← Clinical Trials
Aggressive Non-Hodgkin Lymphoma Phase 1/2 Completed NCT02348216

ZUMA-1: Complete Statistical Analysis of Axicabtagene Ciloleucel in Refractory Aggressive Non-Hodgkin Lymphoma

An independent statistical analysis of the phase 1/2 ZUMA-1 study evaluating axicabtagene ciloleucel in adult participants with refractory or relapsed aggressive non-Hodgkin lymphoma, including registered efficacy and safety endpoints and the statistical framework appropriate for their interpretation.

Trial start: April 21, 2015  ·  Primary completion: July 27, 2023  ·  Sponsor: Kite, A Gilead Company
Scope of this record

This page separates reported registry information from statistical interpretation. The ClinicalTrials.gov record identifies six registered primary endpoints and indicate that results are posted, but contain no formal statistical analyses. Numerical treatment-effect estimates, confidence intervals, and p-values are therefore not presented here unless they are explicitly contained in the ClinicalTrials.gov record.

Registry note: This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record.

1. Trial at a Glance

ZUMA-1 was a non-randomized, sequential phase 1/2 treatment study evaluating axicabtagene ciloleucel in adult participants with refractory or relapsed aggressive non-Hodgkin lymphoma. The registry lists 307 participants enrolled, seven arms, no masking, and treatment as the primary purpose.

307
Enrollment
Registry enrollment
7
Arms
Sequential design
1/2
Phase
Phase 1/2
6
Primary endpoints
All binary
FeatureZUMA-1
Trial nameZUMA-1
ClinicalTrials.gov identifierNCT02348216
Therapeutic areaHematology
PhasePhase 1/2
StatusCOMPLETED
AllocationNON_RANDOMIZED
Design modelSEQUENTIAL
MaskingNONE
Primary purposeTREATMENT
Enrollment307.0
Arms7
Lead sponsorKite, A Gilead Company
Sponsor typeINDUSTRY

2. Clinical Question

The registry describes ZUMA-1 as a study evaluating the safety and efficacy of KTE-C19, now represented in the registry intervention list as axicabtagene ciloleucel, in adult participants with refractory or relapsed aggressive non-Hodgkin lymphoma. The statistical structure is different from a conventional randomized comparative trial because allocation is explicitly recorded as non-randomized and the design model as sequential.

Population

Adult participants with refractory diffuse large B cell lymphoma, relapsed diffuse large B-cell lymphoma, transformed follicular lymphoma, primary mediastinal B-cell lymphoma, or high grade B-cell lymphoma.

Intervention

Axicabtagene ciloleucel, with the registry also listing fludarabine, cyclophosphamide, levetiracetam, tocilizumab, dexamethasone, high-dose methylprednisolone, bendamustine, rituximab, doxorubicin, prednisone, vincristine, and ifosfamide among the interventions.

Comparator

No randomized comparator is identified in the ClinicalTrials.gov record. Allocation is recorded as NON_RANDOMIZED.

Primary question

How safe and effective is axicabtagene ciloleucel in the study populations represented by the sequential phase 1/2 program?

3. Trial Design

01
Enrollment307 participants
02
Sequential studyNon-randomized allocation
03
AxicabtageneStudy intervention
04
AssessmentEfficacy and safety endpoints
05
Follow-upPrimary endpoints through registry-defined time frames
Allocation
NON_RANDOMIZED. The registry does not describe randomized assignment between treatment groups.
Design model
SEQUENTIAL. The study is structured as a phase 1/2 program with distinct cohorts and safety-management components.
Masking
NONE. The ClinicalTrials.gov record identifies the study as unmasked.
Primary purpose
TREATMENT. The study evaluates treatment safety and efficacy.
PHASE 1

Phase 1 study

  • Axicabtagene ciloleucel is the identified study intervention.
  • The registered primary endpoint concerns adverse events defined as dose-limiting toxicities.
  • The registry time frame begins at the first infusion date of axicabtagene ciloleucel.
PHASE 2

Pivotal and safety-management cohorts

  • The registry identifies a pivotal study with Cohorts 1 and 2.
  • The safety-management study includes Cohorts 3, 4, 5, and 6.
  • Primary endpoints differ between the pivotal and safety-management components.

