ITC Analysis - Multi-Agent Evidence Synthesis Orchestration

Design and execute a complete indirect treatment comparison analysis for: $ARGUMENTS

Thinking

This workflow orchestrates multiple specialized agents to conduct rigorous ITC analysis following NICE DSU guidance and tidy modelling principles. The approach emphasizes:

• Phase-based coordination: Each phase builds upon previous outputs with clear handoffs
• Method-appropriate specialists: Matching the right expert to the analytical task
• Assumption transparency: Explicit checking and documentation of all assumptions
• Reproducibility: Complete documentation for replication
• Sensitivity analysis: Multiple approaches to validate findings

The multi-agent approach ensures each aspect is handled by domain experts:

• ITC Architect selects methodology
• Appropriate specialists execute analysis
• Code Reviewer validates implementation

Phase 1: Evidence Assessment & Method Selection

<Task> subagent_type: itc-architect prompt: | Assess the evidence structure and recommend ITC methodology for: $ARGUMENTS

Deliverables:

1. Evidence mapping:

   • Available studies and their characteristics
   • IPD vs AgD availability
   • Network structure and connectivity
   • Treatment comparisons of interest

2. Population assessment:

   • Key population characteristics across studies
   • Potential effect modifiers
   • Population differences requiring adjustment

3. Assumption evaluation:

   • Transitivity assessment (for NMA)
   • Effect modifier identification (for MAIC/STC)
   • Data requirements for each method

4. Method recommendation:

   • Primary recommended method with justification
   • Alternative methods for sensitivity
   • Package recommendations (meta, netmeta, maicplus, stc, multinma)

5. Analysis plan outline:

   • Key analyses to perform
   • Sensitivity analyses needed
   • Reporting requirements

Provide clear rationale for method selection referencing NICE DSU TSD guidance. </Task>

Phase 2: Data Preparation

<Task> subagent_type: itc-code-reviewer prompt: | Review and prepare data for ITC analysis based on methodology: {phase1.itc-architect.output}

Data preparation tasks:

1. Data validation:

   • Check IPD structure and completeness
   • Validate AgD format and required fields
   • Identify missing values and handling strategy

2. Effect size calculation (if needed):

   • Calculate study-level effect sizes
   • Variance estimation
   • Transformation to common scale

3. Covariate preparation:

   • Identify effect modifiers to include
   • Calculate centering values from AgD
   • Prepare covariate summaries

4. Data quality checks:

   • Outlier detection
   • Consistency checks
   • Documentation of data decisions

Provide tidy R code following TMwR patterns with validation functions. Action: review_only (document issues and recommendations) </Task>

Phase 3: Primary Analysis

Execute primary analysis based on recommended method from Phase 1.

If Pairwise Meta-Analysis:

<Task> subagent_type: pairwise-meta-analyst prompt: | Conduct pairwise meta-analysis based on: {phase1.itc-architect.output} Using prepared data from: {phase2.itc-code-reviewer.output}

Analysis requirements:

1. Primary analysis:

   • Fit both fixed and random effects models
   • Report effect estimate with 95% CI
   • Include prediction interval

2. Heterogeneity assessment:

   • Q statistic and p-value
   • I², τ², H² with confidence intervals
   • Interpretation of heterogeneity level

3. Publication bias (if ≥10 studies):

   • Funnel plot
   • Egger's/Peters' test
   • Trim-and-fill if indicated

4. Visualizations:

   • Forest plot with study-level results
   • Funnel plot

5. Sensitivity analyses:

   • Leave-one-out analysis
   • Fixed vs random effects comparison
   • Risk of bias subgroup (if available)

Use meta or metafor package. Provide complete, reproducible R code. </Task>

If Network Meta-Analysis:

<Task> subagent_type: nma-specialist prompt: | Conduct network meta-analysis based on: {phase1.itc-architect.output} Using prepared data from: {phase2.itc-code-reviewer.output}

Analysis requirements:

1. Network assessment:

   • Network diagram
   • Evidence distribution
   • Multi-arm trial identification

2. Primary NMA:

   • Fit consistency model (random effects)
   • All pairwise comparisons
   • League table

3. Consistency assessment:

   • Global consistency test
   • Node-splitting for closed loops
   • Net heat plot

4. Treatment rankings:

   • P-scores or SUCRA
   • Ranking probabilities
   • Cumulative ranking plots

5. Sensitivity analyses:

   • Fixed vs random effects
   • Excluding specific studies
   • Frequentist vs Bayesian comparison

Use netmeta (frequentist) and/or gemtc (Bayesian). Provide complete R code. </Task>

If MAIC:

<Task> subagent_type: maic-specialist prompt: | Conduct MAIC analysis based on: {phase1.itc-architect.output} Using prepared data from: {phase2.itc-code-reviewer.output}

Analysis requirements:

1. Weight estimation:

   • Define matching covariates
   • Center IPD on AgD targets
   • Estimate propensity weights
   • Bootstrap for variance

