The TOGETHER Trial and the Mysterious Endpoint

To what end was a pointless endpoint chosen?

This article is part of a series on the TOGETHER trial. More articles from this series here.

The primary endpoint a trial will use is perhaps one of the most important design decisions. Trials usually optimize for an endpoint that is simple, unambiguous, and in common use. They do this so that their results will not be challenged after the fact with arguments that focus on the arbitrariness of the endpoint.

It appears that the TOGETHER trial did not heed this advice. It used an endpoint that was both composite and never-before used in a trial. The endpoint is the sum of all patients that had an “event,” where an event means either a hospitalization for COVID-19, or an observation in an Emergency Room for more than six hours.

In my original article, before the TOGETHER paper was released, I highlighted the concern that the endpoint was uncommon, which made the results harder to interpret and trust.

As the suspended—but never forgotten—@Covid19Crusher had articulated at the time:

By way of explanation, principal investigator Dr. Edward Mills, in response to a reviewer comment at GatesOpenResearch.org wrote:

The composite endpoint addresses both hospitalization and a proxy for hospitalization, retention in a COVID-19 emergency setting, as many patients who would be hospitalized were prevented from admission due to hospital over-capacity during peak waves.

The key word to focus on here is “proxy.” So what is a proxy in this context? According to a definition by The Center for Government Excellence (GovEx) at Johns Hopkins University:

A proxy is an indirect measure of the desired outcome which is itself strongly correlated to that outcome.

So, we would expect the ER observation for more than six hours to be strongly correlated to hospitalizations. Is that true? Let’s have a look:

Well, it appears that if you are on placebo, you are about 1/3 as likely to be observed in the ER for over six hours as you are to be hospitalized for COVID. If you are on ivermectin, you’re close to 1/2 as likely to be observed in ER for > 6 h than be hospitalized. If you are on fluvoxamine however, you are 1/10 as likely to have an ER Observation for > 6 h as you are to be hospitalized for COVID.

In other words, it appears that the relationship between extended ER observations and hospitalizations for COVID is anything but reliable. In fact, it appears that what fluvoxamine stands out for is avoiding extended ER observation, whereas its advantage in avoiding hospitalization is quite uncertain.

This may be why the NIH, in its evaluation of the TOGETHER trial wrote:

The >6-hour emergency setting observation endpoint has not been used in other studies of interventions for nonhospitalized patients who are at high risk for hospitalization and death.

and also that…

It is difficult to define the clinical relevance of the >6-hour emergency setting observation endpoint and apply it to practice settings in different countries.

In all this, it doesn’t seem that anyone has noticed that the initial protocol defined the primary endpoint as ER observation for more than 12 hours, which was then changed two months into the trial to ER observation for more than six hours. Changing primary endpoints is never a good look, though it can some times be justified. However, simply changing the duration of the observation must surely raise an eyebrow.

Here is what the endpoint change in Clinicaltrials.gov looks like:

Looking at Clinicaltrials.gov one more sleight of hand becomes apparent: The authors represented the two primary outcomes as separate. And yet, the paper presents them as a single combined measure, just like @Covid19Crusher feared at the time:

In Need of Explanation

To conclude, we are looking at a primary endpoint that was highly unusual to begin with, and in addition was modified in the midst of the trial to reduce the time of observation from 12 to 6 hours. It was then modified again to blend ER observations with hospitalizations into a single endpoint. Each of these items on its own is cause for concern. The fact that they are all occurring in the same trial should make everyone demand of the authors a complete and convincing explanation for this design decision. To date, such an explanation has not been forthcoming.

This article is part of a series on the TOGETHER trial. More articles from this series here.