The registry lists seven arms overall, but the registry-reported trial extraction does not provide a complete arm-by-arm treatment description or allocation count for all seven arms. The analysis therefore avoids assigning participants to treatment groups that are not explicitly identified in the ClinicalTrials.gov record.

4. Trial Timeline and Follow-Up

2015-04-21

Trial start

The registry lists April 21, 2015 as the study start date.

2023-07-27

Primary completion

The registry lists July 27, 2023 as the primary completion date.

Completed

Current registry status

the ClinicalTrials.gov record in the ClinicalTrials.gov record lists the study status as COMPLETED.

5. Registered Primary Endpoints

ZUMA-1 has six registered primary endpoints in the ClinicalTrials.gov record. All six are classified as binary endpoints in the extraction. The endpoints are distributed across the phase 1 study, the phase 2 pivotal study, and four safety-management cohorts.

Study componentRegistered primary endpointTime frame
Phase 1 Study Number of Participants Experiencing Adverse Events (AEs) Defined as Dose Limiting Toxicities (DLTs) First infusion date of axicabtagene ciloleucel up to 30 days
Phase 2 Pivotal Study (Cohorts 1 and 2) Overall Response Rate (ORR) as Assessed by Investigator Per Revised International Working Group (IWG) Response Criteria for Malignant Lymphoma First infusion date of axicabtagene ciloleucel up to last follow-up visit; maximum duration: 7.7 years
Phase 2 Safety Management Study (Cohort 3) Percentage of Participants With Treatment-Emergent Cytokine Release Syndrome (CRS) and Neurologic Toxicities by Severity Grades First infusion date of axicabtagene ciloleucel up to last follow-up visit; maximum duration: 6.8 years
Phase 2 Safety Management Study (Cohort 4) Percentage of Participants With Treatment-Emergent CRS and Neurologic Toxicities by Severity Grades First infusion date of axicabtagene ciloleucel up to last follow-up visit; maximum duration: 5.4 years
Phase 2 Safety Management Study (Cohort 5) Percentage of Participants With Treatment-Emergent CRS and Neurologic Toxicities by Severity Grades First infusion date of axicabtagene ciloleucel up to last follow-up visit; maximum duration: 4.4 years
Phase 2 Safety Management Study (Cohort 6) Percentage of Participants With Treatment-Emergent CRS and Neurologic Toxicities by Severity Grades First infusion date of axicabtagene ciloleucel up to last follow-up visit; maximum duration: 4.1 years

Endpoint definitions

The phase 1 DLT endpoint is defined using axicabtagene ciloleucel-related events with onset within the first 30 days following infusion. The registry definition includes grade 4 neutropenia lasting > 21 days, grade 4 thrombocytopenia lasting > 35 days from day of cell transfer, any axicabtagene ciloleucel-related adverse event requiring intubation, and other specified grade 3 or grade 4 toxicities.

For the phase 2 pivotal component, ORR is defined as the percentage of participants achieving either a complete response or partial response, assessed by study investigators using revised IWG Response Criteria for Malignant Lymphoma. The registry-reported definition describes complete response as complete disappearance of all detectable clinical evidence of disease and disease-related symptoms, with required regression of lymph nodes and nodal masses.

For the safety-management cohorts, treatment-emergent adverse events are defined as adverse events with onset on or after the start of treatment. CRS is graded using the Lee et al 2014 framework. The registry definition distinguishes grades 1 through 4 according to increasing intervention requirements and severity, with grade 4 described as life-threatening symptoms and requirements for ventilator support.

6. Results Status

the ClinicalTrials.gov record states that results are posted and that 28 outcome measures are posted. However, the registry-reported extraction contains no formal statistical analyses. That distinction is important: an outcome can have posted registry data without the registry extraction supplying a formal between-group hypothesis test, effect estimate, confidence interval, or p-value.

Numerical-results rule: No numerical ORR estimate, DLT comparison, CRS estimate, neurologic-toxicity estimate, confidence interval, or p-value is included in the posted results.

Because the design is non-randomized and the registry-reported statistical-analysis field is empty, conventional comparative quantities such as a randomized-treatment hazard ratio or randomized-group risk ratio cannot be presented as if they were documented analyses of this registry record.

