Surgeons believe that some operations spill live cancer cells, which then implant and regrow. The evidence for this in kidney and ureter cancer is strong on paper. Laparoscopic removal of the kidney and ureter, diagnostic ureteroscopy before that surgery, and removing the end of the ureter through the urethra are each linked to more tumors reappearing in the bladder afterward. Almost all of that evidence, however, comes from non-randomized comparisons. The patients who received the "risky" manipulation may simply have had more aggressive or more widespread disease. We propose a way to separate these two explanations with data that already exist. The approach borrows the logic Mendelian randomization uses to catch invalid genetic instruments. Mendelian randomization uses inherited genetic variants as natural experiments to estimate causal effects.
The central idea is anatomical. If a manipulation causes harm by flushing cells downstream in the urine, it should raise recurrence only where that urine goes, the bladder. It should not raise recurrence at sites the urine never reaches. If the association instead reflects which patients were chosen for the manipulation, recurrence should rise at those other sites too. We treat recurrence at those sites as negative-control outcomes. We then pool the resulting "downstream minus control" contrasts across several different manipulations, each chosen for different clinical reasons, using the robust pooling estimators developed for Mendelian randomization. The result is a seeding estimate that stays valid even if some of the manipulations are badly confounded.
Core Claim
In upper tract urothelial carcinoma (UTUC; cancer of the lining of the kidney pelvis and ureter), the excess intravesical recurrence (IVR; new tumors in the bladder after surgery) attributed to seeding can be tested against two specific confounding stories.
The first is confounding by tumor aggressiveness. Sicker patients get more manipulation and do worse everywhere.
The second is urothelial field effect. Patients whose whole urinary lining is prone to cancer are both more likely to be investigated endoscopically and more likely to grow new tumors anywhere in that lining.
Each story has its own negative control. Distant metastasis tracks aggressiveness but cannot be reached by urine flow. Recurrence in the opposite, untouched upper tract tracks field effect, but it sits upstream of the bladder and outside the manipulated lumen. A real seeding effect predicts a positive downstream contrast against both controls. It also predicts that the contrast agrees across manipulations whose selection mechanisms differ. Neither confounding story predicts both patterns.
The mapping to Mendelian randomization sensitivity analysis is deliberate:
- Each manipulation plays the role of a genetic instrument.
- A confounding pathway that reaches recurrence without passing through cell spillage plays the role of horizontal pleiotropy, the case where a variant affects the outcome through a route other than the exposure.
- The weighted median estimator stays consistent if fewer than half of the instruments (by weight) are invalid. Here it becomes a way to pool manipulations so that a minority with unusual selection bias cannot drive the answer.
Where This Sits Relative to Prior Work
The motivating evidence comes from three sources:
- Tsuboi et al.'s meta-analysis pooled 31 studies and 20,048 patients. It reported laparoscopic versus open removal of kidney and ureter at HR 1.28 for bladder recurrence, early ureter ligation at HR 0.64, and bladder cuff excision through the urethra versus from inside the bladder at HR 1.64. It relied on 29 retrospective studies and made no adjustment for unmeasured confounding.
- The CROES-UTUC ureteroscopy study found pre-operative ureteroscopy associated with IVR at HR 1.705 (95% CI 1.082–2.688) in 269 patients. Two-year IVR-free survival was 52% with ureteroscopy versus 64% without. Its multivariable Cox model cannot address why those patients were scoped.
- Work on onco-surgical principle violations links containment breaches to worse survival in the same disease, with the same observational limits.
We reuse the effect estimates and cohorts from these studies. The step none of them takes is testing whether the association is anatomically specific.
The research file's section 2 identifies the obvious intersection: Mendelian randomization of host inflammatory or protein exposures on post-surgical recurrence. We rejected it, and one paper constrains that choice decisively. Yang et al. found across seven biobanks that none of 888 known susceptibility variants predicted post-diagnosis survival. Direction agreement between susceptibility and survival effects was 54%, essentially chance. They also showed that selecting only people who already have the disease distorts genetic associations, a problem called index-event bias.
A germline Mendelian randomization study of seeding would therefore need a genome-wide association study of bladder recurrence after nephroureterectomy. None is reported, and it would be small, weakly heritable, and exposed to index-event bias. MRdb offers 27,000+ exposures and 48,507 outcomes, but it carries no post-surgical seeding endpoint. The only paper in the file that places Mendelian randomization next to minimally invasive surgery, Kang et al.'s editorial, keeps the two strands separate and applies Mendelian randomization only to cancer incidence.
We therefore take the methods of Mendelian randomization sensitivity analysis, not its genetic data:
- the weighted median, weighted mode, and Egger-type intercept, as used in applied studies such as the metformin drug-target analysis;
- the insistence of Korf et al. on naming the estimand and its assumptions explicitly.
