Learn target trial emulation
A short list rather than a bibliography. Everything below is worth reading end to end; the corpus this site indexes is under Studies.
Start here
The paper that named the framework. Start here: it lays out the protocol components and why specifying them before analysis is what prevents the classic errors.
The design errors the framework exists to prevent, each shown as a violated trial component. The most practically useful single paper for someone reviewing an observational study.
A four-page summary for clinical readers. The right thing to hand a collaborator who has not met the framework.
A worked example in a setting where eligibility and time zero are genuinely hard to define.
Benchmarking against trials
The largest prospective benchmarking programme: 32 trials emulated in claims data, with agreement declared in advance. The single biggest contributor to this database.
Context for what "agreement" has meant in earlier meta-epidemiological work, and why the metric you choose changes the answer.
Software
Sequential trial emulation with inverse-probability weighting, including the expansion of person-time into nested trials.
For sustained treatment strategies, where standard adjustment cannot handle time-varying confounding affected by prior treatment.
The protocol components
Specify each of these for the trial you would have run, then say how you emulate it. Most published failures are a component left implicit.
| Component | What goes wrong when it is implicit |
|---|---|
| Eligibility criteria | Conditioning on events that happen after time zero, which selects on the outcome. |
| Treatment strategies | A strategy nobody could actually follow, or one whose adherence is undefined. |
| Assignment procedure | Confounders left unadjusted because no one wrote down what assignment depends on. |
| Follow-up start (time zero) | Immortal time bias: follow-up begins before treatment is assigned, so the treated cannot die early by construction. |
| Outcome | An outcome measured differently in the data than in the trial, making the comparison meaningless. |
| Causal contrast | Intention-to-treat and per-protocol effects reported interchangeably. |
| Analysis plan | Time-varying confounding handled with standard adjustment, which biases rather than corrects. |