8 Decision record

Every decision that could change a result is recorded here, with the alternatives considered and who made it. Entries are never deleted; a reversed decision gets a new entry that supersedes the old one. Timing says whether any results had been seen when the decision was made: a change made after seeing results is post hoc, and is marked as such here and in the analysis plan. This file is a hand-written log, so its numbers are typed; the report chapters compute theirs.

8.1 D-001: Primary outcome is any sore throat at rest at PACU 30 min

Date: 2026-10-16
Decided by: PI, confirmed on a call with Matt Mulvahill on 2026-10-16 (notes/2026-10-16-pi-call.md). The intake meeting on 2026-10-10 left the definition open.
Decision: The primary outcome is any sore throat at rest 30 minutes after PACU arrival, defined as pacu30min_throatPain > 0 on the 0-10 scale.
Why: It is the PI’s primary question (“did it reduce sore throat afterwards?”). On the call the PI chose “any” (score > 0) with no higher threshold: “any is what patients notice”.
Timing: At planning, before any outcome was examined. The data file has not been opened.
Record updated: 2026-10-16, after plan review, to cite the call note and correct the attribution; the decision itself is unchanged.
Alternatives: A higher threshold on the 0-10 score (raised at intake; not chosen); the 0-10 score itself, analysed as ordinal (deferred; see the intake notes).

8.2 D-002: “Does it last” is answered by an odds ratio at each time point

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst); the PI agreed to the population-averaged (population-level) odds ratio on a call on 2026-10-16 (notes/2026-10-16-pi-call.md)
Decision: Secondary outcomes are any sore throat at rest (score > 0) at PACU 90 min, 4 h after surgery and the first postoperative morning. “Does it last past the recovery room?” is answered by a separate treatment odds ratio at each of the four time points. One common odds ratio across the four time points is reported as a summary. The odds ratios are population-averaged (marginal). There is no formal test of a treatment-by-time interaction.
Why: An odds ratio at each time point answers the client’s question directly; a common odds ratio is a useful one-number summary.
Timing: At planning, before any outcome was examined.
Alternatives: A formal treatment-by-time interaction test (not chosen); later time points only, without a summary.
Record updated: 2026-10-16, after plan review, to cite the call note; the decision itself is unchanged. See D-013 and D-014 for how the per-time and common odds ratios are interpreted.

8.3 D-003: No multiplicity adjustment

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: No adjustment for multiple comparisons.
Why: There is one primary comparison; the secondary analyses are descriptive.
Timing: At planning, before any outcome was examined.
Alternatives: Holm or Bonferroni adjustment across the four time points.

8.4 D-004: Cough is not analysed

Date: 2026-10-16
Decided by: PI, on a call on 2026-10-16 (notes/2026-10-16-pi-call.md)
Decision: Cough (at extubation and at each postoperative assessment) is not an outcome of this analysis.
Why: The call note records “cough: not interested, drop it”.
Timing: At planning, before any outcome was examined.
Alternatives: Cough as a secondary outcome.
Record updated: 2026-10-16, after plan review, to cite the call note; the decision itself is unchanged.

8.5 D-005: Intention to treat, patient as unit, keep preoperative sore throat

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: The analysis population is all randomized patients, analysed in the group they were randomized to. The unit of analysis is the patient; the common odds ratio uses one row per patient per assessment, clustered on patient. Patients with a preoperative sore throat (preOp_pain = 1) are kept.
Why: Intention to treat preserves randomization, and the file has no adherence or protocol-deviation variable for a per-protocol analysis. Excluding patients with a preoperative sore throat would change the population and could break randomization unless the protocol specified it. Whether the protocol excluded them is an open question.
Timing: At planning, before any outcome was examined.
Alternatives: Per-protocol analysis (not possible from this file); excluding patients with a preoperative sore throat.

8.6 D-006: Report proportions and risk differences beside odds ratios

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: For the primary outcome and each assessment, report the proportion with any sore throat in each arm and the risk difference (licorice minus sugar water) with its 95% CI, beside the odds ratio.
Why: When the outcome is common, the odds ratio overstates the relative risk; readers need the absolute difference.
Timing: At planning, before any outcome was examined.
Alternatives: Odds ratios only.