Finally, the randomized SEIPLUS lavage trial supplies a calibration setting. There, a randomized seeding-directed intervention produced no significant change in peritoneal recurrence-free survival (70.0% vs 66.6%, p = 0.54).
Research Program
Estimand. For manipulation $j$, the target is the seeding contrast
$$\delta_j = \log HR_j^{\text{bladder}} - \log HR_j^{\text{control}}$$
In words, $\log HR_j^{\text{bladder}}$ is the log hazard ratio of the manipulation for bladder recurrence. $\log HR_j^{\text{control}}$ is its log hazard ratio for a negative-control outcome, either distant metastasis or contralateral upper tract recurrence.
Identifying assumption. $\delta_j$ equals the causal seeding effect under one condition, which is the analog of the exclusion restriction. Any confounder must shift the bladder hazard and the control hazard by the same log amount.
How violations bias the estimate. Violations run in two directions:
- Toward the null (the conservative direction). A manipulation might truly spread cells to distant sites, for example through fluid backflow into veins during ureteroscopy. The contrast then understates seeding, which only makes a positive finding harder to obtain.
- Toward a false positive (the dangerous direction). A confounder might shift the bladder hazard more than the control hazard. This is exactly what field effect does to the distant-metastasis control, and it is why the second, contralateral control exists.
Pooling across manipulations. Across the $J$ manipulations, we pool the $\delta_j$ in three ways:
- Inverse-variance weighting, as the analog of the standard inverse-variance weighted Mendelian randomization estimator.
- A weighted median, which stays valid if manipulations carrying less than half the total weight have their own bias. The manipulations are ureteroscopy, laparoscopic approach, bladder cuff route, late ureter ligation, and extravesical cuff. They are selected for different reasons: diagnostic uncertainty, surgeon preference and body habitus, institutional habit, and training.
- An Egger-style regression of $\log HR_j^{\text{bladder}}$ on $\log HR_j^{\text{control}}$. A nonzero intercept indicates a seeding effect common to all manipulations. A slope different from one indicates that confounding reaches the two outcomes unequally.
With roughly five manipulations the Egger fit will be imprecise. We report it as a diagnostic, not as the primary estimate.
Mechanistic dose ordering. Before any outcome data are opened, we pre-register an ordering of the manipulations by expected urinary spillage. The ordering draws on the mechanistic literature, including the in vitro floater work showing that cells shed into saline stay viable. A seeding explanation predicts that $\delta_j$ rises monotonically along this ordering. Confounding has no reason to.
Data. Two data sources feed the analysis:
- Reconstructed patient-level data. We rebuild individual patient data from the published Kaplan–Meier curves (plots of survival over time) for all three outcomes, using the reconstruction approach Tsuboi et al. used for bladder recurrence. This works for the subset of the 31 studies that report distant metastasis-free and contralateral upper tract recurrence curves.
- Registry data. We re-analyze CROES-UTUC at the patient level, with the ureteroscopy contrast computed separately for each outcome.
Simulation. A simulation study will generate cohorts with known seeding effects, aggressiveness confounding, field-effect confounding, and a fraction of manipulations carrying manipulation-specific bias. This measures bias and confidence-interval coverage for every estimator before we trust any of them on real data.
Baselines
Naive reference. The first reference point is the naive pooled association reported in the literature. For the laparoscopic contrast, this is the meta-analytic bladder recurrence HR of 1.28 (95% CI 1.06–1.54) from Tsuboi et al.. For ureteroscopy, it is the CROES multivariable HR of 1.705 from the ureteroscopy study. We recompute both on exactly the reconstructed and registry data we use, so that differences reflect the method and not the sample.
Strongest standard method. The strongest conventional competitor is doubly robust adjustment on measured covariates. We use inverse probability of treatment weighting (reweighting patients so the treated and untreated groups look alike on measured factors) and targeted maximum likelihood estimation. We report both in the transparent format of Korf et al. and give them the same covariates, the same patients, and the same model-tuning budget as our estimators. In a benchmark where no unmeasured confounding was plausible, Korf et al. report these estimators agreeing closely, with average treatment effects of 3.318 to 3.448. That agreement is exactly what we expect to break down under unmeasured confounding.
What counts as beating the baselines. No reported result was found on Lacuna for any negative-control recurrence outcome in UTUC. So "beating" is defined by the simulation and by agreement with randomized evidence, not by matching a published number. Our estimator wins in simulation if, under confounding that inverse probability weighting cannot see, it keeps 95% interval coverage near nominal while the baseline's coverage collapses. It wins on real data if its estimate shifts relative to the adjusted HR in the direction the confounding story predicts, and its diagnostics show why.
Ablations. Two ablations isolate the key ingredients:
- The per-manipulation contrast $\delta_j$ without cross-manipulation pooling tests what robust pooling adds.
