Advanced Methods II: Target Trial Emulation for Causal Inference in Practice

1. Course Aim

Using routinely collected data like electronic health records or claims data to answer questions about (long-term) ‎effectiveness or safety of drugs or preventive measures bears many sources of potential bias, over and above ‎confounding. These have been described as “self-inflicted” biases because they can often be avoided or ‎minimised by careful study designs. Target trial emulation (TTE) is an increasingly popular principle to facilitate ‎the formulation of suitable designs. However, lack of familiarity with the statistical techniques and relevant ‎software may deter researchers from actually using this valuable approach. This course therefore aims to address ‎and demonstrate practical aspects and implementation of TTE.‎

2. Requisites Statement

Advanced level course – prior knowledge of basics on causal inference (e.g., attendance of the unit ‘Introduction ‎to causal inference’) will be beneficial. ‎

3. Course Objectives

‎1.‎ Create an awareness of the most common design-related biases and an ability to identify these as ‎potential threats to the validity of naïve analyses.‎
‎2.‎ Develop a basic ability to interpret and apply the techniques of “sequential target trials” and “clone-‎censor-weighting” to avoid, e.g., prevalent user and similar types of bias. This includes knowledge on ‎the required data management aspects.‎
‎3.‎ Understand examples of dynamic treatments where the analysis calls for advanced techniques such as ‎the g-formula.‎
‎4.‎ Use and interpretation of the main techniques to check the plausibility of typical causal assumptions ‎underlying the validity of a TTE analysis.‎

4. Syllabus Outline

1. Principles of target trial emulation
We will revisit the key elements of a target trial and key aspects of its emulation. The importance of the ‎alignment of the assessment of eligibility, assignment to treatment arms, start of follow-up (aka ‘time-‎zero-alignment’) will be illustrated with real-data examples. This will highlight the potential for bias in ‎many traditional analyses.‎

2. Choosing a baseline and sequential trials ‎
Unlike RCTs, routinely collected (observational) data do not necessarily have a natural ‘baseline’ (or ‎time-zero). When this is not uniquely given by the research question there are statistical techniques to ‎deal with this ambiguity: for instance, it is often suitable to create a sequence of trials with many ‎baselines. Students will see practical examples including software code demonstrating this approach; ‎examples of sensitivity analyses will also be covered.‎

3. Sustained / dynamic treatment strategies and ‘cloning’‎
Finally, we will consider more advanced analysis techniques for sustained and dynamic treatment ‎comparisons. Here, and additional problem that can occur is that patients are then not uniquely ‎assignable to a single trial arm; this can be addressed by ‘cloning’ or the g-formula as will be illustrated. ‎Students will see further practical examples including software code demonstrating this approach; ‎practical implementation of the relevant sensitivity analyses will also be covered.‎