ADTTE: Time-to-Event in ADaM

Picture a quality control review of an overall survival analysis in an Analysis Data Model time-to-event dataset, where the Kaplan–Meier curves cross and the hazard ratio favors the control arm. The statistician asks one question: which CNSR value do the death records carry? In the Analysis Data Model, CNSR equals one means censored and CNSR equals zero means event — that rule is fixed, never inferred from how yes or no flags behave elsewhere. If death rows carry CNSR equals one, the analysis models censoring instead of the event, so the curves flip, the hazard ratio reverses, and every conclusion reverses while the Kaplan–Meier plot still looks normal. Today we cover the time-to-event dataset anatomy, the statistical analysis plan rules, the censoring cascade, real rows from our demo study, and the two quality control listings that catch a broken dataset.

Here is the anatomy of an Analysis Data Model, ADaM, Time-to-Event dataset, or ADTTE. The grain is one record per unique subject identifier, USUBJID, per analysis, with parameter code PARAMCD and parameter PARAM. One study can run overall survival, OS, and progression-free survival, PFS, off the same machinery. The start date variable STARTDT is the time-to-event origin, analysis date ADT is the event or censoring date, censor CNSR is the 0 or 1 flag, and analysis value AVAL is the analysis duration per the statistical analysis plan, SAP. Event Description, EVNTDESC, and Censoring Date Description, CNSDTDSC, state what the event or censoring was and why a censored date was chosen. SRCDOM, SRCVAR, and SRCSEQ, source domain, variable, and sequence, record the source for ADT for audit, while treatment and population flags merge from ADSL, the Subject-Level Analysis Dataset.

Before any ADaM, Analysis Data Model, code is written, the three decisions that shape a time-to-event dataset already live in the SAP, the Statistical Analysis Plan. What counts as an event, when to censor, and which date starts the clock are SAP content, not programmer preference. A typical SAP says the event is the first occurrence per the adjudicated domain, captured in ADT, the analysis date. If there is no event, censor at last known contact and set CNSR, the censoring flag, to one; the origin date, first dose, is carried in STARTDT. Every ADT, CNSR, and AVAL, the analysis value, traces to SRCDOM, SRCVAR, and SRCSEQ, the source domain, variable, and sequence number. Never invent a censoring rule, a made-up rule is an inspection finding, so when the SAP is silent, raise a spec query instead.

When a subject has no event, the cascade assigns an analysis date, ADT — the first applicable rule wins. Priority one: an event on record gives ADT as the event date, sets the censor flag, CNSR, to zero, and uses the event record as source. Priority two: discontinued the study gives ADT as discontinuation date, CNSR one, reason from the Disposition domain, DS. Priority three: known alive at last contact means ADT is the last known alive date from the Subject-Level Analysis Dataset, ADSL, variable LSTALVDT, with CNSR one. Priority four: otherwise, ADT is the data cut date, CNSR one, reason data cut. This enforces one censoring date, one reason, and one traceable source row using source domain, SRCDOM, source variable, SRCVAR, and source sequence, SRCSEQ — and the priority order is statistical analysis plan, SAP, content.

CNSR, the censoring indicator, is an ADaM convention — Analysis Data Model — where zero marks an event and one marks censoring. In the derivation, state it explicitly: cnsr = 0; on the event branch, and cnsr = 1; on the censoring branch. Reverse that coding and you instead model the hazard of censoring, so the curves flip, the hazard ratio reverses direction, and every conclusion reverses with it. At the analysis step, declare the semantics as well: time tte * cnsr(1); tells PROC LIFETEST, the Lifetest procedure, which value is censored. PROC PHREG, the proportional hazards procedure, carries identical semantics. Duration runs from STARTDT, Start Date, to ADT, Analysis Date, and a mandatory guard rejects negative duration before analysis.

Now we turn to partial dates at the event. In an Analysis Data Model, or ADaM, time-to-event dataset, a partial event date must be imputed according to the Statistical Analysis Plan, the SAP, and that imputation must be flagged. The reason is a time-to-event, or TTE, wrinkle: the imputation direction can change whether the record counts as an event at all, or it can push ADT, analysis date, before STARTDT, study start date. Those consequences affect CNSR, the censoring indicator, and AVAL, the analysis value, so they belong to data review, not to a fallback branch that the programmer invents. For the extract supplied with this lesson, no partial-date rows exist, so no USUBJID, unique subject identifier, data row is shown for this rule.

