Lead Investigator: Tobias Polak, Erasmus Medical Center
Title of Proposal Research: Tositumomab: statistical modeling of expanded access programs and conventional trials
Vivli Data Request: 6477
Funding Source: Health~Holland Funding from Dutch Government. Grant number: EMCLSH20012
Potential Conflicts of Interest: None
Summary of the Proposed Research:
Patients suffering from seriously debilitating or life-threatening conditions who are not eligible for further treatments or any clinical trials, may resort to ‘expanded access’: pre-approval access to investigational treatments. Expanded access (EA), also known as early access, pre-approval access or compassionate use (Kimberly et al., 2017), is the formal regulation adopted by the Food and Drug Administration (FDA) in 1987 (Young et al., 1988), propelled by the HIV/AIDS crisis. In the United States (US) the FDA regulates this process of formalized non-clinical trial access whereas in the European Union (EU) the responsibility lies with individual member states (Darrow et al., 2015). The numbers of requests for EA are growing (Jarow et al., 2017) and state and federal legislation, such as Right-to-Try laws in the US, stress the need and interest of patients in having earlier access to medicines that are still under clinical investigation.
EA data can be seen as a type of real-world data (RWD). RWD are ‘information on health care that is derived from multiple sources outside typical clinical research settings’ (Sherman et al., 2016). Recent publications and regulatory frameworks have boosted the promise of RWD (Jarow et al., 2017; Stower, 2019). It can come in many forms and shapes, such as electronic health records, social media or claims databases.
EA programs also form a source of RWD. Historically though, EA programs were only deemed fit for treatment and not for research. Although the primary purpose of EA is treatment, scholars have argued that there is a moral obligation to collect outcome data in all cases where patients are treated with investigational medicines (Bunnik, Aarts and van de Vathorst, 2018; Fountzilas, Said and Tsimberidou, 2018; Chapman et al., 2019). The debate on combining data collection and EA has substantially increased in recent years (S. Usdin, no date; Sutter, no date), with FDA-officials confirming beginning 2018 their willingness to review data from EA programs to support drug applications (Chapman et al., 2019). Considering the increasing interest in both expanded access and real-world data, the question arises whether alternative ways of access to novel treatments can provide clinical information and impact regulatory decision making. (Polak, van Rosmalen and Uyl – de Groot, 2020) have shown that there is an increasing trend in the use of RWD from EA by regulators and industry.
There has been a growing interest to include RWD in the medical domain, both from a regulatory and reimbursement perspective. Although randomized controlled trials (RCT) remain the gold standard for evidence generation, results are often obtained within highly controlled and regulated settings. This might limit the generalizability of findings. Another limitation of RCTs is that the sample sizes are often relatively small, especially for rare diseases and in disease areas where patients are reluctant to be randomized. The need for higher external validity of RCT results has led to an increased call for the use of RWD (FDA, 2018). Often, RWD are analyzed separately from the data derived in conventional clinical trials. Analysis of (single arm) RWD – and all other data sources where control groups are absent – has to be done with care. The absence of randomization introduces inherent confounding. Confounding can be partly overcome by statistical techniques, such as propensity score matching (Garrido et al., 2014), instrumental variables or adding the suspected confounding variables in the analysis.
Our research will help clarify the usability of these data. Awareness of the potential value of RWD from EA should facilitate that these data are incorporated in decision making whenever this is feasible and appropriate, and this may impact traditional clinical development. For patients, this better use of available data can result in speedier access to more diverse treatments. This project will yield direct insights, via scientific publications and statistical methodology, into how to best combine data from expanded access programs with data of (randomized) controlled trials. This is directly beneficial to regulators, drug developers and (bio)statisticians.
To illustrate our research, we will apply our methodology on the compound ‘tositumomab’, a monoclonal CD20-antibody that is available as treatment for non-Hodgkin lymphoma’s. We will be jointly working on 5 studies on tositumomab, all in slightly different indications, and the expanded access program that encompasses indications that match the previous 5 studies. These data sets are made available for research via the Vivli Center for Global Clinical Research data (https://vivli.org) and have been made available by the marketing authorization holder GlaxoSmithKline.
Statistical Analysis Plan:
We aim to develop and evaluate methods for (i) dynamic borrowing of other data scores and (ii) methods for attenuating bias by confounding. We here provide further background on these methods and our plans.
In the analysis of clinical trials, there is often data from relevant previous studies available. A common example is the availability of data from a previous RCT in which the control arm patients received the same treatment as in the new (current) study. Provided that the studies are sufficiently similar in terms of setting and outcomes, the data of the historical controls can be incorporated into the analysis of the current trial. Due to the possibility of between-study heterogeneity (e.g. differences in study populations between trials), a naïve pooling of historical controls and the current controls is rarely acceptable. Statistical methods have been devised to appropriately downweight the historical data, based on their characteristics and the observed differences with the current controls. A well-known technique for downweighting the historical controls is the power prior 13–15.