7. Planned Analysis

The registry identifies six binary primary endpoints. For binary outcomes, the central statistical task is to estimate the proportion of participants experiencing the prespecified event and to quantify uncertainty around that proportion. In a non-randomized sequential study, those estimates are primarily descriptive unless the protocol specifies a formal benchmark, historical control, or other comparison.

Phase 1 dose-limiting toxicities

The phase 1 primary endpoint is the number of participants experiencing adverse events defined as DLTs during the first 30 days after the first axicabtagene ciloleucel infusion. A typical analysis would summarize the number and percentage of participants meeting the DLT definition, with an exact or otherwise appropriate binomial confidence interval when inferential precision is desired.

Binary endpoint framework
p = x / n

where x is the number of participants meeting the endpoint definition and n is the relevant analysis population. The registry extraction in the ClinicalTrials.gov record does not provide the corresponding formal statistical-analysis estimate.

Overall response rate

ORR is a binary response endpoint because each participant is classified according to whether a complete response or partial response was achieved under the registered response criteria. The natural descriptive estimand is therefore a response proportion. If a formal comparison were specified, the analysis would depend on the design and prespecified comparator; because allocation is non-randomized and no formal statistical analysis is reported here, a randomized treatment-effect interpretation would not be appropriate.

CRS and neurologic toxicities

The four safety-management primary endpoints concern the percentage of participants with treatment-emergent CRS and neurologic toxicities by severity grade. The statistical description should preserve the ordinal severity information rather than collapsing all grades into a single binary outcome when the purpose is to understand the safety profile. Grade-specific counts and percentages can show how the distribution of toxicity changes across severity categories.

What can be estimated

For each binary endpoint, the observed event proportion can be calculated from the appropriate analysis population. Confidence intervals can quantify sampling uncertainty around those proportions.

What cannot be assumed

A non-randomized cohort proportion is not automatically a causal treatment effect. Without a randomized comparator or prespecified external benchmark, differences across cohorts can reflect differences in study populations or procedures.

8. Statistical Methodology

Binary endpoint analysis

Each registered primary endpoint is classified as binary in the ClinicalTrials.gov record. A binary endpoint records whether an individual participant meets a predefined condition, such as experiencing a DLT or achieving an overall response. The basic estimand is the event probability within the defined analysis population.

Observed proportion
p̂ = x / n

The observed proportion is straightforward, but its interpretation depends critically on the denominator. In a sequential study with different cohorts and different follow-up windows, the denominator and analysis population must be stated explicitly.

Confidence intervals for proportions

For a binary endpoint, a confidence interval around the observed proportion communicates statistical precision. Exact binomial intervals are one common approach, particularly when event counts are small. The registry extraction does not identify a particular confidence-interval method, so no specific interval procedure is attributed to the ZUMA-1 protocol on this page.

Response-rate analysis

ORR combines complete response and partial response into a single binary efficacy endpoint. Statistically, that simplifies the endpoint to responder versus non-responder, but the clinical meaning still depends on the underlying response criteria. The registry specifically identifies investigator assessment according to revised IWG Response Criteria for Malignant Lymphoma.

Safety-event analysis

Safety endpoints require careful attention to exposure and observation time. A treatment-emergent event is defined in the registry as an adverse event with onset on or after the start of treatment. CRS and neurologic toxicities are further characterized by severity grade. A useful safety analysis therefore reports both how often events occur and how severe they are.

Non-randomized interpretation

Randomization creates a design-based mechanism for balancing measured and unmeasured prognostic factors in expectation. ZUMA-1 is explicitly classified as non-randomized. Consequently, an observed difference between sequential cohorts cannot automatically be interpreted as the causal effect of axicabtagene ciloleucel in the same way as a difference generated by randomized treatment assignment.

Sequential design

A sequential design can allow information from earlier stages or cohorts to inform subsequent study development or safety management. Statistically, however, sequential observation creates additional opportunities for selection, changing eligibility, changing management procedures, and evolving event ascertainment. Those features need to be understood before combining cohort-level estimates into a single summary.

9. Statistical Methods Explained

Why is ORR a binary endpoint?

ORR classifies each participant as having achieved either a complete response or a partial response, versus not achieving either response category. This creates a proportion that is easy to summarize, but it discards the additional information contained in the distinction between complete and partial response and in the timing and duration of response.

Why does the non-randomized design matter?

