CDISC Pilot 01 — SDTM to ADaM in SAS and R

Double-programming an FDA-grade analysis package on a publicly redistributable clinical-trial dataset

Author

Paulina Del Mundo Del Fierro, MD, MPH

Published

June 12, 2026

TipAbstract

Every clinical trial submitted to the FDA arrives as a precisely-structured package of datasets, and pharma companies typically write that package twice — once in SAS (the industry-standard tool) and once independently in another language — then compare the two outputs to catch programmer errors before the trial database is locked. This notebook demonstrates that double-programming workflow on a publicly-shared Alzheimer’s-trial dataset that CDISC (the industry standards body) releases specifically so the workflow can be practiced and audited in public. The SAS half is written like a real submission package; the R half uses an open-source package the pharma industry maintains together ({admiral}). Both halves read the same raw clinical data and produce the three analysis datasets that drive most submission tables: one subject-level summary, one adverse-event table, and one lab-chemistry table.

The R half executes on render: it reads the CDISC Pilot SDTM domains from the {pharmaversesdtm} package, derives ADSL, ADAE, and ADLBC with {admiral}, and produces the ICH E3 demographics table below — all from data shipped in the package, with no external download. The one piece that does not run here is the cross-language reconciliation step, which diffs the SAS-derived and R-derived outputs against CDISC’s reference answer key: it needs the SAS OnDemand outputs, which are not part of this repo. Read this as an executed R derivation alongside a complete SAS submission package, with the byte-for-byte SAS-vs-R reconciliation specified but not run in this environment.

NoteStatus — the R derivation executes; SAS-vs-R reconciliation is the remaining step

What runs on render. The R half executes end-to-end: it reads the CDISC Pilot SDTM domains from {pharmaversesdtm}, derives ADSL, ADAE, and ADLBC with {admiral}, writes a v5 transport XPT, and renders the ICH E3 §14.2.1 demographics table — all from data shipped in the package, no external download.

What’s still a specification. The SAS submission package (sas/adsl.sas, sas/adae.sas, sas/adlbc.sas, sas/t_14_2_01_demog.sas) is written and embedded below, but the SAS programs are not executed here, so the closing cross-language reconciliation — diffing the SAS-derived, R-derived, and CDISC-reference ADSLs byte-for-byte — does not run (it needs the SAS OnDemand outputs). That, plus a define-XML-driven {xportr}/{metacore} conformance layer, is the remaining work to make this a fully self-checking submission package.

1 Why this analysis

Every interventional clinical trial submitted to the FDA, EMA, or PMDA arrives as a CDISC package: collected data conforms to SDTM (Study Data Tabulation Model, IG v3.4), and analysis-ready data conforms to ADaM (Analysis Data Model, IG v1.3). Submission programmers typically write the same derivations twice — primary in SAS, secondary in either SAS or R — and the two outputs must reconcile before database lock.

This notebook is that double-programming exercise on a public dataset:

  1. Read the CDISC Pilot SDTM domains (DM, AE, EX, LB) from the {pharmaversesdtm} package.
  2. Derive ADSL, ADAE, and ADLBC using ADaM IG v1.3 conventions — twice, once in SAS, once in R/{admiral}.
  3. Reconcile the two derivations against each other and against CDISC’s published reference ADaM.
  4. Report an ICH E3 §14.2.1 demographics table and the reconciliation result.

The CDISC Pilot is a fictional Alzheimer’s trial (placebo / Xanomeline low dose / Xanomeline high dose) released by CDISC under a permissive license expressly so the standards can be taught and audited in public; the {pharmaversesdtm} package ships its SDTM domains as R data. See https://github.com/cdisc-org/sdtm-adam-pilot-project.

Having a colleague read your code catches typos and obvious logic errors, but a reviewer reads along the same logical track you wrote on: if your mental model of the derivation is wrong, theirs tends to bend the same way. Double-programming is stronger because it is replication, not review. Two programmers, ideally in two languages, start from the same spec and the same raw data and derive independently. A bug in the SAS branch is unlikely to recur as the identical bug in the R/{admiral} flow, so a byte-level diff of the two outputs surfaces exactly the discrepancies a shared-track review misses.

The failure mode it catches is the plausible-but-wrong line. A derivation that types EXDOSE > 0 where the spec meant EXDOSE >= 0 compiles, runs, passes a Pinnacle 21 conformance check, and is silently wrong by a handful of subjects: invisible to any single-implementation check, visible the moment two independent derivations disagree. What byte-level reconciliation does not catch is a specification-level error. If both implementations are written to the same wrong define-XML spec, they agree with each other and are both wrong. Reconciliation tests the programming, not the specification.

