PCP-combination

Imported from Russlewisbo/PCP-combination (AGENTS.md). Install upstream with npx skills add Russlewisbo/PCP-combination. Copyright stays with the author.

PCP Combination — Project Memory & Session Log

How to use this file: Read the Current status and Next steps sections first to re-orient. At the end of every working session, append a row to the Session log and refresh Current status / Next steps. Anything provisional should be flagged so it can be reconciled later.

Bayesian meta-analysis of combination/adjunctive therapy for Pneumocystis jirovecii pneumonia (PCP/PJP) in HIV-negative / immunocompromised adults (PROSPERO-registered). Primary question: comparative effectiveness/safety of combination/adjunctive regimens vs monotherapy (TMP-SMX). Primary outcome: all-cause mortality (30-day and/or in-hospital), OR with 95% credible intervals. Plan: Bayesian random-effects meta-analysis (NMA if data permit).

Live site: https://Russlewisbo.github.io/PCP-combination/ (published from the gh-pages branch via quarto publish gh-pages --no-prompt).

Current status (as of 2026-08-06)

• Verified primary analysis complete: adjunctive echinocandin/caspofungin + TMP-SMX vs TMP-SMX monotherapy, all-cause mortality.
• Headline: primary (k=3) OR 0.78 (95% CrI 0.31–1.93); sensitivity (k=6) OR 0.76 (0.43–1.37). Suggestive of benefit but inconclusive (CrI includes 1).
• Timing subgroups added (per Yang et al. 2026 Table 1): initial (k=3) OR 0.74 (0.34–1.62); mixed (k=3) OR 0.78 (0.31–1.93). No salvage-vs-monotherapy study available on our side.
• Benchmarked against the Yang et al. 2026 echinocandin meta-analysis (see below).

Figure style (ALWAYS APPLY)

All figures must be JAMA style with a blue-grey palette. source("analysis/jama_style.R"), then add theme_jama() and use jama_blue_grey / jama_tier for colours. No panel gridlines; classic axes; sans font.

Environment / gotchas

• Stan is broken on this machine (rstan TBB linker error; no cmdstanr). Fit Bayesian models with bayesmeta (semi-analytic, no MCMC). rjags / JAGS are NOT installed (no jags binary; runjags also absent) — do not rely on them. The Model-fit / MCMC page now gets genuine sampler diagnostics from a self-contained base-R random-walk Metropolis sampler (fit_mcmc() in analysis/meta_helpers.R), which cross-checks bayesmeta.
• Priors used throughout: mu ~ N(0, 1.5) on log-OR; tau ~ half-normal(0.5).
• Helper fit_bm() wraps bayesmeta() with these priors (defined in the analysis script).

Key files

• analysis/pcp_meta_analysis.R — reproducible meta-analysis script (data, log-ORs, models, subgroups).
• analysis/pcp_primary_meta_dataset.csv — verified analysis dataset (has timing column).
• analysis/jama_style.R — figure theme + palette.
• analysis/export_figures.R — regenerates every site figure as a 300 dpi PNG in figures/.
• analysis/prisma2020_flow.R + analysis/prisma2020_data.csv — PRISMA2020-package flow diagram.
• goodness-of-fit.qmd — fit/heterogeneity, influence (Baujat), posterior predictive checks.
• PCP_combination_extraction_database_v0.2_machine-draft.xlsx — extraction DB (MACHINE FIRST-PASS, unverified; needs dual human verification).
• Source PDFs: ~/main/PCP_combo_thesis/Review_PDFs/; plain text in ~/main/PCP_combo_thesis/scratch/pdf_text/ (used to verify counts against source).
• STATUS.md — older running log (superseded by this file for analysis work).

Reading the Excel DB

Sheets have a title row above the header → read_excel(path, sheet, skip = 1). First data row (Smith2019) is an EXAMPLE row — drop it. Key sheets: 1_Study_characteristics, 2_Arms_covariates, 3_Outcomes, 4_Beta_D_glucan, 5_Corticosteroids, 8_Study_index (eligibility).

Locked analysis decisions

• Primary contrast: adjunctive echinocandin/caspofungin + TMP-SMX vs TMP-SMX monotherapy.
• Effect measure: odds ratio. Timepoint: 30-day preferred, else in-hospital.
• Duplicate Qi2025 == Yanmeng2026 (byte-identical, same PUMCH cohort) → keep Qi2025 only.
• Synthesis is PAIRWISE (echinocandin+TMP-SMX vs TMP-SMX). NMA pre-specified "if data permit" but NOT supportable (single connected contrast, k=6) → descoped. PRISMA meta-analysis box reconciled to 6 studies / 6 reports.

Study inclusion (after SOURCE verification)

Verified included set — primary = Qi2025, Xu2025, Lu2017; sensitivity adds Jin2019, Qi2023, Li2024a. All extracted counts confirmed accurate against source text.

