Clinical Trial Secondary Analysis: SAR, TLF Tables, and CONSORT Workflow

Secondary analysis of clinical trial individual patient data demands an auditable ICH E3 Statistical Analysis Report alongside manuscript-ready CONSORT figures and interaction-tested subgroup forest plots.

HumanPen Team

· 9 min read

Short Answer: Separating Internal SAR Audit Trails from Journal Deliverables

A secondary analysis of clinical trial individual patient data requires two distinct deliverables: an internal, auditable Statistical Analysis Report containing all ICH E3 tables, listings, and figures across intention-to-treat and safety populations, and a separate manuscript package with a validated CONSORT flow diagram, baseline Table 1, and interaction-tested subgroup forest plots. Decoupling regulatory audit tables from journal figures prevents methodological discrepancies, ensuring verifiable reproducibility before submitting to peer-reviewed journals.

A published medical manuscript typically features three to four condensed summary tables, but peer reviewers and regulatory bodies evaluate the rigor of a secondary analysis through its full audit trail. If you cannot trace every randomized participant from baseline assignment to final follow-up across both Intention-to-Treat and Per-Protocol sets in an underlying statistical report, any isolated subgroup finding risks immediate dismissal as post-hoc data dredging.

Clinical fellows and biostatisticians analyzing public repositories—such as the International Stroke Trial (IST) on Edinburgh DataShare—often jump straight to running regression models on filtered subsets. When journals request the full Statistical Analysis Plan (SAP) or complete baseline distributions of excluded participants, researchers scramble to reconstruct their data lineage. Building an internal Statistical Analysis Report (SAR) with complete Tables, Listings, and Figures (TLF) first establishes a single source of truth, making manuscript tables trivial to extract and defend.

ICH E3 Statistical Analysis Report vs Journal Results: Core Dimensions

The International Council for Harmonisation guideline ICH E3 defines the structure and content of comprehensive clinical study reports, while the EQUATOR Network CONSORT Statement governs the transparent reporting of randomized trials in peer-reviewed journals. Because regulatory auditors and journal readers operate under different time constraints and technical requirements, your secondary analysis package must satisfy both standards without conflating them.

Structural DimensionInternal Statistical Analysis Report (SAR)Journal Manuscript Results Package
Primary Objective & AudienceComplete regulatory and methodological audit trail for study biostatisticians and auditorsClear, concise clinical narrative and effect-size summary for editors and clinicians
TLF Inventory ScopeComprehensive repository of 20 to 50 formatted tables, individual patient listings, and diagnostic figures3 to 5 high-impact tables and figures (Table 1 baseline, Table 2 primary outcomes, CONSORT diagram, forest plot)
Analysis PopulationsParallel evaluation across Intention-to-Treat (ITT), Modified ITT, Per-Protocol (PP), and Safety setsPrimary narrative focused on ITT population, with Per-Protocol reported as sensitivity analysis
Model TransparencyFull regression outputs, baseline diagnostics, coefficient tables, standard errors, and fit metricsSummary Odds Ratios (OR) or Hazard Ratios (HR) with 95% confidence intervals and exact p-values
Protocol Deviations & Missing DataExhaustive listings of inclusion violations, loss to follow-up, and missingness imputation modelsSummary narrative in Methods/Results describing attrition and primary imputation handling

Notice that the manuscript package is not a substitute for the SAR; it is an executive distillation of it. When reviewers raise questions about potential confounding, baseline imbalances, or selective attrition, having an existing 30-to-40-page SAR allows you to cite specific internal TLF tables in your response letter immediately, rather than having to re-engineer ad-hoc scripts under stressful revision deadlines.

CONSORT Flow Accounting, Covariate Adjustment, and Subgroup Interaction Tests

The CONSORT 2025 Statement requires an uninterrupted participant accounting pipeline across four sequential phases: enrollment, intervention allocation, follow-up status, and final statistical analysis. In complex trials involving factorial or multi-arm designs, every branch must balance mathematically to zero unaccounted cases. Aligning your manuscript with CONSORT clinical trial reporting standards prevents editorial rejections rooted in sample size discrepancies between text and figures.