This is also why pharma pays the upfront cost of CDISC at all. Agreeing on SDTM and ADaM before a single analysis runs is the same investment a research group makes when it writes a data dictionary before fieldwork: cost paid early to make a whole class of downstream operations (pooling, review, reconciliation, audit) mechanical rather than bespoke. The transferable discipline is writing every derivation against a specification auditable by someone who was not in the room when the program was written.

Note

The .qmd source for this page, the SAS programs, and the YAML define-XML extract live on GitHub.

2 Setup

Show code 
suppressPackageStartupMessages({
  library(admiral)      # ADaM derivations
  library(xportr)       # writes v5 transport, validates against CDISC CT
  library(haven)        # reads/writes XPT
  library(metacore)     # reads define-xml -> R object
  library(metatools)    # applies metacore to a dataset
  library(dplyr)
  library(tidyr)
  library(gtsummary)
  library(here)
})

DATA_SDTM <- here::here("traces", "applied", "03-cdisc-adam-pilot", "data", "sdtm")
DATA_REF  <- here::here("traces", "applied", "03-cdisc-adam-pilot", "data", "adam_reference")
DATA_OUT  <- here::here("traces", "applied", "03-cdisc-adam-pilot", "data", "adam_r")
SAS_OUT   <- here::here("traces", "applied", "03-cdisc-adam-pilot", "sas", "output")
dir.create(DATA_OUT, showWarnings = FALSE, recursive = TRUE)

3 Read SDTM

On the pathway → 01 · Measurement: data standards and provenance, the CDISC SDTM structure a datapoint inherits before any analysis touches it.

Show code 
# SDTM domains from the CDISC Pilot (Xanomeline) study, supplied as R data by the
# pharmaverse {pharmaversesdtm} package — no external download or `make data` step.
dm <- pharmaversesdtm::dm
ae <- pharmaversesdtm::ae
ex <- pharmaversesdtm::ex
lb <- pharmaversesdtm::lb

list(DM = dim(dm), AE = dim(ae), EX = dim(ex), LB = dim(lb))
$DM
[1] 306  28

$AE
[1] 1191   35

$EX
[1] 591  17

$LB
[1] 59580    23

4 Derive ADSL (R / {admiral})

On the pathway → 00 · Framing the study: defining the analysis populations (ITT, Safety, Per-Protocol, mITT) before anything is estimated.

ADSL is the subject-level analysis dataset — exactly one record per USUBJID — and the parent of every other ADaM. The derivations below follow ADaM IG v1.3 §3.1 and the CDISC Pilot 01 ADSL spec.

A clinical trial doesn’t have one denominator — it has four, each suited to a different question, and every analysis must say which population it’s run on:

Population Flag Who’s in Answers
Intent-to-Treat (ITT) ITTFL Every randomised subject, analysed as randomised “What is the effect of being assigned the treatment?” — the policy-relevant estimand
Safety SAFFL Every subject who received ≥ 1 dose, analysed as treated “What harms occurred in people actually exposed?” — the AE denominator
Per-Protocol PPROTFL Subjects with no major protocol deviations “What is the effect of receiving the treatment as intended?”
Modified ITT MITTFL Pre-specified narrower ITT — e.g., ≥ 1 dose AND a baseline measurement “ITT with a minimum-quality data filter applied”

ITT is the regulatory default for efficacy because it preserves randomisation. Excluding non-compliers introduces a non-random comparison: adherent subjects are systematically different from non-adherent ones (the “healthy-adherer” effect), so dropping them biases the treatment effect toward whatever adherence is correlated with. ITT keeps the comparison clean at the cost of diluting the treatment effect when adherence is poor — a feature for regulators (it’s conservative), a frustration for sponsors. Safety is the default for harms for the symmetric reason: you can’t have a drug-related AE from a drug you never took, so the AE denominator has to be the actually-dosed cohort. In a hypothetical 254-subject trial, 250 might receive at least one dose, 238 of those complete the protocol cleanly, and 3 randomised subjects might withdraw before dosing — so efficacy reports \(n = 254\) (ITT), the AE table reports \(n = 250\) (Safety), and the dose-response analysis reports \(n = 238\) (PP). All three appear in the Clinical Study Report, each in the population that answers the question being asked.

