[M5][CAUSAL] Define the longitudinal causal-inference protocol for treatment strategies

[andreazedda]

Objective

Create a causal-analysis protocol that is explicitly separate from mechanistic what-if simulation and can support future estimation of dynamic treatment effects from suitable longitudinal data.

Scope

• define target-trial emulation templates;
• specify eligibility, treatment strategies, time zero, follow-up, outcomes, censoring, competing events, and estimands;
• represent time-varying treatment, toxicity, disease state, supportive care, and confounding;
• support marginal structural models, parametric g-formula, longitudinal TMLE, and other justified methods behind versioned interfaces;
• define positivity, consistency, exchangeability, measurement, and interference assumptions;
• implement diagnostics for overlap, weight instability, informative missingness, and model misspecification;
• keep causal estimates and mechanistic counterfactual trajectories as separate output classes.

Required output contract

causal question
estimand
population
strategies
outcome and horizon
causal graph
adjustment strategy
identification status
estimation method
diagnostics
sensitivity analysis
allowed conclusion

Acceptance criteria

• No causal estimate is produced without a documented identification argument.
• Time-varying confounding affected by prior treatment is handled explicitly.
• Mechanistic simulator outputs cannot be relabelled as causal effects.
• Positivity and weight diagnostics are mandatory.
• Sensitivity to unmeasured confounding and measurement error is reported.
• Target-trial specifications are frozen before estimation.
• Negative and non-identifiable results are valid outcomes.
• Causal outputs have distinct schemas, labels, and model cards.

Non-goals

This issue does not guarantee that current datasets identify any patient-specific or population treatment effect.