He Named My Health Economics Model After Himself in the NICE Submission Then the Appraisal Panel Required My PSI Accreditation
The survival curve overlay print was lying flat on the R Studio workstation desk.
It was an A4 printout.
The solid lines represented the original Kaplan-Meier survival analysis from the Phase III oncology trial.
The dashed lines represented the adjusted survival curves — corrected for the time-varying confounding caused by disease progression.
The divergence between the solid and dashed curves began at the Year 1 mark.
By Year 3, the difference was stark.
Nadia had drawn a double-headed arrow between the Year 3 survival probability coordinates in red ink.
She had written: “23% HR inflation” in the space between the curves in her small, precise handwriting.
Junior biostatistician Dmitri was sitting at the secondary monitor, sorting through the SAS log output files.
He was twenty-seven.
He had spent three weeks formatting the patient switching logs under her direction — scrubbing the clinical trial database, aligning the cross-over dates, checking the dose adjustment variables.
He was thorough.
“Dmitri,” Nadia said.
Dmitri turned from his screen.
She pointed to the red annotation on the overlay.
“The time-varying confounding by disease progression is the source,” she said.
“Patients in the control arm who progressed switched to the active treatment.”
“The unadjusted model counts their survival outcomes under the control arm.”
“It attributes their extended survival to the control drug.”
“The hazard ratio is artificially inflated.”
“The unadjusted model shows a hazard ratio of 0.62.”
Dmitri looked at the curves.
“The unadjusted ratio would pass the efficacy threshold,” he said.
“Yes,” Nadia said.
“But the unadjusted model is biased.”
“The actual hazard ratio is 0.81.”
“When we apply the inverse probability of treatment weighting, the switching bias is removed.”
“At 0.81, the efficacy is below the threshold for unconditional regulatory approval.”
“It changes the entire regulatory pathway.”
“The drug will require a conditional approval with post-market commitment studies.”
She opened the statistical analysis plan directory.
She logged the final run outputs.
The file carried her PSI accreditation signature code in the header: PSI-AS-NV-8812.
She picked up the survival curve overlay print.
She placed it on the corner of her desk.
It would stay there as the reference document for the Phase III statistical reanalysis — the physical evidence of the 23% correction.
She read the EMA regulatory submission confirmation four weeks later, in her office.
The confirmation document was titled: “Castellano Efficacy Reanalysis — statistical reanalysis conducted under CMO M. Castellano’s oversight.”
She scrolled to page 18.
“Biostatistics support: Dr. Nadia Volkov, PSI-AS-NV-8812.”
She looked at the survival curve overlay print on her desk.
The solid Kaplan-Meier curves.
The dashed IPTW-adjusted curves.
The red “23% HR inflation” text.
She opened the R Studio project file.
She checked the log.
Hazard ratio: 0.81.
Baseline hazard ratio: 0.62.
She closed the project file.
She opened the next statistical design brief on her screen.
Three weeks before the submission was filed, Chief Medical Officer Dr. Marco Castellano had come to her office.
He was fifty-six.
He managed the company’s clinical development portfolio and held the primary regulatory submission authority for the oncology programme.
He had picked up the survival curve print from her desk.
He looked at the solid and dashed lines.
He looked at the red arrow.
He looked at the “23% HR inflation” text.
She said: “The IPTW adjustment is the necessary correction.”
“The unadjusted hazard ratio of 0.62 is biased because forty-eight percent of the patients in the control arm crossed over to the active treatment after disease progression.”
“The EMA will reject the unadjusted analysis if they perform a subgroup survival review.”
“The SAP must specify the marginal structural model to justify the conditional approval pathway.”
“The analysis is certified under my PSI Accredited Statistician registration.”
He held the print.
He said: “This is the corrected picture the EMA needs to see.”
She said: “The SAP carries my PSI registration, PSI-AS-NV-8812.”
“The registration should be cited in the Competent Person declaration for the biostatistics module.”
He said: “Good. Excellent work, Nadia.”
He set the print back on her desk.
He left her office.
She looked at the print.
She noted: “the EMA needs to see.”
The biostatistics had made it visible.
Her biostatistics.
She opened the PSI accreditation portal.
She verified her registration details.