Show code 
adsl <- dm |>
  mutate(TRT01P = ARM, TRT01A = ACTARM) |>
  # Treatment start/end from EX (first/last dosing record)
  derive_vars_merged(
    dataset_add = ex,
    by_vars     = exprs(STUDYID, USUBJID),
    new_vars    = exprs(TRTSDT = convert_dtc_to_dt(EXSTDTC)),
    filter_add  = EXDOSE > 0 & !is.na(EXSTDTC),
    order       = exprs(EXSTDTC),
    mode        = "first"
  ) |>
  derive_vars_merged(
    dataset_add = ex,
    by_vars     = exprs(STUDYID, USUBJID),
    new_vars    = exprs(TRTEDT = convert_dtc_to_dt(EXENDTC)),
    filter_add  = EXDOSE > 0 & !is.na(EXENDTC),
    order       = exprs(EXENDTC),
    mode        = "last"
  ) |>
  derive_vars_duration(new_var = TRTDURD, start_date = TRTSDT, end_date = TRTEDT) |>
  mutate(
    SAFFL  = if_else(!is.na(TRTSDT), "Y", "N"),
    ITTFL  = if_else(ARM != "Screen Failure", "Y", "N"),
    AGEGR1 = case_when(
      AGE <  65 ~ "<65",
      AGE <= 80 ~ "65-80",
      TRUE      ~ ">80"
    ),
    AGEGR1N = case_when(AGEGR1 == "<65" ~ 1, AGEGR1 == "65-80" ~ 2, TRUE ~ 3)
  )

cat(sprintf("ADSL: %d subjects x %d variables\n", nrow(adsl), ncol(adsl)))
ADSL: 306 subjects x 37 variables
Note

The full derivation (~25 variables) lives in the source .qmd; the chunk above shows the patterns. {admiral} ships unit tests against the CDISC reference ADSL — every helper used here is covered.

4.1 Same derivation in SAS

The primary submission program — same logic, expressed in SAS.

sas/adsl.sas
/******************************************************************************
 * Program  : adsl.sas
 * Purpose  : Derive ADSL (Subject-Level Analysis Dataset) per ADaM IG v1.3 §3.1
 *            and the CDISC Pilot 01 ADSL specification.
 * Inputs   : SDTM.DM, SDTM.EX, SDTM.SV, SDTM.SUPPDM
 * Output   : ADAM.ADSL
 * Notes    : One record per USUBJID. TRT01P / TRT01A from DM.ARM / DM.ACTARM.
 *            Treatment dates come from EX (first non-zero dose / last non-zero
 *            dose). Population flags follow the Pilot SAP §6.
 *****************************************************************************/

%include "setup.sas";

/*----- Treatment epoch from EX -------------------------------------------*/
proc sql;
  create table _trtdt as
    select STUDYID, USUBJID,
           min(input(EXSTDTC, e8601da.)) as TRTSDT format=date9.,
           max(input(EXENDTC, e8601da.)) as TRTEDT format=date9.
    from sdtm.ex
    where EXDOSE > 0
    group by STUDYID, USUBJID;
quit;

/*----- Build ADSL --------------------------------------------------------*/
data adam.adsl (label="Subject-Level Analysis Dataset");
  merge sdtm.dm (in=in_dm)
        _trtdt;
  by STUDYID USUBJID;
  if in_dm;

  /* Treatment variables — IG v1.3 §3.1.1 */
  length TRT01P TRT01A $40;
  TRT01P = ARM;
  TRT01A = ACTARM;
  TRT01PN = case
              when ARM = "Placebo" then 0
              when ARM = "Xanomeline Low Dose"  then 54
              when ARM = "Xanomeline High Dose" then 81
              else .
            end;
  TRT01AN = TRT01PN;

  /* Treatment duration in days (inclusive) */
  if not missing(TRTSDT) and not missing(TRTEDT) then
    TRTDURD = TRTEDT - TRTSDT + 1;

  /* Population flags — IG v1.3 §3.1.4 */
  length SAFFL ITTFL $1;
  SAFFL = ifc(not missing(TRTSDT), "Y", "N");
  ITTFL = ifc(not missing(ARMCD) and ARMCD ne "Scrnfail", "Y", "N");

