Variants of Randomized Trials (Cochrane Handbook Ch 23)

Your Role

Guide handling of non-standard RCT designs in SRs: cluster RCTs, crossover trials, stepped-wedge designs, equivalence/non-inferiority trials, and multi-arm trials. Each requires different data extraction, effect measure computation, and analysis.

---

Prerequisites

• Included studies list with designs identified (from sr-extraction)

---

Workflow

Step 1: Classify Each Study's Design

ASK the user for each included study:

1. Was randomization at individual or group level? (cluster)
2. Did patients receive both treatments in sequence? (crossover)
3. Was there a sequential roll-out of the intervention? (stepped-wedge)
4. Is the trial testing equivalence/non-inferiority (rather than superiority)?
5. Are there >2 intervention groups? (multi-arm)

Step 2: Cluster RCTs

Detection: Individual randomization not stated → suspect clustering. Check for: ICC reported, unit of analysis error, multilevel analysis.

Extraction: Identify if trial correctly accounted for clustering. If not → extract raw data and compute effective sample size.

ICC imputation (if not reported):

python3 scripts/cluster_handler.py --study Smith2021 --type extract --n 200 --m 20

Effective sample size (ESS):

ESS = N / (1 + (m - 1) × ICC)
where m = average cluster size

Analysis: Use ESS in meta-analysis (round down). Sensitivity: exclude cluster trials and compare.

RUN cluster-adjusted MA:

Rscript scripts/cluster_ma.R --data extraction.csv --icc 0.05

Step 3: Crossover Trials

Detection: Same patients receive both treatments in random order with washout period.

Appropriateness: Crossover is appropriate for stable chronic conditions, NOT for acute EM conditions (sepsis, trauma, cardiac arrest).

Extraction: Extract paired data:

• Mean difference (MD) between treatments
• SD (or SE) of the paired differences
• OR correlation coefficient between paired measurements

Analysis:

• If paired data available: use paired analysis
• If only unpaired data reported: treat as parallel-group (conservative), or impute correlation

python3 scripts/crossover_extract.py --study Jones2020 --type impute --correlation 0.5

Flag: carryover effects possible. If washout <5 half-lives → high risk of carryover.

Step 4: Stepped-Wedge Trials

Detection: All clusters start in control, sequentially cross to intervention at random time points. Common in EM quality improvement studies.

Extraction: Extract from appropriate analysis (mixed-effects model with fixed time effect + random cluster effect).

Include only if: The analysis accounted for time effects. If analyzed as simple cross-sectional pre-post → exclude or flag as high risk.

Analysis: Use effect estimate directly from the stepped-wedge analysis. Do NOT treat as parallel-group.

Step 5: Equivalence and Non-Inferiority Trials

Detection: Trial declares a non-inferiority margin (delta). Reported as "non-inferior" if CI excludes delta.

Extraction: Extract BOTH ITT AND per-protocol results (both required per CONSORT). Note the pre-specified margin.

Analysis:

• For NON-INFERIORITY: include in MA. If converting OR to RR, check margin compatibility.
• For EQUIVALENCE: include in MA if compatible with superiority trials.
• SENSITIVITY: exclude NI/equivalence trials and compare.

FLAG: "The non-inferiority margin was [delta]. The trial's finding [does/does not] apply to superiority questions."

Step 6: Multi-Arm Trials

Extraction: Identify all eligible intervention arms and comparator arms.

Analysis:

• If 1 intervention + 2 comparators → include only the relevant comparator
• If 2 interventions + 1 comparator → include both interventions:
  • Option A: combine intervention arms (pool across arms)
  • Option B: split control group (floor(N_c / k) per comparison)
  • Option C: select the most relevant intervention arm

Option B is generally recommended (Ch 23.3.4):

python3 scripts/split_control.py --n-control 120 --k-arms 2 --method floor

---

Scripts

scripts/cluster_handler.py

Extract or impute ICC, compute effective sample size, assess unit-of-analysis error. Usage: python3 cluster_handler.py --study Smith2021 --type impute --outcome mortality --icc-estimate 0.05

scripts/cluster_ma.R

Meta-analysis incorporating cluster trials via ESS or multi-level model. Usage: Rscript cluster_ma.R --data extraction.csv --icc 0.05 --method ess

scripts/crossover_extract.py

Extract paired data from crossover trials. Impute correlation if needed. Usage: python3 crossover_extract.py --study Jones2020 --md 2.5 --sd-paired 3.2

scripts/split_control.py

Split the control group N across multiple comparisons from multi-arm trials. Usage: python3 split_control.py --n-control 120 --k-arms 2 --method floor

---

Assets

assets/rct-variant-classifier.md

Decision tree for identifying variant RCT designs from study methods text.

assets/icc-reference-table.md

Published ICC values for common EM/CC outcomes from the literature (e.g., ICU LOS ICC ≈ 0.05-0.15, mortality ICC ≈ 0.01-0.05).

assets/non-inferiority-margins-em.md

Commonly used non-inferiority margins in EM/CC trials: mortality RR margins 1.10-1.50, mortality RD margins 5-10%, intubation success RD margin 5%.

---

Guardrails

1. "Cluster trials analyzed without adjusting for clustering commit a unit-of-analysis error — ALWAYS adjust (Ch 23.1.3)."
2. "Crossover trials are INAPPROPRIATE for acute EM conditions that change over time (sepsis, trauma). Check period-by-treatment interaction before including (Ch 23.2.4)."
3. "Do NOT combine equivalence/non-inferiority trials with superiority trials without justification and sensitivity analysis."
4. "For stepped-wedge trials: only include if analysis accounted for time effects (fixed + random time)."
5. "If cluster trial does not report ICC, impute from similar trials and test sensitivity in meta-regression."
6. "For multi-arm trials, splitting control groups is conservative — verify with sensitivity analysis combining arms."

Handoff

→ sr-extraction (variant-adapted extraction), → sr-synthesis (adjusted MA with ESS/nested models)