This checkpoint reviews core concepts in the Analysis Data Model (ADaM) time-to-event analysis dataset, known as ADTTE. For the censor indicator, CNSR, if CNSR is 0, the correct interpretation is A: the subject had the event of interest, and ADT, the analysis date, is usually the event date, because CNSR equals 1 means the observation is censored. For the ADTTE quartet of STARTDT, ADT, AVAL, and CNSR, the correct choices are A, B, C, and D: STARTDT is the time-to-event origin date, ADT holds the event date when CNSR equals 0 or the censor date when CNSR equals 1, AVAL is the analysis value representing time from STARTDT to ADT in the analysis time unit, and CNSR marks whether the observation is censored. When a subject discontinued without the event of interest, the usual next cascade rule is B: set ADT to the date the subject discontinued and set CNSR to 1, because STARTDT is the origin date, not the censor date. For censoring rules, the Statistical Analysis Plan (SAP) defines the rules and priority order for choosing ADT, while the ADTTE programmer implements them; if the SAP is silent, the programmer must not improvise and should ask the study statistician or SAP owner for clarification, obtain an agreed rule, and document it.

This slide shows three overall survival, or OS, rows, taken verbatim from the time-to-event analysis data set, ADTTE, extract, and no patient values are invented. Row one is censored: its censor indicator, CNSR, equals one, and its analysis value, AVAL, is about 0.4928, with AVALU, the analysis value unit, set to months. Its start date, STARTDT, is 2019-05-08, and its analysis date, ADT, is 2019-05-22, sourced from the subject-level analysis dataset, ADSL, variable LSTALVDT, the last known alive date. The two death rows, CNSR equals zero, both point to the analysis data set for response, ADRS, with source variable SRCVAR of ADT and source sequence number SRCSEQ of four. Together with SRCDOM, the source domain, these columns act as audit instructions pointing to the upstream record behind each date.

You'll do this one hands-on on the website at jaimeyan.com/learn. The exercise asks you to read the censoring indicator (CNSR) on each overall survival (OS) row for the demo subject, decide whether it is an event or a censoring, and pick the matching event description (EVNTDESC) or censoring description (CNSDTDSC). You'll also trace each source domain (SRCDOM), source variable (SRCVAR), and source sequence number (SRCSEQ) trio back to its source record, then apply the two quality control (QC) listing rules to see whether any supplied row would be flagged, so give it a try right after the video.

Quality control, or QC, listings catch broken time-to-event analysis datasets, ADTTE. Listing A flags rows where ADT, the analysis date, is before STARTDT, the time-to-event origin, with CNSR, the censoring indicator, equal to zero. That means an event before origin, so every hit is a misassigned origin or source-date error; Listing B finds censored subjects where CNSR equals one but SRCDOM, SRCVAR, or SRCSEQ, source domain, variable, and sequence, are blank or dangling. For example, USUBJID, the unique subject identifier, 74001-002, has a death row pointing to an ADT record 4.0. Both listings catch defects reaching statisticians, and no structural conformance tool checks the cascade against statistical analysis plan. In the modern workflow, ADTTE builds late from the subject-level analysis dataset plus event-bearing data, and the SRC variables let independent QC re-derive ADT and diff it against the submission copy.

Welcome to the closing check, where we will read a real Analysis Dataset for Time-to-Event row and test three key points. For the checkpoint, inspect the Overall Survival, or OS, row in the Analysis Dataset for Time-to-Event, or ADTTE. The correct answer is A: Analysis Date, or ADT, Censor, or CNSR, Event Description, or EVNTDESC, and the source domain, variable, and sequence variables, or SRCDOM, SRCVAR, and SRCSEQ, together establish the analysis time contribution, event-versus-censor status, a clinical description of that status, and traceability to the source data record. If Censor, or CNSR, is accidentally flipped from 0 to 1 for an event observed on Analysis Date, or ADT, the correct answer is A: the subject's row no longer contributes an observed event to the numerator at the event time and is instead handled as censored, so the Kaplan-Meier, or KM, output may still run but the hazard interpretation changes. For the Quality Control, or QC, listings, the correct answers are A and B: the first listing targets event rows where Analysis Date, or ADT, occurs before Start Date, or STARTDT, the analysis origin, and the second targets censored rows with Censor, or CNSR, equal to 1 that have no matching source domain, variable, and sequence record to justify the censoring reason.

This closes the Analysis Data Model, or ADaM, line: ADTTE, the Time-to-Event Analysis Dataset, is one row per subject per analysis. STARTDT, the Time-to-Event Origin Date, sets time zero; ADT, the Analysis Date, holds the event or censoring date; and CNSR, the Censor variable, carries the semantics. CNSR equals zero for an event, one for censored; reversing those values silently reverses the hazard ratio while output still looks normal. The censoring cascade gives every censored subject one date, one reason, and one traceable source row in Statistical Analysis Plan, or SAP, priority order. Two Quality Control listings — events before STARTDT, and censored without a source row — catch what conformance tools cannot. As of 2026-08-30, verify which CNSR value means event, then hand-trace Source, or SRC, records; this closes the line from subject-level to basic and occurrence structures to time-to-event.