Fix variant phase assignment with per-location approach

[kathsherratt]

Summary

• Replace global median-date hack in variant phase classification with correct per-location approach
• Aggregate variant data across sources (ECDC TESSy/GISAID) before dominance assessment, fixing spurious results from small-sample sources
• Require >50% variant share and enforce chronological ordering to guard against noisy surveillance data
• Fix Sys.Date() to study end date 2023-03-17 for reproducibility
• Regenerate GAMM results (output/results.rds) and diagnostic plots

Key results change

With per-location variant phases providing better covariate adjustment, Method and CountryTargets effects remain near-zero — consistent with, and slightly strengthening, the original finding of no systematic performance difference between model structures.

Known limitation

Hungary (HU) starts with Delta phase instead of Alpha due to sparse early variant surveillance data (no weeks with Alpha >50% in ECDC). This is acceptable given the data resolution.

Test plan

• classify_variant_phases() runs without errors
• All 32 locations have variant phase assignments with no NAs
• Omicron arrival dates differ by location (GB earliest Dec 11, SK/BG latest Jan 15)
• GAMM fits successfully for both cases and deaths
• quarto render report/results.qmd produces complete HTML

🤖 Generated with Claude Code

[kathsherratt]

Prompts used to generate this PR

This PR was generated and iteratively refined using Claude Code (claude-opus-4-6). Below are the user prompts that drove the work, in chronological order.

Session 1 — Planning

each top-level bullet point in Plan.md indicates a discrete pull request. plan your approach to implementing these changes, for now ignoring the bonus features section.

In PR2, we will need to check data sources for variant phases in Switzerland and the UK, and ensure they are processed consistently. include this in the plan.

(During planning, user was asked about variant fix approach and chose "Switch to per-location (Recommended)")

Session 1 — PR 2 implementation

The initial per-location variant phase fix was implemented as planned: replacing the median-date hack with per-location dominant variant phases based on when each variant first exceeded 50% of sequenced samples. Commit: 2f6da1c.

Session 1 — Follow-up: methods text and Supplement

Before continuing the current plan, let's continue to improve PR#144. First, re-render the Supplement so that it shows the changes to variant phases, and save as a PDF. Then, suggest text to add to the methods section of the main manuscript to reflect how variant phases were calculated. If this needs more than 2-3 sentences to communicate, instead summarise in 1 sentence and add the detail to the Supplement (with a note in the main manuscript). When this is complete, stop and confirm before proceeding with the current plan.

yep

(Approved the manuscript text change)

Session 1 — Debugging CH/GB variant data

re variant phases - something wrong with swiss data processing; currently not showing variant phases for ba.2 onwards. check what this is and if it's relevant to gb data which also looks off.

(When asked whether to add to PR #144 or create a new PR, user chose "Add to PR #144". Claude investigated and identified two bugs: CH deduplication discarding hospital data for BA.2+ variants, and mean() instead of sum() preventing GB from detecting BA.4/5 dominance. User approved the fix plan.)

Session 2 — Rebase after PR #143 merge

PR 143 has been merged - check you are up to date with main.

(Approved force-push after clean rebase)

Session 2 — Hungary variant data fix

let's fix the issue with variant data in Hungary. The current data are missing (Hungary did not report variant data until later), which means the backfill method gives the Delta phase from the start of the timeseries.

I've identified that Delta started spreading from 23 July 2021 from this source: https://abouthungary.hu/news-in-brief/delta-and-gamma-variant-identified-in-hungary

Update the code so that Hungary is in the Alpha variant phase until Delta starts in the week following 23 July 2021. Add a comment in the code referencing the source. Then update the manuscript/supplement with appropriate text.

[kathsherratt]

Noting that this was initially prompted by following Plan.md. And code manually reviewed at each stage as per prompt history above.