From Mock Shell to RTF

Table 14.1.2 just came back from Quality Control, or QC, with a pass two finding: the column denominator doesn't match the population statement on the mock shell. Our program counted subjects with study exposure, but the shell was written against the safety population. Both numbers can be defensible, yet only one is the deliverable. That is the whole game of Tables, Listings, and Figures, or TLF, outputs: the mock shell is a contract, and most late-stage QC findings are contract violations, not math errors. Map the shipping path: read the mock shell, write the program plan, build the table with PROC REPORT and the Output Delivery System, or ODS, to Rich Text Format, or RTF, then run the four-pass QC with discrepancy records. This lesson uses simulated teaching data, and every row shown traces to the subject-level analysis dataset extract.

A mock shell is not decoration; it is a contract you can be checked against. It fixes the title block, column headers, a sample row or two, and the footnotes. The output identifier ties the shell to its filing location and the Statistical Analysis Plan (SAP) traceability. The title block commits you to exact wording and a named population. Each denominator N tells quality control (QC) which N feeds every percentage, and if the header N is wrong, every percentage beneath it inherits that defect. Footnotes carry methods, dictionary versions, and SAP section references, and a wrong dictionary version is the same severity as a wrong percentage. Most late-stage QC findings are contract violations, not arithmetic errors.

Before you write any code, read the shell as a program plan with six decisions in a fixed order. Step one is the population flag in the ADSL, the Subject-Level Analysis Dataset, choosing between the Safety Analysis Set, or SAF, the Intention-to-Treat, or ITT, population, or completers, and you record the population statement verbatim. This is the highest-risk decision because skipping it produces the classic denominator-class defect, the most common serious discrepancy on a first Quality Control, or QC, cycle. Denominators always come from ADSL under that flag, never from the event dataset, because an event-dataset count silently means subjects with events. Then fix row structure, column blocks, statistics, formats, footnotes, and pagination in that order. Remember, decimal precision is a shell property, not a programmer preference, so set it when reading the shell, not when formatting the output.

Let's check your understanding of the shell contract for tables, listings, or figures, or TLF. First, when you review a TLF shell, which element is the contract for the output file location and traceability back to the Statistical Analysis Plan, or SAP? The correct answer is B, the Output ID, because the Output ID carries the filing location and points back to the SAP item; titles, footnotes, and row labels define content or formatting, not the file-name or location contract. Second, for the safety-population table in the demo, event counts come from the adverse-event analysis dataset, or ADAE, and each column percent is n (%) subjects with the event. The correct denominator source is B: subject-level analysis dataset, or ADSL, records with the safety population, or SAF, flag, SAFFL='Y', grouped by the same treatment variable used in each column, because safety-population percentages must not be derived from event rows or from all randomized subjects regardless of the population flag. Third, when you read a TLF shell into a program plan before writing code, the correct choices are A, B, and D: confirm the ADSL population flag such as SAFFL='Y' before subsetting event records; decide whether zero-count or missing row categories must still be displayed as part of the shell contract; and translate the shell's row order and column layout into the sort keys and column mapping for PROC REPORT. Choosing a SAS version or an Output Delivery System, or ODS, Rich Text Format, or RTF, bookmark style is an environment or formatting decision, not a first-read shell-contract decision.

Five rows from the ADSL, the Subject-Level Analysis Dataset, become this table header. SAFFL, the safety population flag, is Y on all five rows, so the safety analysis set denominator is N equals 5. Treatment column denominators come from TRT01PN: subjects 001 through 003 give N equals 3, subject 004 gives one, subject 005 gives one, so three, one, one, not a pooled number. Read straight from the rows, SEX M counts three and SEX F counts two; AGEGR1 greater than or equal to 65 counts two, below 65 counts three. Now the trap: filtering DLTEVLFL equals Y, the dose-limiting toxicity evaluable flag, keeps only subjects 004 and 005, so N equals 2, which is defensible for a DLT-evaluable population but not the safety deliverable your shell names.

This scene shows the table pattern: PROC REPORT with pre-built cells. Analysis blocks are stacked, then one PROC REPORT lays them out by block, treatment arm, and column stat. For categorical variables, SQL builds the N per arm and level plus the denominator from the Subject-Level Analysis Dataset, ADSL; the cell is always a count over that treatment denominator, never another dataset. For continuous variables, PROC MEANS per arm writes mean, SD, median, min, max; the canonical macro loops AGE, HEIGHT, WEIGHT, but our extract has only AGE and SEX. Because numbers ship as pre-built strings, what Quality Control, QC, reads in the Rich Text Format, RTF, is exactly what the program wrote — no surprises. The title block inside the Output Delivery System, ODS, is the shell contract in code, including the simulated teaching data note and the table title.

