Who Lives Longer but Feels Worse Sooner? Semi-Competing Risks for Patient-Reported Deterioration

Many treatments improve the number doctors track, such as tumor control, survival, or tremor amplitude, and at the same time make patients feel worse. Larger ablation lesions suppress tremor more but cause speech and balance problems. Opioids lower pain scores but sedate. Prostatectomy and radiotherapy cure equally well but cause different long-term harms. Trials usually report these two sides separately. One side is a survival curve. The other is a quality-of-life score averaged over whoever is still answering questionnaires. That second number is distorted twice. Patients who die stop reporting, and so do patients who quit treatment because they feel terrible. The average score therefore looks better than the patients' actual experience.

We propose to treat "clinically meaningful worsening on a patient-reported questionnaire" as a time-to-event outcome. Death can cut it off, and symptom-driven dropout can hide it. We then estimate a quantity that trials do not currently report: the fraction of patients for whom a treatment delays death but brings forward quality-of-life deterioration. We also estimate how large that trade-off is within this group. Answering this requires three pieces together. A "semi-competing risks" model handles an intermediate event that death can pre-empt but not the reverse. A dropout model uses each patient's last reported symptoms. A discrete-time hazard model matches the fact that questionnaires arrive at scheduled visits, not in continuous time.

Core Claim

The obvious analysis at this intersection fits a Cox proportional hazards model (a regression for the instantaneous event rate, which assumes treatment multiplies that rate by a constant factor over time) to time-to-deterioration. It handles death in one of two ways. Either death is censored, which assumes the dead would have behaved like the living. Or death is folded into a composite "deterioration-free survival" endpoint. Neither can answer the clinically important question. A composite hazard ratio near 1 is consistent with "no effect" and also with "half the patients trade quality for quantity and half the reverse."

Our claim is that this trade-off can be stated as a principal stratum, meaning a subgroup defined by how a patient would respond under both arms. With potential deterioration time $D(z)$ and death time $T(z)$ under arm $z \in \{0,1\}$, and $D(z)=\infty$ when death comes first, the target is

$$\pi_{\text{trade}} = P\big(T(1) > T(0),\; D(1) < D(0)\big),$$

the share of patients who live longer but deteriorate sooner on treatment. Alongside it we estimate the average difference $E[D(0)-D(1)]$ within that stratum.

This quantity cannot be estimated from data alone. Each patient is seen under only one arm, so any estimate depends on an assumption about how a person's outcomes under the two arms relate. We therefore estimate it under explicit structural assumptions and report how it moves as a sensitivity parameter for that unobservable link varies. A further claim is that symptom-driven dropout must enter as dependent censoring, modeled from each patient's own last observed scores. Otherwise both the stratum sizes and the within-stratum effects are biased toward the more toxic arm, because its sickest patients vanish from follow-up.

Research Program

The method supplier is Causal inference for semi-competing risks data. That paper defines latent strata such as "always-diseased" and "disease-harmed" over bivariate potential event times. It replaces strict monotonicity with a weaker order-preservation assumption and fits illness-death models, which track the transitions healthy→ill, healthy→dead, and ill→dead. A shared frailty links those transitions. A frailty is a per-patient random effect that captures unmeasured frailness affecting all of a patient's transition rates. That paper's intermediate event is an objective diagnosis, Alzheimer's onset, and its "treatment" is a fixed genotype with no dropout. We keep the illness-death and frailty machinery and change three things.

First, the intermediate event becomes confirmed patient-reported deterioration. This means a drop beyond a prespecified minimal important difference that is confirmed at the next visit, so that one bad week does not count. The definition will be preregistered per instrument. Examples are PROMIS Physical Function, which Patient reported outcomes can improve performance status assessment: a pilot study found tracks wearable step counts (r = 0.57) where clinician ECOG ratings did not, or EPIC domains as used in Patient-reported outcomes after surgery or radiotherapy for localised prostate cancer: a retrospective study.