PSI-AS-NV-8812.
Accredited Statistician.
Nadia Volkov.
She closed the browser.
She went back to the IPTW model calculations.
The Phase III trial had enrolled four hundred and eighty patients across twenty-two clinical sites in Europe.
The database had been locked in October.
For the first two weeks after lock, the clinical data had been processed using standard survival models.
The clinical affairs team had celebrated when the initial Kaplan-Meier output showed a hazard ratio of 0.62.
It was a clean, statistically significant result that would support a rapid regulatory filing.
It looked like a blockbuster drug.
Nadia had looked at the treatment crossover files on the third day of analysis.
She had noted that eighty-four patients in the control arm had been switched to the active treatment within twelve days of disease progression.
The protocol allowed crossover for ethical reasons.
But the statistical consequences of that crossover had been ignored in the primary clinical report.
She had spent four days explaining to the clinical operations lead why the 0.62 ratio was biased.
“If we ignore the crossover, we are comparing the active drug against a control group that is actually receiving the active drug for half of their trial duration,” she had said.
“The treatment effect is confounded.”
“The survival benefit is inflated.”
Dmitri had begun the data preparation for the marginal structural model the next morning.
It was tedious work.
They had to construct a person-month dataset — creating twelve thousand rows of data, calculating the time-varying probability of treatment switching for each patient at each month, and estimating the inverse probability weights.
Every weight had to be stabilized to prevent extreme values from distorting the survival estimates.
Dmitri had verified the weights twice.
Nadia had checked the stabilized weight distribution against the trial parameters.
The R Studio workstation had run the final weighted Cox regression for six hours.
The output was the survival curve overlay that now lay on her desk.
The Annual Congress of the European Society for Clinical Pharmacology was held at the Congress Center in Vienna.
Four days.
Three thousand oncology specialists, clinical trialists, and pharmaceutical executives.
Dr. Marco Castellano presented in the afternoon session of the third day.
His presentation was titled: “Correcting Treatment Switching Bias in Phase III Oncology: The Survival Reanalysis.”
His slide 18 was her survival curve overlay print.
The solid Kaplan-Meier curves.
The dashed IPTW-adjusted curves.
The red annotation was visible in the center of the slide — “23% HR inflation” was displayed in bold, black typeface.
He had replaced her handwriting with a marketing agency’s clean sans-serif font, but the data points were unchanged.
He said: “Our statistical reanalysis approach corrected the time-varying treatment switching confounding in the control arm.”
He said: “This methodology produced a corrected hazard ratio of 0.81.”
He said: “This adjustment confirms the therapeutic benefit of the active drug while adjusting for crossover bias.”
He said “our statistical reanalysis approach.”
He did not name the IPTW marginal structural model.
He did not name the PSI Accredited Statistician designation.
He did not name PSI-AS-NV-8812.
He did not name Dr. Nadia Volkov.
In the fourth row, Julian Ward — the company’s Vice President of Regulatory Affairs — watched the presentation.
He had been the one who had submitted the final Phase III reanalysis package to the EMA.
He had circulated the internal summary email: “Castellano’s oncology survival breakthrough.”
He had not looked at the signature block on the Statistical Analysis Plan document in the submission appendix.
He had looked at the CMO’s signature on the main regulatory cover sheet.
He watched the slide and thought: Castellano has saved the filing path.
The EMA contact email arrived on a Wednesday morning at 10:22.
The sender was Dr. Petra Holst, EMA Scientific Committee Lead Statistician.
Subject: “EMA Scientific Committee — Oncology Product Assessment — Statistical Certification Review.”
Nadia opened the email.
“Dr. Volkov — I am writing on behalf of the Committee for Medicinal Products for Human Use (CHMP) regarding the pharmacovigilance review of the oncology submission. The post-market safety and efficacy data show worse survival outcomes in clinical practice than the trial’s submitted hazard ratio of 0.62 projected. The CHMP requires a formal review of the marginal structural model used for the conditional approval reanalysis. Under EMA biostatistics guidelines, all complex Statistical Analysis Plans involving time-varying covariate adjustments must be certified by an accredited statistician. The submission identifies the analysis as ‘Castellano’s Efficacy Reanalysis.’ The analysis metadata contains the PSI registration number PSI-AS-NV-8812. We require: one, verification that PSI-AS-NV-8812 is your PSI accreditation; two, the raw SAS/R logs for the IPTW weights generation; three, your attendance at the Scientific Committee review hearing in London. Please respond within 48 hours.”