  /* Age groupings per Pilot SAP */
  length AGEGR1 $5;
  if      AGE <  65 then do; AGEGR1 = "<65";   AGEGR1N = 1; end;
  else if AGE <= 80 then do; AGEGR1 = "65-80"; AGEGR1N = 2; end;
  else if not missing(AGE) then do; AGEGR1 = ">80";  AGEGR1N = 3; end;

  format TRTSDT TRTEDT date9.;
  keep STUDYID USUBJID SUBJID SITEID
       AGE AGEGR1 AGEGR1N SEX RACE ETHNIC COUNTRY
       ARM ARMCD ACTARM ACTARMCD
       TRT01P TRT01PN TRT01A TRT01AN
       TRTSDT TRTEDT TRTDURD
       SAFFL ITTFL;
run;

proc sort data=adam.adsl; by STUDYID USUBJID; run;

/*----- Quick sanity log --------------------------------------------------*/
proc freq data=adam.adsl;
  tables TRT01A * SAFFL / nocum nopercent;
run;

5 Derive ADAE and ADLBC

On the pathway → 01 · Measurement: operationalizing the endpoints, the treatment-emergent flag and the baseline-adjusted change variables.

ADAE is occurrence-level (one record per AE per subject). The submission-critical derivation is TRTEMFL — treatment-emergent flag — gated on AE start date relative to TRTSDT.

ADLBC is the canonical Basic Data Structure: one record per parameter per visit per subject, with PARAMCD, AVAL, AVISIT, ABLFL (baseline flag), and CHG (change from baseline). The lab unit conversions key off the CDISC LB controlled terminology.

A patient reports a headache on day 14 of dosing. Did the drug cause it? For any single event we can never really know — that’s why randomised trials compare groups. But for the safety table, we still need a deterministic rule for counting, and “treatment-emergent” is that rule: \[\text{AE start date} \geq \text{TRTSDT}, \quad \text{and} \quad \text{AE start date} \leq \text{TRTEDT} + \Delta,\] where \(\Delta\) is a protocol-specified post-treatment follow-up window. Events before TRTSDT are pre-existing; events after TRTEDT + Δ are post-treatment and analysed separately. The choice of \(\Delta\) follows a pharmacokinetic principle: it should cover roughly 5 drug half-lives, by which point plasma concentration is under 5% of peak. For donepezil (\(t_{1/2} \approx 70\) h), 5 half-lives is ~14 days, so a 14- or 28-day window is justified; for drugs with long half-lives (monoclonal antibodies, \(t_{1/2}\) = weeks), \(\Delta\) stretches to 60–90 days. Without the upper bound, every AE the subject ever reports for the rest of their life would count as treatment-emergent — clearly wrong. Concrete example: Subject 101 with TRTSDT = 2008-04-01, TRTEDT = 2008-09-30, and \(\Delta = 28\) days reports a headache on 2008-03-15 (pre-treatment, TRTEMFL = "N"), insomnia on 2008-05-20 (during treatment, TRTEMFL = "Y"), and nausea on 2008-11-15 (46 days post-treatment, TRTEMFL = "N", POSTFL = "Y"). Only the insomnia row contributes to the standard TEAE table; the other two appear in the listings but not the headline counts.

A clinical lab value like serum creatinine has a baseline distribution spanning roughly 0.5–1.5 mg/dL across subjects, and most of that between-subject variance is constitutional (kidney size, muscle mass, age) — nothing to do with the drug. Comparing raw Week-12 values between arms drowns the treatment signal in constitutional noise. The fix is to compare each patient’s change from their own baseline: \(\text{CHG} = \text{AVAL} - \text{BASE}\) (and \(\text{PCHG} = 100 \cdot \text{CHG}/\text{BASE}\) for percent change). Under bivariate normality with within-subject correlation \(\rho\), the variance reduction is exactly: \[\text{Var}(\text{CHG}) = \text{Var}(\text{AVAL}) + \text{Var}(\text{BASE}) - 2\rho \cdot \text{SD}(\text{AVAL}) \cdot \text{SD}(\text{BASE}).\] For lab markers with typical \(\rho \approx 0.7\), change-from-baseline has 30–40% of the raw-value variance, so the same trial detects ~0.6× the effect size at the same power. Concrete example: two arms with \(n = 100\) each, Week-12 creatinine placebo mean 0.95 (SD 0.20) vs active 0.85 (SD 0.20). The raw two-sample t has SE 0.028 and standardised effect 3.5; with \(\rho = 0.7\), Var(CHG) = 0.024, SE = 0.022, standardised effect 4.5. Same data, sharper conclusion. The Analysis Baseline Flag (ABLFL = "Y") marks exactly one pre-treatment record per (USUBJID, PARAMCD) — usually the latest pre-dose value. ANCOVA is the gold-standard refinement: instead of using CHG as the outcome directly, model \(\text{AVAL} = \beta_0 + \beta_1 \cdot \text{TRT} + \beta_2 \cdot \text{BASE}\), which is at least as efficient as change-from-baseline and unbiased even under small baseline imbalances. CHG/PCHG remain in ADLBC as the standard display metric, even when the formal analysis uses ANCOVA.