In a randomized trial, treatment groups are created by the randomization mechanism. Here, the registry explicitly records allocation as non-randomized. Therefore, if two cohorts have different response proportions, the difference can have several explanations besides treatment itself, including differences in participant characteristics, disease status, enrollment period, or clinical management.

Why are confidence intervals important for a response rate?

An observed response percentage is an estimate of an underlying event probability. A confidence interval communicates how precisely that probability has been estimated under the selected statistical model. A wider interval indicates greater uncertainty; a narrow interval indicates greater statistical precision. Neither interval describes the response probability of every individual participant.

Why should CRS be reported by severity grade?

Severity contains information that a simple yes/no indicator loses. Two cohorts could have the same overall incidence of CRS while having very different distributions of grade 1, grade 2, grade 3, and grade 4 events. The registered endpoint therefore appropriately specifies toxicity by severity grades.

Why must the analysis denominator be explicit?

A percentage has meaning only relative to its denominator. ZUMA-1 contains multiple study components and cohorts, while the primary endpoints use different populations and follow-up definitions. Combining denominators across cohorts without a prespecified rationale could produce a number that does not correspond to any registered endpoint.

Can a posted registry result be treated as a formal hypothesis test?

No. Posting an outcome measure and supplying a formal statistical comparison are different things. The ClinicalTrials.gov record say that results are posted, but no formal statistical analyses were posted. Consequently, this page does not infer a p-value, confidence interval, or comparative effect estimate from the mere existence of posted outcome data.

10. Safety Results

Study component / cohortParticipants with serious adverse events / at risk
Phase 1 Study: Axicabtagene Ciloleucel5/7
Phase 2 (Pivotal Study): Cohort 140/77
Phase 2 (Pivotal Study): Cohort 214/24
Phase 2 (Safety Management Study): Cohort 326/38
Phase 2 (Safety Management Study): Cohort 424/41
Phase 2 (Safety Management Study): Cohort 527/50
Safety denominator caution: These serious-adverse-event values are cohort-level affected/at-risk counts, not randomized treatment comparisons. The registry-reported final safety-management Cohort 6 field is incomplete and therefore is not assigned a numerical value on this page.

The safety results illustrate why adverse-event interpretation should remain tied to the study component and denominator. The value 40/77, for example, describes affected participants relative to the registry-reported at-risk count for that specific pivotal cohort; it does not establish a causal difference from the 14/24 value in another cohort.

11. Clinical Biostats Interpretation of the Available Safety Data

How to read affected / at-risk counts

A serious-adverse-event value such as 5/7 means that the registry field identifies 5 affected participants among 7 participants at risk in that study component. It is a descriptive frequency, not a hazard ratio or a randomized treatment effect.

Why cohort comparisons require caution

The phase 1 study and phase 2 cohorts are components of a sequential, non-randomized program. Comparing their percentages as though they were parallel randomized treatment groups could confound treatment effects with differences in cohort composition, study stage, eligibility, follow-up, and safety-management procedures.

What the numerator does not mean

The number of affected participants does not by itself establish severity, duration, causality, or whether an event was related to axicabtagene ciloleucel. Those questions require the corresponding adverse-event definitions and detailed safety tables.

Why the denominator matters

Safety proportions are conditional on the population at risk. A value calculated using a different denominator can answer a different question. This is particularly important when multiple cohorts have different sample sizes and different registered follow-up windows.

12. Primary Endpoint Interpretation

The primary endpoints address three distinct statistical questions within the ZUMA-1 program: early treatment-related toxicity in phase 1, tumor response in the pivotal phase 2 study, and treatment-emergent CRS and neurologic toxicity in the safety-management cohorts.

Endpoint familyStatistical quantityInterpretive question
DLTProportion with a registered DLTHow frequently did participants meet the prespecified dose-limiting-toxicity definition during the first 30 days?
ORRProportion achieving CR or PRHow frequently did participants achieve a response under the registered response criteria?
CRSPercentage by severity gradeHow frequently did treatment-emergent CRS occur, and how was severity distributed?
Neurologic toxicitiesPercentage by severity gradeHow frequently did treatment-emergent neurologic toxicities occur, and how was severity distributed?

These endpoints should not be treated as interchangeable measures of a single underlying quantity. DLT is a prespecified early safety construct, ORR is an efficacy construct, and CRS and neurologic toxicity are treatment-emergent safety constructs. Each has its own analysis population and time frame.