She read “PSI registration number PSI-AS-NV-8812.”
She read “EMA biostatistics guidelines.”
She read “accrued statistician certification.”
She looked at the survival curve overlay print on her desk.
The solid KM curves.
The dashed IPTW-adjusted curves.
The red “23% HR inflation” annotation.
She picked up the print.
She held it to the light.
She looked at the red text.
She set it down.
She did not call Castellano.
She opened a reply to Dr. Holst.
She confirmed that PSI-AS-NV-8812 was her registration.
She confirmed her availability for the EMA Scientific Committee review hearing.
She wrote: “I am preparing the full Statistical Analysis Plan, the R code scripts for the inverse probability weighting, and the model diagnostic logs for the committee file.”
She sent the email.
She began preparing the documentation package.
The R code scripts for the stabilized weights — four separate model scripts.
The patient-month covariate dataset files — three hundred megabytes of time-to-event logs.
The model diagnostic outputs — showing the mean stabilized weights were close to 1.0, verifying the model specifications.
Her PSI Accredited Statistician certificate — showing PSI-AS-NV-8812, valid, with the PSI stamp.
The compilation took four hours.
She sent the complete documentation package to Dr. Holst.
She returned to her next project — a Phase II immunotherapy trial in kidney cancer, a different patient population, a different covariate structure.
Castellano received the formal EMA CHMP review notice in his office that afternoon.
The email was marked high priority by the regulatory affairs department.
He read the heading: “EMA CHMP Notification — Pharmacovigilance Review — Statistical Methodology Audit.”
He read the request for the certified Statistical Analysis Plan.
He called Julian Ward.
Julian was direct.
“The EMA guidelines are strict. A clinical trial SAP involving marginal structural models must be certified by an accredited statistician. Your medical degree covers clinical design, not mathematical statistics. You cannot be examined on the stabilization parameters or the time-varying confounder calculations. You do not hold the PSI credential.”
Castellano said: “Has Nadia been notified?”
“The EMA statistician emailed her this morning,” Julian said.
“She has already responded. She confirmed her registration. She sent the complete model scripts and weight calibration datasets directly to the CHMP. She did not contact regulatory affairs before submitting.”
Castellano did not say anything for a moment.
He looked at the EMA submission confirmation folder on his desk.
“Castellano Efficacy Reanalysis.”
He looked at the CHMP notification letter.
“PSI credential required.”
The Vienna presentation hall had been crowded with five hundred delegates during his session.
The slides had been prepared by the media agency in Basel — dark grey backdrops, gold lines for the survival curves, the 23% figure highlighted in a bright amber circle.
Julian Ward had spent the hour after the session distributing the print summary booklets to the research directors who had gathered at the company’s reception desk in the foyer.
The reception desk had been packed.
The clinical investigators had been asking about the covariate models.
Julian had handed them the booklet, pointing to the statistical design page.
“The entire correction is detailed in the Castellano reanalysis,” Julian had said.
“Dr. Castellano resolved the crossover bias that was masking the efficacy.”
“The booklet explains the regulatory submission framework.”
Julian had not read the statistical analysis plan that was archived in the clinical trials directory on the server.
He had not read the certification sheet where PSI-AS-NV-8812 was stamped in black ink.
He had not read the name: Dr. Nadia Volkov.
He sat in his office on the fifth floor of the Basel headquarters building until ten o’clock that night.
The corridors were empty.
The medical affairs staff had left at six.
The clinical development team had gone home.
Only the security guard remained on the ground floor, his footsteps echoing faintly in the atrium far below his door.
He had the EMA CHMP notification letter on his desk.
Beside it lay the printout of the regulatory submission.
The survival curve overlay was printed on page 18 — the solid and dashed lines showing the 23% correction.
He had highlighted the Year 3 divergence with a yellow marker when he first reviewed the filing copy.
He had been Chief Medical Officer for nine years.
He had signed the regulatory submissions for four separate oncology trials.