Show code 
adae <- ae |>
  derive_vars_merged(
    dataset_add = adsl,
    by_vars     = exprs(STUDYID, USUBJID),
    new_vars    = exprs(TRTSDT, TRTEDT, TRT01A)
  ) |>
  mutate(
    ASTDT = convert_dtc_to_dt(AESTDTC),
    AENDT = convert_dtc_to_dt(AEENDTC)
  ) |>
  derive_var_trtemfl(
    new_var        = TRTEMFL,
    start_date     = ASTDT,
    end_date       = AENDT,
    trt_start_date = TRTSDT,
    trt_end_date   = TRTEDT
  )

adlbc <- lb |>
  filter(LBCAT == "CHEMISTRY") |>
  derive_vars_merged(adsl, by_vars = exprs(STUDYID, USUBJID),
                     new_vars = exprs(TRTSDT, TRT01A)) |>
  mutate(PARAMCD = LBTESTCD, AVAL = LBSTRESN, AVISIT = VISIT, ABLFL = LBBLFL) |>
  derive_var_base(by_vars = exprs(STUDYID, USUBJID, PARAMCD)) |>
  derive_var_chg() |>
  derive_var_pchg()

cat(sprintf("ADAE: %d records (%d treatment-emergent) | ADLBC: %d chemistry records\n",
            nrow(adae), sum(adae$TRTEMFL == "Y", na.rm = TRUE), nrow(adlbc)))
ADAE: 1191 records (855 treatment-emergent) | ADLBC: 32740 chemistry records

5.1 Same derivations in SAS

sas/adae.sas
/******************************************************************************
 * Program  : adae.sas
 * Purpose  : Derive ADAE (Adverse Events analysis dataset) per
 *            ADaMIG-OCCDS v1.1.
 * Inputs   : SDTM.AE, ADAM.ADSL
 * Output   : ADAM.ADAE
 * Notes    : Occurrence-level structure (one record per AE per subject).
 *            TRTEMFL is the treatment-emergent flag — gated on AESTDT
 *            relative to TRTSDT/TRTEDT.
 *****************************************************************************/

%include "setup.sas";

proc sql;
  create table _ae as
    select a.*,
           input(a.AESTDTC, ?? e8601da.) as AESTDT format=date9.,
           input(a.AEENDTC, ?? e8601da.) as AEENDT format=date9.,
           s.TRTSDT, s.TRTEDT, s.TRT01A, s.SAFFL
    from sdtm.ae a
         left join adam.adsl s
           on a.STUDYID = s.STUDYID and a.USUBJID = s.USUBJID;
quit;

data adam.adae (label="Adverse Events Analysis Dataset");
  set _ae;

  /* Treatment-emergent: AE start on/after TRTSDT and on/before TRTEDT+30d */
  length TRTEMFL $1;
  if not missing(AESTDT) and not missing(TRTSDT) then do;
    if AESTDT >= TRTSDT and (missing(TRTEDT) or AESTDT <= TRTEDT + 30)
      then TRTEMFL = "Y";
    else TRTEMFL = "N";
  end;

  /* Analysis study day */
  if not missing(AESTDT) and not missing(TRTSDT) then
    ASTDY = AESTDT - TRTSDT + (AESTDT >= TRTSDT);
  if not missing(AEENDT) and not missing(TRTSDT) then
    AENDY = AEENDT - TRTSDT + (AEENDT >= TRTSDT);

  keep STUDYID USUBJID AESEQ
       AETERM AEDECOD AEBODSYS
       AESEV AESER AEREL AEOUT
       AESTDT AEENDT ASTDY AENDY
       TRT01A TRTEMFL SAFFL;
run;

proc sort data=adam.adae; by STUDYID USUBJID AESEQ; run;

sas/adlbc.sas
/******************************************************************************
 * Program  : adlbc.sas
 * Purpose  : Derive ADLBC (Laboratory Chemistry, Basic Data Structure)
 *            per ADaM IG v1.3 §4 (BDS).
 * Inputs   : SDTM.LB, ADAM.ADSL
 * Output   : ADAM.ADLBC
 * Notes    : One record per USUBJID per PARAMCD per AVISIT.
 *            ABLFL = "Y" on the last non-missing record on/before TRTSDT.
 *            CHG and PCHG are derived from the baseline AVAL.
 *****************************************************************************/

