Reading focus
Syllabus: What Would Justify a Causal Claim?
Connect each protocol decision to the corresponding information in an observational study.
Why you’re reading it. Turn causal assumptions into a concrete design whose gaps can be inspected.
A question to consider. Using records of students who choose to attend a study group, specify eligibility, comparison strategies, starting time and an outcome; what important information might still be missing?
Why this reading
Hernán and Robins make the intended comparison explicit before analyzing an observational database. Specifying the trial one would like to conduct reveals decisions about eligibility, treatment strategies, timing and outcomes that an analysis can otherwise leave hidden.
Before you read
A target trial is the explicit experiment an observational analysis aims to emulate. Time zero is the start of follow-up, aligned with eligibility and assignment to strategies. An intention-to-treat effect concerns assignment; a per-protocol effect concerns following the specified strategy. These contrasts can require different assumptions and analyses.
Read for
- Compare the protocol components with the information an observational database would need to supply.
- Examine how misaligned starting times or selecting people after treatment begins can distort a comparison.
- Keep specifying the target separate from successfully emulating it with available measurements.
Consider afterwards
- An analysis compares people who complete six months of therapy with everyone who never starts. How would defining eligibility, assignment and follow-up at a common starting point change the question?
Editorial context
Emulation organizes design choices; it does not randomize existing observations or recover unmeasured confounders. The authors discuss data limitations, treatment strategies and adherence as substantive obstacles. A later clarification by Hernán and colleagues (2025), “The target trial framework for causal inference from observational data,” distinguishes avoidable design problems from limitations of the available data. The framework’s name is not evidence that an application satisfies its assumptions.