AI Agent Workflow Collection

Use after estimating an observational target trial emulation to stress-test whether randomization was plausibly emulated — negative control outcomes/treatments/populations, covariate balance, positivity / weight diagnostics, and benchmarking against a randomized trial. Falsification, not verification.

Use when the outcome can be precluded by another event — e.g. death before the outcome of interest, or any "competing event" / "truncation event" — and you must choose and justify a causal estimand before analysis. Also use when loss to follow-up is present and you need to frame it as a joint intervention on treatment and censoring. Forces an explicit estimand and flags the common controlled-direct-effect trap.

Use when deciding how to adjust for confounding in a target trial emulation — eliciting the set of baseline confounders, choosing an adjustment method (matching, standardization/g-formula, or IP weighting), and judging whether emulating randomization is even reasonable (vs. unmeasured / intractable confounding that may call for instrumental-variable approaches).

Use when the treatment is sustained over time rather than a one-time decision at baseline — e.g. "initiate and continue therapy unless contraindicated", treatment with a duration, dose strategies, or strategies defined by WHEN/HOW LONG to treat. Decides static vs dynamic, whether time-varying confounding / treatment-confounder feedback requires g-methods, and whether the strategies are distinguishable at time zero (→ cloning) and need a grace period.

Use when a user wants to specify, sharpen, or critique a causal question as a target trial for target trial emulation (TTE) with observational (or trial) data — e.g. "I want to estimate the effect of X on Y", "is this a good causal question?", "set up a target trial for ...", "specify the protocol", or before running any TTE analysis. Interactively elicits the eight protocol elements and pushes back on common, avoidable design errors before any estimation.

Use when defining or auditing "time zero" (start of follow-up) in a target trial emulation — when an analysis needs to align eligibility, treatment assignment, and start of follow-up, when there is risk of immortal-time or prevalent-user selection bias, when individuals are eligible at multiple times, or to decide between a single emulated trial and sequential (nested) trials.

Use to generate and run the R analysis for a (specified) target trial emulation — estimating the causal effect after the protocol, time zero, and confounders are set. Dispatches by OUTCOME TYPE (time-to-event, continuous, binary) and ADJUSTMENT METHOD (none/RCT, matching, standardization, IP weighting) and writes a readable, course-faithful R script that calls the TTECC helpers.

Use to draft the write-up of a target trial emulation — the protocol table, a CONSORT-style participant-flow diagram, methods, and results — with disciplined causal language (causal terms in Title/Intro/Methods/Discussion, never in Results). Produces a manuscript/methods skeleton.