Contact ACOMED statistics (Tel.: 0341/3910195, E-mail:info@acomed-statistik), if you need support with the conception, planning and analysis of your study regarding Estimands.

The following links are helpful. In our opinion, the documents, especially the training materials, are also easy to understand for non-statisticians.

• ICH Guideline EP9 (Statistical principles for clinical trials, here: no Information on the topic of Estimands): link
• Addendum on estimates and sensitivity analysis: link
• ICH training material on the topic of Estimands: link

Why were Estimands introduced?

Clinical trials investigating the effectiveness of treatments assume an ideal situation: patients are exposed to the treatment or observation without any deviations occurring over time. However, such events ("intercurrent events") are frequently observed.

• insufficient therapy adherence,
• an interruption of observation,
• a discontinuation of treatment due to a success
• a discontinuation or interruption of treatment due to lack of effect, side effects or organizational reasons,
• the interim treatment with another, known effective drug (rescue medication),
• The onset of another, more serious illness,
• Death of the patient,
• uvam

Integrating these events into statistical analysis is complex. A key starting point is the intended-to-treat principle: all randomized patients are included in the analysis. Often, this principle is somewhat relaxed: all randomized patients who receive at least one treatment are included. Previous strategies, with the aforementioned deviations, range from ignoring or excluding the affected patients to conducting numerous sensitivity analyses under various assumptions. This leads to statistical analyses that fail to address the actual objective of the study and can result in erroneous conclusions. The situation is also challenging from the perspective of regulatory authorities, who must evaluate new solutions proposed by sponsors for each study. Therefore, a structured, systematic approach was necessary.

The introduction of Estimands now provides such a structured framework for the process.

Intercurrent events

ToDo: Apprildung



Estimands

These questions are addressed in the Estimand concept (ICH E9(R1) Addendum, final version of January 30, 2020). This addendum provides a structured framework for linking study objectives (Purpose of the estimate: What is to be estimated?) with a suitable study design, study implementation, and instruments for analysis (Estimation method: How should the estimate be made?) before.

An estimand is understood to be The precise description of the treatment effect in relation to the clinical question, which is defined by the clinical question. It summarizes, at the level of specific populations, how the treatment outcome for the same patients appears under different, comparable treatment conditions. In our view, this definition is not yet helpful when it comes to the concrete implementation of study planning. It becomes clearer when one considers (1) the attributes that contribute to the establishment of an estimand, and (2) the strategies that serve to address the aforementioned deviations.

The protocol must specify a primary estimand and, if applicable, further secondary estimands. The following must be defined when designing estimands:

• Attribute (Population, Treatment/treatment methods, endpoint, intercurrent event, outcome at population level)
• Strategies to describe the clinical question in light of events
• Dealing with missing values.

Estimands Attribute

• population
• Treatment/Treatment methods
• Endpoint variable
  This could – with the same goal – be, for example, a continuously scaled variable (change vs. baseline) for one estimand, and a binary scaled variable (response y/n) for another estimand.
• Specifications such as events that affect the measurement or its interpretation (intercurrent events) are handled.
  For example, the occurrence of emergency medication could be counted as "non-response" regardless of the subsequent course.
• Summary of results at the population level
  e.g. baseline-adjusted mean difference of change vs. baseline for one estimand, baseline-adjusted response difference for another estimand

Estimands: missing valuesel

The ICH E9 (R1) Addendum calls for a more precise definition of what constitutes "missing values." Missing values—even independently of the estimands concept—represent a limitation of the data that must be addressed in the protocol, defined in detail in SAP, and considered in the statistical analysis. Missing values do not only occur monotonically after a specific time point before the end of the study until the end of the study, but also sporadically (non-monotonic), for example, if a participant is unable to attend a scheduled visit or if individual data points are missing.

The handling of missing values should be specified during study planning depending on (1) the type of absence (MAR - missing at random: The absence can be attributed to known variables and thus described - the values can be appropriately replaced (su); MNAR - missing not at random: The absence cannot be attributed to known variables), (2) the pattern of missing data (monotonic, non-monotonic) and, if possible, (3) the reason for the missing values.

Dealing with missing values practically means replacing missing values (ImputationIn this approach, multiple imputation is usually chosen, meaning that missing values are repeatedly replaced from a suitably selected random sample, and the study is analyzed for each replacement. Finally, the individual results are averaged.

The random sample can be constructed according to the following approaches, for example:

Reference-based imputation: The substitution is performed under the MAR assumption, i.e., data from similar patients (same age, gender, baseline values) are imputed from the control group.