%include "setup.sas";

proc sql;
  create table _lb as
    select l.STUDYID, l.USUBJID, l.LBSEQ,
           l.LBTESTCD as PARAMCD,
           l.LBTEST   as PARAM,
           l.LBSTRESN as AVAL,
           l.LBSTRESC as AVALC,
           l.LBSTRESU as AVALU,
           l.VISIT    as AVISIT,
           l.VISITNUM as AVISITN,
           input(l.LBDTC, ?? e8601da.) as ADT format=date9.,
           s.TRTSDT, s.TRTEDT, s.TRT01A, s.SAFFL
    from sdtm.lb l
         left join adam.adsl s
           on l.STUDYID = s.STUDYID and l.USUBJID = s.USUBJID
    where l.LBCAT = "CHEMISTRY";
quit;

proc sort data=_lb; by STUDYID USUBJID PARAMCD ADT; run;

/*----- Baseline flag: last record on/before TRTSDT ----------------------*/
data _bl;
  set _lb;
  by STUDYID USUBJID PARAMCD;
  retain _last_pre _last_aval;
  if first.PARAMCD then do; _last_pre = .; _last_aval = .; end;
  if not missing(ADT) and not missing(TRTSDT) and ADT <= TRTSDT
     and not missing(AVAL) then do;
    _last_pre  = ADT;
    _last_aval = AVAL;
  end;
  if last.PARAMCD then output;
  keep STUDYID USUBJID PARAMCD _last_pre _last_aval;
run;

data adam.adlbc (label="Laboratory Test Results - Chemistry");
  merge _lb (in=in_lb)
        _bl (rename=(_last_pre=BLDT _last_aval=BASE));
  by STUDYID USUBJID PARAMCD;

  length ABLFL $1;
  if not missing(ADT) and not missing(BLDT) and ADT = BLDT and AVAL = BASE
    then ABLFL = "Y";

  if not missing(AVAL) and not missing(BASE) then do;
    CHG  = AVAL - BASE;
    if BASE ne 0 then PCHG = (CHG / BASE) * 100;
  end;

  keep STUDYID USUBJID PARAMCD PARAM AVAL AVALC AVALU
       AVISIT AVISITN ADT ABLFL BASE CHG PCHG
       TRTSDT TRT01A SAFFL;
run;

proc sort data=adam.adlbc; by STUDYID USUBJID PARAMCD AVISITN; run;

6 Write conformant XPTs

Show code 
# Write a v5 SAS transport (XPT) file. A full submission package drives
# variable-level metadata (length, label, type, format) from the define-XML
# spec via {xportr} + {metacore}; that conformance layer is the next step once a
# define-XML extract is wired in. Here we write the analysis-ready ADSL directly.
haven::write_xpt(adsl, file.path(DATA_OUT, "adsl.xpt"), version = 5)
cat("Wrote", file.path(DATA_OUT, "adsl.xpt"), "\n")
Wrote C:/Users/pauli/PaulinaDelMundoMD/paulinadelmundomd/traces/applied/03-cdisc-adam-pilot/data/adam_r/adsl.xpt

7 Demographics — ICH E3 §14.2.1

On the pathway → 06 · Recommendation: reporting standards, the conventions (here ICH E3) that make a result auditable rather than taken on faith.

ICH E3 — Structure and Content of Clinical Study Reports (CPMP/ICH/137/95) is the harmonised template FDA, EMA, and PMDA all expect for the Clinical Study Report. Section 14.2.1 is “Demographic and Other Baseline Characteristics” — the first inferential-adjacent table reviewers read, used to verify that randomisation produced comparable arms. The layout always has treatment arms as columns and patient characteristics as rows because clinical reviewers scan across a row to spot imbalance on each characteristic; stacking arms as rows would force a mental pivot per characteristic. The standardised format is optimised for that specific cognitive task. Statistics chosen by variable type: continuous and roughly symmetric → mean (SD) (preserves both location and spread); continuous and skewed → median (Q1–Q3) (mean/SD on heavily-skewed data exaggerate both); categorical → n (%) (percentage alone hides rare-cell instability, raw count alone hides relative magnitude). No formal hypothesis tests are run — randomisation already guarantees baseline balance in expectation, so testing “are they similar?” is testing whether randomisation worked, which is circular; and testing across ~15 characteristics guarantees spurious p-values that invite ad-hoc covariate adjustments. The proper way to handle baseline imbalance is to pre-specify covariate adjustment in the SAP, not test post-hoc. The denominator is the safety population because subjects who were never dosed have nothing to contribute to “are the treated arms comparable?” A typical layout for this trial would show, say, Age (mean SD) 75.2 (8.1) / 74.8 (7.9) / 75.5 (8.4) across the placebo and two Xanomeline arms — well within trivial noise; reviewer scans, confirms randomisation worked, and moves to the efficacy tables.

