Active Comparators and Clinical Equipoise

Why the Comparator Is Part of the Causal Question

An emulated trial does not ask whether a treatment is beneficial in the abstract. It compares specified treatment strategies for a specified population, beginning at a specified time. Choosing an active comparator is therefore part of defining the causal question: the comparison is between initiating one treatment and initiating another, not between treatment and an unspecified absence of treatment.

A useful active comparator is a treatment that could plausibly have been chosen for the same patients at the same treatment decision. This can make the comparison more clinically interpretable and can improve comparability of treatment groups. It does not, by itself, remove confounding. The comparator must be selected with attention to clinical equipoise, indication, and availability in the target population.

Definition: An active-comparator design compares initiation of one active treatment with initiation of another active treatment. Clinical equipoise is genuine uncertainty within the relevant clinical community about which of the compared strategies is preferable for the patients and outcomes at issue. It is a justification for a meaningful comparison, not a claim that the treatments are identical or that each clinician is personally indifferent.

Equipoise is specific to the population and decision. Two drugs may both be accepted treatments for a broad diagnosis, while only one is appropriate for a subgroup with a particular contraindication. Conversely, a comparison can be scientifically useful even if clinicians expect one treatment to be better, provided there is still a meaningful question about the size or balance of benefits and harms. Equipoise is not established by observing similar treatment frequencies or by obtaining a nonsignificant statistical test.

Aligning Indication, Population, and Treatment Availability

Start by specifying the treatment decision the emulated trial represents. Ask whether both treatments address the same clinical problem, are considered at the same stage of care, and could reasonably be initiated for eligible patients. If one treatment is normally reserved for severe disease and the other for mild disease, a comparison of all users may primarily contrast disease severity and treatment pathway rather than the treatments themselves.

A shared indication does not require identical clinical roles. It does require a sufficiently clear population in which each strategy is a plausible option. The target population may therefore be narrower than the population with the diagnosis: for example, patients without contraindications to either option, treated in settings where both are available, and at a common point in care.

Availability has several dimensions. A treatment may be approved but not stocked locally, unaffordable for some patients, inaccessible under insurance rules, or unavailable during part of the study period. These constraints affect who could receive each strategy. Restricting the target population to settings and patients for whom both treatments are genuinely options may improve the meaning of the comparison, but it also changes whom the result describes.

Name the decision.
Specify the clinical problem, point of care, treatment initiation strategies, and follow-up question.

Assess shared indication.
Ask whether both treatments are plausible responses to that problem for the proposed target population.

Check equipoise and availability.
Describe why the comparison is still clinically relevant and whether both strategies can actually be chosen in the relevant settings.

State the resulting population.
Make clear which patients and care settings the contrast covers, including any exclusions imposed by contraindications or lack of access.

Worked Example: Choosing Between Two Treatments for the Same Decision

Suppose an emulated trial will compare initiation of Drug A with initiation of Drug B for adults with newly diagnosed condition X. Both drugs are accepted first-line options in the participating clinics. The proposed eligibility criteria exclude patients with a documented contraindication to either drug, and both drugs are covered during the study period.

Decision and strategies. The decision is initial pharmacologic treatment after diagnosis. The strategies are initiate Drug A and initiate Drug B at that decision. The comparison is not “Drug A versus no care.”

Indication and equipoise. Since both drugs are used for the same condition at the same treatment stage, the clinical indication is shared. If the relevant clinical community is uncertain about their comparative balance of effectiveness and adverse effects in this population, the comparison has clinical equipoise. A belief that one may be more effective does not by itself invalidate the question.

Target population. The estimand applies to eligible adults in clinics where both options are available and neither is contraindicated. It does not automatically apply to patients who can receive only one drug, or to settings where one drug is not accessible. Those exclusions should be made explicit rather than concealed by describing the result as applying to all adults with condition X.

