Chemotherapy Changes and Survival in Cancer

Why Chemotherapy Survival Is a Longitudinal Problem

Chemotherapy decisions are often revised as a patient’s cancer and general condition change. Performance status and tumor response can predict death, but they can also influence whether the next chemotherapy cycle is given, delayed, changed, or stopped. If earlier treatment affects these measurements, they are time-varying confounders affected by prior treatment. Simply adjusting for their latest values in a conventional outcome model can fail to estimate the total effect of a treatment strategy.

The aim here is to estimate marginal survival under specified chemotherapy strategies: the survival that would be observed in the study population if everyone followed each strategy. This is a case-study application of inverse-probability weighting. The key practical detail is that treatment decisions exist only while a patient is alive and in care. A weight must therefore stop at death, while including the treatment decision made in the interval in which death occurs.

Define the Strategies, Timeline, and Target

Use discrete intervals, such as successive chemotherapy cycles. Let $T$ be the interval of death, with $T=\infty$ for a patient who does not die during follow-up. At the start of interval $j$, a patient is eligible for a treatment decision if $T\ge j$. Assume the decision $A_j$ is made before death during interval $j$ is counted. Thus a patient who dies in interval $j$ has reached the decision $A_j$, but does not reach any later decision.

Let $L_j$ denote the measured clinical history just before decision $j$, including performance status and tumor response when available. Write $H_j=(V,\bar L_j,\bar A_{j-1})$ for the history available at that decision, where $V$ denotes baseline covariates. A chemotherapy strategy $g$ specifies a treatment $g_j(H_j)$ at every decision reached. It may be sustained treatment, such as continuing a specified regimen, or a dynamic strategy, such as changing regimen when response is inadequate. Let $T^g$ be the death interval that would occur under strategy $g$, and define marginal survival through interval $k$ by $S^g(k)=P(T^g>k)$.

Definition: For a patient following strategy $g$, define the stopped treatment weight through horizon $k$ as $$ W_k^g = \prod_{j=1}^{\min(k,T)} \frac{I\{A_j=g_j(H_j)\}} {P\{A_j=g_j(H_j)\mid H_j,T\ge j\}}. $$ The product includes the treatment decision in the death interval when $T\le k$, and includes no decisions after death. For a patient who survives beyond $k$, it includes decisions $1,\ldots,k$. The weight is defined only for decisions actually reached; it does not require post-death treatment values.

In the denominator, the treatment probability is evaluated under the observed treatment process among patients alive at the start of the decision interval. The indicator makes the weight zero if the observed treatment at a reached decision does not follow $g$. This version uses unstabilized weights to keep the change-of-measure argument transparent. Strategies must specify treatments with positive probability in the relevant histories; a deterministic regimen with no observed support cannot be identified by weighting.

Assumptions for a Causal Survival Interpretation

Consistency: For a patient whose observed treatment history follows $g$ up to a given time, the observed clinical history and death outcome through that time equal the corresponding outcomes under $g$.
Sequential exchangeability: At each reached decision $j$, conditional on $H_j$ among patients with $T\ge j$, the treatment choice is independent of the relevant future potential outcomes under the strategies being compared. This rules out unmeasured confounding of each treatment decision, conditional on the recorded history.
Positivity: For histories that can occur under $g$ among patients alive at decision $j$, $P\{A_j=g_j(H_j)\mid H_j,T\ge j\}>0$. For stable estimation, these probabilities must not be too close to zero.
Well-defined treatment and no interference: Treatment versions are sufficiently specified for consistency, and one patient’s regimen does not change another patient’s outcome.
Outcome observation: Death is ascertained through the horizon. If follow-up can be lost before death status is known, assumptions and methods for censoring must also be addressed; the treatment weight alone does not correct informative loss to follow-up.

These assumptions distinguish the statistical reweighting identity from its causal interpretation. The identity changes the treatment assignment law in the observed data. Consistency and sequential exchangeability are what allow that intervention distribution to represent the counterfactual outcomes under a chemotherapy strategy.

Why the Stopped Weight Works

At each decision reached, inverse-probability weighting cancels the observed probability of the treatment actually prescribed by $g$. The cancellation must stop when the patient dies: later treatment probabilities are neither observed nor defined. The resulting identity applies to the entire observed path only up to the earlier of the horizon and death.

Theorem: Suppose the relevant treatment probabilities are positive. For any bounded measurable function $f$ of the clinical and treatment path through $\min(k,T)$, and the death status through $k$, $$ E\!\left[W_k^g f\right]=E^g[f], $$ where the right-hand side is expectation under the intervention that assigns $g_j(H_j)$ at each decision reached. Under consistency and sequential exchangeability, this is the corresponding expectation of the counterfactual path under $g$.

Factor the observed path.
Conditional on each history, the observed law factors into a treatment assignment probability and the subsequent law of the next clinical history and death, up to the next reached decision or death.

