TTE Design Skill

• (A) An existing RCT I want to emulate (provide NCT ID, trial name, or reference)
• (B) A defined research question with PICO elements
• (C) An early-stage idea — help me define everything

Entry point determines which steps below are pre-filled vs. walked through:

• Entry A: Steps 1-2 may be partially pre-filled from the RCT; Step 4 starts with the provided RCT
• Entry B: Step 1 may be partially pre-filled from the stated question; Steps 2-4 walked through
• Entry C: All steps walked through interactively

Step 1: Define the Causal Question

Skip or abbreviate if Entry A or B provides a clear causal question.

Help the user formulate an explicit causal question:

• What causal effect are you trying to estimate?
• Is this about effectiveness, safety, or both?
• Frame as: "What would happen if we randomly assigned patients to [intervention] versus [comparator]?"

The question must be answerable in principle by a randomized trial. If it isn't, help the user refine until it is. A poorly defined causal question leads to ambiguous design choices downstream.

Read ../tte-foundation/references/framework.md for the full Step 1 guidance.

Step 2: Classify the Scenario

Present the three scenario types and help the user classify their study. Read ../tte-foundation/references/scenarios.md for the full definitions and sub-scenarios.

Use the decision logic from ../tte-foundation/references/scenarios.md:

1. Does a relevant RCT exist? → If no, likely Scenario 3
2. Does the RCT fully address the clinical question? → If yes, Scenario 1; if no, Scenario 2

Recommend a classification based on context, and ask the user to confirm. Also identify the specific sub-scenario (e.g., 1d, 2a, 2c) — this affects downstream design choices.

Step 3: Specify PICO

Walk through each element one question at a time. Partially pre-fill if Entry A or B provides elements.

• Population: Who is eligible? What clinical setting and time period? How is time zero determined?
• Intervention: What treatment strategy? Be specific about dosing, timing, duration.
• Comparator: What is the comparison? Suggest active comparator when possible (39.2% of published studies did not use one — this is a common weakness). If no-treatment comparator is used, flag the need for sequential design.
• Outcome: Primary and secondary outcomes. Time horizon for assessment.

The PICO must be precise enough to operationalize with real-world data.

Step 4: Review Existing RCTs

The approach depends on the scenario classified in Step 2.

For all scenarios: Search for relevant RCTs. If the ClinicalTrials.gov MCP tools are available, use search_trials with the PICO elements to find candidate trials. Present key details (NCT ID, title, enrollment, status, primary endpoints) and let the user select which to reference.

Scenario 1 (Replication/Prediction):

• Identify the RCT to emulate
• Use Cochrane Risk of Bias criteria to evaluate candidate RCTs when multiple exist
• The selected RCT's protocol becomes the basis for the TTE protocol
• Deviations from the RCT protocol must be explicitly documented

Scenario 2 (Extension):

• Recommend the dual-protocol approach:
  • Protocol A (Validation): Emulate a known RCT to verify data source and methods produce comparable estimates
  • Protocol B (Extension): Modify Protocol A to address the actual research question (broader population, new comparator, longer follow-up, etc.)
• Document which elements change between Protocol A and B, and why

Scenario 3 (No RCT Feasible):

• Document the absence of relevant RCTs
• Justify the target trial design via:
  • Clinical rationale and treatment guidelines
  • Biological plausibility
  • Expert consensus
• Flag that without an RCT benchmark, sensitivity analyses become even more critical

Step 5: Build the Protocol

Walk through each of the 7 core methodologic components. Read ../tte-foundation/references/components.md for definitions, quality standards, and common pitfalls for each component.

For each component, explain what the paper recommends, ask the user for their choices, and flag potential issues.

5.1 Eligibility Criteria

• Suggest criteria based on the RCT (if Scenario 1/2) and PICO
• Confirm all criteria use baseline information only — using post-baseline information introduces selection bias and immortal time bias (15.2% of published studies made this error)
• Ask whether participants could become eligible at multiple time points → if yes, recommend sequential trial emulation (only 14.8% of studies used this)
• Suggest including a participant flowchart

5.2 Treatment Strategies

• Define intervention and comparator in operational terms (drug names, doses, CPT codes, etc.)
• Classify the treatment strategy type:
  • Point treatment: Single treatment decision at time zero (e.g., medication initiation)
  • Static sustained: Continuous treatment per a fixed rule (e.g., remain on medication)
  • Dynamic: Treatment decisions depend on evolving patient status (e.g., treat if biomarker exceeds threshold)
• Recommend active-comparator new-user design for medication studies (Rec 4.1-4.2)
• For sustained strategies: define treatment windows, adherence criteria, and grace periods
• If using a no-treatment comparator: recommend sequential design (Rec 4.3)

5.3 Assignment Procedures

Recommend methods based on treatment strategy type:

Describe the approach to emulate randomization and specify balance diagnostics (e.g., SMD < 0.1).

