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

TTP Meta-Analysis Project Memory

Project Overview

Bayesian meta-analysis investigating the association between blood culture time to positivity (TTP) and clinical outcomes in bacteremia. Short TTP (reflecting higher bacterial inocula) is hypothesized to predict worse outcomes.

Primary outcomes:

1. Mortality (all-cause, typically 30-day or in-hospital)
2. Persistent bacteremia (microbiological clearance failure)

Data Files

Main Database

• File: TTP_MetaAnalysis_Extraction_Complete.xlsx
• Format: Excel workbook with relational tables linked by study_id

Sheets/Tables

Protocol Document

• File: TTP_meta-analysis_database_and_packages.md
• Contains database architecture, R package recommendations, prior specifications, and analytical notes.

Data Loading

library(tidyverse)
library(readxl)
library(janitor)

path <- "/Users/russelllewis/Desktop/ttp metaanalysis/TTP_MetaAnalysis_Extraction_Complete.xlsx"

tbl_study      <- read_excel(path, sheet = "tbl_study") |> clean_names()
tbl_population <- read_excel(path, sheet = "tbl_population") |> clean_names()
tbl_ttp        <- read_excel(path, sheet = "tbl_ttp") |> clean_names()
tbl_outcomes   <- read_excel(path, sheet = "tbl_outcomes") |> clean_names()

Data Cleaning Notes

Key Cleaning Steps

1. Remove [NON-PICO] rows — flagged in notes column (13 rows)
2. Harmonize outcome types:
   • mortality, in_hospital_mortality → "mortality"
   • persistent_bacteremia, bacteremia_clearance, microbiological_clearance → "persistent_bacteremia"
3. Effect size source classification:
   • reported_adjusted — preferred (adjusted OR/HR/RR with CI)
   • reported_unadjusted — unadjusted OR/HR/RR with CI
   • two_by_two — computed from 2×2 table via metafor::escalc()
   • continuous_converted — OR per hour TTP (different scale)
   • insufficient_data — excluded (22 rows)

Deduplication Rules

• Prefer adjusted over unadjusted when both exist (Cilloniz2017, Hou2023)
• Remove Melling2019 row 2 (extreme sparse-cell OR = 1722)
• Keep Marco2025 duplicate (two genuinely different subgroups)

Moderator Join Issue (RESOLVED)

• arm_id values in tbl_outcomes don't match tbl_population (which uses "all")
• Solution: Join on study_id only (one population row per study)

Final Dataset (tbl_es)

• 42 effect sizes from 40 unique studies
• 33 mortality (median OR = 2.64)
• 9 persistent bacteremia (median OR = 2.02)
• 2 continuous TTP effects (OR per hour) — kept separate

Moderator Completeness

Key Results Summary

Mortality (k=33)

• P(OR > 1) = 100% — posterior certainty short TTP increases mortality
• I² = 99.3% — substantial heterogeneity
• Egger's test p < 0.0001 — significant funnel asymmetry
• Prior sensitivity: OR range 1.99–2.15 (±4%) — insensitive to prior choice
• Meta-regression: No moderator explained heterogeneity (pathogen class, S. aureus, ROB, blood culture system)

Persistent Bacteremia (k=7-9)

• P(OR > 1) = 97.9% — strong but not decisive
• Estimate heavily influenced by Melling2019 (OR=37) and Kassis2009 (OR=75) — sparse 2×2 cells
• Prior sensitivity: OR range 1.38–5.84 — more sensitive than mortality (small k)

Pathogen-Specific Analysis (COMPLETED 2026-02-14)

Stratified mortality analysis by pathogen class:

Key Finding: No significant difference between pathogen classes (ratio of ORs = 1.05, 95% CrI: 0.61–1.83). TTP is a pathogen-agnostic prognostic marker that works equally well for Gram-positive and Gram-negative BSI.

Clinical Implication: TTP can be used immediately upon culture positivity as an early warning sign, even before organism identification is complete.

Bayesian Model Specification

# Primary model (brms)
fit <- brm(
  yi | se(sei) ~ 1 + (1 | study_id),
  data = es_data,
  prior = c(
    prior(normal(0, 1), class = "Intercept"),
    prior(cauchy(0, 0.5), class = "sd")
  ),
  backend = "rstan",  # or "cmdstanr" if available

  chains = 4, iter = 4000, warmup = 1000,
  control = list(adapt_delta = 0.95)
)

Prior Rationale

• Intercept: Normal(0, 1) — 95% prior mass OR 0.14–7.4
• tau: Half-Cauchy(0, 0.5) — moderate heterogeneity allowed

Key Fitted Model Objects (in R session)

Report Status

• File: TTP_Meta_Analysis_Report.qmd
• Status: INCOMPLETE — truncates at line 340 mid-sentence in moderator completeness table
• Completed sections:
  • Introduction & Background ✓
  • Data Loading & Preparation ✓
  • Effect Size Computation ✓
  • Descriptive Summary (partial)
• Missing sections:
  • Primary Meta-Analysis (Bayesian + Frequentist pooled estimates)
  • Forest Plots
  • Publication Bias Assessment
  • Sensitivity Analyses
  • Meta-Regression
  • Discussion/Conclusions

Publication Bias: Bayesian Selection Models

NEW ANALYSIS (completed 2026-02-13): Fully Bayesian selection models using Stan

Frequentist vs Bayesian Selection Models

Key Findings

1. Frequentist step-function shows extreme adjustment — δ = 0.037 means non-significant studies have only 3.7% publication probability; adjusted OR = 1.09 (non-significant)

2. Bayesian models are robust to publication bias — Even with skeptical priors, pooled OR remains ~2.14 with P(OR>1) = 100%

3. Why the discrepancy?

   • Only 3/33 studies (9%) are non-significant
   • Frequentist MLE pushes δ to extreme values with sparse data
   • Bayesian priors regularize estimates
   • 91% significance rate exceeds expectation even for OR=3.0 (~68% expected)

4. Interpretation: Some publication bias likely exists. True effect probably lies between OR 1.50 (trim-and-fill) and 2.14 (unadjusted). The frequentist step-function (1.09) likely overadjusts.

5. Conclusion: Short TTP robustly associated with mortality across all Bayesian approaches, though exact magnitude uncertain due to probable publication bias.

Stan Model Implementation

Custom Stan model implements step-function selection:

• Prior on μ: Normal(0, 1)
• Prior on τ: Half-Cauchy(0, 0.5)
• Prior on δ: Beta(2, 5) for skeptical version (mean = 0.29)
• Selection mechanism: target += log(delta) for non-significant studies

Packages Required

# Core
library(tidyverse)
library(readxl)
library(janitor)
library(metafor)
library(brms)

# Bayesian diagnostics
library(posterior)
library(bayesplot)
library(ggdist)

# Visualization
library(patchwork)

# Stan (for custom selection models)
library(rstan)

Next Steps / TODO

1. URGENT: Complete Quarto report — file still truncated at line 340

   • Fix moderator completeness table
   • Add all primary analysis sections (pooled estimates, forest plots)
   • Add publication bias section with ALL models (funnel, Egger, trim-and-fill, frequentist selection, Bayesian selection)
   • Add sensitivity analyses
   • Add meta-regression results
   • Add discussion/conclusions

2. Consider Bayesian selection model ✓ COMPLETED — implemented via Stan

3. Dose-response modeling for continuous TTP studies (currently excluded from pooling)

4. Manuscript preparation — once report is complete, extract key tables/figures for publication