Replace treatment assignment.
At a reached decision $j$, multiplying by $I\{A_j=g_j(H_j)\}/P\{A_j=g_j(H_j)\mid H_j,T\ge j\}$ removes the observed treatment probability and puts probability one on the treatment prescribed by $g$.

Stop at death or the horizon.
For a path dying in interval $t\le k$, the factorization has treatment terms only through $t$; for a path surviving beyond $k$, it has terms through $k$. Summing or integrating these reweighted path laws gives the intervention law for every bounded $f$, proving the identity. Under the causal assumptions, that law is the counterfactual law under $g$. $\square$

For survival, take $f=I(T>k)$. A survivor through $k$ has reached all $k$ decisions, so its contribution uses $W_k^g$; a patient who dies earlier has weight factors only through the death interval, and contributes zero to $I(T>k)$. This is precisely why extending every patient’s weight mechanically through $k$ would be invalid.

Worked Examples

Worked Example: A Patient Who Dies During Follow-Up

Consider the strategy “give regimen R at every cycle reached.” A patient has probability $0.50$ of receiving R at cycle 1 and receives R. The patient is alive at cycle 2, has a poor response, and the probability of receiving R at cycle 2 given this history is $0.25$. The patient receives R at cycle 2, then dies during that interval. Calculate the weight through horizon $k=4$.

The patient reached decisions 1 and 2, so the product stops at $\min(4,T)=2$. Both treatment indicators equal one. Therefore $$ W_4^g=\frac{1}{0.50}\times\frac{1}{0.25}=2\times4=8. $$ There are no factors for cycles 3 and 4: the patient died before those decisions. The contribution to survival beyond cycle 4 is $8\times0=0$, since $I(T>4)=0$. Assigning hypothetical treatments after death would add undefined probabilities and would not be part of the weight.

Worked Example: A Small Weighted Survival Estimate

Suppose four patients are evaluated under the strategy “give chemotherapy at each reached decision.” For this illustration, use the empirical Hájek estimate, the ratio of the weighted survival total to the total weight. Patient 1 receives chemotherapy at cycle 1 with probability $0.50$, survives that cycle, then receives chemotherapy at cycle 2 with probability $0.80$ and survives beyond cycle 2. Patient 2 receives chemotherapy at cycle 1 with probability $0.50$, survives, but does not receive it at cycle 2; the probability of chemotherapy at cycle 2 given this patient’s history was $0.25$. Patient 3 does not receive chemotherapy at cycle 1, whose probability was $0.50$. Patient 4 receives chemotherapy at cycle 1 with probability $0.50$, then dies during cycle 1.

Patient 1 follows the strategy through cycle 2, so $W_2^g=(1/0.50)(1/0.80)=2.5$, and contributes $2.5$ to both the weight total and the weighted survival total. Patient 2 has a zero weight because the cycle-2 treatment indicator is zero; its cycle-2 factor is $0/0.25=0$. Patient 3 has a zero weight at cycle 1 because its treatment indicator is zero. Patient 4 dies in cycle 1, so the weight stops there: $W_2^g=1/0.50=2$. Its survival indicator through cycle 2 is zero.

The weighted survival estimate is therefore $$ \widehat S^g(2)= \frac{2.5(1)+0(0)+0(0)+2(0)} {2.5+0+0+2} =\frac{2.5}{4.5} \approx 0.556. $$ This tiny, illustrative sample is not a reliable study estimate. Its purpose is to show how a death in cycle 1 contributes a finite, stopped weight to the denominator, but not to survival beyond cycle 2.

Worked Example: Extreme Weights and Effective Sample Size

Suppose four patients who follow a strategy have weights $1.25, 2, 5,$ and $0.8$, reflecting different treatment probabilities along their histories. Their total weight is $1.25+2+5+0.8=9.05$. The usual weight-based effective sample size is $$ n_{\mathrm{eff}}= \frac{(\sum_i W_i)^2}{\sum_i W_i^2} = \frac{9.05^2}{1.25^2+2^2+5^2+0.8^2} = \frac{81.9025}{31.2025} \approx 2.62. $$ Although four patients are present, the weight concentration gives information roughly comparable to only 2.62 equally weighted observations by this diagnostic. The weight of 5 deserves investigation: it may reflect a genuinely uncommon treatment choice in that history. Truncating it can reduce instability, but changes the estimator and may introduce bias; results with and without a prespecified truncation should be compared rather than treating truncation as a cure for poor overlap.