5.4 Follow-up

• Define time zero explicitly — eligibility, treatment assignment, and follow-up start must align at the same point. Misalignment is the most common source of immortal time bias.
• Specify a look-back period for baseline covariate assessment
• Recommend a design diagram (visual timeline) — only 16.9% of published studies included one
• Define censoring events (death, disenrollment, end of study, treatment switching if relevant)
• Discuss potential for immortal time bias if there's a grace period

5.5 Outcomes

• Prespecify primary and secondary outcomes with clinical definitions
• Define time frames and measurement methods
• Document data sources for outcome identification
• Suggest outcome quality assessment (validation algorithms, chart review, validated phenotyping)
• Only 15.6% of published studies assessed outcome quality — this is a common gap

5.6 Causal Contrast

Help the user choose and justify the estimand:

• 32.1% of published studies did not state their causal contrast — always specify explicitly
• Designate primary and secondary contrasts
• Justify the selection based on treatment strategy type and clinical relevance

5.7 Statistical Methods

Align methods with the causal contrast:

• ITT: Standard regression, weighted models, propensity score methods
• PP: Inverse probability of treatment weighting, marginal structural models for time-varying treatments
• Dynamic strategies: Clone-censor-weight + G-methods or doubly robust estimators

Specify:

• Primary and secondary analyses
• Variance estimation method (robust SE, bootstrap)
• Missing data approach
• Pre-specified subgroup analyses (if any)

Step 5.8: Confounder Sub-Workflow

This is a critical part of protocol development. Follow the 5-layer workflow from ../tte-foundation/references/confounders.md:

Layer 1 — RCT-Informed (high confidence): Pull baseline covariates from RCTs identified in Step 4. These are evidence-based confounders. Present them to the user: "Here are confounders from related trials: [list]. These form the backbone of the adjustment set."

Layer 2 — LLM-Suggested (medium confidence, with guardrails): Based on disease area and intervention, suggest additional candidate confounders. Each suggestion must:

• Be labeled: "AI-generated suggestion, not evidence-based"
• Include a brief rationale
• Be explicitly accepted or rejected by the user before entering the list

Present grouped by category (demographic, clinical history, medications, lab values, socioeconomic, lifestyle).

Layer 3 — DAG Prompting (user-driven): Ask structured questions to elicit domain knowledge:

• "Are there variables that affect both treatment choice AND the outcome?"
• "Are there prognostic factors for the outcome that may differ between groups?"
• "Are there variables that are consequences of treatment but precede the outcome?" (potential colliders — avoid adjusting for these)

Organize into: treatment-outcome confounders, treatment-selection confounders, outcome-related prognostic factors.

Layer 4 — Classification (conceptual): For each confounder: classify as measured / proxy available / unmeasured. This is conceptual — actual data verification is /tte-data's job.

Layer 5 — Sensitivity Analysis Plan: Based on Layer 4 classifications:

• Unmeasured: recommend E-values, instrumental variables, negative/positive controls
• Proxies: recommend measurement error methods, sensitivity to proxy quality
• All studies: DAG-guided minimal adjustment set, varying assumptions

Document the sensitivity analysis plan as part of the statistical analysis section.

Step 5.9: Causal Assumptions

Prompt the user to consider and document three causal identification assumptions:

1. Exchangeability: Conditional on measured covariates, treatment groups are comparable (only 63.7% of studies discussed this)
2. Positivity: All covariate strata have non-zero probability of each treatment (only 37.1% discussed this)
3. Consistency: The intervention is well-defined — same treatment leads to same outcome regardless of how it was assigned (only 5.5% discussed this)

For each, discuss plausibility in the context of this specific study.

Output

Generate a complete TTE protocol document with two sections:

Section 1: Narrative Protocol

Full protocol following the narrative format from ../tte-foundation/references/protocol-template.md, covering all 7 components plus the confounder plan and causal assumptions.

Section 2: Tabular Protocol

Compact two-column table (Target Trial Specification | Emulation Approach) following the tabular format from ../tte-foundation/references/protocol-template.md.

Save the protocol to the project directory as tte-protocol.md.

End with:

Run /tte-assess to verify the methodologic quality of this protocol. Run /tte-data when ready to check data feasibility for implementation.

Reference Files

The following files in ../tte-foundation/references/ contain detailed guidance:

• framework.md — 5-step design framework
• scenarios.md — 13 sub-scenarios across 3 types, with decision logic
• components.md — 7 core components with quality standards and pitfall rates
• confounders.md — 5-layer confounder identification workflow
• protocol-template.md — Narrative and tabular protocol templates

Read scenarios.md during Step 2. Read components.md during Step 5. Read confounders.md during Step 5.8. Read protocol-template.md when generating output.