I’ve looked through previous discussions but haven’t found one that directly addresses this issue. I’ve come across arguments suggesting that composite outcomes, especially those including all-cause mortality (ACM), should always be preferred in drug trials. The rationale seems to be that patients should not be given treatments that might reduce their lifespan, even in ways unrelated to the primary purpose of the drug.
As a Master’s student in Statistics, I find myself not appreciating this very much. I’m more interested in understanding the treatment’s effect on the specific outcome the drug was developed to address. Focusing on broader outcomes like ACM could dilute important information, especially if those outcomes are not of direct relevance to the patient or are not meaningfully affected by the intervention.
I’m curious to hear what medical statisticians think about this. Is it truly necessary, or even advisable to require evidence of reduced all-cause mortality across multiple trials before recommending a drug? Are there strong methodological or ethical arguments in favor of this stance?
If there are any papers or reviews that delve into this debate, I’d be very grateful for references.
It’s wrong to sweep deaths under the rug, so deaths must be included in endpoints
If death is very rare, an ordinal outcome with death as the worst category will account for death but deaths won’t count very much against a treatment
An ordinal endpoint that is dominated (because of few deaths) with a main outcome of interest thought to be related to the treatment will be driven by that outcome
Much has been written about the inability of reliably classifying deaths as unrelated to treatment
I might be misunderstanding your question, but I’m wondering where you heard/read that only drugs with a proven mortality benefit should/will be approved (?) This is not the case. In fact, very few of the drugs physicians prescribe every day (e.g., antacids, pain medications, antidepressants, inhalers, antibiotics…) will have a proven mortality benefit. Medications for many acute or chronic conditions (e.g., antihistamines for hives, NSAIDs for musculoskeletal pain, loperamide for IBS-associated diarrhea) would not be expected to reduce mortality, since the underlying conditions they are meant to treat are not inherently life-threatening. Therefore, the trials that led to their approval would have been unlikely to include mortality as an endpoint.
Specifying composite endpoints will increase the number of “events of interest” that are captured in each treatment arm during a trial, thereby increasing the trial’s power to detect an effect of therapy, if it exists. But including mortality in a composite endpoint would only make sense in specific clinical contexts. This is because only a subset of all trials would be expected to record more than a very small number of deaths (if any). Unless trials involve very sick people or enrol very large numbers of subjects or follow subjects for a very long time, the number of deaths recorded during a trial is going to be small.
It’s challenging to interpret the effect of a treatment on mortality when death is included as part of a composite endpoint but the trial captures only a small number of deaths. This issue was discussed in a thread started during the pandemic:
It seems like it would be most important to get a handle on a therapy’s effect on mortality if one or more things are true:
The sponsor is trying to show that the treatment reduces mortality;
The therapy has recognized, potentially life-threatening toxicities related to its mechanism of action (e.g., immunosuppression, bleeding, clotting, hemodynamic or metabolic/electrolyte effects);
The therapy is going to be used in a medically precarious population (i.e., a population already at increased risk of death by virtue of their illness), especially if use is expected to be prolonged (e.g., therapies for heart failure, cognitive impairment, cancer).
If a trial doesn’t record enough deaths to allow us to infer a mortality benefit reliably, sometimes the best we can do is to ensure that there’s no adverse mortality signal (i.e., a between-arm imbalance in mortality that favours placebo/standard of care over the new treatment). Such an imbalance would always be a red flag for drug regulators.
I agree with parts of that but not others. Restricting the case to a trial not intended to show a mortality benefit but experiencing a few deaths, also assume that death prevents the main endpoint from being assessed. Though the trial will not be able to make a statement about the benefit of treatment on mortality, mortality is a bad event and should not be covered up as some sponsors have done. We can generalize the question from “does the treatment improve Y” to “does the treatment improve Y’ “ where death is at the worst value of Y’. This gets at the fundamental clinical question “do patients taking drug B have better outcomes than patients taking A?”.
For example consider a study whose main goal is to assess treatment benefit on patients’ functional status and B=new drug and A=control. If there are 4 deaths in B and 1 in A, B should be penalized. After penalizing for death, suppose that we find P(treatment benefit | data, prior) = 0.97. We could interpret this as saying that the study provided strong evidence for benefit of treatment B, penalized for death or after accounting for death.
Oooh- this thread has become very interesting It gets at an important question: What should regulators do with a drug trial that records a small number of deaths but happens to show a mortality imbalance favouring placebo/comparator (e.g., 4 vs 1 death; 5 vs 1 death; 6 vs 1 death;…)? Consideration of clinical context/cause of death would probably become very important in this situation.
The regulator’s concern would be whether the imbalance seen with a small number of recorded deaths reflects a “fluke” (e.g., flipping tails 4 times on a row with a fair coin) or whether it would have persisted if the trial had been larger or lasted longer and many more deaths had been recorded (e.g., flipping tails 40 times in a row- do we still think the coin is fair?), perhaps suggesting a true adverse effect of the therapy.
I would imagine that this is a common scenario faced by Drug Safety Monitoring Boards when they do interim analyses of ongoing trials (?) Not having sat on such a board, I’d be interested to hear how these discussions unfold. I would imagine that there’s a detailed clinical review of the cause(s) of death of the participants and that it would be reassuring to a board if causes of death were clearly unrelated to the treatment being tested (e.g., passenger in a car involved in an accident, anaphylactic reaction after consumption of a known allergen) (?)
I could imagine a DSMB allowing a trial to continue and the trial ending with a mortality imbalance favouring placebo/comparator but with few recorded deaths. I could also imagine an outcry about the imbalance after publication of the trial, with sponsors arguing that the imbalance was a fluke and at least some clinicians being reluctant to prescribe the drug. I’m trying to imagine whether hesitant clinicians would be placated by the new drug having been “penalized” in the manner that you describe (i.e., by having the deaths work against the new drug in the assessment of its efficacy for improving the main outcome(s) of interest), given that the penalization wouldn’t really address their concern about a possible treatment-emergent mortality signal (?)
Good thoughts. Rather than agonizing over whether a mortality imbalance is a fluke, I would just explicitly penalize for death in the main analysis and do a separate subjective assessment of mortality on its own.
Your question reminds of a concern I have for the composite endpoint of Progression Free Survival (PFS) in intervention oncology trials - typically as the primary endpoint. As an advocate for patients I’ve yet to be persuaded that overall survival should be lumped with time to progression as one score. Equally concerning: why comparing HRQoL (a signal for living better while on treatment and in follow up) is not required in RCTs using PFS (a surrogate for living longer) as the primary endpoint - particularly for study drugs given until progression or unacceptable toxicity.
I cannot think of an intervention study in oncology that included all cause mortality - presumably because there are so many confounding variables, such as age, fitness, prior and following treatments, disease heterogeneity at the genetic level.
Good points. An ordinal multistate model can handle this, and it would be optimal to elicit patient utilities to apply to the different overall states so that treatments could be compared on the basis of expected utility computed from the ordinal model.
If you are certain about cause of death, non-related causes could just be censored.
It gets at an important question: What should regulators do with a drug trial that records a small number of deaths but happens to show a mortality imbalance favouring placebo/comparator (e.g., 4 vs 1 death; 5 vs 1 death; 6 vs 1 death;…)? Consideration of clinical context/cause of death would probably become very important in this situation.