Different estimands require different methods of estimation, and different estimators may require more or less stringent assumptions. Our work focuses on identifying appropriate estimators for different settings.

What are estimators and why are they relevant?

An estimator is the method we use to analyse our data in order to compute an estimate of the treatment effect. Different estimators may answer different research questions (i.e. estimands) and may require different assumptions in order to be valid. It’s therefore important to use an appropriate estimator for a given setting, as inappropriate estimators may answer the wrong question or be at risk of bias.   

Some estimands, such as those based on a treatment policy strategy, can be simple to estimate, for instance using a standard intention-to-treat approach. However, estimation becomes more complicated when there is missing data, or if different estimands are of interest. For instance, instead of asking the question “What is the effect of assigning a treatment?”, we might be interested in answering “What is the treatment effect in those patients who adhere?”, or “What is the effect if patients hadn’t been allowed to switch treatment arms?” These questions are harder to answer, because occurrences like non-adherence or treatment switching may differ between treatment arms, which can break randomisation. Thus, more advanced statistical methods, often based on causal inference, are required.

A hypothetical estimand can be estimated using inverse probability of censoring weights. This is illustrated on a left-to-right time-line showing two individuals who don’t adhere (marked X) and so are replaced in the analysis by increasing the weights of similar individuals who do adhere.

What have we done?

Our work focuses on identifying the best estimators in different settings. We have:

  • Described a simple ‘Modified Intention-To-Treat’ approach to estimate the principal stratum effect in patients who would actually start on treatment
  • Described methods to estimate hypothetical strategies for treatment switching or non-adherence
  • Evaluated methods to estimate hypothetical strategies for non-adherence in non-inferiority trials
  • Developed methods to estimate treatment policy and hypothetical strategies with missing data, including in trials affected by pandemics

How will this make a difference?

Our work provides practical guidance on which estimators should be used in practice. Better choice of estimator can lead to more relevant questions being answered and reduced risk of bias.