Target Trial Emulation

Defines the exact inclusion and exclusion rules that would govern a hypothetical trial, then applies them to observational data at a specific baseline time zero. This prevents immortal time bias and prevalent user bias by ensuring patients are not enrolled after the outcome has already occurred.

• Key principle: Criteria must be assessable at the moment of treatment assignment
• Common pitfall: Using future information to define eligibility
• Example: Requiring patients to have a diagnosis code for diabetes before initiating metformin, not after

Establishes the causal contrast by defining the active comparator and control group as they would exist in a randomized trial. This involves specifying the grace period for treatment initiation and handling treatment switching through an intention-to-treat or per-protocol emulation approach.

• Active comparator design: Comparing a new drug to an existing standard-of-care rather than non-use
• Grace period: A clinically meaningful window (e.g., 30 days) after eligibility to initiate treatment
• Cloning technique: Creating duplicate patient records to prevent immortal time bias when comparing sustained strategies

Defines the primary endpoint and the follow-up period using the same rigor as a clinical trial. The outcome must be measured from time zero and assessed identically in both arms to prevent differential misclassification. Follow-up typically ends at the earliest of outcome occurrence, death, loss-to-eligibility, or administrative censoring.

• Blinded assessment: Using objective endpoints (e.g., mortality) minimizes detection bias
• Competing risks: Death before the outcome must be handled explicitly
• Example: Measuring 5-year cardiovascular event risk starting from the date of statin initiation versus non-initiation

Pre-specifies the covariates to be balanced between arms, mirroring the role of randomization. This typically uses propensity score methods or inverse probability weighting to control for measured confounders. The key is that covariates are measured at baseline, not after treatment initiation.

• Propensity score matching: Pairing treated and untreated patients with similar estimated treatment probabilities
• Stabilized weights: Reducing variance in inverse probability-weighted estimators
• Negative control outcomes: Testing for residual confounding by examining associations known to be null

Addresses deviations from the assigned treatment strategy, such as non-adherence or treatment discontinuation. The emulation distinguishes between the intention-to-treat effect (the effect of assignment) and the per-protocol effect (the effect of sustained adherence), often using inverse probability of censoring weighting to adjust for informative loss to follow-up.

• Censoring at deviation: Patients are artificially censored when they stop adhering
• Time-varying confounding: Factors like side effects that influence both adherence and outcome must be adjusted
• Example: A patient who stops taking a prescribed drug is censored in the per-protocol analysis but retained in the intention-to-treat analysis

A foundational method for emulating the randomization of a target trial. PSM estimates the probability of receiving the treatment based on observed baseline covariates. Matching treated and control units with similar scores creates balanced comparison groups, mimicking the exchangeability expected in a randomized experiment. This directly addresses confounding by indication, a primary threat to validity in observational studies.

A technique to handle informative censoring—a form of selection bias where loss to follow-up is related to the outcome. In target trial emulation, IPCW assigns higher weights to uncensored individuals who resemble those who were censored. This preserves the trial's intention-to-treat principle by adjusting for deviations from protocol that would break randomization, ensuring the analysis reflects the full assigned groups.

The analytical principle that compares outcomes based on the initial assigned treatment strategy, regardless of adherence. Emulating this principle is critical in observational studies. It requires defining time zero (eligibility and assignment) precisely and using methods like cloning and censoring to handle treatment switching, ensuring the estimate reflects the real-world effectiveness of a decision to treat, not just the biological effect of perfect adherence.

A methodological approach to handle treatment strategies that diverge after baseline. Each eligible individual is cloned and assigned to each strategy at time zero. Clones are then artificially censored when their observed data deviates from their assigned protocol. This design, combined with IPCW, explicitly emulates a multi-arm randomized trial and prevents the immortal time bias that plagues naive comparisons.

A visual tool for encoding assumptions about the causal structure of the problem before emulation begins. A DAG identifies confounders, mediators, and colliders. This step is essential for selecting the correct adjustment set for methods like PSM or IPCW. It ensures the observational analysis explicitly controls for the right variables and avoids introducing collider bias by conditioning on common effects of treatment and outcome.

A class of models that estimates the causal effect of a time-varying treatment in the presence of time-dependent confounding. In the target trial framework, MSMs are often fit using inverse probability of treatment weighting (IPTW). This allows for the estimation of parameters like the Average Treatment Effect (ATE) in complex longitudinal settings where treatment and confounders evolve together, closely mirroring a pragmatic trial.