Every submission had been compiled under his direction.
That was the structure — the CMO signs the clinical efficacy summary.
He was the CMO.
He had understood that authority clearly for nine years, and he had never looked at it closely enough to see the statistical boundary that had now appeared in front of him.
He could explain the clinical rationale.
He could explain the patient enrollment criteria.
He could explain the safety profile — the adverse events, the grade 3 toxicities, the dosing interruptions.
He had read enough statistical summaries to speak confidently about Cox proportional hazards and progression-free survival in front of regulatory advisory panels.
He could not explain the inverse probability weighting diagnostics.
If the EMA panel asked him: how did you adjust for the time-varying selection bias in the treatment crossover arm?
He would have no answer.
If they asked: what was the specification of the logistic regression used to generate the monthly treatment assignment probabilities?
He would have no answer.
If they asked: why is the stabilized weight mean of 0.99 the correct verification of the model parameters, and how did you handle patients with extreme weights?
He had no answer.
He could not defend IPTW marginal structural model calculations he did not conduct.
PSI-AS-NV-8812 was on the original Statistical Analysis Plan.
PSI-AS-NV-8812 was Dr. Nadia Volkov’s registration.
His MD was a medical credential.
He was not a PSI-accredited statistician.
He was a medical officer who managed clinical trials.
He looked at the submission folder.
“Castellano Efficacy Reanalysis.”
His name at the top.
Her name in the appendix.
“Biostatistics support: Dr. Nadia Volkov, PSI-AS-NV-8812.”
He had not written the appendix.
His regulatory affairs team had formatted the submission from his drafts.
He had reviewed it.
He had signed the submission cover sheet.
He had read “Castellano Efficacy Reanalysis” and understood it as the project’s output — the analytical tool his biostatistics group had deployed — not as a statement about who had designed the marginal structural model.
He had been wrong to understand it that way.
He had not known he was understanding it wrongly until this night.
The specific moment was a Tuesday afternoon in March, eighteen months ago.
He had been in his office at his desk.
He had the survival curve print in his hand — she had just placed it on his desk.
She had said: “The IPTW adjustment is the basis of the conditional approval path.”
He had said: “This is the corrected picture the EMA needs to see.”
He had been looking at the dashed lines.
He had been calculating — without calculating formally — how the corrected hazard ratio of 0.81 would support the conditional filing, how the post-market commitment studies would be structured, how the Phase III trial’s safety data would be integrated into the new pathway.
He had understood “the corrected picture the EMA needs to see” as a regulatory milestone.
He had not understood, in that moment, that identifying the time-varying confounding and the 23% correction — the work that changed the drug’s approval pathway — was not regulatory framework execution.
He had not understood that the biostatistical correction was itself the discovery.
He had never examined whether “the execution was the invention.”
He had never examined whether the IPTW model was the output of his clinical program or the product of her professional expertise.
He had said: “Good. Excellent work, Nadia.”
He had set the print back on her desk.
She had left his office.
He had the biostatistics report in the submission file.
He had scheduled the safety reviews for the next trial stage — twelve months away.
He had not escalated the statistical validation to immediate independent audit.
He had understood “0.81 hazard ratio” as a model parameter.
He had understood it within the context of clinical trial reports.
The EMA pharmacovigilance review had been opened eighteen months after that conversation.
He picked up his phone.
He opened the regulatory affairs draft folder.
He typed her name.
“Statistical Analysis Plan by Dr. Nadia Volkov, PSI Accredited, PSI-AS-NV-8812.”
He was beginning to understand that there had been a specific moment when he could have looked at this directly.
That he had been holding the survival curve print in his hand when that moment arrived.
That he had said “excellent work, Nadia” instead.
The EMA Scientific Committee review hearing was held at the EMA headquarters in Amsterdam on a Thursday morning.
The room was large, bright, and quiet.
Eight chairs on one side of the mahogany table for the CHMP Scientific Committee panel.
Two chairs on the other side for the technical witnesses.
The panel consisted of Dr. Petra Holst as the lead statistician, two CHMP Clinical Assessors, and the committee secretary.
Castellano sat in the row of observer chairs along the wall.
He had said to the panel when the session opened: “Dr. Volkov is the PSI-accredited statistician who designed the IPTW model. The statistical methodology questions are for her.”