8.7 D-007: Models for the primary, per-assessment and common odds ratios

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst). The PI agreed to unadjusted logistic regression at 30 minutes as the primary model on a call on 2026-10-16 (notes/2026-10-16-pi-call.md); the remaining models are the analyst’s call.
Decision: The primary odds ratio and the odds ratio at each later assessment come from separate unadjusted logistic regressions of any sore throat on treatment, one per assessment. The common odds ratio comes from a GEE (logit link; treatment and assessment as a factor; exchangeable working correlation; robust standard errors; geepack). Wald 95% CIs throughout.
Why: With treatment as the only term, the logistic odds ratio is the marginal odds ratio, so each separate fit estimates the agreed estimand (D-002) as a plain two-group comparison. Separate fits keep earlier assessments from informing later ones through the working correlation when data are missing. The GEE targets the marginal common odds ratio.
Timing: At planning, before any outcome was examined.
Alternatives: One GEE with a treatment-by-assessment interaction for the per-assessment odds ratios; independence working correlation; profile-likelihood CIs.
Record updated: 2026-10-16, after plan review, to cite the call note for the PI’s agreement to the primary model; the decision itself is unchanged.

8.8 D-008: Also fit a random-intercept mixed model, labeled conditional

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: Fit a logistic mixed model with the same terms as the GEE and a random intercept for patient (lme4). Report its treatment odds ratio beside the GEE common odds ratio, clearly labeled as a patient-level (conditional) odds ratio. The GEE odds ratio remains the summary estimate.
Why: Reviewers will ask for a mixed model. Showing both, labeled, is how we stop the larger conditional odds ratio being quoted as the effect.
Timing: At planning, before any outcome was examined.
Alternatives: Not fitting the mixed model (the earlier draft of the plan).

8.9 D-009: Covariates for the adjusted sensitivity analysis

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: The adjusted sensitivity analysis adjusts for sex, age, BMI, ASA status, smoking and surgery size. Preoperative sore throat is included unless so few patients have it that its coefficient would be unstable; that is decided from the baseline count when the data arrive, without looking at outcomes, and logged. Mallampati class is not included. The adjusted odds ratio is reported as a conditional odds ratio, with an explanation of why it differs from the unadjusted one.
Why: Covariates are chosen on prognostic grounds, before any data are seen. Mallampati predicts a difficult intubation rather than sore throat. The analysis is a sensitivity check, so the conditional odds ratio is reported as such rather than standardized.
Timing: At planning, before any outcome was examined. The data file has not been opened.
Alternatives: Standardization (g-computation) to a covariate-adjusted marginal odds ratio (not chosen); including Mallampati class.

8.10 D-010: Missing outcomes: complete case, with bounds on the primary

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: Every analysis uses the patients observed for the outcome at that assessment (complete case). For the primary outcome, also report best- and worst-case bounds. Any missing score is treated as missing; there is no variable for “not assessable”.
Why: At intake the client thought a couple of patients might have no postoperative data, though they were not sure. If missingness is that limited, complete case with bounds is enough; the bounds show the range the missing patients could move the primary result.
Timing: At planning, before any outcome was examined.
Alternatives: Multiple imputation; bounds for every assessment.

8.11 D-011: Adjusted analysis: scope, covariate forms, preop sore throat rule

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: Refines D-009. The adjusted sensitivity analysis is for the primary outcome only. Age and BMI enter as linear terms; ASA status, smoking and surgery size as factors. Preoperative sore throat is left out of the adjusted model if fewer than 10 patients have it, counted from baseline data only.
Why: The cut-off fixes in advance the rule in D-009 (leave it out if so few patients have it that its coefficient would be unstable), so the choice does not depend on judgement after the data arrive. The covariate forms were proposed in drafting and confirmed by the analyst, so that they are fixed before any data are seen.
Timing: At planning, before any outcome was examined. The data file has not been opened.
Alternatives: Adjusted analyses at every assessment; splines for age and BMI; a judgement call on preoperative sore throat at data receipt.