- Robust pooling of the raw bladder recurrence HRs without negative-control subtraction tests what the anatomical control adds.
Center-preference instrument. We also run an instrumental-variable analysis that uses each CROES center's ureteroscopy rate as the instrument. This baseline is described further below.
Why not just…?
Why not just run a Mendelian randomization study of IL-6 or other host proteins on post-surgical recurrence? No recurrence-after-nephroureterectomy genome-wide association study is reported in the file. The largest relevant evidence, Yang et al., shows post-diagnosis outcomes are nearly uncorrelated with susceptibility genetics and are exposed to index-event bias. Even if such a study existed, it would answer a different question, namely which patients make good soil for implanted cells. It would not tell us whether a given surgical act spills seeds. We treat host-factor Mendelian randomization as a follow-up only if a large post-surgical genome-wide study appears.
Why not just adjust better, with propensity scores plus a sensitivity bound for unmeasured confounding? Propensity methods balance only measured covariates. Field effect and occult aggressiveness are rarely measured well in registries. A generic sensitivity bound tells you how strong a hidden confounder would need to be. It cannot tell you whether one exists. Our negative controls provide empirical evidence on that question. The simulation will show whether adjustment alone recovers the truth when field-effect confounding is present; we expect it will not.
Why not just use one negative-control outcome and skip the pooling machinery? This is a real threat. A single well-chosen control applied to ureteroscopy might carry most of the value. The first ablation tests exactly this. If pooling adds no robustness in simulation when some manipulations carry their own bias, and no change in real estimates, we will report the simpler two-outcome design as the recommendation. We will not keep the heavier method for its own sake.
Why not just use surgeon or center preference as an instrument? This is a credible competitor and is included as a baseline. Its weakness is that center preference is correlated with case mix, referral patterns, and surgical skill, all of which affect recurrence directly. That is the same exclusion-restriction problem, only harder to check. With 29 centers, the instrument may also be weak. If the preference-instrument estimate and our negative-control estimate agree, each strengthens the other. If they disagree, the negative controls let us check which estimate's assumptions fail visibly.
Why not just wait for randomized trials? Randomizing patients to known-hazardous choices, such as omitting early ligation, is not acceptable. Also, SEIPLUS shows that even a completed 662-patient trial can leave the seeding question unresolved.
Proposed Experiments and Decisive Checks
Simulation. The first experiment is the simulation study described above. We vary the seeding effect, the strength of aggressiveness and field-effect confounding, the fraction of manipulations that carry their own bias, and cohort sizes matched to CROES (about 270 patients) and to the reconstructed meta-analysis. For every estimator we report bias, root mean squared error, and interval coverage.
Real data. The second experiment applies all estimators to the CROES ureteroscopy contrast and to the reconstructed multi-study data for each manipulation. We report per-manipulation contrasts against both controls, the pooled weighted-median and inverse-variance-weighted estimates, the Egger-style intercept and slope, and a test of the pre-registered dose ordering.
Decisive outcomes. Three results would each settle the question:
- Seeding is supported if contrasts against both controls are positive, agree under weighted-median pooling, and rise with the pre-registered spillage ordering.
- Confounding by aggressiveness is supported if the distant-metastasis HRs match the bladder recurrence HRs, so that the contrasts collapse to zero.
- Field effect is supported if the contrast survives the distant-metastasis control but vanishes against the contralateral upper tract control.
Calibration against randomized data. Through collaboration with the SEIPLUS investigators, we compute the analogous peritoneal-versus-distant contrast on the randomized arms. Because randomization removes confounding, the contrast should reproduce the trial's null. If our method manufactures a seeding effect where randomization shows none, the method fails.
Pre-specified failure condition. The approach fails if too few studies report negative-control curves to reconstruct them. We will know this after the reconstruction step and before any estimation.
References
- Differential effect of surgical technique on intravesical recurrence after radical nephroureterectomy in patients with upper tract urothelial cancer: a systematic review and Meta-analysis
- Impact of Pre-Operative Ureteroscopy on Bladder Recurrence Following Nephroureterectomy for UTUC
- Violation of onco-surgical principles is associated with survival outcomes in upper tract urothelial carcinomas after radical nephroureterectomy
- Limited overlap between genetic effects on disease susceptibility and disease survival
- MRdb: a comprehensive database of univariate and multivariate Mendelian randomization with large-scale GWAS summary data
- Editorial: Minimally invasive therapies and molecular mechanisms related to recurrence in hepatobiliary and gastric cancers
- Evaluating the Impact of Putative Metformin Targets on Cancer Outcomes: A Drug‐Target Mendelian Randomization Study
- Causal clarity in statistical software
- Three-year outcomes of the randomized phase III SEIPLUS trial of extensive intraoperative peritoneal lavage for locally advanced gastric cancer
- In vitro assessment of intra-operative and post-operative environment in reducing bladder cancer recurrence