He had stepped back.
He sat down.
He did not speak again.
Nadia was at the witness table.
She had brought the survival curve overlay print from her office.
She placed it flat on the table, directly beside the unsigned EMA biostatistics declaration form that the CHMP administrative team had printed from the regulatory submission folder.
The laminated print.
The solid KM curves.
The dashed IPTW-adjusted curves.
The red “23% HR inflation” annotation in her handwriting.
Beside it, the blank signature line on the declaration form.
Dr. Holst looked at both of them.
She said: “Dr. Volkov. For the review record, please confirm the statistical accreditation under which the Phase III oncology trial’s reanalysis was certified.”
“PSI Accredited Statistician,” Nadia said.
“PSI-AS-NV-8812.”
“The PSI accreditation in the pharmaceutical industry requires demonstrated competence in time-to-event statistical analysis, clinical trial reporting, and regulatory biostatistics compliance.”
“The registration is logged in the Statistical Analysis Plan metadata and certified in the report index.”
Dr. Holst wrote in her file.
The committee secretary typed.
Dr. Holst said: “Please explain the statistical mechanism that corrected the 23% hazard ratio inflation bias.”
“The primary trial analysis was biased due to treatment switching,” Nadia said.
“Forty-eight percent of patients in the control arm crossed over to the active treatment after disease progression.”
“Because their post-progression survival was extended by the active drug, the unadjusted Kaplan-Meier curves attribute this survival benefit to the control arm.”
“The treatment effect is underestimated.”
“The unadjusted hazard ratio was 0.62.”
“Our marginal structural model uses inverse probability of treatment weighting to remove this bias.”
“The model assigns time-varying weights to each patient based on their progression status and treatment history.”
“This weighting creates a pseudopopulation where treatment switching is independent of progression.”
“The corrected hazard ratio is 0.81.”
“The 23% difference in the hazard ratio represents the inflation bias resolved by this methodology.”
The first CHMP Clinical Assessor said: “The original submission was certified by Dr. Lars Svensson’s equivalents — in this case, Dr. Marco Castellano under his medical authority. On what basis do you say a medical qualification is insufficient to certify this Statistical Analysis Plan?”
“EMA biostatistics guidelines require the statistician of record to hold the relevant professional accreditation for complex statistical reanalyses,” she said.
“Dr. Castellano is a medical doctor.”
“His medical qualification covers clinical trial design, patient safety, and medical monitoring.”
“It does not cover time-varying covariate adjustment or weight stabilization diagnostics.”
“These methods require calibration of the propensity score models, verification of the positivity assumption, and checking of the stabilized weight distribution.”
“These tasks fall under the discipline of Accredited Statistics.”
“Dr. Castellano did not design the IPTW models.”
“He did not write the R scripts.”
“He does not hold the PSI credential.”
“He cannot certify the biostatistics.”
“The guidelines require the PSI accreditation for this reanalysis.”
The second CHMP Clinical Assessor said: “Can you verify that the inverse probability weights did not violate the positivity assumption?”
“Yes,” Nadia said.
“The positivity assumption requires that the probability of treatment switching remains between zero and one for all patient subgroups.”
“We verified this by plotting the stabilized weight distribution.”
“The mean of the stabilized weights was 0.99 with a range of 0.12 to 3.42.”
“There were no extreme weights that would destabilize the model.”
“The weight distribution logs are detailed in Appendix C of the documentation package.”
“The positivity checks were completed prior to running the final weighted Cox regression.”
The first CHMP Assessor said: “The review panel must confirm if there was any selection bias in the treatment switching probability models. How did you select the covariates for the propensity score?”
“The covariates were selected based on clinical relevance and association with disease progression,” she said.
“We included Eastern Cooperative Oncology Group performance status, tumour burden at baseline, time-varying neutrophil count, and time-varying alkaline phosphatase level.”
“The selection was pre-specified in the Statistical Analysis Plan before the database lock.”
“We did not adjust the covariate selection post-hoc to manipulate the hazard ratio.”
“The covariate models are reproducible.”
“The model scripts are in the database.”
Dr. Holst looked at the overlay.
She looked at the red text.
