29  Exercise 10: The Analysis Data Reviewer’s Guide

29.1 Goal

Close the conformance finding the session reports, at its source rather than in the output, and then write the one narrative ADRG section this course can honestly ask you to write, because you have the material for it.

29.2 Setup

Work from the project root:

library(tidyverse)
library(xml2)
library(openxlsx)
library(defineR)

adrg <- list.files("data/adrg", full.names = TRUE) |>
  set_names(~ basename(.x)) |>
  map(read_csv, show_col_types = FALSE)

29.3 Task

1. (Guided) Close the rule-73 finding. The session’s section 6.2 reports seventeen variables with Origin Type="Derived" and no MethodOID. Thirteen are dataset-level and four are value-level.

   Fix it where it belongs. Work on a copy of data/spec/SDTM_METADATA.xlsx, add the missing entries to COMPUTATION_METHOD, reference them from COMPUTATIONMETHODOID in VARIABLE_METADATA, regenerate with R/build_define.R, and re-run the check.

   The derivations must describe what R/build_sdtm.R actually does. Read it. Do not write “derived per SAP” seventeen times; a method that says nothing is worse than an honest finding, because it suppresses the finding without supplying the information.

   Then answer: LBDRVFL is one of the seventeen and it is null on every record in the study. What should its computation method say, and is Derived even the right origin for it? Justify from the Define-XML origin definitions rather than from convenience.

2. (The trap) Regenerate everything and find what moved. Re-run R/build_adrg_components.R after your fix.

   Section 6.2 should now be clean. Check whether anything else changed, and explain each difference you find. In particular, look at the standards inventory. It reads the Define-XML version out of the regenerated file rather than restating it, which is either a safeguard or a trap depending on what you changed.

   Then answer: the session argues the ADRG’s generated tables should never be hand-maintained. You have just regenerated six tables from one workbook edit. What would have gone wrong if section 6.2 had been a paragraph someone typed?

3. (Semi-guided) Write section 3.4. Visit windowing and record selection is a narrative section, and it is the one narrative section you have everything you need for, because ADLB established the facts and argued the decision.

   Write it. Real prose, aimed at a reviewer who does not know this study. It must cover:

   • Whether visit windowing was used, and what AVISIT actually is in GLPX-1
   • How the baseline record is selected, stated precisely enough that a reader could reproduce it
   • That ABLFL disagrees with SDTM’s LBLOBXFL on 342 subject-parameter pairs, why, and what it does to the headline result

   Aim for 200–300 words. Then answer: which sentence in what you wrote could a script have generated, and which could not? That split is the whole session.

29.4 Acceptance criteria

• All seventeen findings closed: thirteen COMPUTATION_METHOD entries referenced from VARIABLE_METADATA, and four from VALUELEVEL_METADATA, each describing the actual derivation in R/build_sdtm.R
• Regenerated define.xml still validates against the 2.0 schema
• The rule-73 check returns zero findings
• The LBDRVFL origin question answered from the Define-XML origin definitions
• All six ADRG component tables regenerated, with every change accounted for
• A 200–300 word section 3.4 covering windowing, baseline selection and the ABLFL/LBLOBXFL divergence with its numbers
• The generated-versus-written sentence split identified

TipHint

Task 1: the thirteen dataset-level findings are mostly DM reference dates and LB standardised results; the four value-level ones are the LBSTRESN definitions, and they are fixed in VALUELEVEL_METADATA rather than VARIABLE_METADATA. R/build_sdtm.R has the logic; study_day() in particular is worth quoting rather than paraphrasing. For LBDRVFL, re-read what Findings established about why it is null, then check the Define-XML Type definitions - Derived means a value “calculated by an algorithm or reproducible rule”, and a variable with no values has not been calculated by anything.

Task 2: R/build_adrg_components.R reads the Define-XML version from the MetaDataVersion element. If your regeneration changed it, the inventory changes with it and that is the safeguard working. If you edited DEFINE_HEADER_METADATA you may have changed something you did not intend.

Task 3: the numbers are all in ADLB, the 342 pairs, the ADY <= 1 rule, and the -1.116 versus -1.138 comparison. The Completion Guidelines section 3.4 lists the questions the section must answer; answer those, in order, in prose. Resist writing a paragraph about why baseline definitions matter in general: a reviewer wants this study’s rule, not a tutorial.

Solutions are in the paid tier.