The canonical example is the estimation of progression in oncology trials. This is traditionally handled by defining progression free survival as progression, using the composite strategy for the intercurrent event of death, and 'making the IE part of the endpoint'
You can also, should you be so inclined, take a very different philosophical view from what is written in E9 (r1), and to some extent above, and say:
"I am interested in the treatment policy estimand. Some people have died before progression and therefore cannot experience the event. I still want to use the information from them in the estimation of the treatment policy estimand. As I have time to event data, I can assume they have had the event or they did not, and I can pick some "time from randomisation until censoring or event" to include in the analysis. I will consider that they did have the event (as both progression and death are bad). It makes no sense to impute a time greater than that observed in the trial, and I also don't really believe they progressed before they died. Therefore I will choose the time of death as the 'time' I include in my model. I can also impute different times for these patients, to see the effect on my estimation."
I struggle to see how such a statement of what I am doing isn't estimating, in a reasonable way, the treatment policy estimand. And of course, the two analyses are identical.
Question: I'm curious about 'take a different philosophical view from what is written in E9 (R1)'. What do you regard as the addendum's philosophical view? As far as I can tell it is 'define your estimand'.
Response: My post says one cannot conceive of what a treatment policy estimand is in the presence of truncating events. You said suppose you are interested in a treatment policy estimand and then outlined an imputation strategy. But we cannot get to the imputation strategy because my point still stands: we can't conceive of what a treatment policy estimand is, either mathematically or in English.
Hi Tim, I've been a bit busy but wanted to come back to this. The problem is to some extent, that in my toy philosophical example, I want to apply the "strategy" to the concept of the "estimand", whereas what is being argued is that different classes of intercurrent event require a different strategy. Furthermore, it is not possible to apply a particular strategy to a particular class. I think the logical next step is that the strategy is an inherent property not of the estimand in and of itself, but of the intercurrent event class. When viewed through this lens, the concept of the treatment policy estimand in and of itself makes no sense from a philosophical point of view, because we are trying to ascribe a single class attribute to something made up of multiple components that might belong to different classes.
This actually helps! We can then more properly define the treatment policy estimand as "the estimand that treats all intercurrent events, regardless of what they are, under the assumption of treatment policy" - our phrase "treatment policy estimand" is just a shorthand, and a somewhat lazy one at that, for the actual thing we are interested in. We can now also point out the issues with terminal events - they form a class of intercurrent event for which treatment policy strategy is not possible, and thus our shorthand "treatment policy estimand" makes no sense.
You can also take this further , and explicitly handle different intercurrent events in different ways outside of terminal events. The addendum touches on this in A.3.4. This has actually been proposed in e.g. the EU Diabetes guideline, which mentions treating some intercurrent events using a hypothetical strategy, and some using a treatment policy. What is the "name" of this estimand? Hypolicy? Polithetical? And why should being able to give it a name matter? As long as you can describe what you are doing, the assumptions you make, and can estimate it in a reasonable fashion, this should be enough.
I actually think the addendum is clear on this, specifically in Section A.3.2. It also seems to go out of its way to use the phrase "estimand based on treatment policy" as opposed to "treatment policy estimand" which again is a better shorthand for "estimand based on treatment policy for all intercurrent events". We have already got lazy as a community with our shorthand descriptions and it makes things more complicated and not less in some situations. Nevertheless, we need simple and understandable concepts so we can communicate across disciplines and if MOST of the time, this laziness actually helps, maybe we should just accept it, and the price we pay is the difficult edge cases such as the ones you highlight where we have to row back and be precise about what we really mean.
Thank you for taking the time to clarifying this! I now understand what you meant – apologies for the initial misunderstanding.
I really like the addendum's phrasing of 'based on a treatment policy', and suspect this was so that the ICE strategy doesn't dominate the other attributes too much.
I also like its approach of taking each ICE separately, and think we can define the strategy for a given ICE when we're not feeling lazy… but your mashup terms above have opened a whole world for me. Imagine the fun when we have four different types of intercurrent events!
Enjoying your posts here! Re: direct effect I think the main distinction between a mediator event vs truncation event is that the mediator doesn't preclude the outcome from occurring...so direct effect is just portion of total effect that doesn't go through mediator (yes cross world counterfactuals are involved but let's not go there)..it's definitely hurting my head to think about competing/truncating events as akin to mediation where outcome isn't observed
Thanks! Yes, absolutely agree about that distinction. Cross-world counterfactuals always hurt my head. I'm convinced when other people say how useful they can be but I'm not smart enough.
