Tables, Listings, and Figures (TLFs) for RWE Submissions

Tables, Listings, and Figures (TLFs)

are the headline evidence artifacts of an FDA submission: tables summarize baseline characteristics, primary and secondary endpoints, and safety; listings dump subject-level records showing every observation behind table cells; figures display Kaplan-Meier curves, forest plots, swimmer plots, and CONSORT-style flow diagrams. Together they turn the ADaM analysis package into the clinical study report (CSR) body and the regulatory evidence package a review team actually reads.

Why they matter for RWE

In a trial, TLF shells follow a protocol-specified statistical analysis plan. In an RWE package, the TLFs must additionally make the design legible: the cohort-attrition flow figure (how many patients survived each eligibility criterion — the observational analogue of CONSORT), baseline tables that reveal covariate imbalance driving any weighting, weighted vs unweighted sensitivity columns, and KM figures whose numbers-at-risk tables expose censoring patterns. Reviewers of external-control arguments scrutinize TLFs for exactly the things observational designs can hide: immortal-time artifacts in time-zero alignment, differential depletion, positivity violations visible as extreme weights or empty strata.

Production discipline

• Shells first. Mockup every table aligned to the SAP/population definitions before programming; late shell changes cascade through validation.
• Double programming. The primary programmer produces each output; an independent validator re-produces it from the specifications; outputs reconcile cell-by-cell. Pinnacle 21 and ADRG cross-checks supplement but do not replace this.
• Macro standardization. Population filters come from ADSL flags only — never hard-coded subsets inside table macros, which is how population drift sneaks in.
• Numbering and cross-referencing. Consistent numbering (Table 14.x.x conventions), footers stating population, cut date, and analysis dataset versions.

RWE-specific content requirements

• Cohort disposition figure: eligibility funnel from source population to analysis cohorts, with counts and reasons at each step.
• Baseline tables by cohort and by weighting scheme: unweighted, IPTW-weighted (with standardized mean differences), and any matched sample side-by-side.
• Follow-up and censoring displays: KM curves with numbers at risk, censoring marks, and median follow-up by arm.
• Sensitivity tiers: pre-specified alternative definitions (on-treatment vs ITT-analogue, different censoring rules, outcome windows) presented consistently.

Pros, cons, and trade-offs

• vs ad hoc summary tables: TLFs carry SAP alignment, double-programming verification, and reconciliation pedigree; ad hoc outputs do not survive regulatory scrutiny.
• vs figures alone: curves communicate but hide cell counts; tables carry the denominators reviewers check. Use both, cross-referenced.
• Trade-off: completeness vs readability — every output must serve the CSR narrative; redundant cuts multiply validation cost without evidentiary gain.

When NOT to use

Exploratory sensitivity runs stay in the statistical appendix or off-submission environment. Never present a TLF whose population, censoring rules, or windows differ from the locked SAP without a documented amendment.

Decision diagram

flowchart LR
  ADaM[ADaM analysis datasets<br/>ADSL, ADTTE, ADEFF, ADAE] --> T1[Table 1: Baseline characteristics]
  ADaM --> Disp[Disposition table]
  ADaM --> PE[Primary endpoint table]
  ADaM --> Sec[Secondary endpoint tables]
  ADaM --> Saf[Safety summary tables]
  ADaM --> List[Subject listings]
  ADaM --> Fig[Figures: KM curves, forest plots]
  ADaM --> ARM[ARM XML extension to define.xml]
  T1 --> CSR[Clinical study report]
  Disp --> CSR
  PE --> CSR
  Sec --> CSR
  Saf --> CSR
  List --> CSR
  Fig --> CSR
  ARM --> CSR

Worked example

Scenario

We need a Table 14.2.1 (Primary Endpoint) and a Table 14.1.1 (Baseline Characteristics) for an RWE submission comparing Drug A vs Drug B with a time-to-first-hospitalization outcome. Both come from the locked ADaM package; the ARM extension in define.xml carries the Population filter (ITTFL='Y'), the Analysis (ADTTE PARAMCD=TTFHOSP), and the Programming Statement that produced each cell.

Dataset

Headline TLF summaries from the locked ADaM ADTTE.

Steps

Result

Two RTF tables delivered: Table 14.1.1 (baseline characteristics by ACTARM) and Table 14.2.1 (primary endpoint summary with events and person-time by ACTARM). The ARM XML in define.xml carries the Output definitions, Analysis references, Filters, and Programming Statements. A reviewer opens the RTF, clicks the ARM reference, and re-derives any cell from the ADaM records.

Trade-offs

Runnable example

# Toy example: ADaM ADSL + ADTTE -> Table 1 baseline summary + primary endpoint table.
# In a real submission these are validated functions with locked source; the toy shows the
# cell-level trace-back discipline.
import pandas as pd

def table_1(adsl: pd.DataFrame, value_vars=("AGE", "SEX", "RACE")) -> pd.DataFrame:
    # Baseline characteristics summary, stratified by ACTARM. Each cell links to a row.
    rows = []
    for v in value_vars:
        summary = adsl.groupby(["ACTARM", v]).size().unstack(fill_value=0)
        for arm in summary.index:
            for val, n in summary.loc[arm].items():
                rows.append({"Variable": v, "Arm": arm, "Value": val, "N": int(n)})
    return pd.DataFrame(rows)

def primary_endpoint_table(adtte: pd.DataFrame) -> pd.DataFrame:
    # Hazard ratio + 95% CI for the first PARAMCD; in production use lifelines or statsmodels.
    # Toy: just count events and person-time by ACTARM.
    out = (adtte.groupby("ACTARM")
                 .agg(events=("CNSR", lambda s: int((s == 0).sum())),
                      censored=("CNSR", lambda s: int((s == 1).sum())),
                      person_days=("AVAL", "sum"))
                 .reset_index())
    out["person_years"] = (out["person_days"] / 365.25).round(1)
    return out

library(dplyr)

table_1 <- function(adsl, value_vars = c("AGE", "SEX", "RACE")) {
  adsl |>
    tidyr::pivot_longer(cols = dplyr::all_of(value_vars), names_to = "Variable", values_to = "Value") |>
    dplyr::count(ACTARM, Variable, Value) |>
    tidyr::pivot_wider(names_from = ACTARM, values_from = n, values_fill = 0)
}

primary_endpoint_table <- function(adtte) {
  adtte |>
    dplyr::group_by(ACTARM) |>
    dplyr::summarise(
      events = sum(CNSR == 0),
      censored = sum(CNSR == 1),
      person_days = sum(AVAL),
      person_years = round(person_days / 365.25, 1)
    )
}

/* Toy: ADaM ADSL -> Table 1 baseline characteristics summary. */
proc freq data=work.adsl noprint;
  tables ACTARM * AGE / out=work.age_freq;
  tables ACTARM * SEX / out=work.sex_freq;
run;

Citations

1. [1]R Consortium R for Clinical Study Reports and Submission Working Group. TLF Overview. r4csr.org.
2. [2]CDISC. Analysis Results Metadata (ARM) v1.0 for Define-XML v2.0. CDISC standards.
3. [3]U.S. Food and Drug Administration. Study Data Technical Conformance Guide — Technical Specifications Document.

1. [4]Yan F, Zhang Y, Tian Y. Evidence Behind the Automation of Clinical Trial Statistical Programming: A Scoping Review of Technology Adoption, Validation Frameworks, and AI/ML Integration (2020-2025). medRxiv. 2025.