13. Time Frames and Estimands

The registered time frames are statistically important because the same event definition can produce different estimates when observation windows differ. ZUMA-1 uses a 30-day window for the phase 1 DLT endpoint, while the phase 2 ORR and safety-management endpoints extend from first infusion to the last follow-up visit, with maximum durations ranging from 4.1 to 7.7 years depending on the endpoint.

EndpointRegistered observation windowPrimary statistical implication
Phase 1 DLTFirst infusion through 30 daysEarly toxicity probability within a fixed short-term window
Phase 2 pivotal ORRFirst infusion through last follow-up; maximum 7.7 yearsResponse proportion accumulated over the registered follow-up period
Safety-management Cohort 3First infusion through last follow-up; maximum 6.8 yearsLongitudinal ascertainment of treatment-emergent CRS and neurologic toxicity
Safety-management Cohort 4First infusion through last follow-up; maximum 5.4 yearsLongitudinal ascertainment of treatment-emergent CRS and neurologic toxicity
Safety-management Cohort 5First infusion through last follow-up; maximum 4.4 yearsLongitudinal ascertainment of treatment-emergent CRS and neurologic toxicity
Safety-management Cohort 6First infusion through last follow-up; maximum 4.1 yearsLongitudinal ascertainment of treatment-emergent CRS and neurologic toxicity

For a fixed-window binary endpoint, the estimand can be stated as the probability that a participant experiences the event during the specified window. For an endpoint assessed through last follow-up, interpretation is more dependent on follow-up duration and event ascertainment. These are not necessarily the same statistical estimand simply because both are reported as percentages.

14. Missing Data, Censoring, and Follow-Up

The ClinicalTrials.gov record does not specify a formal missing-data or imputation strategy, nor do they provide a detailed censoring rule for the registered binary endpoints. It would therefore be inappropriate to attribute a particular imputation method, missing-as-nonresponse rule, or censoring convention to the ZUMA-1 statistical analysis from the ClinicalTrials.gov record alone.

Binary efficacy endpoints

For ORR, missing response assessments can affect the denominator and classification of participants. The prespecified analysis population and handling of unevaluable participants should therefore be established before interpreting a response percentage.

Safety endpoints

For treatment-emergent toxicities, incomplete follow-up can affect the opportunity to observe events. Event definitions and observation windows are therefore essential to interpreting the reported percentage.

15. Randomization and Causal Interpretation

ZUMA-1 is explicitly classified as NON_RANDOMIZED. This is one of the most important statistical characteristics of the study. In a randomized controlled trial, the randomization mechanism supports a causal interpretation of an appropriately analyzed between-arm difference. In a non-randomized study, that design-based protection is absent.

This does not make the observed response or safety data statistically uninformative. It changes the question they can answer. A well-defined cohort can provide evidence about the frequency of responses or adverse events in the studied population. A causal comparison against another treatment requires additional design or analytical assumptions that are not reported in the ClinicalTrials.gov record here.

Descriptive versus comparative interpretation
Observed cohort rate ≠ automatic causal treatment effect

A cohort response rate describes the observed study population. To interpret a difference between cohorts causally, investigators would need a defensible comparison framework and assumptions addressing why the cohorts can be compared.

16. Multiplicity and Interim Analysis

The ClinicalTrials.gov record identifies a phase 1/2 sequential design and six registered primary endpoints, but they do not provide a formal multiplicity strategy, alpha-spending plan, interim-analysis boundary, or prespecified hypothesis hierarchy.

That absence matters because the six endpoints are not six interchangeable repetitions of the same test. They correspond to different study components and different clinical questions. A formal multiplicity interpretation would require knowing which hypotheses were confirmatory, whether they shared a familywise error budget, whether any analyses were exploratory, and how sequential decisions were prespecified.

What can be said

The registry lists six primary endpoints across phase 1 and phase 2 study components. The design is sequential and non-randomized.

What cannot be inferred

The ClinicalTrials.gov record does not establish an alpha allocation, multiplicity adjustment, interim boundary, or confirmatory hierarchy.

17. Bayesian Methods

No Bayesian statistical method is identified in the ClinicalTrials.gov record. The registry extraction does not provide a Bayesian prior, posterior probability, Bayesian decision rule, or credible interval. Consequently, Bayesian interpretations are not attributed to the ZUMA-1 analysis on this page.