Show code 
adsl |>
  filter(SAFFL == "Y") |>
  select(TRT01A, AGE, AGEGR1, SEX, RACE, ETHNIC) |>
  gtsummary::tbl_summary(
    by = TRT01A,
    statistic = list(
      all_continuous()  ~ "{mean} ({sd})",
      all_categorical() ~ "{n} ({p}%)"
    ),
    digits = list(all_continuous() ~ 1)
  ) |>
  gtsummary::add_overall() |>
  gtsummary::modify_caption("**Table 14-2.01** — Demographic and baseline characteristics, safety population")
Table 14-2.01 — Demographic and baseline characteristics, safety population
Characteristic Overall
N = 1681		 Xanomeline High Dose
N = 721		 Xanomeline Low Dose
N = 961
Age 75.0 (8.1) 73.8 (7.9) 76.0 (8.1)
AGEGR1


    <65 19 (11%) 11 (15%) 8 (8.3%)
    >80 47 (28%) 12 (17%) 35 (36%)
    65-80 102 (61%) 49 (68%) 53 (55%)
Sex


    F 90 (54%) 35 (49%) 55 (57%)
    M 78 (46%) 37 (51%) 41 (43%)
Race


    AMERICAN INDIAN OR ALASKA NATIVE 1 (0.6%) 1 (1.4%) 0 (0%)
    BLACK OR AFRICAN AMERICAN 15 (8.9%) 9 (13%) 6 (6.3%)
    WHITE 152 (90%) 62 (86%) 90 (94%)
Ethnicity


    HISPANIC OR LATINO 9 (5.4%) 3 (4.2%) 6 (6.3%)
    NOT HISPANIC OR LATINO 159 (95%) 69 (96%) 90 (94%)
1 Mean (SD); n (%)

8 Reconciliation: SAS vs R vs CDISC reference

On the pathway → ∗ · Defend it: double-programming reconciliation, an independent re-derivation that tries to break the result rather than confirm it.

The closing chunk pulls (a) the ADSL produced by sas/adsl.sas, (b) the ADSL produced above in R, and (c) the reference ADSL CDISC published with the pilot, sorts all three on USUBJID, normalizes column order via the metacore spec, and asserts they are identical. It carries eval: false here because it needs the SAS OnDemand outputs, which are not part of this repo; once those are wired in, a non-empty diff would fail the render.

Show code 
adsl_sas <- haven::read_xpt(file.path(SAS_OUT,  "adsl.xpt"))
adsl_r   <- haven::read_xpt(file.path(DATA_OUT, "adsl.xpt"))
adsl_ref <- haven::read_xpt(file.path(DATA_REF, "adsl.xpt"))

normalize <- function(d) d |> arrange(USUBJID) |> select(sort(names(d)))

stopifnot(
  identical(normalize(adsl_sas), normalize(adsl_r)),
  identical(normalize(adsl_r),   normalize(adsl_ref))
)

9 Pinnacle 21 conformance

In a full submission package, the {xportr} writers enforce variable-level metadata (length, type, label, format) against the metacore spec — the same rule families Pinnacle 21 community edition checks — and the export is then run through Pinnacle 21 for a conformance report. That metadata-driven conformance layer is the next build here (it needs a define-XML extract wired into {metacore}); the XPT written above is the analysis-ready data, not yet a metadata-validated submission file.