8.12 D-012: Report missingness by arm; revisit the approach if it is large

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: Report the number of patients missing each outcome, by arm, in the data chapter. If missingness is more than the client expected, revisit the missing-data approach (D-010) before fitting any model, with a decision entry.
Why: D-010 rests on the client’s belief that only a couple of patients lack postoperative data; this checks that belief before any result is seen.
Timing: At planning, before any outcome was examined.
Alternatives: Proceed with complete case regardless of the amount missing.

8.13 D-013: Durability is judged from per-time estimates, without testing

Date: 2026-10-16
Decided by: PI (per-time results descriptive); Matt Mulvahill (analyst; interpretation rule)
Decision: No formal testing across time points; D-003 stands. Whether the effect lasts is judged from the per-time odds ratios, risk differences and their 95% CIs. A non-significant result at a later time point is not evidence that the effect wore off.
Why: The PI wants the per-time results descriptive. The plan review pointed out that, without a stated interpretation, a null result on the first postoperative morning would be read as the effect wearing off, when it is only absence of evidence.
Timing: At planning, after plan review, before any data were seen.
Alternatives: Fixed-sequence testing (30 min, 90 min, 4 h, POD1, stopping at the first non-significant result), proposed in the plan review.

8.14 D-014: The common odds ratio assumes one common effect

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: Keep the exchangeable GEE common odds ratio (D-007) as a summary across all four time points, including 30 minutes. The plan states that it assumes one effect common to all time points and does not answer “does it last”; the per-time estimates do. The same applies to the mixed-model common odds ratio (D-008).
Why: The plan review noted that if the effect changes over time the model is misspecified and the summary is a weighted average that leans toward the early, better-observed time points, so it could be quoted as a lasting effect.
Timing: At planning, after plan review, before any data were seen.
Alternatives: Drop the common odds ratio; an independence working correlation with stated weights; excluding 30 minutes from the summary.

8.15 D-015: Missing data: MCAR assumption and a numeric revisit trigger

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: Supersedes the trigger in D-012. The complete-case analyses assume outcomes are missing completely at random (MCAR). If more than 5% of patients are missing the outcome at any time point, the missing-data approach is revisited before any model is fitted, with a decision entry. Best- and worst-case bounds remain on the primary outcome only (D-010). Missingness by arm is still reported (D-012).
Why: Complete case targets all randomized patients only under MCAR, so the assumption is stated. The review noted that “more than the client expected” left the trigger to be set after missingness was seen; a number fixes it in advance.
Timing: At planning, after plan review, before any data were seen.
Alternatives: A trigger on the between-arm difference in missingness; bounds at later time points too; naming a fallback such as multiple imputation now.

8.16 D-016: Sparse covariate categories merged; Mallampati reason restated

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: Adds to D-009 and D-011. In the adjusted model, any covariate category with fewer than 10 patients is merged into its neighbour, decided from baseline counts only. The adjusted odds ratio stays conditional. The reason for leaving out Mallampati class is restated: it acts through intubation difficulty, which is not measured, and leaving it out keeps the model small.
Why: The review noted roughly 11 parameters with no rule for sparse categories, risking separation and unstable estimates. The original Mallampati reason (“predicts a difficult intubation rather than sore throat”) was argued backwards, since difficult intubation plausibly causes sore throat.
Timing: At planning, after plan review, before any data were seen.
Alternatives: Standardization (g-computation) to a covariate-adjusted marginal odds ratio, proposed again in the review (not chosen); including Mallampati class.

8.17 D-017: Verification additions

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: When data are simulated from a random-intercept model, the true marginal odds ratio for the recovery tests is computed by numerical integration over the random-intercept distribution. The Wald CI for the risk difference gets a known-answer test. Wald CIs for the risk difference are kept.
Why: The marginal truth is not a model parameter, so the GEE recovery test needs it computed. Every calculation in the report needs a test.
Timing: At planning, after plan review, before any data were seen.
Alternatives: Newcombe intervals for the risk difference, suggested in the review for proportions near 0 or 1.

