I have read through this observational data emulated target trial and would be interested in what the methodologists here think of the study. I personally like it and only found small quibbles with it. I think the conclusion is right and I suspect it will bear out in the prospective trials that are currently running with HCQ not showing any benefit in the end. In this article, it seems like the authors were pretty careful about how they set their analysis up.
Thanks!
I was going through posts on target trial emulation and thought I should flag that this was not a target trial emulation. It was called that on the preprint but on eventual publication in the BMJ “target trial emulation” was dropped from the title. We have a preprint that explains what a target trial emulation is, and by extension, what is not and uses a clear example of the Stanford Heart Transplant Study.
Moved to another discussion.
The main aim of the post was to call out the target trial issue and I think we should stick with that in this thread. Interestingly, these authors allowed a “grace period” of 48 hours after admission for initiation of hydroxychloroquine (HCQ) but the start of follow-up (baseline or time zero) for each patient was the time of admission to hospital. This leads to what is called a “from threshold” analysis (the threshold being hospital admission) which is also the eligibility. We call this a type 3 failure of emulation where the treated are immortal till they get HCQ while the controls could die within this 48h period. Additionally there were people who received HCQ after 48h and these were excluded which is again a problem and they dealt with this by two sensitivity analyses: “Firstly, mimicking an intention-to-treat analysis: all patients eligible for the study were analysed, and those who received hydroxychloroquine after 48 hours were analysed in the control group. Secondly, mimicking an as-treated analysis: patients who received hydroxychloroquine after 48 hours were analysed in the treatment group.” This again is incorrect so there were many issues here of failure to emulate a target trial and your comments on this aspect would be appreciated. You can move the other comment on pathways to our previous thread so that more discussion on that aspect may ensue.
Brilliant point that eligibility and time zero are the “natural pair”. Indeed that the natural pair is actually biological state and biological time, with eligibility and time zero serving only as observable proxies.
Minimum requirement (target-trial framework)
T0 = E
Everyone becomes eligible and enters follow-up at the same procedural time.
This requirement addresses temporal biases arising from misalignment of eligibility and follow-up. It is fundamentally an administrative or methodological requirement rather than a biological one.
Stronger requirement
T0 = E = B
where (B) is a biologically meaningful state.
Under this framework, eligibility and follow-up begin at a biologically relevant point rather than at an administrative event. Merely synchronizing on a procedural clock does not guarantee biological coherence. For example, a 48-hour grace period represents approximately two rotations of the Earth. This interval is defined by convention rather than by disease biology.
In rapidly evolving infections, such delays cannot automatically be regarded as trivial administrative accommodations. In our studies of Ebola viral load and antibody time series, substantial biological evolution occurred over comparable periods. Viral replication, host immune activation, and tissue injury may change markedly during such intervals. A synchronization point defined solely by elapsed clock time therefore risks grouping together patients occupying very different biological states.
Strongest requirement
T0 = E = B(t)
where eligibility and follow-up are synchronized not only on the same biological process but also on approximately the same biological time within that process.
Under this formulation, patients are aligned not merely because they share a diagnosis or biological pathway, but because they occupy comparable positions along the trajectory of that pathway. Two patients with the same infection may differ substantially in biological time despite sharing the same hospital admission date, positive test date, or severity threshold.
Achieving this level of synchronization may require analysis of relational biological time series, including pathogen burden, immune response trajectories, biomarker evolution, physiological adaptation, and other dynamic features that more faithfully represent progression through the disease process than calendar time alone.
The progression from:
T0 = E.
to
T0 = E = B
to
T0 = E = B(t)
represents a shift from procedural validity, to biological validity, to biological-temporal validity.
I agree, B(t) is important. To align B(t) with T0=E means that “severity” at say admission needs to be accounted for and thus I would consider B(t) a confounder (and we can adjust if there is a measurable marker for B(t) at baseline). However if A happens at variable times post admission in only the treated group, no direct adjustment is possible to align A and that is where the difficulty arises.
I got a query about why we did not reference a recent (2026) paper by Fu, Hernan et al in our preprint. They talk (like us) about the types of target trial emulation designs and provide three options using an example of metformin initiation and mortality, In target trial type 1 they have eligibility (E) as diagnosis of type 2 diabetes in previous month, treatment strategies were start metformin and continue use vs never start metformin and T0 as time of assignment to a treatment strategy. They have E in target trial 2 as diagnosis of type 2 diabetes in the past 3 months, the same treatment strategies as target trial 1 and the same T0 (time of assignment to a treatment strategy). They finally have a third design that aligns with our preprint.
We did not cite this paper because we initially (see addendum 2 below) did not agree with target trial design types 1 & 2 and, of course, we believed that target trial design type 3 needed broader analytical solutions while Fu et al tied it down to a design with a single analytical fix.
Our initial problem (see addendum 2 below) was that trial type 1 synchronizes treatment assignment by restricting the treated group to those who initiate in the eligibility month, forcing A=T0=E at the cost of enormous data loss and misclassification of all late initiators as untreated until they initiate, whereupon they are censored. In Trial Type 2, based on how Fu et al. worded their description, we assumed they were advocating moving T0 and E forward to each initiation time sequentially, but by moving T0 and E to A rather than fixing them at the true eligibility event, it collapses the pre-treatment period, making it invisible and embedding lead time bias and selection/survival collider bias within the sequential structure. After reflecting on this for some time, we have reached the conclusion that this is not what Fu et al meant and this is discussed in addendum 2 below.
Finally, our preprint now references their paper with our clarification and expands on what the type 3 design means and how it can be delivered while Fu et al focus on only one analytical solution for this design.
Addendum
I posted a rapid response to their paper on the BMJ site: Rapid Response to Fu et al.: Starting Right, But How Far Right?
Addendum 2
This is our update in our preprint to the target trial type description by Fu et al based, not on what they describe, but our reinterpretation of what they mean:
There has been a recent three-target-trial design type classification where all three trial types share the same foundational design requirement that E = T0 are aligned by design and both precede A. The three types differ in how treatment initiation relates to E = T0 and how delayed treatment initiation is accommodated analytically: Type 1, A coincides with E=T0 by clinical necessity; therefore no pre-treatment follow-up exists and no additional analytic alignment is required; Type 2, E = T0 is defined as the time individuals first satisfy a biologically specified time rule (e.g., entering an age or gestational interval). Sequential target trials repeatedly instantiate E = T0 over calendar time by applying the same rule at successive time points, generating distinct risk sets of individuals who satisfy it at each time. Within each trial, A may occur after E = T0 but only within a prespecified window, so all initiation times within that window define a single intervention strategy rather than distinct time-varying contrasts. The clinical question is whether initiating within the window improves outcomes versus not initiating or initiating an alternative strategy within the same window; Type 3, A may occur at any time after E = T0, with treatment timing determined by evolving patient characteristics, disease progression, and clinical decision-making. Consequently, treatment initiation arises during follow-up and requires explicit analytical methods to correctly handle pre-treatment person-time and avoid its misclassification. Thus, proper emulation first requires synchronizing the natural pair of time zero (T0) with eligibility (E) by design and ensuring that both precede treatment (A).