What the Comparison Identifies

The causal interpretation can be stated using baseline covariates $L$, treatment $A$, and potential outcomes $Y^1$ and $Y^0$, where $A=1$ denotes initiation of the first active treatment and $A=0$ initiation of the comparator. Let $E=1$ indicate membership in the target population, and let $Y$ be an outcome measured at the prespecified horizon. The following standard identification result clarifies what a carefully chosen comparator does and does not accomplish.

Theorem: Suppose, among eligible individuals $E=1$, (i) consistency holds, so $Y=Y^a$ when $A=a$; (ii) conditional exchangeability holds, $Y^a\perp A\mid L,E=1$, for $a\in\{0,1\}$; and (iii) positivity holds, $P(A=a\mid L=l,E=1)>0$ for each $a$ and almost every $l$ in the target population. Then $$ E[Y^a\mid E=1] = \int E[Y\mid A=a,L=l,E=1]\,dF(l\mid E=1). $$ Thus the mean contrast between the strategies is identified by standardizing each treatment-specific conditional mean over the same target distribution of $L$.

Proof. By iterated expectation, $$ E[Y^a\mid E=1] = \int E[Y^a\mid L=l,E=1]\,dF(l\mid E=1). $$ Conditional exchangeability gives $E[Y^a\mid L=l,E=1]=E[Y^a\mid A=a,L=l,E=1]$. Consistency gives $E[Y^a\mid A=a,L=l,E=1]=E[Y\mid A=a,L=l,E=1]$. Positivity ensures that this observed conditional mean is defined for almost every $l$ in the target population. Substitution proves the formula. Applying it for $a=1$ and $a=0$, then subtracting, gives the standardized mean difference. $\square$

The theorem makes two points important for comparator selection. First, an active comparator can help define a more focused treatment decision, but exchangeability remains an assumption: patients may still choose or be assigned treatments for reasons related to their outcomes. Second, the target distribution is not automatically the distribution of every person with the diagnosis. It is the distribution in the population specified by eligibility and availability.

Worked Example: Standardizing Over the Same Indication Mix

Consider two active treatments, A and B, and a binary baseline severity measure $L$. Suppose the eligible target population is half low severity and half high severity. The observed risks, assumed here to equal the corresponding potential-outcome risks under conditional exchangeability and consistency, are:

\begin{aligned} P(Y=1\mid A=1,L=\text{low})&=0.02, & P(Y=1\mid A=1,L=\text{high})&=0.12,\\ P(Y=1\mid A=0,L=\text{low})&=0.03, & P(Y=1\mid A=0,L=\text{high})&=0.15. \end{aligned}

Standardize both treatments to the same severity distribution. For A, the target risk is $0.5(0.02)+0.5(0.12)=0.07$. For B, it is $0.5(0.03)+0.5(0.15)=0.09$. The standardized risk difference, A minus B, is $0.07-0.09=-0.02$, or 2 percentage points lower risk under A.

Now suppose 90% of A initiators are high severity, but only 10% of B initiators are high severity. Their crude risks are $0.1(0.02)+0.9(0.12)=0.11$ for A and $0.9(0.03)+0.1(0.15)=0.042$ for B. The crude difference is $0.11-0.042=0.068$, apparently favoring B. The reversal occurs because the groups have different severity distributions. An active comparator did not prevent this imbalance; measured baseline severity must be handled in the analysis, and unmeasured confounding may remain.

Availability, Positivity, and the Limit of the Comparison

Clinical equipoise and statistical positivity are related in practice but are not the same condition. Equipoise concerns whether the comparison is a relevant clinical question. Positivity concerns whether both treatments occur with positive probability across the covariate profiles in the target population. A trial question can be clinically reasonable while the available observational data contain too few patients receiving one option in important subgroups.

Result: If treatment is deterministic at a covariate value $l$—for example, $P(A=1\mid L=l,E=1)=1$—the observed data at $l$ do not identify the outcome mean under $A=0$ without additional assumptions.