  • Factorial accounting integrity: In a 3×2 factorial trial, verify participant allocations across all treatment combinations (such as active drug with placebo, active drug with active comparator, or double placebo) so that row totals match the grand randomized sample.
  • Prespecified covariate adjustment: Primary efficacy analyses must prespecify key baseline prognostic factors (such as baseline stroke severity, age, or time from onset to treatment) to guard against post-hoc model hunting, while always presenting both unadjusted and adjusted effect sizes.
  • Mandatory interaction test reporting: Never conclude that a treatment works differently across patient strata based merely on separate within-group p-values (for instance, observing `p = 0.02` in female patients and `p = 0.18` in male patients); journals demand a formal test of treatment-by-subgroup interaction (`p_interaction`).
  • Forest plot visual standards: Present point estimates, 95% confidence intervals, and subgroup-specific event counts alongside a clear column of interaction p-values, ensuring that the null vertical line (`OR = 1.0` or `HR = 1.0`) is distinctly marked.

Prestigious clinical journals like The Lancet, JAMA, and The BMJ systematically flag subgroup claims that lack interaction statistics, categorizing them as exploratory hypothesis-generating observations rather than confirmed findings. When condensing these complex multivariable results into tightly constrained abstract formats, adhering to established conventions for shortening structured medical abstracts ensures that confidence intervals, absolute risk reductions, and interaction boundaries remain crystal clear.

Dual-Engine Verification: Reproducing IPD Event Counts in Python and R

To ground secondary analyses in reproducible evidence, biostatisticians turn to landmark open-access trial datasets. A prime benchmark is the International Stroke Trial (IST), published by the IST Collaborative Group in The Lancet (1997, DOI `10.1016/S0140-6736(97)04011-7`), which randomized 19,435 acute stroke patients across 467 hospitals in 36 countries. The verified dataset `IST_corrected.csv`, archived on Edinburgh DataShare (`10283/124`), contains granular event timelines that require dual-engine scripted execution to reproduce exact historical counts.

  • Granular patient accounting: Ingest `IST_corrected.csv`, verify that all 19,435 randomized cases are mapped to their factorial treatment assignments (`RXASP` for aspirin, `RXHEP` for heparin), and confirm that no observations are dropped during initial preprocessing.
  • Event count reproduction: Recompute the 6-month primary poor outcome (death or dependency, `OCCODE` categories), reproducing the historical Lancet totals: 62.2% in the aspirin cohort versus 63.5% in the non-aspirin cohort.
  • Dual-engine modeling parity: Fit binary logistic regression models in Python using `statsmodels.api.Logit` and in R using `glm(..., family = binomial())`, verifying that the unadjusted association (`p = 0.072`) and the prespecified covariate-adjusted model yield identical point estimates (`OR = 0.93`, `95% CI 0.87–0.99`) across both programming environments.
  • Single-truth artifact generation: Export all model parameters, sample counts, and p-values into a machine-readable JSON dictionary (`results.json`) that feeds directly into LaTeX, Word, and SVG plot templates, eliminating manual transcription errors across deliverables.
Discrepancies between Python and R in logistic regression almost never stem from underlying mathematical differences. Instead, they arise because default implementations differ: `scikit-learn` applies an L2 penalty by default unless explicitly configured with `penalty=None`, while R `glm()` performs unpenalized maximum likelihood estimation. Always enforce unpenalized estimation when auditing regulatory trial endpoints.

Streamlined Secondary Analysis Delivery in HumanPen Start Project

Handling clinical trial secondary analysis requires an environment where data ingestion, statistical modeling, and document compilation execute deterministically. In HumanPen Start Project, the entire process from raw CSV upload to final manuscript delivery is structured under a verified, auditable protocol. Attach your data and papers, describe what you need, approve the plan before work starts, and get the reviewed report, charts, and code ready to preview online or download.

