Comparative effectiveness
Answer trial-like causal questions and generate real-world evidence from longitudinal registry data.
Frameshift applies the target trial framework to comparative-effectiveness questions in Danish registry data. The protocol of the hypothetical randomised trial is written in full and fixed before the outcome is examined, and each component is then emulated in the observational data. We also run the statistics on randomised trials, which is where the design discipline described below comes from.
The discipline is in the sequence. Eligibility, treatment assignment and the start of follow-up are aligned at a single time zero, preventing immortal-time and related design biases. Where treatment initiation is the decision of interest, a new-user design avoids conditioning the comparison on prior treatment survival or tolerance. Where the available data cannot faithfully reproduce a component of the target trial, we state the departure explicitly and assess its likely impact, quantitatively where possible.
Protocol
The seven target-trial protocol components, with time-zero alignment shown separately because it is critical to the emulation. Each is fixed for the hypothetical trial and mapped to what the registers can actually support.
| Component | Target trial | Emulation in registry data |
|---|---|---|
| Eligibility | Assessed once, before assignment, on information available at baseline. | Applied using pre-baseline register history, with the lookback window stated and held constant across arms. |
| Treatment strategies | The regimens compared, including duration, switching and adherence rules. | Reconstructed from dispensing and procedure records, with an explicit grace period for initiation. |
| Assignment | Randomised, so the arms are exchangeable in expectation. | Not randomised. Conditional exchangeability targeted on measured covariates, sequentially on covariate history where treatment and confounders vary over time, and named as an assumption. |
| Time zero | Eligibility, assignment and start of follow-up coincide. | Aligned to the same instant. Misalignment here is a major source of immortal-time and related design bias. |
| Follow-up | From time zero until outcome, loss to follow-up or administrative end. | Defined from the same time zero using national longitudinal follow-up, with censoring and competing events handled according to the estimand. |
| Outcome | Prespecified outcome definition, ascertainment and assessment window. | Mapped to prespecified registry definitions, with differences in ascertainment, coding and measurement made explicit. |
| Causal contrast | Intention-to-treat effect, per-protocol effect, or both. | The observational analogue of each, stated explicitly. The analogue of intention to treat is the effect of treatment initiation, and unlike its trial counterpart it is not protected by randomisation, so it still requires full baseline confounding control and is not the conservative option. Per-protocol effects may require artificial censoring and weighting. |
| Analysis plan | Prespecified estimator, adjustment set and subgroup analyses. | Prespecified before the outcome is examined, with the estimator chosen to match the causal contrast. |
Departures from the target trial are documented as part of the deliverable.
Method
The analysis follows from the causal contrast, not from what is convenient in the data.
The default starting point.
A new-user restriction addresses prevalent-user bias on both of its mechanisms: the depletion of susceptibles that arises when a comparison conditions on having already tolerated treatment, and the adjustment for covariates that treatment has itself already altered. It works only if the lookback window is long enough to separate true initiators from re-initiators, which register history bounds. Choosing an active comparator that shares the indication reduces confounding by indication, to the extent that both arms genuinely sit at the same decision point. That equipoise is an argument to be made per question, not a property of the design, and it does not address channelling by severity within the indication.
For strategies not assignable at baseline.
Where a strategy is defined by what happens after time zero, each individual whose baseline data are compatible with more than one strategy is assigned to every such strategy, so the arms are identical at time zero by construction. Clones are censored when their observed treatment history departs from the assigned strategy. That censoring is non-random with respect to prognosis, so inverse-probability-of-censoring weights are fitted on the time-varying covariates that predict it. Validity rests on those covariates being sufficient, with no unmeasured common cause of deviation and outcome, and on positivity holding among those still uncensored. That is the strongest and least testable assumption in the procedure, and we state it. Because cloning duplicates individuals, confidence intervals are obtained by bootstrapping over persons.
Matched to the estimand.
We use inverse probability of treatment weighting, standardisation, or g-methods where treatment and confounders vary over time. We report the weight distributions and the covariate balance the weights actually achieve.
When the patient dies first.
Death may preclude the outcome of interest and should be handled according to the estimand rather than treated mechanically as censoring. Cause-specific hazards, cumulative incidence functions and other competing-event quantities answer different questions, and we state which one the clinical question requires.
Where the estimate earns trust.
We run negative control outcomes and exposures, alternative adjustment sets and lookback windows, and quantitative bias analysis to establish how strong unmeasured confounding would need to be to explain the result.
Questions it answers
Best suited to comparative questions where randomisation is unavailable or impractical, or where a trial's follow-up is too short to reach the outcome that matters.
Deliverables
Each deliverable below maps onto a part of the target-trial protocol.
Assess whether the available Danish register data can support a credible emulation of the target trial, and identify the main data and design limitations before a full study begins.
Prespecify eligibility, treatment strategies, assignment procedure, time zero, follow-up, outcomes, causal contrast and analysis plan.
Translate the protocol into operational cohort, exposure, covariate and outcome definitions using the available longitudinal data.
Estimate the prespecified contrast with appropriate adjustment, diagnostics and sensitivity analyses.
Deliver tables, figures, technical reporting and publication-ready methods and results, with assumptions and departures from the target trial documented explicitly.
Scope and limits
Every effect estimate from observational data rests on identifying assumptions: conditional exchangeability given the measured covariate history, positivity within every stratum of the covariates conditioned on, and consistency, which requires the treatment strategies to be sufficiently well defined. No interference between individuals is assumed alongside consistency. Emulation makes these assumptions explicit rather than implicit, and makes some of them checkable: violations of positivity from sparsity can be seen directly in the data, though structural violations cannot be distinguished from sparsity at any sample size and remain a subject-matter judgement. Violations of exchangeability can sometimes be detected with negative controls and their consequences quantified through bias analysis. Exchangeability and consistency themselves cannot be verified. Emulation makes them visible; it does not discharge them.
We report the assumptions each estimate depends on, evaluate observable implications and diagnostics where possible, quantify sensitivity to unmeasured confounding, and say plainly when a question cannot be answered credibly in the registers available.
Get in touch
Tell us the strategies you want to contrast and the outcome that matters. We will assess whether the available registers can support a credible emulation, what assumptions it would require, and where the important limitations lie.
