Network Meta Analysis

Key Teaching Points:

• Direct evidence: A vs B from head-to-head trials
• Indirect evidence: A vs B inferred through common comparator C
• Network geometry: How treatments are connected through trials

Socratic Questions:

• "If we have trials comparing A vs C and B vs C, can we learn something about A vs B?"
• "What assumptions must hold for indirect comparisons to be valid?"
• "Why might indirect evidence differ from direct evidence?"

2. The Transitivity Assumption

Critical Concept: For indirect comparisons to be valid, studies must be similar enough that patients could have been enrolled in any of them.

Factors to Assess:

• Patient populations (age, severity, comorbidities)
• Intervention definitions (doses, durations)
• Outcome definitions and timing
• Study design and risk of bias

Teaching Framework:

Transitivity Check:
┌─────────────────────────────────────────┐
│ Could patients in A vs C trials have    │
│ been enrolled in B vs C trials?         │
│                                         │
│   YES → Transitivity likely holds       │
│   NO  → Indirect comparison may be      │
│         biased (effect modification)    │
└─────────────────────────────────────────┘

3. Network Geometry

Network Plot Elements:

• Nodes = treatments (size ∝ sample size)
• Edges = direct comparisons (thickness ∝ number of studies)
• Closed loops = allow consistency checks

Types of Networks:

4. Statistical Models

Frequentist Approach (netmeta package):

library(netmeta)

# Create network meta-analysis
nma <- netmeta(
  TE = effect_size,
  seTE = standard_error,
  treat1 = treatment1,
  treat2 = treatment2,
  studlab = study_id,
  data = mydata,
  sm = "OR",           # Effect measure
  random = TRUE,       # Random effects
  reference.group = "placebo"
)

# View results
summary(nma)
forest(nma)
netgraph(nma)

Bayesian Approach (gemtc/BUGSnet):

library(gemtc)

# Define network
network <- mtc.network(data.ab = arm_level_data)

# Run model
model <- mtc.model(network, type = "consistency")
results <- mtc.run(model, n.adapt = 5000, n.iter = 20000)

# Results
summary(results)
forest(relative.effect(results, t1 = "placebo"))

5. Consistency Assessment

What is Inconsistency?

• Disagreement between direct and indirect evidence
• Suggests violation of transitivity
• Must be assessed in closed loops

Methods to Assess:

1. Loop-specific: Compare direct vs indirect in each loop
2. Node-splitting: Separate direct and indirect for each comparison
3. Global: Design-by-treatment interaction model

R Code for Node-Splitting:

# Node-splitting analysis
netsplit(nma)

# Interpretation:
# p < 0.05 suggests inconsistency for that comparison

6. Ranking Treatments

SUCRA (Surface Under Cumulative Ranking):

• Ranges from 0% to 100%
• Higher = more likely to be best
• Accounts for uncertainty

P-scores (frequentist equivalent):

# Get rankings
netrank(nma, small.values = "good")

# SUCRA-like plot
plot(netrank(nma))

Caution: Rankings have high uncertainty - always report with confidence intervals!

7. Presenting Results

League Table:

# Create league table
netleague(nma, digits = 2)

Forest Plot of All Comparisons:

# Forest plot vs reference
forest(nma, reference.group = "placebo")

Network Graph:

# Network visualization
netgraph(nma, 
         plastic = FALSE,
         thickness = "number.of.studies",
         multiarm = TRUE)

Assessment Questions

1. Basic: "What is the difference between direct and indirect evidence?"

   • Correct: Direct comes from head-to-head trials; indirect is inferred through common comparators

2. Intermediate: "What is the transitivity assumption and why is it important?"

   • Correct: Studies must be similar enough for indirect comparisons to be valid

3. Advanced: "How would you interpret a significant node-splitting test?"

   • Guide: Suggests inconsistency between direct and indirect evidence for that comparison; investigate sources of heterogeneity

Common Misconceptions

1. "NMA always gives better estimates than pairwise MA"

   • Reality: Only if transitivity holds; otherwise can introduce bias

2. "Treatment rankings are definitive"

   • Reality: Rankings have high uncertainty; focus on effect estimates

3. "More connections = better network"

   • Reality: Quality of evidence matters more than network complexity

Example Dialogue

User: "I have 15 trials comparing 5 antidepressants. Some are head-to-head, some vs placebo. How do I analyze this?"

Response Framework:

1. Acknowledge NMA is appropriate
2. Ask about outcome type and effect measure
3. Discuss transitivity assessment
4. Guide through network visualization
5. Explain consistency checks
6. Discuss ranking with appropriate caveats

References

• Cochrane Handbook Chapter on NMA
• Salanti G. Indirect and mixed-treatment comparison. Lancet 2012
• Rücker G, Schwarzer G. netmeta package documentation
• PRISMA-NMA extension statement

Adaptation Guidelines

Glass (the teaching agent) MUST adapt this content to the learner:

1. Language Detection: Detect the user's language from their messages and respond naturally in that language
2. Cultural Context: Adapt examples to local healthcare systems and research contexts when relevant
3. Technical Terms: Maintain standard English terms (e.g., "network meta-analysis", "SUCRA", "transitivity") but explain them in the user's language
4. Level Adaptation: Adjust complexity based on user's demonstrated knowledge level
5. Socratic Method: Ask guiding questions in the detected language to promote deep understanding
6. Local Examples: When possible, reference studies or guidelines familiar to the user's region

Example Adaptations:

• 🇧🇷 Portuguese: Use Brazilian clinical guideline examples (CONITEC evaluations)
• 🇪🇸 Spanish: Reference PAHO/OPS treatment recommendations
• 🇨🇳 Chinese: Include examples from Chinese NMA publications

Related Skills

• meta-analysis-fundamentals - Basic concepts prerequisite
• heterogeneity-analysis - Understanding between-study variation
• bayesian-meta-analysis - Alternative modeling approach
• grade-assessment - Rating certainty of NMA evidence