8.18 D-018: Constant-effect simulation for the common OR; test the bounds

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: Adds to D-017. The recovery test for the common odds ratio (GEE and mixed model) simulates a treatment effect that is constant over time. The best- and worst-case bounds procedure gets a known-answer test.
Why: The common odds ratio assumes one common effect (D-014); only with a constant simulated effect does it have a single true value to recover. The bounds are analysis logic and need a test like any other (CLAUDE.md rule 3). Both were raised in the plan review.
Timing: After the plan was locked at plan-v1, before any data were seen. Not post hoc: no outcome or other data value had been examined. Changes verification only; no estimand, outcome or model changes.
Alternatives: Simulating a time-varying effect and checking the weighted-average target; no dedicated test for the bounds.

8.19 D-019: The delivered file is read-only and checksum-guarded

Date: 2026-10-16
Decided by: Matt Mulvahill (analyst)
Decision: data/raw/licorice_gargle.csv, the first delivery (received 2026-10-16 from the PI’s fellow; the public medicaldata version), is the analysis’s raw data, stored unchanged. Its MD5 (fe23d586f6bdc92a5db1342f8e41a283) is recorded in data/raw/README.md and in scripts/02_make_analysis_data.R, which refuses to run if the file changes. Every cleaning step happens in code. A corrected or fuller file is saved as a new file with a new entry here.
Why: Project rule (CLAUDE.md rule 2). Every result must trace back to bytes that can be shown to be the ones received. The file disagrees with the codebook’s enrollment count (Chapter 1, open questions), so a second delivery is plausible and should not silently replace this one. The file is public and contains no identifiable data, so it is kept in data/raw/ and committed. data/phi/ is not used.
Timing: At data receipt, after the plan was locked at plan-v1. Only the structure, the baseline codes, the validity of the outcome codes and the outcome missingness had been examined. No outcome distribution or comparison between arms had been seen. Not post hoc, and no change to the plan.
Alternatives: Asking the client for a corrected file before starting; reading the data straight from the medicaldata package (rejected: the package version can change underneath the analysis).

8.20 D-020: Mixed model fitted by adaptive quadrature (9 points)

Date: 2026-10-17 (proposed 2026-10-16)
Decided by: Matt Mulvahill (analyst), confirming Claude’s proposal on 2026-10-17, before any model was fitted to the trial data.
Decision: The random-intercept mixed model (D-008) is fitted with lme4::glmer(nAGQ = 9), adaptive Gauss-Hermite quadrature with 9 points, instead of the default Laplace approximation (nAGQ = 1). The estimand, terms and software are unchanged.
Why: With four binary outcomes per patient, Laplace underestimated the patient SD in simulation (1.40 against a true 1.5, from 20,000 simulated patients). On a simulated trial of the real size it also gave a convergence warning (max|grad| 0.08); 9-point quadrature fitted the same data without one. Quadrature is exact enough for a single scalar random intercept and costs about a quarter of a second per fit.
Timing: After the plan was locked at plan-v1, while writing the model functions, from simulated data only. No model had been fitted to the trial data, and no outcome distribution or arm comparison had been seen. Not post hoc. Changes the fitting method only, not the plan.
Alternatives: Laplace (the lme4 default); more quadrature points; a different package (GLMMadaptive).

8.21 D-021: Extend the worst-case bound to enrolled patients absent from the file

Date: 2026-10-17
Decided by: Matt Mulvahill (analyst), after the analysis review.
Decision: Adds to D-010. A further worst-case bound for the primary odds ratio. Every enrolled patient missing from the delivered file (the codebook’s enrollment count minus the rows in the file) is added to the licorice arm with a sore throat, on top of the in-file worst case. The existing best- and worst-case bounds are unchanged. The report now says that they cover only the patients in the file.
Why: The analysis review noted that “patients missing the primary outcome cannot change this conclusion” covered only the patients with no post-operative data in the file. The patient the codebook counts as enrolled but the file lacks was not covered. Whether that patient was randomized is an open question for the client (Chapter 1).
Timing: Post hoc. Added after all results had been seen, including the primary estimate and the in-file bounds. It adds a supplementary analysis. It does not change the primary outcome, the primary model or any estimand.
Alternatives: Rewording only, with no extended bound (the review’s minimum); an extended best case as well; placing the absent patient in the sugar-water arm, which is less extreme for licorice.