Methodology: Benefit Risk Assessment when Risk and Benefit are only observed in different timing

In the FDA benefit risk guidance (2021) the following point was mentioned: “The time course over which the benefits and risks occur (e.g., considering adverse events that may occur shortly after initiation for benefits that may take years to accrue).” How would you suggest to do the benefit risk assessment in this case reflecting the different timing on onset of risk Vs observation of benefit? Is there any existing methodology we can follow or to adapt from? What are the key considerations we should put in place to make this BRA reasonable?

Maybe we can think of a hypothetic example. Let’s say in a study where we are comparing Treatment A with the stand of care (SOC)—Treatment A is 6-month in total, SOC is 12-month in total.

Let’s say Treatment A might be associated with one particular risk (let’s say Risk One), and Risk One is normally observed in the first 8 weeks. If the subject did not experience Risk One in the first 8 weeks, they are very unlikely to have it after 8 weeks of treatment. But once they had Risk One in the first 8 weeks, they are unable to finish the 6-month treatment and will have to be withdrawn, thus will be claimed as treatment failure due to safety reasons. For those who were able to complete the treatment with Treatment A, their cure rate is very high. So the benefit for Treatment A is the high cure rate as well as a much shorter treatment duration, as long as they can pass the first 8 weeks (Risk One free).

Can you please come up with recommendations for this scenario? Feel free to add more details to the made-up example above if this helps, thank you!

In Health economic modeling we would deal with sort of issue through a discount rate, which tries to directly account for the present value of benefits realized in the future (usually in terms of QALYs). Could something like that apply here?

Interesting thought! I am not familiar with the QALYs but I think I see where you are coming from. For each of the individual patient, the outcome of the benefit is “all or none” at the predefined endpoint, while for the Risk One, it is also “yes or no”. But I assume you are talking about the discount rate from a population level so we can aggregate the benefit and risk in a meaningful way for comparison purpose. Is that correct?

QALYs are just a way to adjust a year of life lived for the quality of life in that year so for individuals who have Risk One maybe they have like earn 0.8 QALYs for that year and then go on to live for 10 years so 10.8 QALYs accrued in total (ignoring for simplicity that QALYs reduce with age and disease). Then you apply a discount rate so that life years earned farther in the future are valued lower than those today.

The only other benefit risk type work I’ve done has been in multi-criteria decision analysis and haven’t really considered discounting in those applications. I guess you could do something similar by having a preference weight for avoiding Risk One and a preference weight for life years lived and then apply discounting to life years?

If we just stuck within the binary yes/no Risk One/Cure framework then you could maybe argue that clinician/patient elicited preference weights already account for delay in realization of the benefit? I suppose you could maybe investigate with some sort of discrete choice experiment where you estimate utility of cure that is realized at different points in the future?

When the risk and benefit time frames are not terribly different, a state transition model may be the most well-defined way to handle treatment withdrawal. For example one may estimate these probabilities:

  • P(bad clinical event by time t)
  • P(moderately bad clinical event or worse by time t)
  • P(clinical event or need for treatment withdrawal by time t)

When the time frames are much different can we really do a risk-benefit assessment? Suppose that we estimate P(safety problem by 2y) and benefit is only monitored for 8w. To do a proper risk-benefit tradeoff would we not need the benefit to be measured as of 2y?

I think this is fair. A common example I can think of is neonatal interventions to prevent something like sepsis, newborn lung disease, or bowel infections with the long-term outcome being development at 2-3 years of age. The famous example is post-natal steroids which are very effective at reducing lung disease of prematurity but increase the risk for developmental delay. I can see an argument that you could measure quality of life or something similar at 2-3 years but clinicians think in terms of “What did it cost me to reduce this NICU complication.”

1 Like