She looked at the dashed curves.
She said: “Dr. Volkov. Your PSI accreditation and your IPTW SAP methodology are the statistical foundation of this pharmacovigilance review.”
The committee record that the secretary was building read: “PSI Accredited Statistician: Dr. Nadia Volkov, PSI-AS-NV-8812. IPTW marginal structural model. 23% HR inflation correction. SAP certified by Dr. Volkov. Submission amended to name Dr. Volkov.”
There were three other company executives in the room as observers.
The first — a woman in her forties who was the company’s Director of Pharmacovigilance — looked at the laminated survival curve print on the table.
She had seen many EMA reviews.
She had never seen one resolve a treatment switching bias issue this cleanly.
She leaned forward and took a photograph of the print with her phone.
She sat back.
The second observer — a male clinical research director, late fifties — had read the “Castellano Efficacy Reanalysis” when it was submitted.
He had noted “Castellano” as the certifying officer.
He was now looking at Dr. Nadia Volkov answering the biostatistics questions.
He wrote the PSI-AS-NV-8812 registration number in his notebook.
Julian Ward, the VP of Regulatory Affairs, was in the observer row next to Castellano.
He had been the one who had distributed the “Castellano breakthrough” booklet.
He was watching Dr. Volkov explain the stabilized weight distribution.
He did not look at Castellano.
After the review hearing closed, Castellano walked out to the taxi stand.
He did not wait for Nadia.
He called her that evening.
“The review outcome is satisfactory,” he said.
“Your IPTW model was the technical basis.”
“I’ve already filed the amended EMA submission — your name and PSI registration going forward.”
“And I’m implementing a company protocol requiring PSI-accredited statistician named authorship on all EMA Statistical Analysis Plans involving complex modelling.”
She said: “The SAP methodology was complete.”
He said: “Yes.”
A moment of silence.
He said: “Excellent work, Nadia.”
She said: “Yes.”
She set the print on her desk.
She opened the next project file.
The second time he had said “excellent work.”
The first CHMP Assessor reviewed the methodology report.
He looked at the censoring parameters.
“Dr. Volkov,” he said.
“The EMA guidelines require full disclosure of any informative censoring bias.”
“Did the treatment switching create any differential drop-out rates between the trial arms?”
“No,” Nadia said.
“We adjusted for informative censoring by including censoring weights in the stabilized weight model.”
“The probability of remaining uncensored at each month was estimated using a pooled logistic regression model.”
“The covariates in the censoring model matched the treatment switching models.”
“This adjustment ensures that the weighted survival curves are unbiased by differential drop-out rates.”
“The censoring weight logs were verified by the independent biostatistics audit.”
“The statistical proof is in Appendix D of the submission file.”
He nodded.
He noted the response on his sheet.
The secretary’s typing was a steady clicking in the background.
Dr. Holst said: “The EMA product assessment record will be updated to include the following statement: ‘The statistical reanalysis described in the efficacy assessment is based on, and fairly represents, the Statistical Analysis Plan certified by Dr. Nadia Volkov, a PSI Accredited Statistician.'”
“Is that statement acceptable to you, Dr. Volkov?”
“Yes,” Nadia said.
“That statement is correct.”
Six weeks after the review hearing, she was back in her biostatistics office.
It was a new project — a Phase III immunotherapy trial in kidney cancer, a different patient population, a different treatment crossover rule, a different time-varying covariate structure.
Dmitri was at the data preparation station, organizing the new trial’s time-to-event dataset — cleaning the patient files, setting up the censoring matrices, checking the baseline demographic tables.
The survival curve overlay print was on the corner of her desk.
She had not moved it since she had returned from Amsterdam.
She had used it every morning.
She picked up the print and held it to the light — the solid Kaplan-Meier curves, the dashed IPTW-adjusted curves, the red “23% HR inflation” annotation.
She used it as a calibration reference: comparing the new trial’s treatment switching pattern against the inverse probability weighting approach she had developed for the previous oncology model, confirming that the methodology transfer was appropriate.
The new trial’s protocol allowed crossover at progression, but the timing was different.
The smectite-equivalent statistical traps were different.
But the weighting logic was the same.
The propensity scores.
The stabilized weights.
The Cox regression.