Second, time is discretized to the visit schedule. Each transition gets a discrete-time hazard with interval-specific baseline terms, fitted with the two-step estimator of PyDTS: A Python Package for Discrete Time Survival Analysis with Competing Risks via the PyDTS resource. Treatment coefficients are allowed to differ between an early and a late window. Proportional hazards then becomes a testable restriction rather than an assumption. This matters because acute toxicity that peaks and subsides is exactly the case where a constant hazard ratio fails.

Third, dropout and treatment discontinuation become a third competing exit. Their hazard depends on the most recent patient-reported scores and their slope. This yields inverse-probability-of-censoring weights for the deterioration and death processes. A tilt parameter then lets dropout depend on the unobserved next score beyond what the observed scores explain. The motivation is that the symptoms that drive people out are routinely recorded as minor in trial registries, as Frequent discussion of insomnia and weight gain with glucocorticoid therapy: an analysis of Twitter posts documents.

The output is a two-dimensional sensitivity map. One axis is the cross-world dependence between $D(0)$ and $D(1)$. The other is the dropout tilt. Each cell gives $\pi_{\text{trade}}$ and the within-stratum deterioration shift. This plays the same role as the bias functions in A Surrogate Endpoint Based Provisional Approval Causal Roadmap. That roadmap rejects the idea of a perfect surrogate by attaching explicit transport-bias functions. We borrow the stance of reporting estimates as a function of an untestable bias, not its estimators. Its target is a serum surrogate for a binary risk, whereas ours is a time-to-event trade-off between two patient outcomes.

Inference on Controlled Effects for Assessing Immune Correlates of Protection Based on a Cox Model asks a related question: what happens to clinical risk if a biomarker is set to some level? We do not treat the questionnaire score as a surrogate to be validated in that sense. Our question is about two co-primary outcomes in tension, not whether one predicts the other.

Baselines

The naive reference point is the analysis most quality-of-life sections already contain. It is a standard Cox model for time to confirmed deterioration, with death and dropout both treated as uninformative censoring. We run it on exactly the same deterioration definition, visit grid, and covariates as our method, so differences reflect only the handling of death and dropout. It cannot produce $\pi_{\text{trade}}$ at all. Its role is to show how wrong the deterioration hazard ratio becomes under symptom-driven dropout, measured as bias and confidence-interval coverage against simulation truth.

The second baseline is a Cox model on the composite endpoint "deterioration or death, whichever first." This is the strongest simple competitor for testing whether arms differ, because it removes the censoring-by-death problem without any latent strata. We will beat it only in a specific sense. In scenarios where the composite hazard ratio is near 1 but a substantial trade-off stratum exists, our method must recover $\pi_{\text{trade}}$ within its interval while the composite test has no power. In scenarios with no trade-off, we must lose little efficiency relative to it on the overall comparison.

The strongest existing method is the semi-competing risks estimator of Causal inference for semi-competing risks data. We apply it as published, with deterioration as the non-terminal event, continuous time, a single proportional treatment effect, and no dropout model. On its own Adult Changes in Thought (ACT) application the paper reports about 35% always-diseased and 53% never-diseased, and roughly a 2-year earlier diagnosis for APOE ε4 carriers in the always-diseased stratum. Our re-implementation must reproduce these numbers before any comparison counts. No reported result found on Lacuna for this or any method on a patient-reported deterioration outcome. We give it the same frailty family, the same covariates, and the same tuning budget for frailty variance and baseline-hazard smoothness. Beating it means lower bias and nominal coverage for stratum proportions and within-stratum effects when dropout is symptom-driven or when the treatment effect is time-varying, while matching it when neither holds.

Three ablations isolate our ingredients. The first removes the dropout model and keeps everything else. The second forces a single time-constant treatment coefficient. The third replaces discrete visit intervals with continuous time. If removing the dropout model does not degrade results under informative-dropout scenarios, the core ingredient is not earning its place.

We do not use DeepSurv (DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network) as a main comparator. It learns a single nonlinear risk for one event and recommends the arm with lower hazard, which is the single-endpoint view we are arguing against. Its reported METABRIC C-index (0.867 vs 0.816 for linear Cox) measures ranking of a single event. Following Position: Stop Chasing the C-index when Evaluating Survival Analysis Models, which shows that the C-index ignores calibration and is over-optimistic under dependent censoring, we do not use the C-index as an endpoint.

