard ratios
- Double Variable Importance Matching to Estimate Distinct Causal Effects on Event Probability and Timing
- Disentangling Whether from When in a Neural Mixture Cure Model for Failure Time Data
- Semiparametric Transformation Models for Survival Data With a Cure Fraction
- Improved Mixture Cure Model Using Machine Learning Approaches