European Statistical Forum - Programma

Clinical trials aim to provide reliable evidence of efficacy and safety to support approval of experimental medicines and to inform prescribers as to their use. To generate reliable data, designs aim to minimise bias, including through pre-specification of methods for analysis. Pre-specification has come to include not just identification of a primary endpoint, but families or hierarchies of secondary endpoints that are subject to study-wise type I error control. Also, investigations that aim to explore limitations in the data and to gain a full understanding of the dataset. The benefits of pre-specification are clear but, given the complexities of a clinical trial dataset, are there also risks?

Interpretation of clinical trial data can rely on statistical rules and algorithms, with only pre-specified analyses subject to error control being considered reliable for decision making and for labelling. Information on secondary endpoints outside of the hierarchy (or analysed after a test in the hierarchy has failed) and other, scientifically justifiable, investigations identified only post hoc can be dismissed as unreliable. Is this too restrictive, potentially ignoring what can be learned from a full interrogation of the dataset, from other trials, and from a biological and pharmacological understanding of the experimental treatment? The talk will consider whether a strict application of statistical rules preclude a well-informed, proportionate interpretation of clinical trial data with potentially detrimental consequences for drug approvals and labelling.

The role of biostatistics in the planning and conduct of a clinical study is established and pretty undisputed today.

Biostatistician’s role in the interpretation of results is however still undervalued. With the availability of more and more simple to use tools, however, correct planning and execution of a clinical trial becomes easier and less of an issue and problems in the interpretation of study results more dominant. This presentation will focus on issues in the interpretation of study results, for example the challenges between predefined controlled analyses versus the use of exploratory analyses or the need to understand the data in detail before the interpretation or the many different aspects to be considered when interpreting study results; and it will highlight the value of biostatistics there.

The talk will discuss the different aspects of clinical trial interpretation in the example of external control analyses in more details.

From a traditional perspective, multiplicity in clinical trials has related to advanced testing procedures in confirmatory trials with respect to considering a hierarchy or multitude of endpoints. In recent years it has become obvious that multiplicity in clinical development may have very different meanings ranging from indications, populations, and treatment groups up to having combined criteria for (GoNoGo) decision making and number of trials. The common theme is that statistical reasoning needs to play a key role in considering, evaluating and judging upon these aspects both from a companies’ and a regulators perspective.

In this presentation we will address and present some typical examples of multiplicity in clinical development and trial design. A special focus will be given to two areas. The first one is multiplicity in early phase trials and in particular combined decision criteria considering multiple endpoints and populations. Here examples and principles will be presented on how to address this topic in the context of GoNoGo decision making. This will also be contrasted against the classical way of dealing with this aspect in confirmatory settings.

The second area to be discussed is examples for designing Phase 3 programs with multiple pivotal trials, endpoints, doses and/or interim analyses. Here complications can arise for example, from pooling data across trials for testing secondary endpoints or from changing testing strategies after interim analyses.

Traditionally, the planning of clinical trials has been based on considerations of the power of a test of a given alternative hypothesis. Power as we understand it was based on ideas introduced by Neyman and Pearson in 1933. In 1939, Jeffreys pointed out that if the true value was unknown, so was the power and he suggested to understand the true power of a study the conditional power values should be averaged with respect to their prior probabilities or uncertainties, leading to an unconditional probability ot power. More recently O’Hagan and Stevens introduced the concept of assurance, again an unconditional probability.

In this talk I review recent ideas in the use of these unconditional probabilities in planning trials.

The small sample sizes in rare disease settings is a challenge for clinical trial design.

Bayesian adaptive trial methods are often the only rescue by allowing the cautious incorporation of auxiliary data, such as registry data,and other relevant information such as natural history to inform the trial itself. First a strategy with a one-arm trial augmented by the participants’ own natural history data will be envisaged. From such one-arm trial the predictive distribution of the future course of the disease in the absence of intervention will be derived. Patient response is then be defined by the degree to which post-intervention observations are unlikely with the predicted disease trajectory. Such one-arm trials offer obvious advantages in efficiency and ethical hazard but they cannot offer a protection against bias arising from the presence of “placebo effect,”. In the second part we’ll present the impact of various scenarios of placebo effects on one-arm responder studies, as well as two-arm versions that incorporate a small concurrent placebo group but still borrow strength from the natural history data and auxiliary data.

We will propose Bayesian changepoint models that specify a parametric functional form for the patient’s post-intervention trajectory, which in turn allow quantification of the treatment benefit in terms of the model parameters. Operating characteristics of the different scenarios will be presented.

It suggests that the two-arm responder and changepoint methods can offer protection against placebo effects, improving power while protecting the trial’s Type I error rate.

Patient-centric medicine is in full swing, yet there are no statistical methods that allow patients to make individualized treatment decisions based on all relevant efficacy and toxicity outcomes. There has been much recent work on the new statistical method of “generalized pairwise comparisons” (GPC) to allow formal decisions to be made on the totality of the available information in a rigorous way. With GPC, all efficacy, toxicity and quality of life data from patients enrolled in clinical trials comparing various interventions can be used to analyse any number of prioritized outcomes of any type (binary, continuous, time to event, etc.) The method compares all possible pairs of patients formed by taking one patient from the experimental group and one patient from the control group of a randomized trial. A measure of the overall treatment effect, called the “Net Treatment Benefit” (NTB), is the difference between the probability that a patient taken at random in the experimental group has a better outcome than a patient taken at random in the control group. The interpretation of the NTB is simple, and as such it may facilitate patient-centric treatment choices. Other measures of treatment benefit have been proposed, including the win ratio and the win odds. All these measures of treatment benefit can be expressed as functions of standard measures in the case of a single variable and under some distributional assumptions (e.g., normality for continuous outcomes or proportional hazards for times to event). Several applications will be used to illustrate the interest of GPC over a wide range of clinical situations.

In classical adaptive designs of RCTs with time-to-event (TTE) endpoints the conditional power (CP) is routinely used for making interim decisions such as sample size re-estimation, dose selection, population enrichment, etc. This CP is calculated under the assumption of constant and proportional hazard rates. Under non-proportionality and/or time dependent hazard, the CP can easily mis-inform decision makers and thus increase the chance of an incorrect interim decision. The use of Bayesian predictive power (PP) or the probability of success (PoS) have recently gained popularity due to the flexibility with respect to such assumptions. However, the complexity of calculations limits their use, especially when carrying out simulations to establish the operating characteristics of such adaptive designs. We discuss different ways and simplifications for the calculation of PP for TTE and discuss applications of PP in Bayesian as well as Hybrid adaptive designs. We will make a general comparison of different methods for PP calculations along with comparisons with the traditional CP under non-proportionality of hazards.