ADaM OCCDS: Treatment Flags in ADAE

Welcome to Day 3 of Quality Control (QC). The safety table reports 312 treatment-emergent adverse events, but the independent QC program counts 315. That three-event gap is our story. All three disputed rows have onset before first dose, yet they were flagged treatment-emergent because the flag compared onset against the randomization date instead of the treatment start date variable, TRT01SDT. This defect is invisible in the code but visible in the counts, which is why reviewers run date-comparison listings, not code reviews alone. Our agenda covers the Analysis Data Model (ADaM) Occurrence Data Structure (OCCDS) grain, a disciplined adverse events to Subject-Level Analysis Dataset (ADSL) merge, the treatment-emergent flag, TRTEMFL, anchored to TRT01SDT, partial-date rules, and the QC checks that catch each defect.

Now look at the Occurrence Data Structure, or OCCDS. In the Analysis Data Model, or ADaM, we work with three grains. The Subject-Level Analysis Dataset, or ADSL, holds one row per subject, while the Basic Data Structure, or BDS, holds one row per subject-parameter-timepoint. The Adverse Events Analysis Dataset, or ADAE, is an OCCDS dataset, so it carries exactly one row per adverse event; five events from one subject give five rows. Collapsing ADAE to one row per subject, or per body system, is the defining OCCDS mistake; those shapes belong in summary tables derived from ADAE, never inside ADAE itself.

Building ADAE, the Adverse Events Analysis Dataset, means merging adverse event data from SDTM, the Study Data Tabulation Model, with ADSL, the Subject-Level Analysis Dataset. The AE domain sits on the left with many rows per subject, and ADSL sits on the right with one row per subject; they join BY the unique subject identifier, or USUBJID. In a Statistical Analysis System, or SAS, merge, ADSL supplies TRT01SDT, TRT01EDT, TRT01A, and SAFFL onto every AE row, so the first-dose date arrives through TRT01SDT. Then parse character dates to numeric with `aestdt = input(aestdtc, yymmdd10.);` but length-check the string first, because a partial date parsed as complete is a silent defect. A subject in AE but not in ADSL is a data-management listing, never rows to drop silently; subjects in ADSL with no events are normal and contribute only to denominators.

The treatment-emergent flag, TRTEMFL, answers a treatment question, not a randomization question: did the adverse event, or AE, begin on or after first dose? Under the Statistical Analysis Plan, or SAP, convention we compare ASTDT to TRT01SDT, and a grace window is added only when the SAP defines one explicitly. The anchor is first dose, never randomization, so an AE starting between randomization and first dose is not treatment-emergent, which is exactly the defect from Scene 1. In SAS, the canonical flag block reads: if aestdt >= trt01sdt and (aeendt = . or aestdt <= trt01edt + 30) then trtemfl = 'Y'; else trtemfl = 'N';. The first treatment-emergent event flag and the first preferred term occurrence flag both come from sort order: by usubjid descending trtemfl aestdt aeseq; then if first.usubjid and trtemfl='Y' then aoccfl='Y';.

AESTDTC is text in the International Organization for Standardization 8601 format, so a partial day or month will not parse as a date. Before the treatment-emergence comparison, impute an analysis start date into ASTDT exactly as your Statistical Analysis Plan prescribes, and record that decision in an AESTDTCF-style flag. Impute first, then compare; comparing raw partial text to TRT01SDT is a parse defect. Plans legitimately differ: some accept the earliest plausible date, others require a complete date before calling an event treatment-emergent. Transcribe the rule you were given; do not choose one. Finally, the Medical Dictionary for Regulatory Activities terms, AEDECOD, AEBODSYS, and the rest of the hierarchy, ride through from Adverse Events untouched, with one dictionary version per data-cut snapshot.

Welcome to this checkpoint on Occurrence Data Structure (OCCDS), merge direction, and the treatment-emergent anchor in Analysis Data Model (ADaM) adverse-events datasets. In the adverse-events analysis dataset (ADAE), the record grain is one row per adverse event occurrence, so if one subject has five adverse event (AE) records in Study Data Tabulation Model (SDTM) AE, how many ADAE rows will that subject contribute before filtering or merging with Subject-Level Analysis Dataset (ADSL)? The correct answer is B: five rows, because ADAE retains one row per adverse event occurrence, and Unique Subject Identifier (USUBJID) is not unique in ADAE. You build ADAE by merging ADSL with an adverse-event input by USUBJID, but a subject appears in the adverse-event input with no matching ADSL row; assuming ADAE must contain AE records only for subjects represented in ADSL, which Statistical Analysis System (SAS) IN= condition prevents the unmatched AE rows from entering ADAE? The correct answer is B: require both ADSL and the AE input with if inADSL and inAE;, because the unmatched AE rows have inADSL=0 and inAE=1, so they are rejected. Which statements correctly describe treatment-emergent flag (TRTEMFL) derivation and partial-date discipline in ADAE? The correct answers are A and D: for TRTEMFL, Analysis Start Date (ASTDT) is compared with Treatment Start Date (TRT01SDT), and an AE is treatment-emergent when the AE analysis start date is on or after the first treatment date, not randomization date, because a subject can be randomized before the first dose; and partial dates require a planned, documented imputation rule established by the study statistician or lead statistical programmer and recorded in the ADaM specification or metadata. Ad hoc per-record choices are not reproducible, and an AE beginning before TRT01SDT is not treatment-emergent.