The EMA CHMP pharmacovigilance record was filed in the regulatory database.
“PSI Accredited Statistician: Dr. Nadia Volkov, PSI-AS-NV-8812. IPTW marginal structural model. 23% HR inflation correction. SAP certified by Dr. Volkov. Submission amended to name Dr. Volkov as certifying biostatistician. EMA CHMP Scientific Committee pharmacovigilance audit completed. Efficacy reanalysis certified under PSI accreditation rules.”
That record was permanent.
Dmitri said, from the data station: “PSI-AS-NV-8812 is in the CHMP record.”
She said: “Yes.”
Dmitri said: “The 23%.”
She said: “The 23% correction is in the record.”
Dmitri looked at the print.
He went back to the dataset formatting.
The original EMA public product assessment report was still on the EMA website — a ninety-two-page PDF document detailing the conditional approval decision, containing “Castellano’s reanalysis” in the efficacy summary section.
The document had been downloaded by hundreds of medical directors and pharmaceutical analysts since the conditional approval was announced.
It was not recalled.
It remained in the public regulatory archive.
She had the EMA product number: EMA/H/C/005824.
The amended submission was in the CHMP review file.
The public product assessment page had not been changed.
Both were permanent records.
The company’s new JORC-equivalent biostatistics protocol had arrived as a PDF attachment from the CMO’s office: “Protocol STAT-EMA-2024-12: PSI Accredited Statistician certification requirement for all regulatory SAP submissions involving complex marginal structural models.”
She had read it.
She had filed it in the department files.
The new trial brief had arrived from Castellano on a Friday afternoon.
Subject: “Statistical Analysis Plan — Phase III Kidney Cancer — Weighting Models.”
It read: “Statistical Analysis Plan — Dr. Nadia Volkov, PSI lead.”
She had opened it.
She had begun the model design for the hectorite-equivalent statistical check.
Dmitri had the first batch of covariates ready.
He said: “The dataset is loaded.”
She loaded the model scripts.
She looked at the blank axes on her monitor — the space where the survival curves would appear when the R scripts finished running.
She looked at the print in her hand.
She looked at the solid and dashed lines.
The Phase III kidney cancer trial was statistically complex.
The crossover was influenced by both disease progression and time-varying toxicity.
Patients who experienced grade 3 adverse events switched treatment arms earlier than those who progressed without toxicity.
This was not a simple treatment crossover problem.
Toxicity and progression acted as joint time-varying confounders.
If the toxicity was associated with early progression, the crossover rate would be high, but the treatment effect would be masked.
If the toxicity was independent of progression, the adjustment would be straightforward, but the stabilized weights would be highly variable.
She needed to specify these joint propensity models.
She would run 1,200 simulations across the ten main clinical sites to check the model’s stability.
Dmitri would handle the covariate coding, the data preparation, and the weight distributions.
It would take four weeks of model design.
The office was cool.
The computer fans hummed behind her desk, maintaining a steady temperature as the processor handled the simulation runs.
The server connection status indicator flashed green, regulating the data transfer from the clinical trial database in Frankfurt.
She looked at the blank software screen.
The calibration of the propensity scores for the joint confounders was the critical gate.
Toxicity and progression had overlapping timelines — both occurred post-treatment, but toxicity had distinct acute phases that the model must separate from the chronic progression indicators.
She had spent three days calibrating the covariate definitions using pure clinical indicators from the pilot study’s reference database.
The model profile was saved as “Kidney-Immunotherapy-v1.4.lib.”
She looked at the survival curve overlay print on her desk.
The monazite-equivalent trial had been simple compared to this.
The previous trial had a single crossover trigger, easily adjusted by the progression indicator.
The joint confounders would require complex multi-state models and careful deconvolution of the toxicity-progression ratios.
But the physical reality remained.
The survival curves on the screen were a reflection of the actual treatment effect.
The trial data did not know who had signed the EMA declaration.
The survival probability would decrease according to the biological response, regardless of whether Castellano’s name or her name was printed on the EMA submission.
The hazard ratio was locked in the clinical trial database.
She looked at the print.
The solid KM curves.
The dashed IPTW-adjusted curves.
She placed the overlay on the desk.
She pointed to the dashed curves.