10 Limitations

  • Cross-language reconciliation is specified, not run. The R derivations execute on render against {pharmaversesdtm}, but the closing reconciliation chunk — which would diff the SAS, R, and CDISC-reference ADSLs — carries eval: false because it needs the SAS OnDemand outputs, which are not part of this repo. Until those are wired in, the SAS half is a complete submission-style specification rather than executed output, and the byte-for-byte SAS-vs-R check is the remaining deliverable.
  • Fictional data. No efficacy or safety claim is made about Xanomeline from this analysis; the CDISC Pilot is a teaching artifact, not a real trial.
  • One TLF, not the full SAP. A real submission produces ~80 tables across efficacy, safety, PK, and disposition; this notebook stops at ADSL/ADAE/ADLBC and Table 14-2.01.
  • No PGx layer in v1. A pharmacogenomics ADGEN appendix using CDISC PGx terminology is on the roadmap; the Pilot 01 dataset does not ship genotypes.

11 References

12 Appendix — SAS source

All six SAS programs embedded verbatim, in the run order documented in sas/README.md. Click any filename heading to open the file on GitHub.

12.1 sas/setup.sas

sas/setup.sas
/******************************************************************************
 * Program  : setup.sas
 * Purpose  : Define libnames and the project format catalog for CDISC Pilot 01.
 * Inputs   : ../data/sdtm/, ../data/adam_r/ (R reference output)
 * Outputs  : LIBNAME SDTM, LIBNAME ADAM, LIBNAME RREF, FMTLIB.PILOT
 * ADaM IG  : —
 *****************************************************************************/

%let proj = /home/&sysuserid/cdisc-pilot;   /* SAS OnDemand layout */

libname sdtm xport "&proj/data/sdtm.xpt"  access=readonly;
libname adam      "&proj/adam";
libname rref xport "&proj/data/adam_r.xpt" access=readonly;

proc format library=adam.fmtlib;
  value $sex   "M" = "Male" "F" = "Female";
  value $arm   "Placebo"          = "Placebo"
               "Xanomeline Low Dose"  = "Donepezil 5 mg"
               "Xanomeline High Dose" = "Donepezil 10 mg";
  value agegrn 1 = "<65" 2 = "65-80" 3 = ">80";
run;

options fmtsearch=(adam.fmtlib);

12.2 sas/t_14_2_01_demog.sas

sas/t_14_2_01_demog.sas
/******************************************************************************
 * Program  : t_14_2_01_demog.sas
 * Purpose  : Table 14-2.01 — Demographic and Baseline Characteristics
 *            (ICH E3 §14.2.1), safety population, by treatment arm.
 * Inputs   : ADAM.ADSL
 * Output   : output/t_14_2_01_demog.rtf
 *****************************************************************************/

%include "setup.sas";

ods listing close;
ods rtf file="output/t_14_2_01_demog.rtf" style=journal;

title1 "CDISC Pilot 01 — Donepezil in Mild to Moderate Alzheimer's Disease";
title2 "Table 14-2.01  Demographic and Baseline Characteristics  (Safety Population)";

proc tabulate data=adam.adsl (where=(SAFFL = "Y")) format=8.1 missing;
  class TRT01A AGEGR1 SEX RACE;
  var   AGE;
  table (AGEGR1='Age group (years)'  all='Total')
        (SEX  ='Sex'                 all='Total')
        (RACE ='Race'                all='Total'),
        (TRT01A='' all='Overall') * (n='n' colpctn='%' * f=5.1)
        AGE  =' ' * (mean='Mean' std='SD' median='Median' min='Min' max='Max');
run;

ods rtf close;
ods listing;

12.3 sas/xpt_export.sas

sas/xpt_export.sas
/******************************************************************************
 * Program  : xpt_export.sas
 * Purpose  : Export ADaM datasets to SAS V5 transport (XPT) for the FDA.
 *            These are the files read by the Quarto reconciliation chunk.
 * Inputs   : ADAM.ADSL, ADAM.ADAE, ADAM.ADLBC
 * Outputs  : output/adsl.xpt, output/adae.xpt, output/adlbc.xpt
 *****************************************************************************/

%include "setup.sas";

%macro export_xpt(ds=);
  libname _x xport "output/&ds..xpt";
  data _x.&ds; set adam.&ds; run;
  libname _x clear;
%mend;

%export_xpt(ds=adsl);
%export_xpt(ds=adae);
%export_xpt(ds=adlbc);

Built with Quarto, {admiral}, and SAS OnDemand for Academics. Source: GitHub.

Note

Two ways to go further. Learn the methods: the From Data to Bedside pathway → walks the theory behind work like this rung by rung, and a free account follows new analyses as they go up. Use them on your data: book a discovery call → for real-world-data analysis, causal evaluation, and submission-grade statistics.