Firstly, we want to consider the patients – who are the target population of interest and what do their treatment journeys look like? In designing clinical trials, we ideally want the scientific questions that we are asking to be directly relevant to real patient experiences. Figure 3 shows a schematic of four potential patient journeys in an oncology setting. After starting study treatment, patients generally receive treatment until disease progression. However, they may discontinue study treatment earlier, e.g. due to adverse events, and they may start a new anticancer therapy afterwards. The example below illustrates how considerations of various patient journeys impact the precise definition of the scientific question of interest and help to identify intercurrent events.
To illustrate the use of the estimand framework in creating a detailed clinical trial objective we shall begin with a naïve objective:
“What is the effect of treatment X on patient’s quality of life?”
It is important to clearly define a specific patient outcome of interest, i.e., to define the “variable of interest” [
]. Simply referring to a “PRO score” is not specific. In our naïve example, we have selected pain severity as the PRO variable of interest.
In addition, one should consider a relevant timeframe to evaluate outcomes when defining the PRO variable of interest. In this simple example, we have suggested a fixed timepoint (6 months post randomisation) at which to evaluate the variable.
Defining the parts of the estimand relating to the population of interest [
] and treatment comparisons [
] are presented next and thus to build this estimand:
The ICH E9 terminology uses the phrase “a summary measure” [
]. It is important to note that in this context the summary measure refers to the statistical summary approach, rather than just the measure or PRO instrument used. The summary measure should be defined; this could be the difference in the mean score between the two groups, or the proportion of patients in each group who have had meaningful improvement in symptoms. In this example, we suggest comparing the difference in mean scores between treatment groups.
Now we turn to intercurrent events, which are at the heart of the estimand framework and involve intricacies that can lead us to define different estimands. For the purposes of this illustrative example, we will consider two potential intercurrent events of interest commonly observed in oncology trials which are also important for PRO endpoints: discontinuation of treatment and patient deaths.
Intercurrent event #1: treatment discontinuation
Different questions arise when considering treatment discontinuation prior to the planned assessment timeframe (e.g. 6 months post-randomization) that correspond to different potential strategies to handle this intercurrent event:
-
Are we interested in the effect of treatment at 6 months post-randomization regardless of whether a patient is still on study treatment? If so, this would be illustrative of using a treatment policy strategy to handle the intercurrent event of treatment discontinuation. Note that this would require collection of the EORTC QLQ-C30 pain scores up to the 6-month period, even after treatment discontinuation (and disease progression and potential subsequent therapy) so that the value of the EORTC QLQ-C30 pain score at 6 months can be directly used in the analysis.
-
Is it relevant to assess the effect only while receiving study treatment? If so, then this is illustrative of using a while on treatment strategy for this intercurrent event. Here, the EORTC QLQ-C30 pain score at 6 months is included in the analysis only if the patient is still on treatment at 6 months. Otherwise, the pain score assessed at the time of treatment discontinuation is used. Therefore, data collection after treatment discontinuation would not be required.
-
Are we interested in the effect in the hypothetical scenario if all patients had received study medication for 6 months? If “yes”, then the approach to handling this would be considered the hypothetical strategy. This can be appropriate strategy if the intercurrent event is not expected to occur in the target population in general. However, this strategy may not be a realistic assumption to be made with regards to PRO data in general.
Intercurrent event #2: death
-
Handling death events with a treatment policy strategy implies that the endpoint is considered regardless of deaths, which is simply impossible and as such would never be a plausible strategy for handling this type of intercurrent event for PRO data.
-
Using a while on treatment strategy to handle deaths as an intercurrent event implies that we are interested in the effect on PROs until death. It is necessary to decide how to use these data, such as using the last observation prior to death if patients die prior to the time of interest for the endpoint (e.g., 6 months).
-
A hypothetical strategy for handling an intercurrent event of death is when we envisage a hypothetical scenario where death would not have occurred and may be considered a reasonable approach for PRO endpoints if the number of patient deaths is expected to be low and that the patients deaths are not expected to be related to treatment (in the timescale of interest for the assessment of PRO endpoint). In this case we can use some assumptions to impute a reasonable value for the patient for the purpose of analysis.
For just these two intercurrent events, strategies for handling intercurrent events could lead to at least five different potentially quite plausible estimands (as treatment policy for death is excluded as implausible). Furthermore, it is conceivable to consider other appropriate strategies for handling these intercurrent events. Both a composite endpoint or a principal stratum approach could also be applicable for a PRO concept of interest [6, 7]. In fact, death is frequently considered as an event in some time-to-deterioration of quality of life analysis implying a composite strategy [8].
In this example we are considering pain, and, therefore, it is also important to consider the impact of pain medication use (if permissible per clinical trial protocol) as a potential intercurrent event. Taking pain medication would not typically be considered as an intercurrent event for most primary efficacy endpoints of an oncology trial as it’s not expected to impact tumor progression. Concomitant pain medication use becomes an important intercurrent event when considering the endpoint of pain as it may impact on the measurement of pain. We would need to define whether we are interested in the effect of our treatment on pain only when no pain medication is taken or regardless of whether patients take pain medication. Current standard practice typically leads to handling pain medications use with a treatment-policy strategy. In this case, we don’t measure just the effect of our treatment on pain, but we collect data that include the effects of treatment and pain medications (if used). As a result, the use of pain medications becomes part of our definition of the “treatment” attribute of the estimand.
Similarly, if patients discontinue treatment and start new anticancer therapy prior to the planned assessment timeline at month 6, this could also be included in the treatment effect planned to be measured. In these cases, it is not enough to define a simple “treatment X” as part of the estimand, rather there is the need to be more specific and define “treatment X followed by any potential subsequent antineoplastic therapy and/or concomitant pain medication (as needed)”.
To return to the example estimand, let us add in potential subsequent treatments and any concomitant treatments to the treatment attribute [
], and then we can clarify our intercurrent event approaches [
]; let us assume a treatment policy approach to handle treatment discontinuation and while on treatment to handle death were used Therefore, our estimand becomes:
It is important to ensure that the final clinical trial objective reflects required specificity and details that explain whether it is intended for the PRO objective to demonstrate statistical superiority or non-inferiority of treatment X compared with treatment Y [9]. Additionally, if there is published or well-accepted minimal threshold for clinically meaningful between-group difference, it should also be stated (e.g. a between-group difference in mean EORTC-QLQ C30 pain sore > 7 points) [10].