Estimating Survival from Weighted Risk Sets

A useful construction estimates each interval’s death hazard under $g$, then multiplies the estimated interval survival probabilities. For a patient at risk at the start of interval $j$, let $W_j^g$ be the weight through decision $j$. It is defined for this risk set because these patients have reached decision $j$. Let $R_j=I(T\ge j)$, and let $D_j=I(T=j)$. The weighted hazard estimate is $$ \widehat\lambda_j^g= \frac{\sum_i W_{ij}^g R_{ij}D_{ij}} {\sum_i W_{ij}^g R_{ij}}. $$

Result: Under the stopped-weight identity, the population weighted risk-set ratio equals the death hazard under strategy $g$: $$ \frac{E[W_j^g I(T\ge j)I(T=j)]} {E[W_j^g I(T\ge j)]} = P^g(T=j\mid T\ge j). $$ Consequently, when death is the event of interest and there is no unaccounted informative censoring, the product-limit construction is $$ \widehat S^g(k)=\prod_{j=1}^{k}(1-\widehat\lambda_j^g). $$

To verify the result, apply the path identity at horizon $j$. The function $I(T=j)$ selects deaths in interval $j$, while $I(T\ge j)$ selects those alive at its start; both functions concern paths whose treatment decision $j$ is reached. Thus the numerator’s population expectation is $P^g(T=j)$, and the denominator’s is $P^g(T\ge j)$. Their ratio is the stated conditional hazard. Finally, $$ P^g(T>k)=\prod_{j=1}^k \{1-P^g(T=j\mid T\ge j)\}, $$ by repeated conditioning on survival to the start of each interval. Replacing each hazard with its weighted estimate gives the product-limit estimator. $\square$

With censoring, risk sets must also account for the probability of remaining observed; simply omitting people lost to follow-up is not justified if that loss is informative. The treatment weight and any censoring adjustment address different processes.

Diagnostics, Alternatives, and Limitations

Before interpreting an estimate, verify the timeline in the data: performance status and tumor response used in a treatment model must be measured before that cycle’s treatment decision. Check that patients who die in an interval contribute treatment information only through the decision in that interval. A treatment probability should be estimated among the people alive and eligible for the corresponding decision, not among a population that includes those already dead.

Inspect treatment probabilities by cycle and clinically important histories. Values near zero mean a strategy is weakly supported by the observed data.
Inspect the distribution of cumulative weights and the effective sample size. A few very large weights can make estimates unstable.
Check whether weighting improves balance in measured pre-decision performance status, response, and other prognostic history among the relevant risk sets.
Compare estimates under reasonable, prespecified weight-truncation choices, and report the resulting sensitivity. Large changes signal dependence on a small number of observations.

A standard time-dependent outcome regression that adjusts for current response and performance status may condition on variables affected by earlier chemotherapy. That can remove part of the treatment effect or create other bias, depending on the model and causal structure. The longitudinal g-formula is an alternative that models the evolving clinical history and outcomes under each strategy, then averages over the histories generated under that strategy. Weighting instead reweights observed histories to represent the treatment assignment specified by $g$. Neither approach fixes unmeasured confounding or absent treatment overlap; their modeling demands differ.

The estimate is only as credible as the recorded treatment, covariate timing, and clinical follow-up. Broad labels such as “chemotherapy” can hide important regimen versions, doses, or delays that undermine consistency. Positivity may also fail if, for example, nearly every patient with very poor performance status stops treatment. No weighting calculation can recover survival under a strategy that is essentially never observed for such patients without additional assumptions or a revised, more supportable target strategy.

Common pitfall: Extending a patient’s treatment weight to every interval in the analysis horizon—even after that patient has died—assumes treatment decisions and their probabilities exist after death. They do not under this timeline. Stop the product at $\min(k,T)$, including the death-interval decision, and define risk-set weights only for patients who reached the decision in question.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A patient is alive at the start of cycle 3 and dies during that cycle. For a horizon of five cycles, which treatment-decision factors belong in the weight?
Why can performance status be both a confounder of cycle-2 treatment and a mediator of cycle-1 treatment’s effect?
Under the sustained chemotherapy strategy, a patient’s probability of receiving the prescribed treatment is $0.40$ at cycle 1 and $0.50$ at cycle 2. The patient follows the strategy at both cycles and dies during cycle 2. What is the stopped weight through cycle 4?
What does a very large cumulative weight suggest, and why is truncation not a complete solution?
In the weighted hazard for cycle $j$, why is it appropriate to define the weight through decision $j$ for patients in the risk set?

Solutions. (1) Include the factors for every decision reached through cycle 3, including cycle 3, and none for cycles 4 or 5. (2) Earlier chemotherapy may change performance status, while that status can affect both the next treatment and future death; it is therefore affected by past treatment and predictive of later treatment and outcome. (3) The patient reaches only cycles 1 and 2, so the weight is $(1/0.40)(1/0.50)=2.5\times2=5$. (4) It often indicates a rare treatment choice or weak overlap in part of the observed history. Truncation can reduce variance but can add bias and cannot create information where the strategy is unsupported. (5) Every person in that risk set is alive at the start of cycle $j$ and has reached its treatment decision; the cycle-$j$ factor is therefore defined, including for patients who die during that cycle.