Introduction to Causal Inference and Causal Data Science

Day 3:
Target Trial Emulation

Bas B.L. Penning de Vries
Dept. Data Science & Biostatistics
Julius Center, UMC Utrecht

Test-your-knowledge quiz

Question 1

Which of the following options best describes the meaning of a potential outcome according to the counterfactual (or potential) outcomes framework?

  1. A possible value of the outcome variable
  2. The outcome of an individual that would be observed had treatment been set (by intervention) to a certain value
  3. The best outcome an individual can achieve
  4. A possible outcome of a study

Learning objectives

By the end of today, you’ll be able to

  • Describe what is meant by a target trial and target trial emulation
  • Identify key components of target trial emulation, including the determination of the start of follow-up (time zero)
  • Recognise a taxonomy of estimands relevant to target trial emulation and distinguish between common targets such as intention-to-treat and per-protocol effects (Friday!)
  • Describe the relevance of target trial emulation in causal inference from observational data
  • Recognise common deviations from a target trial in observational studies
  • Explain the basics of commonly used methods to address these deviations

“Causal inference from observational data can be viewed as
an attempt to emulate a hypothetical randomized trial”

Hernán and Robins, 2020, Causal Inference: What If

Why trials?

Causal inference is about speculating what would happen if …

A causal effect is a contrast between the answers to what-if questions.

Fundamental obstacle: impossible to observe the consequences of ≥ 2 mutually exclusive actions (interventions, treatments, etc.)

Target trial (emulation)

Target trial

A hypothetical trial that — if implemented — would readily allow us to answer our what-if question

  • To help communicate causal estimand (because identification is “straightforward”)
  • To facilitate appraisal of actual research designs (and avoid methodological problems)

Target trial emulation (TTE)

Explicit attempt to address deviations from a target trial, given the (observational) study data at hand

  1. Specify target trial
  2. Emulate it!

Problem 1: Ill-defined or irrelevant interventions

Formulating a target trial helps to communicate the causal estimand and helps to avoid asking vague or “silly” questions (about ill-defined or irrelevant interventions).

  • Eligibility defined by post-baseline events
  • Causal effect of (a reduction/increase in) BMI?
  • “Does water kill?” (Hernán, 2016)
  • Unclear treatment strategies (e.g., stopping rules, dosage, etc.)

Problem 2: Immortal-time bias

Example: do statins prevent cancer?

Problem 3: Including prevalent users

  • Would you consider initiating a treatment regime now (at baseline) for a patient who is already on treatment (prevalent user)?
  • Prevalent users are not part of the target population!
  • Inclusion might result in (selection) bias!

Misalignment of eligibility, treatment assignment, and the start of follow-up can result in time-related bias such as immortal time and selection of prevalent users.

Matthews et al, 2023

Summary

Target trial emulation (TTE)

Explicit attempt to address deviations from a target trial, given the (observational) study data at hand

  1. Specify target trial
  2. Emulate it!

The target trial framework provides an organizing principle for the design of observational studies that leads to clinically interpretable results and analytic approaches that can prevent common biases.

Matthews et al, 2023