Target trial emulation.

Turn a broad question into a trial protocol

“Does the program work?” leaves several decisions unanswered. Who can enter? When is the decision made? What precisely are the alternatives? Which outcome matters, and over what period? Target trial emulation begins by specifying the randomized trial that would answer the question, then mapping its protocol to observational data.

The interactive protocol covers eligibility, strategies, assignment, follow-up, outcome, causal contrast and analysis. A hypothetical trial would randomize the strategies. Its observational emulation uses the treatment choices actually recorded and must address the resulting confounding. Emulation does not manufacture randomization.

Why the date at the top of the record matters

In this invented example, a support program becomes available on a chosen day. People with an earlier unplanned visit cannot start it. A comparison of future starters against never-starters, counted from the initial record date, gives the starter group a guaranteed event-free period. The “A benefit from nothing?” preset produces an apparent benefit even when the programmed effect is exactly zero.

The aligned example instead specifies a trial among people still event-free and otherwise eligible on the program’s start day. Eligibility, strategy classification and follow-up begin together. This is a later-start question with a different eligible population; it is not a universal correction for an effect defined at the initial record date.

Design first, then address confounding

Higher risk people can be more likely to enter the program. Even after aligning time zero, the raw comparison can therefore look harmful when the program reduces risk within both risk groups. Standardization calculates each strategy’s outcome within each risk group, then averages both using the same target population mix.

The mathematical bridge shows the expected group sizes, event counts, risks, risk differences and risk ratios. The adjusted result matches the simulator’s known answer because the example has complete measurement of its only confounder, overlap in treatment choices, full adherence and complete follow-up. Actual studies must assess those assumptions and quantify uncertainty.

From the laboratory to an observational study

Document a protocol, check whether the required information exists, construct the eligible cohorts at the decision point, choose an estimator suited to the causal structure, and investigate measurement, missingness and unmeasured confounding. Questions involving delayed initiation, repeated eligibility or sustained strategies may require designs beyond this simple demonstration.

The lesson includes a downloadable fictional protocol and links to primary methods papers, starting with Hernán and Robins on specifying the target trial and the importance of aligning time zero. Continue with IPTW and matching for adjustment, or intercurrent events for events that change the interpretation of an outcome.