Reject Tied Baseline

Goal: mark the baseline record of each subject and parameter with ABLFL and repeat its value with BASE.

Input: collected laboratory results with test code, collection date, and numeric result (LBTESTCD, LBDTC, LBSTRESN), plus subject treatment start dates from ADSL (TRTSDT).

Variables:

• ADT would contain the collection date, taken from LBDTC.
• TRTSDT would contain the subject's treatment start date, taken from ADSL TRTSDT.
• AVAL would contain the analysis value, taken from LBSTRESN.
• ABLFL would be Y on the baseline record for the subject and parameter: the latest result on or before treatment start, and blank on every other record.
• BASE would contain the baseline record's AVAL on every record for the subject and parameter; missing when no result falls on or before treatment start.

Note: results sharing one collection date can tie for latest, and either candidate gives the parameter a different baseline. The run fails and no dataset is accepted.

How to fix

Decide which of the two draws is the baseline before choosing how to state it. Two results drawn on the same day are usually a sample and its repeat, and the source data system should carry the result the study reports, or a collection time that separates the draws. Correcting it there leaves the rule saying what it means: the latest result on or before treatment start.

When both draws are reportable and the study states a tie-break, choose the record explicitly instead of by date alone. Order the eligible results and mark the first:

- name: ABLRANK
  type: int
  derivation:
    row_number:
      window:
        group_by: [STUDYID, USUBJID, PARAMCD]
        order_by:
          - {variable: ADT, direction: desc}
          - {variable: LBSEQ, direction: desc}
        filter: "ADT <= TRTSDT"

- name: ABLFL
  type: str
  derivation:
    flag: "ABLRANK = 1"

Keep ABLRANK internal by omitting it from output.columns.

Order by the term the study names; the sequence number above stands in for it and is not itself a clinical rule. Do not average the two results into one baseline value, which reports a measurement nobody took.