Why not just…?

Why not just report deterioration-free survival with a Cox model? For a single yes/no comparison of arms this is a real competitor, and we expect it to match or beat us in power when every patient's survival and quality move in the same direction. Its failure is structural. Its hazard ratio averages over patients whose survival gains and quality losses cancel, so it cannot say who pays for longer life with earlier decline. The decisive simulation below sets the composite hazard ratio to 1 while planting a known trade-off stratum. If our estimator cannot recover that stratum with honest intervals, the composite endpoint is all a trial needs and we will say so.

Why not just model the score trajectory and death jointly with a shared random effect? Shared-random-effect joint models of a repeated score and survival also address informative dropout and are a serious alternative. They target a different quantity: the mean score trajectory conditional on the random effect, with no counterfactual pairing of arms. The file contains no reported result for such a model in this setting. To test the threat, we will fit one in the simulations and derive its implied $\pi_{\text{trade}}$ by simulating potential outcomes from the fitted model. If it matches our accuracy, the right contribution is the estimand and sensitivity map, not our estimator, and we would adopt the simpler fitting route.

Why not just collapse quality and survival into quality-adjusted life years? Quality-adjusted survival gives one number. It does so by fixing utility weights that patients do not share, and it again averages away the people on the losing side of the trade. The prostate cohort in Patient-reported outcomes after surgery or radiotherapy for localised prostate cancer: a retrospective study illustrates this: 70% incontinence after surgery against 36.5% rectal bleeding after radiotherapy. Any utility weighting decides between such harms on the patient's behalf. Our output keeps the two axes separate. Utility weighting can still be applied afterward.

Why not just use clinician performance status instead of questionnaires? Clinician ECOG ratings are easy covariates. However, Patient reported outcomes can improve performance status assessment: a pilot study found they did not correlate with measured daily activity (r between −0.08 and −0.38, intervals crossing zero), while PROMIS scores did. We will report every analysis with both definitions of deterioration where both are recorded. Our expectation is that the ECOG version shows a smaller and noisier trade-off stratum. The same paper's finding that the PROMIS–activity link differed by sex leads us to estimate strata separately by sex as a prespecified check.

Proposed Experiments and Decisive Checks

The primary test bed is simulation, because only there are the strata known. We generate two arms from an illness-death process with a shared frailty, with visit-based measurement of a latent score. We vary five factors. The first is the true trade-off share (0%, 10%, 25%). The second is whether the treatment effect on deterioration is early-only, a toxicity pattern, or constant. The third is whether dropout depends on the latest observed score, on the unobserved next score, or on neither. The fourth is the cross-world correlation. The fifth is sample size, from a few hundred, typical of a phase 2 trial, to several thousand.

Metrics are bias, root-mean-square error, and interval coverage for $\pi_{\text{trade}}$, for the other strata, and for within-stratum deterioration shifts. Calibration of predicted cumulative incidence is checked per transition. We also record the width of the sensitivity map, because a method that is unbiased only at the true sensitivity value but uninformative across plausible values is not useful.

The first decisive check is the null-composite scenario described above. The second is a dropout scenario in which the more toxic arm loses its worst-off patients. There, the naive Cox deterioration hazard ratio should flip direction or shrink toward benefit, while our weighted estimator stays near truth. If our estimator also fails here, the dropout model is mis-specified and the proposal's central ingredient is wrong.

On real data, the ACT cohort serves as the reproduction check, using the reported 35%, 53%, and roughly 2-year figures. It also tests whether moving to discrete time and adding a time-varying effect changes those conclusions. SUPPORT, one of the DeepSurv benchmarks, records functional status alongside survival in seriously ill adults. We will use it as a semi-real test bed: functional decline is the non-terminal event where follow-up measurements exist in the release used, and deterioration is overlaid on real covariates and death times where they do not. The analysis compares strata across baseline subgroups rather than a randomized arm, and is labeled as observational.

The intended end use is a randomized trial with scheduled questionnaires and recorded discontinuation reasons. No such dataset appears in the file, so obtaining one under a data-use agreement is a dependency of the program rather than an assumed resource.