Let's walk through the demo subject, who has five adverse event (AE) rows in the extract. The first two, AESEQ 1 and 2, start on 2019-05-20; the last three, AESEQ 3 through 5, start on 2019-05-17. AESEQ 3 through 5 began during treatment with ASTDY 10, while AESEQ 1 and 2 began in safety follow-up with ASTDY 13. All five have no AEENDTC, so the treatment emergent flag, TRTEMFL, is Y for every row, even though AESEQ 1 and 2 start after the treatment end date, TRTEDT, 2019-05-19. The trace is one-to-one: five Study Data Tabulation Model (SDTM) adverse event rows become five Adverse Events Analysis Dataset (ADAE) rows.

This scene is a hands-on exercise, so you will work through it on the website at jaimeyan.com/learn rather than just watching me do it. You will practice completing the Subject-Level Analysis Dataset (ADSL) KEEP list so the MERGE carries the right treatment dates and safety flag, checking that the BY variable is usubjid in both datasets, and then you will build the treatment emergent condition and apply it to the adverse event rows in the extract. Please try it right after this video, because repairing the flag yourself is what turns a count you hope is right into one you can defend.

In the Statistical Computing Environment (SCE), the Adverse Events Analysis Dataset (ADAE) build is spec-driven: the specification owns the treatment-emergent flag (TRTEMFL) name, the emergence convention, and the analysis populations, and define.xml comes from that same metadata. The program transcribes; it does not invent flag logic. MedDRA, the Medical Dictionary for Regulatory Activities, is a versioned artifact: the coding run is recorded, and the ADAE build stamps one version per data-cut snapshot. For quality control (QC), run independent double programming, attach the recurring defect listings to the review, and check structural conformance on every build. A Statistical Analysis System (SAS) habit: run proc freq data=adae_1; tables trtemfl / missing; run; then compare any-AE and treatment-emergent-AE subject counts. In R or Python shops, admiral and admiralpython expose the same occurrence logic — one spec stays the source of truth.

An agent can draft a clean one-pass adverse event to subject-level analysis dataset merge, and that part it does well. What you must not outsource is the emergence convention and the partial-date imputation rule: a one-day grace window, or imputing a month-only date to day-01, both look sane and leave a clean log. Only a diff against the actual sentence in the Statistical Analysis Plan exposes the invention, so ask the agent to state the emergence rule it implemented and its source before anyone reads the code. Then hand-check three borderline subjects: one pre-dose onset, one month-only onset, and one event dated the same day as first dose. The accountability boundary stays put, whoever signs the quality control document owns the flag logic. This level three guidance was last verified 2026-08-30, so re-verify before relying on agent-specific claims.

This checkpoint quizzes you on the Analysis Data Model (ADaM) occurrence data structure (OCCDS) grain, on quality control (QC) catches in the Adverse Events Analysis Dataset (ADAE), and on ownership of treatment-emergence decisions. For the demo subject with five adverse events, the ADaM OCCDS grain gives answer B: one row per adverse event, so five rows in the ADAE. In QC of the ADAE build, the correct choices are A, B, C, and D. A Confusion record with analysis start date (ASTDT) equal to 2019-05-17 and treatment start date (TRT01SDT) equal to 2019-05-08 should have treatment-emergent flag (TRTEMFL) set to 'Y'; a Y flag with ASTDT before TRT01SDT belongs in the date-comparison listing; a missing ASTDT when adverse event start date/time character (AESTDTC) exists belongs in the parse-failure listing; and a subject split between ADAE and Subject-Level Analysis Dataset (ADSL) belongs in an anti-join check. The treatment-emergence and partial-date decisions are owned by the approved Statistical Analysis Plan (SAP), not by artificial intelligence (AI) agent-drafted code, and the person who signs the QC document must verify that the SAP decision rules were faithfully transcribed into the ADAE derivation before accepting the output.

Takeaways: the Adverse Events Analysis Dataset, or ADAE, in the Analysis Data Model Occurrence Data Structure, or OCCDS, keeps one row per occurrence. Treatment emergence compares Analysis Start Date, or ASTDT, against Treatment Start Date, or TRT01SDT, per the Statistical Analysis Plan, or SAP — never randomization; partial dates are imputed per SAP and flagged. Reviewer checks: anti-join ADAE not in Subject-Level Analysis Dataset, or ADSL; date comparison Treatment Emergent Flag, or TRTEMFL, equals Y with ASTDT before TRT01SDT; parse failure missing ASTDT where start date exists; version trace Medical Dictionary for Regulatory Activities, or MedDRA, drift; key uniqueness duplicate rows post-MERGE. Pin one MedDRA version per snapshot; run Quality Control, or QC, as independent double programming. Remember the 312-vs-315 gap: unsupported flags are caught by date listings. Next is Part 9, the Time-to-Event Analysis Dataset, or ADTTE.