We are proposing to extend the ideas of the power prior and other methods for borrowing historical data to RWD data from EA programs. In this case, we downweight the RWD instead of the historical controls. The downweighting serves to account for a) the lower grade of evidence of EA data (compared to RCT data) due to the lack of randomization and the less controlled setting and b) center-specific and study-specific effects that may differ between EA data and RCT data. In the power prior method, this is done by raising the likelihood for the RWD, Λ(θ│Y_RWD ), to a power of α, with α between 0 and 1, whereas the likelihood of the RCT, Λ(θ│Y_RCT ), is not discounted . This leads to
p(θ│α,Y_RCT,Y_RWD )∝Λ(θ│Y_RCT )×Λ(θ│Y_RWD )^α×p(θ)
The power of α can be interpreted as a discounting factor between the data from randomized controlled trials and the data from the real-world. With α = 0, the real-world data is completely ignored and with α=1, the datasets are simply pooled. However, in the above specification, the power α has to be set beforehand. One may resort to Bayesian techniques to estimate α:
p(θ│α,Y_RCT,Y_RWD )∝1/C(α) Λ(θ│Y_RCT )×Λ(θ│Y_RWD )^α×p(θ)×p(α),
where C(α) is a normalizing constant. Estimation of α using the available data leads to adaptive borrowing of information from the RWD: when the RCT data and the RWD are sufficiently similar, α will be estimated as high, and the RWD will be mostly included in the analysis; however when RWD are in conflict with the RCT data, α will be estimated to have a low value, and the RWD will be effectively discarded. This approach is based on the principle that the RCT data represent the highest level of evidence, whereas the RWD have an observational study design with a lower level of evidence.
It will also be necessary to account for differences in disease severity and other patient characteristics between EA data and RCT data. This can be done by combining the power prior approach either with covariate adjustment or with propensity score methods. This has been recently published by 16–18. and an RPackage ‘psrwe’ accompanies their findings. They have not, however, evaluated their methods with extensive simulation. We plan to do this and subsequently apply our findings to a non-simulated dataset.
There are also other methods for discounting historical data, such as methods based on a meta-analysis of all available studies, i.e. the meta-analytic-predictive prior 19,20. Our approach will be compared with these existing alternatives in terms of frequentist characteristics (type I error rate, mean squared error, and statistical power) as well as in terms of performance in real data sets.
Requested Studies:
Expanded Access Study of Iodine I 131 Tositumomab for Relapsed/Refractory Low-Grade and Transformed Low-Grade Non-Hodgkin’s Lymphoma
Data Contributor: GlaxoSmithKline
Study ID: NCT00268203
Sponsor ID: BEX104545
Phase II Study of Iodine-131 Anti-B1 Antibody for Non Hodgkin’s Lymphoma Patients Who Have Previously Received Rituximab
Data Contributor: GlaxoSmithKline
Study ID: NCT00996593
Sponsor ID: 104507
Multicenter, Pivotal Phase III Study of Iodine-131 Anti-B1 Antibody (Murine) Radioimmunotherapy for Chemotherapy Refractory Low Grade B Cell Lymphomas and Low Grade Lymphomas That Have Transformed to Higher Grade Histologies
Data Contributor: GlaxoSmithKline
Study ID: NCT00989664
Sponsor ID: 104504
Phase II a Randomized Study of Iodine-131 Anti-b1 Antibody Versus Anti-b1 Antibody in Chemotherapy-relapsed/Refractory Low-grade or Transformed Low-grade Non-Hodgkin’s Lymphoma (NHL)
Data Contributor: GlaxoSmithKline
Study ID: NCT01573000
Sponsor ID: 104515
A Multi-Center, Randomized, Phase 3 Study of Rituximab Versus Iodine I 131 Tositumomab Therapeutic Regimen For Patients With Relapsed Follicular Non-Hodgkins Lymphoma
Data Contributor: GlaxoSmithKline
Study ID: NCT00268983
Sponsor ID: 393229/028
Summary of Results:
We were unable to bring our analysis to completion due to the following reasons:
- the data were too heterogeneous
- the sample sizes were too different
- we were not able to follow-up on all the answers in time
Although we were unable to complete our statistical analysis plan, we did gain some key learnings from this proposal. We have found that in future studies, it may be beneficial to carefully consider the heterogeneity of the data and sample sizes to ensure a better outcome.