The canonical example is the estimation of progression in oncology trials. This is traditionally handled by defining progression free survival as progression, using the composite strategy for the intercurrent event of death, and 'making the IE part of the endpoint'
You can also, should you be so inclined, take a very different philosophical view from what is written in E9 (r1), and to some extent above, and say:
"I am interested in the treatment policy estimand. Some people have died before progression and therefore cannot experience the event. I still want to use the information from them in the estimation of the treatment policy estimand. As I have time to event data, I can assume they have had the event or they did not, and I can pick some "time from randomisation until censoring or event" to include in the analysis. I will consider that they did have the event (as both progression and death are bad). It makes no sense to impute a time greater than that observed in the trial, and I also don't really believe they progressed before they died. Therefore I will choose the time of death as the 'time' I include in my model. I can also impute different times for these patients, to see the effect on my estimation."
I struggle to see how such a statement of what I am doing isn't estimating, in a reasonable way, the treatment policy estimand. And of course, the two analyses are identical.
Thanks Andrew! A question and a response:
Question: I'm curious about 'take a different philosophical view from what is written in E9 (R1)'. What do you regard as the addendum's philosophical view? As far as I can tell it is 'define your estimand'.
Response: My post says one cannot conceive of what a treatment policy estimand is in the presence of truncating events. You said suppose you are interested in a treatment policy estimand and then outlined an imputation strategy. But we cannot get to the imputation strategy because my point still stands: we can't conceive of what a treatment policy estimand is, either mathematically or in English.
Hi Tim, I've been a bit busy but wanted to come back to this. The problem is to some extent, that in my toy philosophical example, I want to apply the "strategy" to the concept of the "estimand", whereas what is being argued is that different classes of intercurrent event require a different strategy. Furthermore, it is not possible to apply a particular strategy to a particular class. I think the logical next step is that the strategy is an inherent property not of the estimand in and of itself, but of the intercurrent event class. When viewed through this lens, the concept of the treatment policy estimand in and of itself makes no sense from a philosophical point of view, because we are trying to ascribe a single class attribute to something made up of multiple components that might belong to different classes.
This actually helps! We can then more properly define the treatment policy estimand as "the estimand that treats all intercurrent events, regardless of what they are, under the assumption of treatment policy" - our phrase "treatment policy estimand" is just a shorthand, and a somewhat lazy one at that, for the actual thing we are interested in. We can now also point out the issues with terminal events - they form a class of intercurrent event for which treatment policy strategy is not possible, and thus our shorthand "treatment policy estimand" makes no sense.
You can also take this further , and explicitly handle different intercurrent events in different ways outside of terminal events. The addendum touches on this in A.3.4. This has actually been proposed in e.g. the EU Diabetes guideline, which mentions treating some intercurrent events using a hypothetical strategy, and some using a treatment policy. What is the "name" of this estimand? Hypolicy? Polithetical? And why should being able to give it a name matter? As long as you can describe what you are doing, the assumptions you make, and can estimate it in a reasonable fashion, this should be enough.
I actually think the addendum is clear on this, specifically in Section A.3.2. It also seems to go out of its way to use the phrase "estimand based on treatment policy" as opposed to "treatment policy estimand" which again is a better shorthand for "estimand based on treatment policy for all intercurrent events". We have already got lazy as a community with our shorthand descriptions and it makes things more complicated and not less in some situations. Nevertheless, we need simple and understandable concepts so we can communicate across disciplines and if MOST of the time, this laziness actually helps, maybe we should just accept it, and the price we pay is the difficult edge cases such as the ones you highlight where we have to row back and be precise about what we really mean.
https://www.ema.europa.eu/en/documents/scientific-guideline/guideline-clinical-investigation-medicinal-products-treatment-or-prevention-diabetes-mellitus-revision-2_en.pdf
Hi Andrew
Thank you for taking the time to clarifying this! I now understand what you meant – apologies for the initial misunderstanding.
I really like the addendum's phrasing of 'based on a treatment policy', and suspect this was so that the ICE strategy doesn't dominate the other attributes too much.
I also like its approach of taking each ICE separately, and think we can define the strategy for a given ICE when we're not feeling lazy… but your mashup terms above have opened a whole world for me. Imagine the fun when we have four different types of intercurrent events!
Enjoying your posts here! Re: direct effect I think the main distinction between a mediator event vs truncation event is that the mediator doesn't preclude the outcome from occurring...so direct effect is just portion of total effect that doesn't go through mediator (yes cross world counterfactuals are involved but let's not go there)..it's definitely hurting my head to think about competing/truncating events as akin to mediation where outcome isn't observed
Thanks! Yes, absolutely agree about that distinction. Cross-world counterfactuals always hurt my head. I'm convinced when other people say how useful they can be but I'm not smart enough.