Listings and figures are part of the Tables, Listings, and Figures (TLF) deliverable. A listing shows records, not summaries: one record per row, sort order exactly as the shell specifies, and page breaks controlled by the program. Repeated headers on continued pages are expected; reviewers look for the exact (Continued) convention from the shell. For figures, join the analysis data to the Subject-Level Analysis Dataset (ADSL) under the population flag, so only Safety Analysis Set (SAF) subjects reach an axis; place per-group N in the title or legend as the shell directs. Figures often fail Quality Control (QC) on boring details: wrong population, missing N counts, unlabelled groups, or category order that looks reasonable instead of following the shell. Remember, our extract has no adverse event or lab rows, so we teach these patterns as rules without fabricating patient rows.

We deliver through the Output Delivery System, or ODS, into Rich Text Format, or RTF. Titles and footnotes live in the ODS RTF block using title and footnote statements, so they repeat per page as specified. One style template, styles.learn_tlf based on styles.rtf, sets fonts and margins once for every Tables, Listings and Figures, or TLF, output: Courier New 8 point data, Arial 9 point titles. For Page x of y, use an ODS escape character and the PAGEOF construct in a footnote — reliable and safer than counting pages yourself. Precision follows the shell, so a one-decimal mean next to an integer count is a shell property, not a preference — Pass 4 of Quality Control, or QC, verifies that. A footnote with the program path or version makes each RTF traceable to its program and data cut.

The four-pass quality control (QC) order goes cheapest first: format, covering titles, headers, footnotes, and pagination, then data definitions: population flag, denominators, and visit windows per the statistical analysis plan. Each pass gates the next, so never recompute an output that would fail an earlier check. Pass three is the only pass needing real independent work: recompute one row per block, or write a separate QC program. Pass four checks consistency: same N, same terminology, same decimals. Snippet 09-qc-proc-compare sets the release gate: method equals absolute, criterion 1e-9, issue count zero; any nonzero count blocks release, and the log itself is evidence. Record before fixing: output ID, shell reference, observed versus expected, and a named disposition with approver; for agents, require the three closest mismatches they rejected; this was verified 2026-08-30, so treat it as time-sensitive.

This last exercise is hands-on, on the website at jaimeyan.com/learn, so you can pause the video and work through it yourself. You will practice the core industrial judgment of quality control, or QC: given a defect in an output compared with the shell, pick the cheapest QC pass that would actually catch it, whether that pass is format, data definitions, numbers, or consistency. Try it right after the video, and check your reasoning against the pass logic we just walked through.

This is the final check: ship the deliverable. During quality control (QC) review of a tables, listings, and figures (TLF) output made with the Output Delivery System (ODS) in rich-text format (RTF), page 2 repeats the column heading instead of showing a '(Continued)' marker. The answer is A, Pass 1, the format and visual review done before any numeric recomputation, because this is a layout defect, not a data-definition or numeric defect. For a safety-analysis-set (SAF) percentage by treatment group, the correct denominator source is option C: subjects in the subject-level analysis dataset (ADSL) where the safety flag SAFFL = 'Y' in that treatment group. The SAF is a subject-level population, so denominators must count safety-evaluable subjects from ADSL, not events or exposures. In the fixed QC order, Pass 2 checks the output's data definitions against the statistical analysis plan (SAP). The answers are A, B, and C: the SAP population flag such as SAFFL = 'Y' for the SAF, column denominators from the correct ADSL safety population rather than an event dataset, and visit windows and analysis-visit assignments per the SAP. Pass 2 is the data-definition pass; formatting and page breaks belong to Pass 1, and numeric recomputation is later. Before shipping, the fixed quality-control workflow is Pass 1, format and visual review against the shell; Pass 2, data-definition check of population flags, denominators, and visit windows against the SAP; Pass 3, independent numeric recomputation; and Pass 4, final comparison, for example with PROC COMPARE; this order is fixed cheapest-first because you catch cheap format and definition defects before paying for recomputation, and if the shell and code disagree you must first create a discrepancy record containing observed, expected, disposition, and approver.

The mock shell is a contract, so check output ID, titles, column blocks, denominators, footnotes, and the population statement. Pull denominators from the Subject-Level Analysis Dataset, or ADSL, under the population flag, never the event dataset, and honor the shell's precision, including the n (100) convention. Tables ship on pre-formatted PROC REPORT cells; listings on sort, page breaks, repeated headers, and (Continued); figures on the population join and group n. Quality Control, or QC, runs cheapest-first: format, definitions, numbers, consistency; each pass gates the next, never recompute a failed output. Every discrepancy gets a record with shell reference, observed, expected, disposition, and approver; never fix silently. Continue with the paired article From Mock Shell to Rich Text Format (RTF) and its Table, Listing, and Figure (TLF) QC checklist; this lesson used real ADSL-only rows, so listing and figure walks were conceptual.