Estimand framework in Clinical Trial Biostatistics. Edition 1 - Orientation and intercurrent event strategies

Estimands are extremely important in today's Clinical research Biostatistics and Clinical research overall, especially in areas of Clinical trials and Health technology assessment. Further estimand framework is a part of the regulatory frameworks, which makes them even more important.

In this article i will discuss where you may find relevant information from regulatory agencies and also make an orientation about estimands and the intercurrent event strategies which are vital when designing estimands and studies.

Further in this article, i will explain why are estimands very important as systematic descriptions of treatment effect quantities, from biological, statistical and clinical trial perspectives.

The most relevant information may be found in FDA and EMA E9(R1) addendum specifications. Here are some relevant links:

You may find the relevant information in the FDA's E9(R1) addendum: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/e9r1-statistical-principles-clinical-trials-addendum-estimands-and-sensitivity-analysis-clinical (reference 1)

You may also read the relevant information here at the EMA's website E9(R1) addendum : https://www.ema.europa.eu/en/documents/scientific-guideline/ich-e9-r1-addendum-estimands-and-sensitivity-analysis-clinical-trials-guideline-statistical-principles-clinical-trials-step-5_en.pdf (reference 2)

This documents define estimands and the sensitivity analysis in clinical trials. In this article, i will focus on estimands and their systematic description in relation to Biostatistics and Clinical trials, and in some of the next editions on the sensitivity analyses too.

The first focus when discussing estimands is on the word 'systematic'. While researchers frequently use specific definitions for their research, such as defining treatments, population, variables and intercurrent events, systematically relating these, systematically including them in the study design, protocol development and integrating them into the statistical analysis is another level of systematic approach.

An example i like to use before starting explaining estimands is that any result you see in clinical trials, expressed Hazard ratio, Mean difference, confidence intervals, errors etc. is based not just on the numbers as estimates, but also the bound to the overall construct of the study. Estimands are a more systematic way of defining what is actually being estimated, but specifically for that specific study. Behind every statistical or clinical result there is a study construct which defines what the metric and the data reflect and estimands are the main systematic way to define this.

When you compare the results of studies, you are actually comparing their constructs too, which need to be systematically addressed to be accurate in both implementation of the study and interpretation of the results. Estimands are also a great way to make interpretations and comparisons between studies more comparable. For this reason they have a special focus in the regulatory frameworks too.

Its very important that researchers systematically design what is it specifically that the study results will tell.

Estimand framework based on the FDA's and EMA E9(R1) documents (links provided above) is meant to provide a systematic definition of :

1.Intercurrent events (IE) definition (very important).

2. Treatment

3. Population

4. Variable/s of interest

5.Population summary

Having said that, i would like to focus on integration of these systematically, accurately and completely. This is what differentiates the estimand framework from just describing different similar aspects of a study. In an estimand, these segments are intentionally and specifically defined and related to each other

Estimand framework is a big step forward for researchers and statisticians using them, not only for the overall scientific quality and regulatory compliance being added into the research, but also because estimand framework is one of the best ways to connect the biological and statistical segments of research.

Connecting Biological/Clinical and Statistical segments of the research trough defining estimands is probably one of their most important features. In fact, defining estimands in clinical research typically requires having specific communication between clinicians and biostatisticians or sometimes professionals understanding both of these domains.

Now i will discuss one of the most important features to understand when starting to define the estimands, the intercurrent event strategies.

IE strategies are some of the most important segments of an estimand. Estimands segments are interconnected and the intercurrent event strategy is one of the main aspects of integrating clinical and statistical aspects of the framework and the study design overall.

Its important to try to anticipate (based on clinical and biological knowledge and experience form previous studies) and define potential intercurrent event, and how to deal with them. I would like to emphasize the importance of biological/clinical knowledge for Biostatisticians in this segment.

