The Target Trial Protocol as a Causal Question

From a Clinical Question to a Trial Protocol

A clinical question such as “Does treatment improve outcomes?” is not yet a well-defined causal question. It leaves open who is being treated, what treatment means, when treatment begins, which outcome is counted, and over what period. Different answers to those questions can describe different causal effects, even when they use the same treatment labels.

The target trial protocol makes those choices explicit. It describes the hypothetical trial whose causal question an analysis is intended to answer. In observational research, investigators may not be able to conduct that trial, but the protocol still provides a reference for deciding what comparison the observational analysis is meant to emulate.

A compact representation is \(P=(E,G,\mathcal{A},t_0,\tau,Y,\mathcal{C})\), where \(E\) defines the eligible population, \(G\) the strategies, \(\mathcal{A}\) the assignment procedure, \(t_0\) time zero, \(\tau\) the follow-up horizon, \(Y\) the outcome definition, and \(\mathcal{C}\) the planned causal contrast and analysis. This notation is a checklist, not a claim that every study has only one possible choice for each component.

For example, a causal contrast might compare the mean outcome under strategy \(g_1\) with the mean under strategy \(g_0\) among eligible individuals. Writing the contrast requires specifying the population, strategies, outcome, and horizon; the protocol supplies those ingredients. The statistical estimator is a separate object: it is the procedure used to estimate the contrast from available data.

A Protocol Is a Set of Linked Decisions

The components should be specified together. A treatment strategy must be meaningful for the people eligible to receive it; assignment must occur at a defined point; and outcome follow-up must begin from a time that can be applied to both treatment groups. The details below are the protocol-level questions to answer, not substitutes for the more focused treatment of each component.

• Eligibility: Which individuals could enter the trial, and what information determines eligibility? Criteria should be assessable at the point the trial begins.
• Treatment strategies: What are the alternative actions being compared? A label such as “treated” may be insufficient if it hides materially different actions.
• Assignment: How would eligible individuals be allocated to strategies in the hypothetical trial? State whether allocation is randomized and, if so, the allocation probabilities or scheme.
• Time zero: At what instant are eligibility established, assignment made, and follow-up begun?
• Follow-up: How long are outcomes observed, and what ends observation?
• Outcome: What event or measurement counts as the endpoint, and how is it ascertained?
• Analysis: What population-level contrast will be reported, and what analysis will estimate it?

These are not merely reporting details. They determine what the causal contrast refers to. A protocol that says “compare treatment A with treatment B” but omits the population or outcome does not yet define one unique causal question.

Why Protocol Choices Define the Causal Contrast

Potential-outcome notation makes the dependence on the protocol visible. Let \(Y^a(\tau)\) denote the outcome by horizon \(\tau\) that an individual would have under strategy \(a\), and let \(E=1\) indicate membership in the protocol-defined eligible population. For a binary endpoint, one possible target is the risk difference \(E[Y^1(\tau)\mid E=1]-E[Y^0(\tau)\mid E=1]\). Changing \(E\), either strategy, the endpoint, or \(\tau\) changes at least one term in this expression.

Here is the derivation. Let \(A\in\{0,1\}\) be assigned strategy and suppose \(P(A=a\mid E=1)>0\) for \(a=0,1\). Consistency means that among people assigned \(a\), the observed outcome is \(Y=Y^a(\tau)\). Random assignment gives \(A\perp (Y^0(\tau),Y^1(\tau))\mid E=1\). Therefore, for each \(a\),

The first equality uses consistency, and the second uses randomization. Subtracting the equality for \(a=0\) from the equality for \(a=1\) yields

This result concerns the hypothetical randomized trial under the stated conditions. Observational data do not acquire randomized assignment merely because the researcher writes down a target protocol. In an observational emulation, causal interpretation requires additional assumptions and a suitable analysis. \(\square\)

This follows directly from the definition: the contrast at horizon \(\tau\) is \(E[Y^1(\tau)-Y^0(\tau)\mid E=1]\). At another horizon \(\tau'\), it is \(E[Y^1(\tau')-Y^0(\tau')\mid E=1]\). These expressions are equal only if their difference has expectation zero:

There is no general reason this condition must hold. For example, an intervention could delay events without preventing them, producing one contrast at 6 months and a different contrast at 2 years. A horizon must therefore be part of the causal question, rather than an unreported analysis convenience. \(\square\)

Protocol, Emulation, and Analysis Are Not the Same Thing

A target protocol describes the hypothetical trial. An emulation uses available nonrandomized data to approximate specified parts of that trial. An analysis method estimates a chosen contrast from the data. Keeping these roles separate prevents a common mistake: treating the output of an estimator as though it automatically defines the causal question.

For a randomized trial, the protocol’s assignment mechanism is implemented by the trial. In observational data, treatment decisions arise through clinical practice. The analyst must specify the intended protocol and assess whether the observed data and assumptions can support the desired comparison. For example, a marginal structural model may be used in an analysis of longitudinal treatment and confounding, but choosing that model does not resolve an ambiguous eligibility definition, strategy, time zero, outcome, or horizon.

A useful alternative is to begin with the available data and report the comparison they readily support. That can be a valid descriptive analysis, but it may not answer the desired causal question. Protocol-first reasoning instead starts with the question and asks whether the data can emulate it. The choice is not between “protocol” and “model”; the protocol defines the target, while the model is one possible tool for estimation.

Diagnostics, Failure Modes, and Limitations

Before estimation, check whether a reader could independently determine who enters the target population, what each strategy requires, when follow-up starts and ends, and how each outcome is counted. If reasonable analysts could implement materially different versions of the protocol from its description, the causal question remains underspecified.

• Vague intervention: “Receive usual care” may represent different actions across clinicians or settings. Clarify the strategy sufficiently for the intended comparison.
• Unstated population: An average effect among all patients and an average effect among a clinically selected subgroup need not be the same.
• Unstated horizon: A result reported as “risk reduction” is incomplete if the period over which risk is counted is unclear.
• Protocol-data mismatch: A database may not record a required eligibility variable, treatment component, or outcome. This is a limitation on emulation, not something fixed by relabeling the analysis.
• Overclaiming causal interpretation: A precise protocol makes the target clear; it does not itself guarantee exchangeability, positivity, consistency, accurate measurement, or absence of interference.

The protocol is also not a guarantee that the chosen question is clinically important or answerable. Practical constraints may require a different population, endpoint, or follow-up period. Those changes should be treated as changes to the target question and justified on substantive grounds.

Common Mistake: Letting the Data Silently Define the Question

Wrong reasoning: “The database includes everyone who started treatment and records outcomes for one year, so the causal question is the effect of treatment in these people over one year.”

This does not establish that the included people were eligible at a common starting point, that treatment initiation is the intended strategy, or that the comparator and outcome are defined comparably. The data’s available variables and observation window are practical constraints, not a complete causal protocol.

Correction: State the intended protocol first, including the target population, strategies, assignment, time zero, follow-up, outcome, and analysis contrast. Then describe which parts the data can represent and which assumptions are needed to interpret the resulting estimate causally.