9  Research log

Append-only. Newest entries at the bottom. What happened, what we learned, and what it changed. Short is fine. This is a hand-written log, so its numbers are typed; the report chapters compute theirs.

2026-10-16 First data delivery: licorice_gargle.csv, emailed by the PI’s fellow, the public medicaldata version (MD5 fe23d586f6bdc92a5db1342f8e41a283). It has 235 rows and 19 columns, matching the codebook names, and every variable the plan uses is present. There are no duplicate rows and no patient ID column. Disagreement with the notes: the codebook says 236 were enrolled, but the file has 235 rows. That is now an open question in Chapter 1 (D-019 covers how the file is handled).

2026-10-16 Structure check at receipt. Outcome values were checked only for validity and missingness, not summarised. The baseline variables and treat have no missing values and contain only codebook codes. All four planned sore-throat scores are integers from 0 to 10. The same 2 patients are missing every post-operative variable (all cough and pain columns), which confirms the client’s “a couple pts no post-op data”. That is under the 5% revisit trigger in D-015, so complete case plus bounds stands. Only 2 patients have a preoperative sore throat, so under the pre-specified rule in D-011 it is left out of the adjusted model. All other covariate categories in that model have at least 10 patients, so D-016 merges nothing. Mallampati class 4 has 1 patient, but Mallampati is not in any model.

2026-10-16 Cleaning function written (clean_raw(), assessments_long() in R/clean.R) with known-answer tests on a hand-built file of 4 patients. Each test was checked by deliberately breaking the code in three ways: swapped treatment labels, sore throat defined as >= 0, and the whole-number check removed. The tests caught all three. The real file passes validation (235 patients; 940 rows in the long table). Two choices the codebook does not settle: patient_id is the raw row number, since the file has no identifier, and age and BMI get only wide plausibility bounds (age 0 to 120, BMI 10 to 80), because the codebook gives no range for them.

2026-10-17 Model functions written in R/models.R: per-assessment logistic regression with risk difference, exchangeable GEE, random-intercept mixed model, adjusted model with sparse-level merging, and missing-outcome bounds. The simulator is in R/simulate.R, and nothing has been fitted to the trial data. Each recovery test uses 200 simulated trials of 235 patients, 2 of them without post-operative data. The marginal truths (per-assessment OR and RD, the GEE’s common OR, and the within-patient correlation) come from integration over the random intercept. The GEE scenario solves for the licorice intercepts that make the marginal OR constant (D-018). Two findings:

• Laplace fitting of the mixed model underestimated the patient SD and gave a convergence warning on a simulated trial, so the mixed model uses 9-point quadrature (D-020, proposed).
• 13 deliberate breaks were tried. 12 failed tests at once. The independence-GEE break passed every test, because an independence GEE gives the same OR and valid sandwich CIs. The GEE now reports its working correlation, and the test compares that with the integrated truth.
• 90% CIs were caught only by the GEE and adjusted-model coverage checks, so odds_ratio() now has an exact CI check.
• After both changes, all 13 breaks fail at least one test.

2026-10-17 The break harness is now committed as scripts/check_test_breaks.R. D-020 was confirmed by the analyst before any fit to trial data. The final run, against the final tests, lists for each break the tests that failed:

• Mixed model swapped for a logistic regression ignoring patients: the mixed-model recovery test (bias 0.248 against an allowed 0.08; coverage 0.525).
• Mixed model by Laplace: the mixed-model recovery test (mean patient SD off by 0.128 against an allowed 0.1).
• GEE ignoring patients: the GEE recovery test (coverage 0.83; SE ratio 0.70; working correlation off by 0.271, i.e. estimated near zero).
• GEE with independence correlation: the GEE recovery test (no working correlation).
• Per-time fit pooling the assessments: the fit_assessment known answers, plus per-assessment recovery (coverage, SE ratio and sample size).
• RD sign flipped: the RD known answers and RD recovery.
• RD SE using the total n: the RD known answers, plus RD coverage at one assessment.
• 90% CIs: the odds_ratio known answers, plus GEE and adjusted coverage.
• Adjusted model without covariates: adjusted recovery (the large-sample estimate off by 0.171 against an allowed 0.06).
• Sparse levels never merged: the merge and covariate-rule known answers.
• Bounds labels swapped, or both arms filled the same way: the bounds known answers.
• Integration using the variance as the SD: the integration known answers, plus per-assessment and GEE recovery.

2026-10-17 First fit to the trial data, by make results. All 8 fits ran with no warnings: four per-assessment logistic regressions, the GEE, the mixed model, the adjusted model and the bounds.

• Missingness is only the 2 patients with no post-operative data, 1 per arm (0.9% at each assessment). That is under the D-015 trigger.
• Preoperative sore throat (2 patients) is left out of the adjusted model by D-011, and nothing is merged under D-016.
• Primary, at thirty minutes: 42 of 116 with sore throat on sugar water against 22 of 117 on licorice. The marginal OR is 0.41 (0.22 to 0.74), and the bounds run from 0.39 to 0.43, both with CIs below 1.
• The per-assessment ORs stay below 1 through the first postoperative morning (0.39, 0.22 to 0.70).
• GEE common marginal OR: 0.34. Mixed-model conditional OR: 0.14, with a patient SD of 2.76, well above the 1.5 used in the recovery tests.
• I reran the mixed-model and GEE recovery (200 simulated trials each) at SD 2.75. The mixed model had bias 0.001, coverage 0.965, mean SD 2.796 and no warnings. The GEE had bias 0.032, coverage 0.925, and working correlation 0.478 against a true 0.486. Both fits are trustworthy at the trial’s clustering.

2026-10-17 Fixes after the fresh-session analysis review. The review found nothing that changes a conclusion.

1. The SD 2.75 recovery run had not been committed. It is now scenario 4 in test-recovery.R, with fixed seeds; the earlier run was uncommitted, with other seeds.
   • Mixed model: bias -0.010, coverage 0.950, SE/SD 0.99, mean patient SD 2.777, no warnings.
   • GEE: bias -0.017, coverage 0.955, SE/SD 0.95, working correlation 0.481 against a true 0.486.
   • The review’s concern about GEE coverage of 0.925 looks like Monte Carlo noise: on the committed seeds it is 0.955.
   • make test now takes about two minutes.
2. The bounds claims now say they cover only patients in the file. A post hoc worst case (D-021) adds the 1 enrolled patient missing from the file to licorice with a sore throat: OR 0.45 (0.25 to 0.81), matching the review’s check. The enrollment count of 236 is now in raw_codebook(), so the number of absent patients is computed.
3. Every data-dependent claim the review listed now has a computed branch for the other outcome, in Chapters 2, 5, 6 and 7. So does “the arms are similar at baseline”, which the review did not list; it was a judgement on the data and has been removed.
4. D-020’s date corrected.

The summary (index.qmd) is written from computed results, with the open client questions listed.

2026-10-17 Sped up the simulation-recovery tests from about two minutes to under one, at the analyst’s request. The recovery tests are now split across three files that run in parallel (testthat’s parallel mode, slowest first); almost all the time is the mixed model’s adaptive quadrature. The two checks at the trial’s own clustering (SD 2.75) now use 100 simulated trials instead of 200: they are a robustness check on top of the main recovery tests. Every test still passes, and switching the mixed model back to Laplace still fails both mixed-model recovery files. No analysis changed.