Target Trial Emulation Framework

What it is

— The Target Trial Emulation (TTE) framework is a methodological discipline for designing and analyzing observational ("real-world") studies of treatment effects by first writing the protocol of the hypothetical randomized trial you would run if you could, and then emulating each of its components in the available data. It was articulated by Miguel Hernán and James Robins (Harvard) in the 2016 American Journal of Epidemiology paper "Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available," operationalized in the companion 2016 Journal of Clinical Epidemiology paper on preventing immortal time bias, and crystallized into the now-standard seven protocol components in the 2022 JAMA "Target Trial Emulation" guide: (1) eligibility criteria, (2) treatment strategies, (3) assignment procedures, (4) start and end of follow-up (time zero), (5) outcomes, (6) causal contrast (estimand), and (7) the analysis plan. TTE has no formal maintaining organization in the EQUATOR/Cochrane/ISPOR sense — it is a causal framework defined by these canonical papers, not a checklist owned by a society — though it has been woven into FDA and EMA/ENCePP real-world-evidence (RWE) expectations and into reporting/templating guidelines (HARPER, StaRT-RWE) that operationalize it.

When to use

— Apply TTE whenever you are designing, conducting, or appraising a non-interventional comparative-effectiveness or safety study that aims to estimate the causal effect of a treatment strategy from routinely collected data (claims, EHR, registries, linked sources), and especially when that study supports an FDA/EMA submission, an HTA/payer dossier, a peer-reviewed comparative manuscript, a registered protocol, or a post-authorization safety study (PASS). The decision rule versus its siblings: use TTE as the causal design spine that fixes eligibility, treatment strategies, time zero, the estimand, and the analysis; pair it with HARPER or StaRT-RWE when you need a structured protocol template to document those choices; and report the completed study with STROBE (or RECORD/RECORD-PE for routinely collected health data). TTE is not a substitute for these — it is the layer that makes the protocol causally coherent before the template is filled or the report is written. For a regulatory PASS the ENCePP checklist governs procedural completeness; TTE governs whether the design actually identifies a causal effect. For randomized trial protocols use SPIRIT, not TTE.

What it requires

— TTE forces explicit pre-specification of seven things, each of which has a real-world-data failure mode it is designed to prevent. (1) Eligibility must be assessable using only information available at or before time zero, with a documented data-fitness-for-use assessment of whether the source can actually capture those criteria. (2) Treatment strategies must be well-defined and sustained or point-interventions (e.g., "initiate and remain on drug A"), not vague "exposure" definitions. (3) Assignment must specify which baseline covariates are needed to make treatment exchangeable conditional on measured confounders — the explicit confounding-control plan (propensity scores, high-dimensional proxies, or g-methods). (4) Time zero must align eligibility, treatment assignment, and start of follow-up at a single index moment so there is no immortal time and no adjustment for post-baseline variables on the causal pathway. (5) Outcomes must rest on a validated phenotype/algorithm with reported operating characteristics. (6) Causal contrast must name the estimand — intention-to-treat-analogue versus per-protocol, and how intercurrent events (discontinuation, switching, death) are handled. (7) The analysis must match the estimand (PS-based methods for the ITT-analogue; clone-censor-weight or inverse-probability weighting for a sustained per-protocol estimand) and report attrition, balance diagnostics, positivity, and sensitivity/quantitative bias analyses.

When NOT to use — limitations and common misapplications

— TTE is a design framework, not a proof of validity. (1) "Specifying a target trial" does not eliminate unmeasured confounding — emulation aligns time and structure, but the conditional-exchangeability assumption can still fail; a well-specified target trial built on incomplete covariates is still confounded. (2) Estimand bait-and-switch — the default emulation yields an ITT-analogue (effect of initiating a strategy); a per-protocol (sustained-adherence) estimand requires clone-censor-weight or IPW for informative censoring. A common error is claiming a per-protocol effect while actually running an initiation-only analysis. (3) Time-zero misalignment is the dominant practical failure — grace periods, eligibility-time ambiguity, and "exposure defined over a period" all reintroduce immortal time the framework exists to remove. (4) Not for descriptive, single-arm, or hypothesis-generating work — TTE presupposes a comparative causal question; forcing it onto disease-burden or natural-history description is misapplication. (5) Framework-as-theater — listing the seven components in a protocol table without making the causal contrast actually estimable in the data (no defensible comparator, no positivity, an unmeasurable eligibility criterion) is box-ticking. (6) TTE is a design layer, not a reporting checklist and not a risk-of-bias instrument: completing a TTE protocol does not discharge STROBE/RECORD-PE reporting or a formal ROBINS-I appraisal.

How it maps to this catalog

— In this repo, TTE's seven components are implemented by specific concepts a reviewer can hold the protocol against: - The framework itself and its worked emulation: target-trial-emulation. - Treatment-strategy + assignment + time-zero (components 2–4): active-comparator-new-user is the usual analytic core — new-user + active comparator + time-zero alignment maps directly onto trial eligibility and assignment. - Confounding control (component 3): high-dimensional-propensity-score-hdps-rwe for proxy adjustment when key confounders are unmeasured. - Causal contrast / estimand and intercurrent events (component 6): estimands-ate-att-intercurrent-events-rwe. - Outcome and eligibility ascertainment (components 1, 5): diagnosis-phenotype-algorithm-1ip-2op-time-window-rwe and claims-analysis for code-list construction and validation. - Follow-up integrity and attrition (component 4, analysis): attrition-and-loss-to-follow-up-rwe.

Applied note (claims/EHR/registry RWE)

In a claims emulation, time zero is the first qualifying dispensing (NDC + fill date), eligibility is assessed only in the pre-index enrollment window, and continuous medical+pharmacy enrollment across that window is required so "no prior use" is observed rather than missing — Medicare Advantage-only person-time, which lacks fee-for-service claims, must be excluded or it masquerades as a clean washout. In EHR, initiation is the order/administration (prefer linked dispensing to confirm the patient started), and visit-driven capture makes loss to follow-up potentially informative. Registries strengthen indication, severity, and adjudicated outcomes but need claims/death-index linkage for complete exposure and censoring. Across all sources, the discipline is identical: fix one time zero, measure covariates only before it, and apply the same outcome and censoring rules to every arm.