Here are some examples which we will be discussion in the next editions:

• Treatment policy IE strategy. This strategy is very important and is frequently considered for high grade clinical studies. The main aspect of this strategy is reducing the bias that could and frequently in relation to ITT (intention to treat principles). Frequently deployed in RCTs or other randomized trials due to the fact that its important not to break the randomization principles, while making sure we understand and define the intercurrent events as a part of the treatment strategy[1]. A good example of event which would not affect the follow up is the rescue medication as an intercurrent event or potentially a discontinuation. It may be defined as a part of the treatment and ignored as IE that would disable the follow up for such subjects [1].Treatment policy strategies may differ, this is important to note, but in case of the ITT principle being implemented, in my opinion it has the best construct in terms of bias reducing power. Of course, the missing data may be there due to not all participants completing the trials for various reasons, one intercurrent event being mortality which can not be resolved like others (may cause missing data), there is non-adherence to therapy, but still inclusion in the data, all these are something that can be resolved. For phase III trials, this construct would be my favorite.

• Composite strategy As in the Treatment policy the rescue/medication can be defined as the part of the composite variables, but the focus is not on the treatment policy but at the composite treatment itself. In a composite comparison, a typical example would be a composite treatment (main treatment plus or minus another type of medication composite) vs control. A typical composite is treatment effect expressed as categorical variable defined by change from the baseline, based on specific threshold , plus categorical grouping based on the intercurrent event also as a categorical variable [1]. Another composite could be treatment effect change from the baseline expressed as a categorical variable plus rescue medication as a second composite variable. Keep in mind that IE strategy in this case heavily affects the population and population summary segments of the estimands, so the population is defined based on the composite variable strategy itself and the population summary estimate such as rate is also bound in this case to the composite[1]. There are many other ways of making the composite and i will be discussing them in the next editions. While composite strategy is great in terms of the RWE and different combinations used in reality, the preservation of randomization is not as good as in the Treatment policy, ITT based estimands. Also its difficult to differentiate between segments of the composites treatments if the treatments are averaged as average treatment effects as its typically done in the clinical research.
• While one treatment IE strategy. This one has some similarities to composite strategy but also some differences. In this case population summaries such as mean differences, median differences, risk ratios, hazard ratios or other can be expressed as while on treatment, but also until the change introduced but the intercurrent events [3]. Such an approach would tend to differentiate the actual treatment effects from the intercurrent evetns, but its effect on the randomization tends to be present and could result in covariate bias if not adjusted properly. Basically there is a time threshold that can be define in this case.
• Trial product estimand strategyTrial product strategy is frequently a good choice when the confounders have been reduced prior to the study allocation. But always a caution with this principle as we don't always know all the potential confounders. Similar to the while on treatment in terms of trying to direct the main average treatment effects or other treatment effect from the effect of intercurrent events but this time by taking into account only the data which is defined as treatment specifically or untill the intercurrent event and trying to estimate when the data would hypothetically be if the intercurrent event did not occur [3] . For this reason its commonly called as Hypothetical estimand strategy.

There are other estimand strategies of course and i will be discussing some of them in the next editions. In the next 5 editions, i will focus on actual examples of estimands, starting with Treatment policy based estimands examples.

References :

1.https://www.fda.gov/regulatory-information/search-fda-guidance-documents/e9r1-statistical-principles-clinical-trials-addendum-estimands-and-sensitivity-analysis-clinical

2. https://www.ema.europa.eu/en/documents/scientific-guideline/ich-e9-r1-addendum-estimands-and-sensitivity-analysis-clinical-trials-guideline-statistical-principles-clinical-trials-step-5_en.pdf

3. Gogtay NJ, Ranganathan P, Aggarwal R. Understanding estimands. Perspect Clin Res. 2021 Apr-Jun;12(2):106-112. doi: 10.4103/picr.picr_384_20. Epub 2021 Mar 12. PMID: 34012908; PMCID: PMC8112325.

4. Pohl M, Baumann L, Behnisch R, Kirchner M, Krisam J, Sander A. Estimands-A Basic Element for Clinical Trials. Dtsch Arztebl Int. 2021 Dec 27;118(51-52):883-888. doi: 10.3238/arztebl.m2021.0373. PMID: 34857075; PMCID: PMC8962508.