  • Secure sandbox execution: Upload trial datasets (CSV, Excel, Word, PDF, or zip archives up to 2 GB); the system validates data schemas and formulates an analysis plan specifying models, covariates, and endpoints for your explicit confirmation before any code runs.
  • Reproducible dual-engine pipelines: Analysis runs in an isolated Python and R sandbox, executing baseline Table 1 balance tests, Kaplan-Meier curves, and multivariable regressions within a self-contained execution environment.
  • Comprehensive 37-page SAR generation: Compiles an ICH E3-compliant Word report containing 22 publication-grade TLF tables, demographic distributions, subgroup interactions, and protocol deviation summaries.
  • Publication-ready manuscript artifacts: Generates vector and 300 dpi graphics, including a compliant CONSORT participant flow diagram, a 7-subgroup forest plot with explicit interaction p-values, and drafted Methods and Results sections.
  • Publicly inspectable live benchmark: Explore the verified IST secondary analysis example live on the platform to review 55 deliverable files, including `SAR.docx`, `Manuscript_materials.docx`, and the standalone `check_with_R.R` validation script.
Methodological compliance notice: HumanPen delivers reproducible analysis scripts, formatted TLF tables, and manuscript drafts strictly as an academic and methodological reference scaffold. The principal investigator, biostatistician, or clinical research fellow remains responsible for verifying data assumptions, declaring institutional disclosures, and obtaining formal biostatistical sign-off before regulatory submission or journal publication.

Frequently Asked Questions

What is the core distinction between an Intention-to-Treat (ITT) population and a Per-Protocol (PP) population in secondary analysis? The ITT population includes every randomized participant in the arm to which they were originally assigned, regardless of non-compliance, protocol deviations, or premature treatment withdrawal, thereby preserving the structural benefits of randomization and avoiding bias. In contrast, the Per-Protocol population includes only participants who completed the prescribed intervention without major protocol deviations. In regulatory trial reports and top-tier clinical journals, ITT serves as the primary efficacy benchmark, while Per-Protocol analysis is conducted as a sensitivity check to assess the potential impact of non-adherence.

Why do medical journal editors reject clinical subgroup claims that lack formal interaction tests? Clinical researchers frequently observe that a treatment shows statistical significance in one subgroup (such as elderly patients with `p = 0.01`) but not in another (such as younger patients with `p = 0.15`), and erroneously claim that the drug is only effective in the elderly. Statistically, the difference between significant and non-significant is not necessarily statistically significant. Without a formal treatment-by-subgroup interaction term in a unified regression model yielding a significant interaction p-value, claims of differential efficacy represent spurious subgroup post-hoc rationalizations that will be rejected by rigorous peer reviewers.

Can secondary analysis workflows on HumanPen handle trial datasets originally analyzed in SPSS or Stata? Yes. HumanPen Start Project supports uploading dataset files and documentation in Excel, CSV, Word, PDF, and zip formats up to 2 GB total per project. Analyses originally conducted in SPSS or Stata are performed reproducibly in an isolated sandbox using Python or R. Researchers can export legacy SPSS or Stata datasets into CSV or Excel formats alongside their codebook or data dictionary, or bundle files into a zip archive, enabling the automated pipeline to generate consistent baseline tables, multivariable regressions, and subgroup interaction models.

What security and confidentiality safeguards protect clinical trial datasets uploaded for secondary analysis? Data uploaded to HumanPen is transmitted via encrypted channels and executed within single-tenant, isolated execution sandboxes. Files are never shared across accounts and are strictly excluded from model training corpora. Projects and intermediate data are retained for 180 days (6 months) with a 3-day proactive expiration notice, and researchers can permanently delete their data, scripts, and compiled deliverables from the server at any point with a single click.

Sources

KEEP READING