Skip to content

nma_README.md

Latest commit

 

History

History
151 lines (110 loc) · 4.77 KB

nma_README.md

File metadata and controls

151 lines (110 loc) · 4.77 KB

Bayesian Network Meta-Analysis — Advanced NSCLC Overall Survival

A comprehensive R pipeline implementing Bayesian network meta-analysis for health technology assessment, comparing six treatments across ten published studies in advanced non-small cell lung cancer (NSCLC).

Mirrors the indirect treatment comparison methodology required in NICE Single Technology Appraisals when multiple treatments exist but not all have been directly compared in head-to-head randomised controlled trials.

Research context

When a new treatment is assessed by NICE, it often needs to be compared not just to the direct comparator in its pivotal trial, but to all relevant alternatives. Where head-to-head trials do not exist, NMA allows indirect comparison by chaining evidence through common comparators — producing a unified evidence synthesis across the entire treatment landscape.

Evidence network

Six treatments compared across 10 published studies (simulated evidence base modelled on the advanced NSCLC immunotherapy landscape):

Code Treatment
A — Chemotherapy Reference arm (platinum doublet)
B — Nivolumab mono PD-1 checkpoint inhibitor monotherapy
C — Pembrolizumab mono PD-1 checkpoint inhibitor monotherapy
D — Durvalumab + Chemo Immunochemotherapy combination
E — Osimertinib EGFR-targeted therapy
F — Pembrolizumab + Chemo Immunochemotherapy combination

Methods implemented

Method Description
Network plot Evidence map showing direct comparisons and study counts
Fixed effects NMA Assumes homogeneous true treatment effects
Random effects NMA Allows between-study heterogeneity (tau)
DIC model selection Deviance Information Criterion for model comparison
Gelman-Rubin R-hat MCMC convergence diagnostics
Pairwise HRs All treatments vs reference with 95% CrI
League table Full all-vs-all pairwise HR matrix
SUCRA rankings Surface Under Cumulative Ranking — treatment hierarchy
Node-splitting Direct vs indirect evidence consistency test
Heterogeneity (tau) Between-study SD from random effects model

Outputs

File Description
output/network_plot.png Evidence network diagram
output/pairwise_forest_plot.png HRs vs chemotherapy, forest plot
output/sucra_plot.png SUCRA ranking bar chart
output/league_table.csv All-vs-all pairwise HR table
output/pairwise_hrs.csv HRs vs reference with CrI
output/sucra_rankings.csv SUCRA values per treatment
output/model_fit.csv DIC comparison FE vs RE
output/consistency_check.csv Node-split p-values
output/heterogeneity.csv Tau estimate and CrI

How to run

Step 1 — Install JAGS (required) 👉 https://sourceforge.net/projects/mcmc-jags/

Step 2 — Install R packages

install.packages(c("gemtc", "rjags", "ggplot2",
                   "dplyr", "tidyr", "reshape2", "scales"))

Step 3 — Run analysis

source("nma_analysis.R")

MCMC sampling takes approximately 5-10 minutes.

Statistical methods

Model: Bayesian consistency model (Dias et al. 2013 framework) implemented via gemtc with JAGS backend

Likelihood: Normal (log-HR scale)

Priors:

  • Treatment effects: N(0, 100²) — vague
  • Heterogeneity (RE): Half-normal(0, 0.5²) — weakly informative

MCMC: 4 chains, 50,000 iterations, 5,000 burn-in, thinning = 10

SUCRA: Probability of being ranked best, second-best, etc., summarised as area under the cumulative ranking curve (Salanti 2011)

Consistency: Node-splitting (Dias 2010) — tests whether direct and indirect evidence are in statistical agreement

Relevance to NICE submissions

This pipeline implements all components required in a NICE STA NMA:

  • Fixed and random effects models with DIC-based selection
  • Full pairwise league table with credible intervals
  • SUCRA treatment ranking for decision support
  • Consistency assessment as required by NICE DSU TSD 2
  • Heterogeneity estimation and reporting

Key references

  • Dias S, et al. (2013). NICE DSU TSD 2: A generalised linear modelling framework for pairwise and NMA of randomised controlled trials.
  • Salanti G, et al. (2011). Graphical methods and numerical summaries for NMA. J Clin Epidemiol, 64(2):163-71.
  • Higgins JPT, Whitehead A. (1996). Borrowing strength from external trials in a meta-analysis. Stat Med, 15:2733-49.

Dependencies

gemtc     # Bayesian NMA framework
rjags     # JAGS interface for MCMC sampling
ggplot2   # Visualisation
dplyr     # Data manipulation
scales    # Plot utilities

Author

Atrija Haldar LinkedIn MSc Engineering, Technology and Business Management — University of Leeds