A synthetic bioequivalence study, from sample size to CSR text

Stage 1

Design: sample size and randomization

Runs here

What was executed

• PowerTOST 1.5-7, exact method: CV 25%, ratio 95%, limits 80.00% to 125.00%: 28 evaluable for 80% power, 38 for 90%; 5% allowance for non-evaluable participants gives 40 enrolled.
• Sensitivity grid of N and power over CV 15% to 35% and ratio 0.90 to 1.05; pa.ABE() limits of the assumptions.
• blockrand 1.5, seed 2026: permuted blocks of 4, sequences TR and RT 20 : 20; balance check passed (largest running imbalance 2).
• Protocol statistical-section excerpt with ICH M11 headings (estimand, analysis sets, model, missing data, sample size).

Artifacts

• protocol_statistics.htmlM11-headed excerpt, sensitivity tables, schedule check
• randomization_schedule.csv40 sequence assignments
• sensitivity_n.csvN over CV and ratio
• design_parameters.jsonmachine-readable design

Not covered: BOIN/CRM dose-escalation designs, adaptive or group-sequential designs.

Stage 2

Statistical analysis plan and mock shells

Settings drive the code

What was executed

• Estimand with the five ICH E9(R1) attributes; each intercurrent event mapped to a strategy, and why the other four strategies do not apply.
• Analysis sets with their ADaM flags; the ICH M13A model, the tests to report, the decision on unrounded limits.
• BLQ and missing-data rules per ICH M13A; the settings are written to analysis_parameters.json, which the NCA, exclusion and BE programs read.
• Mock shells for the six executed tables.

Artifacts

• sap_excerpt.htmlstatistical methods excerpt
• mock_shells.htmlsix table shells
• analysis_parameters.jsonsettings the code reads

Stage 3

SDTM-like inputs, ADaM, specification and XPT

Runs here

What was executed

• Simulated DM, EX, PC, AE, VS and DS (1185 concentration records at actual times) with 6 documented scenarios: a withdrawal, emesis, a pre-dose carry-over value, a missed sample, a late sample and a BLQ result between quantifiable samples.
• ADSL, ADPC, ADPP and ADAE with admiral 1.5.0 following its ad_adpc.R and ad_adpp.R templates (NFRLT, AFRLT, NRRLT, ARRLT), adaptations listed.
• Specification in Pinnacle 21 layout (xlsx) and JSON, read by metacore::spec_to_metacore(); 24 of 24 metadata checks passed (metatools check_variables, check_ct_data, check_unique_keys, lengths, mandatory values, XPT read-back).
• SAS XPORT v5 with labels and lengths from xportr 0.6.0; Dataset-JSON v1.1 copies from datasetjson 0.4.1, which FDA has not accepted as of Oct 2026.

Artifacts

• ADaM pagedatasets, spec, checks, adaptations
• adam_spec.xlsxP21-layout specification
• adam_spec.jsonmachine-readable specification
• adpc.xptone of four XPT v5 files

Not covered: define.xml and the ADRG are not authored; no Pinnacle 21 or CDISC CORE run; the inputs are SDTM-like, not conformant SDTM.

Stage 4

NCA, bioequivalence analysis and TLFs

Runs here

What was executed

• PKNCA 0.12.1 NCA on actual times with BLQ kept as zero in every position (ICH M13A): 79 profiles. For the profile with a BLQ value between quantifiable samples, AUC(0-t) is 1312.1 h*ng/mL against 1433.0 with the PKNCA default, which drops that value.
• SAP parameter rules through PKNCA exclude(): AUC(0-inf) flagged in 2 profiles (more than 20% extrapolated); the supportive AUC(0-inf) analysis uses 37 participants.
• ICH M13A GLM: Cmax ratio 101.07% (90% CI 98.89 to 103.30), AUC(0-t) ratio 103.99% (90% CI 98.52 to 109.75) in 38 participants; bioequivalence shown.
• Cmax at the first post-dose sample in 2 profiles (ICH M13A 2.1.8.1); a supplementary analysis without those participants.
• Six tables with rtables 0.6.17 and tern 0.9.13, each exported to RTF with formatters::export_as_rtf() and to HTML; three figures; a Quarto results report.

Artifacts

• Analysis pagetables, figures, result files
• report.htmlQuarto results report
• Table 14.4.3.1bioequivalence result (RTF alongside)
• be_results.csvunrounded estimates

Not covered: Phoenix WinNonlin and SAS were not used.

Stage 5

QC: re-derivation by a second method

Same-author re-derivation

What was executed

• 40 of 40 checks passed (5100 values compared with diffdf 1.1.2 at stated tolerances).
• NonCompart 0.8.4 vs PKNCA for every NCA parameter, including the profile with a BLQ value between quantifiable samples, and the SAP parameter-rule flags; the BE model refit as a closed-form period-difference t-test and with nlme::lme().
• ADPC times (including the experimental derive_var_nfrlt()), ADSL population flags and the adverse-event counts rebuilt in base R from the SDTM-like data.
• Controls: a planted 0.1% change is found exactly once, and NonCompart's unadjusted AUC(0-t) differs from PKNCA in exactly the BLQ profile, so the comparison is shown to fail when it should.

Artifacts

• qc_ledger.htmlexecuted ledger with pass/fail counts
• qc_ledger.csvone row per compared variable
• qc_nca_side_by_side.csvboth engines, every profile

Not covered: This is same-author re-derivation: the same person wrote both routes. No second programmer re-derived the results.

Stage 6

CSR section 11-style text, checked against the outputs

Number check

What was executed

• Section 11-style text (data sets analyzed, demographics, compliance, PK and BE results, missing data, conclusions), with every number read from an output file.
• The tlf_interpret number-grounding validator, unchanged: 10 of 10 paragraphs have every number found in the outputs.
• Controls: an altered number is rejected; a wrong statement built from real numbers passes, which is the known limitation; 'well tolerated' is flagged for review.

Artifacts

• csr_section11.htmltext, check results, controls
• csr_section11.txtplain text
• csr_check.jsonvalidator output

Not covered: The validator checks that numbers exist in the outputs, not that statements are correct. No language model runs in this pipeline; the text is filled into fixed sentence templates (written with an AI coding assistant) and still needs a statistician's review.