• Excluded Tian2020 — HIV-POSITIVE cohort (published in HIV Medicine); machine draft misclassified it. (Yang et al. independently excluded it too.)
• Li2024a — confounded (caspofungin bundled with corticosteroid + lower TMP dose); kept in the SENSITIVITY tier per user decision, flagged confounded. Its reference arm is the only steroid-free comparator in the set.
• Qi2025 extraction confirmed: full-cohort Table 4 = 18/35 (caspo) vs 31/79 (mono). Yang et al.'s 18/27 & 62/82 come from the ventilated subgroup (Table 5) + an incorrect pooling of monotherapy+clindamycin deaths — i.e. Yang analysed a subgroup as the whole study.

Sub-analyses NOT estimable

• Corticosteroid presence/absence: near-universal steroid use across studies; only Li2024a has a steroid-free reference arm (single confounded data point).
• β-D-glucan yes/no: only one small study (Lu2017) omitted BDG → no informative contrast.

Comparison with Yang et al. 2026 (SSRN preprint, echinocandin + TMP-SMX vs TMP-SMX)

• We share 9 of their 16 studies; missing 7 (mostly Chinese-language papers / master's theses: Li2015, Liu2015, Wang2021, Wu2023, Xiang2015, Yu2017, Wang-ZG-2019 SLE) because their search included CNKI/Wanfang.
• Yang overall OR 0.93 (0.71–1.21) is null; their initial-strategy OR 0.50 (0.32–0.77).
• Timing does NOT act as an effect modifier in our set. Within-data moderator test: QM(1)≈0.00, p≈0.99; initial 0.70 vs mixed 0.74 (REML) are nearly identical. The earlier "timing explains the gap" note held only under a different regrouping (dropping the mixed Qi2025 → OR 0.59, 0.32–1.09) — it reflects which studies are labelled "initial", not timing modifying the effect. Yang's initial-subgroup signal (0.50) is not reproduced as a subgroup difference on our side.
• Caveats on Yang: frequentist FIXED-effect M-H (narrow CIs); visible extraction errors (Fig 4 ORs 21.32 & 19.00; the Qi2025 subgroup mix-up above). Useful comparator, not ground truth.
• Nuance: Yang plotted Xu2025 as initial (Fig 3A) but Table 1 calls it initial+salvage; we classified it mixed (SOURCE-CONFIRMED: Xu2025 never restricts caspofungin to first-line use, reports no time-to-caspofungin, and flags "selection bias in caspofungin use"). This makes our strict initial subgroup (0.74) less protective than Yang's (0.50).

Retrieval status of the 7 missing studies (checked 2026-08-06)

None are in our local corpus; none could be added to the analysis this session.

• Wang-ZG-2019 (SLE, Medicine Baltimore; PMID 31169741, PMC6571266, open access) — OBTAINED (Europe PMC). INELIGIBLE: single-arm case series (9 patients, all combination, 0 deaths); no monotherapy comparator. Yang's "0/9 vs 2/6" invents a comparator absent from the source — another Yang extraction error. Do NOT add.
• Yu-2017 (Int J Clin Exp Med 2017;10(1):1234–1242) — low retrievability; that journal ceased publication (~2018), no reliable DOI/PMID. Try ijcem.com archive / ResearchGate.
• Chinese-language, CNKI/Wanfang only (need institutional access; 2 are theses): Li-T-2015 (thesis, Capital Medical Univ), Liu-2015 (Chin J Mod Drug Appl 9(17):90–91), Wang-Z-2021 (thesis, Hebei Medical Univ; DOI 10.27111/d.cnki.ghyku.2021.000212), Wu-2023 (J Internal Intensive Med 29(5):393–396), Xiang-2015 (Chin J Clin Rational Drug Use 8(22):69–70). Route: team's Chinese-reading members via cnki.net / wanfangdata.com.cn.

Next steps / open questions

1. Obtain the 5 CNKI/Wanfang Chinese studies via institutional access (team's Chinese readers); Wang-ZG-2019 is ineligible (single-arm) and Yu-2017 is low-accessibility.
2. Complete dual human verification of the extraction DB; have a reviewer sign off the AI disclaimer at the top of pcp_echinocandin_meta_writeup.qmd.
3. Consider a binomial-likelihood hierarchical model (needs a working Stan toolchain).
4. (Optional) Render the write-up to Word/PDF for the thesis team.

Done (2026-08-06)

• Xu2025 timing resolved → mixed (source-confirmed).
• Leave-one-out + prior-sensitivity checks complete (see status below).
• Reproducible JAMA-styled write-up drafted: pcp_echinocandin_meta_writeup.qmd (renders clean).

Session log