18. Stratification and Subgroups

The ClinicalTrials.gov record does not provide randomized stratification factors or numerical subgroup estimates. Because allocation is non-randomized, the usual randomized-trial interpretation of baseline stratification does not apply directly to this record.

For a sequential non-randomized study, subgroup analyses can be useful descriptively, but they introduce another layer of uncertainty. A response percentage in a small subgroup may be highly variable, and an apparent difference between subgroups does not establish a treatment interaction unless the study was specifically designed and analyzed to test that interaction.

19. What a Formal Statistical Analysis Would Normally Add

The ClinicalTrials.gov record contains outcome measures but no formal statistical-analysis entries. For educational purposes, the following elements would normally be expected in a complete statistical report of these endpoint types:

Endpoint typeTypical statistical outputPurpose
DLT proportionEvent count, analysis denominator, percentage, confidence intervalQuantify early dose-limiting toxicity and its precision
ORRResponders, analysis denominator, response percentage, confidence intervalQuantify the observed tumor-response probability
CRS by gradeGrade-specific counts and percentagesDescribe both frequency and severity distribution
Neurologic toxicity by gradeGrade-specific counts and percentagesDescribe frequency and severity distribution
Comparative analysis, if prespecifiedEffect estimate, confidence interval, and hypothesis test appropriate to the designQuantify evidence for a difference when a valid comparator exists

The final row is deliberately conditional. Because the ClinicalTrials.gov record identifies a non-randomized design and do not provide formal statistical analyses, the existence, form, or inferential interpretation of such a comparison cannot be assumed.

20. Limitations

21. Why This Trial Matters Statistically

ZUMA-1 is a useful statistical teaching case because it illustrates a setting in which the main analytical challenge is not simply selecting a hypothesis test. The study combines multiple phases, sequential cohorts, binary efficacy and safety endpoints, different follow-up windows, and a non-randomized allocation structure.

ConceptHow it appears in ZUMA-1
Binary endpointAll six registered primary endpoints are classified as binary in the registry-reported extraction.
Response-rate analysisORR is defined as participants achieving complete or partial response.
Safety endpoint constructionDLT, CRS, and neurologic toxicities use prespecified event definitions and severity classifications.
Sequential designThe registry identifies a sequential phase 1/2 design.
Non-randomized allocationThe allocation field is explicitly recorded as NON_RANDOMIZED.
Multiple cohortsThe primary endpoints cover a phase 1 study, pivotal phase 2 cohorts, and safety-management cohorts.
Different estimandsThe phase 1 endpoint uses a fixed 30-day window, while phase 2 endpoints extend through last follow-up.
Severity gradingCRS and neurologic toxicities are registered by severity grade.
Denominator disciplineSafety results are reported as affected participants divided by participants at risk for individual cohorts.
Registry versus formal analysisResults are posted, but the registry-reported statistical-analysis field contains no formal analyses.

22. Statistical Concepts in This Trial

Learn more about the methods used in this trial:

23. Related Statistical Calculators

24. Sources

Continue through Clinical Biostats

Explore statistical tutorials, calculators, and additional clinical-trial analyses organized around the methods used to design and interpret clinical research.

25. Record Summary

ZUMA-1 is a completed phase 1/2, non-randomized, sequential study enrolling 307 participants and evaluating axicabtagene ciloleucel in adult participants with refractory or relapsed aggressive non-Hodgkin lymphoma. Its six registered primary endpoints span early dose-limiting toxicity, overall response rate, and treatment-emergent cytokine release syndrome and neurologic toxicities by severity grade.

Statistically, the most important feature of the record is the distinction between posted outcomes and formal statistical analyses. The ClinicalTrials.gov record confirms that results are posted and identify 28 posted outcome measures, but they contain no entries in the statistical-analysis field. Accordingly, this page focuses on the estimands, endpoint definitions, denominators, follow-up windows, and methods that would normally be used to analyze these binary outcomes rather than introducing numerical effect estimates that are absent from the ClinicalTrials.gov record.

Clinical Biostats methodology: A trial-results page should not merely repeat registry terminology. The goal is to reconstruct the statistical story of the study while clearly distinguishing documented trial facts from general statistical interpretation and avoiding causal claims that the design does not support.