2. Weight diagnostics:

   • ESS calculation and interpretation
   • Covariate balance assessment
   • Weight distribution summary
   • Extreme weight identification

3. Anchored/Unanchored analysis:

   • Weighted outcome analysis
   • Effect estimates with bootstrap CI
   • Bucher indirect comparison (if anchored)

4. Multiple endpoints (if applicable):

   • Binary, continuous, time-to-event
   • Consistent methodology across endpoints

5. Sensitivity analyses:

   • Different covariate sets
   • Normalized vs unnormalized weights
   • Comparison with STC

Use maicplus package. Provide complete R code with diagnostics. </Task>

If STC:

<Task> subagent_type: stc-specialist prompt: | Conduct STC analysis based on: {phase1.itc-architect.output} Using prepared data from: {phase2.itc-code-reviewer.output}

Analysis requirements:

1. Model specification:

   • Define effect modifiers
   • Center covariates on AgD population
   • Specify regression model with interactions

2. Frequentist STC:

   • Fit outcome regression
   • Extract treatment effect at AgD population
   • Robust standard errors
   • Bucher indirect comparison (if anchored)

3. Bayesian STC (sensitivity):

   • Prior specification
   • Posterior summaries
   • Credible intervals

4. Model diagnostics:

   • Residual assessment
   • Interaction significance
   • Model fit statistics

5. Sensitivity analyses:

   • Different covariate sets
   • With/without interactions
   • Comparison with MAIC

Use stc package. Provide complete R code. </Task>

If ML-NMR:

<Task> subagent_type: ml-nmr-specialist prompt: | Conduct ML-NMR analysis based on: {phase1.itc-architect.output} Using prepared data from: {phase2.itc-code-reviewer.output}

Analysis requirements:

1. Network setup:

   • Set up IPD studies
   • Set up AgD studies
   • Combine network
   • Add integration points

2. Model fitting:

   • Specify regression formula
   • Define priors
   • Fit ML-NMR model
   • Check convergence (R-hat, ESS, traces)

3. Results extraction:

   • Relative effects for all comparisons
   • Treatment rankings
   • Predictions to target population

4. Consistency assessment:

   • Node-splitting
   • Direct vs indirect comparison

5. Sensitivity analyses:

   • Prior sensitivity
   • Integration point sensitivity
   • Covariate inclusion

Use multinma package. Provide complete R code with convergence diagnostics. </Task>

Phase 4: Code Review & Validation

<Task> subagent_type: itc-code-reviewer prompt: | Review all analysis code from Phase 3 for methodological and code quality issues.

Review checklist:

1. Methodological correctness:

   • Appropriate method for data structure
   • Assumptions properly checked
   • Correct effect measures used
   • Proper uncertainty quantification

2. Statistical issues:

   • Correct model specification
   • Appropriate inference methods
   • Multiple testing considerations
   • Sensitivity analyses adequate

3. Code quality:

   • Tidy modelling compliance
   • Reproducibility (seeds, versions)
   • Error handling
   • Documentation

4. Completeness:

   • All planned analyses performed
   • Diagnostics included
   • Visualizations appropriate

Action: review_only Review type: comprehensive

Provide detailed feedback with specific line references and recommendations. </Task>

Phase 5: Final Reporting

<Task> subagent_type: itc-architect prompt: | Synthesize all analysis results and prepare final report structure.

Report components:

1. Executive summary:

   • Key findings
   • Primary effect estimates
   • Strength of evidence

2. Methods summary:

   • Data sources
   • Analytical approach
   • Key assumptions

3. Results synthesis:

   • Primary analysis results
   • Sensitivity analysis results
   • Consistency of findings

4. Limitations:

   • Data limitations
   • Methodological limitations
   • Assumption violations (if any)

5. Conclusions:

   • Summary of evidence
   • Confidence in findings
   • Implications

Provide structured report outline following NICE/PRISMA guidelines. </Task>

Configuration Options

• primary_method: pairwise_ma | nma | maic | stc | ml_nmr | auto (recommended by architect)
• bayesian_analysis: true | false (include Bayesian as sensitivity)
• report_format: comprehensive | summary
• sensitivity_depth: minimal | standard | extensive

Success Criteria

1. Method Selection:

   • Appropriate for data structure
   • Justified with references to guidance

2. Analysis Quality:

   • All assumptions checked
   • Proper diagnostics performed
   • Sensitivity analyses completed

3. Code Quality:

   • Reproducible
   • Well-documented
   • Follows tidy principles

4. Reporting:

   • Complete and transparent
   • Follows PRISMA/NICE guidelines
   • Appropriate interpretation

Final Deliverables

Upon completion, the orchestrated analysis will provide:

• Complete R code for all analyses
• Diagnostic output and visualizations
• Results tables (effect estimates, rankings)
• Sensitivity analysis comparison
• Final report structure
• Reproducibility documentation (session info, seeds)