European Statistical Forum

The field of Real World Evidence is an emerging area which is increasingly applied in medical research. On the other hand, RCTs are considered as the gold standard in regulatory decision making. This talk will shed light on regulatory initiatives and provide perspectives on various aspects that we see in Scientific advice and Marketing Authorisation Applications. In addition, methodological challenges related to causal inference will be discussed.

The individual causal association (ICA) is a surrogacy metric designed to assess the validity of a binary outcome as a putative surrogate for a binary true endpoint. The ICA is based on two pillars:

i) Information theory and

ii) a bivariate causal inference model for a binary surrogate and true endpoint.

The ICA has a simple and appealing interpretation in terms of uncertainty reduction. The identifiability issues inherent to the use of causal inference models are tackled using a two-step procedure.

In the first step, the region of the parametric space of the distribution of the potential outcomes, compatible with the data at hand, is geometrically characterized.

Further, in a second step, a Monte Carlo approach is proposed to study the behavior of the ICA on the previous region. The method is illustrated using data from the Collaborative Initial Glaucoma Treatment Study.

A newly developed and user-friendly R package Surrogate is provided to carry out the evaluation exercise.

When reporting “average” values of continuous endpoints measured at end of trial in a parallel arm RCT the preferred measure is the Least Squares Means. But what is the actual interpretation of this estimated average?

In this talk we argue that Least Squares Means may actually be viewed as estimates of average potential outcomes. Specifically, we show how the least squares means may be identified as such and estimator via the G-computation formula (Robins, 1986).

Besides offering a formalized interpretation of LSMEANS, the above characterization also highlights a somewhat overlooked issue with standard errors of LSMEANS supplied by most standard software. We argue that when models for analysing the endpoint contains covariates like a baseline measurement of the endpoint, then the usual standard error is systematically too low. We finally show how the G-computation framework facilitates correct estimation of the standard error in these instances. All developments are exemplified with a case study assessing QTc prolongation based on LSMEANS.

Robins J. A new approach to causal inference in mortality studies with a sustained exposure period — application to control of the healthy worker survivor effect. Math Model. 1986;7(9–12):1393–1512.

In diabetes trials some patients may require rescue medication during follow-up. If the level of rescue medication use differs between treatment groups, a treatment policy / intention to treat analysis may be difficult to interpret. Here a hypothetical estimand which targets the effect that would have been seen had rescue medication not been available may be of interest to some stakeholders. In this talk I will discuss statistical methods for estimation of such hypothetical estimands. I will first describe hypothetical estimands using the causal inference concepts of potential outcomes, before using the existing causal inference machinery to describe what assumptions are needed to estimate hypothetical estimands. In particular this will allow us to be clear about what variables need to be adjusted for to estimate hypothetical estimands. I will then discuss both ‘causal inference’ and ‘missing data’ methods (such as mixed models) for estimation, and show that in certain situations estimators from these two sets are in fact identical. These links may help those familiar with one set of methods but not the other. They may also identify situations where currently adopted estimation approaches may be relying on unrealistic assumptions, and suggest alternative approaches for estimation.

Questions on the treatment effect in sub-populations defined by post-randomization events are not uncommon in drug development. In the causal inference literature estimands of this type are typically referred to as principal stratum estimands.

As the population is defined based on events that may be influenced by the received treatment, randomization alone can no longer be relied upon for assessment of the treatment effect. Additional assumptions are required to identify the estimand.

In this presentation we will review examples from drug development, where principal stratum estimands could be of interest, review estimation strategies and provide a concrete recent example in a multiple sclerosis development program.

Missing attributes are ubiquitous in causal inference, as they are in most applied statistical studies.

In this work, we consider various sets of assumptions under which causal inference is possible despite missing attributes and discuss corresponding approaches to average treatment effect estimation, including generalized propensity score methods and multiple imputation. Across an extensive simulation study, we show that no single method systematically out-performs others. We find, however, that doubly robust modifications of standard methods for average treatment effect estimation with missing data repeatedly perform better than their non-doubly robust baselines; for example, doubly robust generalized propensity score methods beat inverse-weighting with the generalized propensity score. This finding is reinforced in an analysis of an observational study on the effect on mortality of tranexamic acid administration among patients with traumatic brain injury in the context of critical care management.

Here, doubly robust estimators recover confidence intervals that are consistent with evidence from randomized trials, whereas non-doubly robust estimators do not.