• jump-to-reference (J2R): This assumption implies that the results for patients from the treatment group are obtained after the study dropout from the sample of subjects from the control group.
• Copy reference (CR): This assumption implies that the results after the study dropout gradually converge to those of the subjects in the control group.
• copy increment from reference (CIR): This assumption implies that the results after study dropout show a temporal pattern similar to that of the control group, but that the effect of the treatment up to the time of study dropout lasts until the end of the study (e.g., as one might expect from a treatment that has a lasting effect after study dropout).




worst-case analysis: Replace missing values with the worst values (in Outcome Response)

ToDo. Image

Strategies for dealing with intercurrent events

The ICH E9(R1) Addendum presents strategies for describing the clinical question of interest in relation to intercurrent events (the names of the strategies are for convenience only, their use is not mandatory):

Treatment policy
The occurrence of the intercurrent event is irrelevant for the definition of the treatment effect of interest; that is, the value of the target variable is used regardless of whether the intercurrent event occurred or not. The intercurrent event is considered part of the treatments being compared, which corresponds to the intention-to-treat (ITT) principle defined in ICH E9.
[Prerequisite for using this strategy: "all" data after the occurrence of the intercurrent event must be available. In general, this strategy cannot be applied to terminal intercurrent events (e.g., death) because no values for the target variable are available after the intercurrent event.] (The estimand according to treatment policy is often requested by regulatory authorities in the context of approvals.)

Hypothetical:
One imagines that the intercurrent event would not occur. Many possible hypothetical scenarios are conceivable, but not all are relevant for regulatory decision-making. For example, it may be important to investigate the effect of a treatment under conditions different from those of the feasible study (e.g., when no emergency medication is available).
Data collected after an intercurrent event are considered not meaningful and are therefore not used for analysis. Instead, values are estimated or imputed based on pre-specified assumptions; for example, values from "similar" subjects in the same treatment group can be used.

Composite:
The occurrence of the intercurrent event is added as a component of the variable or integrated into the endpoint. This is particularly relevant when the event itself is the most meaningful outcome that can be observed, e.g., death, emergency medication, discontinuation of treatment due to lack of efficacy, or adverse events. The strategies need not be limited to dichotomous endpoints; that is, if outcomes are measured on a scale or score, subjects experiencing the intercurrent event could receive a poor score. The strategies are especially useful for managing terminal events (such as death). They can also be very useful for secondary estimands. The intercurrent event marks the end of data collection.
[e.g., the intercurrent event "emergency medication" is integrated into the endpoint: a "responder" is defined as a subject with a positive treatment effect at time Z and no intake of emergency medication up to that time. After taking the emergency medication, the patient is counted as a "non-responder", regardless of the treatment effect].

While on treatment (frequently used for safety analyses):
The strategies can influence the definition of the endpoint, for example, by restricting the observation period of interest to the time between baseline and the occurrence of the intercurrent event. The terminology used in the strategy depends on the intercurrent event of interest: e.g., "while on treatment," "while alive," etc. Only data collected before the occurrence of the intercurrent event (e.g., death) are of interest [Note: frequent measurement time points are required]. Particular caution is advised if the occurrence of the intercurrent event differs between the treatments being compared, as it is difficult to derive a reliable estimate for drawing conclusions when follow-up times vary between groups.

Principal stratum:
The population of interest can be limited to the layer/stratum of subjects in whom the intercurrent event did not occur (or did occur). For example, subjects in whom no adverse events occurred that would have led to discontinuation of treatment could be considered the population of interest [Note: it is difficult to estimate how many subjects this will apply to during the planning phase]. The clinical question of interest relates only to the treatment effect within this layer/stratum.
[It is generally not possible to directly identify these subjects prior to the study, as the occurrence of the intercurrent event cannot be perfectly predicted.]

The strategies presented can be used individually or in combination to address a variety of different intercurrent events. The choice of strategies should be the subject of a multidisciplinary discussion, particularly among sponsors, clinicians, statisticians, and regulatory authorities. The preferred strategy for managing each intercurrent event should be clearly defined in the study protocol, along with the rationale for that choice.

Example 1r title

For studies with the indication of pain, but also beyond, the following article (open access) is worth reading:

Cai, X., Gewandter, JS, He, H., Turk, DC, Dworkin, RH, & McDermott, MP (2020). Estimands and missing data in clinical trials of chronic pain treatments: advances in design and analysis. Pain, 161(10), 2308-2320. https://doi.org/10.1097/j.pain.0000000000001937: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7508757/pdf/nihms-1596883.pdf

Example 2l

New text