Proof. At this covariate value, consistency reveals outcomes under $A=1$, but no individual with $L=l,E=1$ is observed under $A=0$. Consider two possible data-generating systems that agree on treatment assignment and on every observed outcome: in both systems, $Y^1$ has the same observed distribution. In the first system, set $Y^0=0$ for everyone with $L=l$; in the second, set $Y^0=1$ for everyone with $L=l$. Because nobody with this $L$ receives $A=0$, these systems imply exactly the same observed data. Yet they imply different values of $E[Y^0\mid L=l,E=1]$. The counterfactual mean is therefore not identified from those data alone. $\square$

Worked Example: When One Treatment Is Not Really Available

A health system considers comparing Drug A with Drug B. For patients in a remote clinic, Drug B is not stocked, so all eligible patients there initiate A. In urban clinics, both drugs are used. A proposed target population includes both clinic types.

Among remote-clinic patients, $P(A=1\mid L=\text{remote},E=1)=1$, where $A=1$ denotes Drug A. The data contain no Drug B outcomes for that stratum. A model can still produce a predicted Drug B risk there, but that prediction relies on extrapolation or additional modeling assumptions; it is not identified by observed treatment comparisons in that stratum.

One defensible design may restrict the target population to clinics where both drugs are available and describe that narrower population. Another may retain all clinics but state the additional assumptions needed to transport or extrapolate results. Simply dropping the remote patients changes the target population; it does not establish an effect for them.

Diagnostics, Alternatives, and Common Pitfalls

Before estimating the effect, examine treatment use across clinically meaningful baseline strata: indication severity, contraindications, care setting, calendar period, and access measures are common starting points. Report strata with absent or very sparse use of one strategy. Estimated treatment probabilities near zero or one, highly variable weights, and large changes in results after reasonable restrictions are warnings that the comparison is weakly supported. These checks diagnose data support; they cannot prove clinical equipoise or rule out unmeasured confounding.

The main alternative is comparison with no treatment, usual care, or a non-active control. That may be the right question when the decision of interest is whether to treat, but it often combines treatment initiation with differences in access, care-seeking, disease severity, and reasons for not treating. An active comparator instead asks which of two feasible treatments is preferable for a shared decision. Neither design is universally superior: the choice follows the clinical question.

A frequent mistake is to treat the existence of two treatments for the same diagnosis as sufficient justification for comparing their users. The correction is to specify the shared decision and target population, assess whether both treatments were genuine options for those patients, and inspect treatment support across baseline profiles. A second mistake is to interpret an active-comparator estimate as automatically causal. The comparator can improve design relevance, but identification still requires assumptions such as conditional exchangeability and positivity.

Comparator choice also limits generalizability. If the analysis excludes patients with contraindications or settings without access to one treatment, its result pertains to the remaining target population. That narrower answer may be more credible and useful than a nominally broader answer that depends on unsupported comparisons.

Key takeaway: Choose an active comparator that represents a real alternative at the same clinical decision, then define the population in which both strategies are relevant and available. Equipoise supports the question; overlap and causal assumptions support its estimation.

Check Your Understanding

Answer each question, then check the brief solutions.

Two treatments are both used for a diagnosis, but one is contraindicated for patients with severe kidney disease. What should be clarified when defining the target population?
Does clinical equipoise imply that treatment positivity holds in the observational data? Explain.
In the standardization result, why are both treatment-specific means averaged over the same distribution of baseline covariates?
A clinic never uses the comparator treatment. Can a flexible regression model alone make the comparator outcome mean in that clinic identified by the observed data?

Solutions. (1) Clarify whether the target population excludes patients with the contraindication or whether the strategies differ by eligibility; the stated comparison should involve treatments that are both clinically plausible. (2) No. Equipoise is clinical uncertainty about the decision, while positivity requires observed treatment variation across target-population covariate profiles. (3) Using the same target distribution compares the two strategies for the same population rather than mixing treatment effects with different covariate compositions. (4) No. It can extrapolate a prediction, but with no comparator users in that clinic the relevant mean is not identified without additional assumptions.