Emulating A Trial Of Chronic Medication Strategies

Why Chronic Medication Strategies Need Careful Protocols

A chronic medication decision is rarely just a choice between two pills on one day. The question may concern the consequences of starting one medication class rather than another, or the consequences of following a longer-term strategy that includes continuing, stopping, switching, or adding treatment. These are distinct causal questions. An emulation must state which one is being asked before outcome data are analyzed.

Consider a hypothetical comparison of two first-line medication strategies for adults beginning long-term treatment for hypertension. A useful observational comparison would identify people at a genuine treatment decision, compare clinically plausible alternatives, and follow outcomes from a common time zero. It would also make explicit how later treatment changes are handled. Without these choices, an apparent comparison may mix different populations, treatment histories, and follow-up periods.

Specify Eligibility, Strategies, and Time Zero

Suppose the target trial concerns adults with diagnosed hypertension who are eligible for either of two medication classes, called X and Y. A protocol might require age 40–75 years, no recorded contraindication to either class, and no dispensing of X or Y during the previous 365 days. It might exclude people already receiving another long-term antihypertensive regimen if that makes them ineligible for the treatment decision of interest. These rules define the study population; they should not be altered after examining which definition gives the preferred result.

The trial's eligibility assessment, strategy assignment, and start of follow-up should occur at the same time, called time zero. In an observational database, one practical time zero is the date of the first qualifying dispensing of X or Y, provided the person meets eligibility criteria on that date. The comparator is an active alternative, not simply “no treatment,” when both medications are reasonable options for the same clinical decision.

“Start X” and “start Y” are not yet complete chronic strategies. One possible sustained strategy is to initiate the assigned class within a short prespecified window and continue it, allowing only specified temporary interruptions or add-on treatments. A different strategy could allow clinicians to add a second medication when a stated blood-pressure threshold is not met. Such a rule is dynamic: what treatment is recommended later depends on information observed during follow-up. The rule must be stated in a way that can be applied to the data.

Define Treatment Changes and the Outcome

For a treatment-policy estimand, participants remain grouped by the medication they initially started. A later switch from X to Y does not change initial assignment; the analysis asks about outcomes under the care patterns that followed each initial choice. This is often relevant to the real-world consequence of choosing an initial treatment, but it does not isolate the effect of sustained use.

For a sustained-strategy estimand, the protocol must define deviations. For example, stopping the assigned class for more than 60 consecutive days, switching to the other class, or adding a prohibited medication might count as a deviation. A prespecified clinical add-on rule might instead be permitted. The data then need dates and sufficient detail to determine whether each person's observed history follows the assigned strategy. If a deviation triggers censoring in an analysis, that censoring is informative unless, conditional on measured history, those censored and uncensored remain comparable for the outcome.

Long-term outcomes also need an operational definition and a time horizon. For example, let \(Y^g(\tau)\) indicate kidney failure by five years under strategy \(g\), before death. Here \(\tau=5\) years, and death before kidney failure is a competing event: it prevents the person from subsequently experiencing kidney failure. The corresponding risk is the probability of kidney failure before death by five years under the strategy. All-cause mortality can be reported as a separate outcome. The endpoint and competing-event handling must match the clinical question.

From an Emulated Trial to an Identified Risk

Let \(A\) denote the initial strategy observed at time zero, with \(A=X\) or \(A=Y\). Let \(L\) denote measured baseline covariates, such as age, baseline blood pressure, kidney function, prior cardiovascular disease, and healthcare use. Let \(Y^g(\tau)\) denote the potential outcome by time \(\tau\) under strategy \(g\). The causal contrast might be the risk difference \(\mathbb{E}[Y^X(\tau)]-\mathbb{E}[Y^Y(\tau)]\), measured in percentage points of five-year risk.

For a baseline treatment-policy comparison, a standard identification argument uses consistency, conditional exchangeability, and positivity. Consistency means that a person's observed outcome equals the potential outcome under the strategy actually received, when the observed treatment history corresponds to that well-defined strategy. Conditional exchangeability means \(Y^g(\tau)\) is independent of \(A\) given \(L\): after conditioning on measured baseline covariates, initial strategy groups have no remaining systematic differences in potential outcomes. Positivity means that each strategy has a positive probability of being received at every covariate pattern represented in the target population. The interpretation also requires no interference between individuals and adequate measurement of exposure and outcomes.

Under these conditions, the risk under strategy \(g\) can be obtained by averaging the conditional observed outcome risk over the baseline covariate distribution in the eligible population. The expectation is important: it applies whether \(L\) is discrete, continuous, or a mixture.

This expression follows by conditioning on \(L\), then using exchangeability to replace the potential outcome risk among people assigned \(g\) with the corresponding risk among those observed receiving \(g\), and using consistency to identify that observed outcome with \(Y(\tau)\). With a sustained strategy, baseline exchangeability alone is generally insufficient: measured time-varying factors that predict both adherence and outcome must also be addressed, and treatment and censoring probabilities must remain positive along relevant histories.

Diagnostics, Alternatives, and Limits

Before estimating effects, check whether the proposed emulation can actually be implemented. Inspect the distribution of baseline covariates across initial strategies, verify that dates and medication classifications are credible, and confirm that both strategies occur across the covariate patterns needed for the target population. Sparse or absent use of one medication among a subgroup signals a positivity problem; extrapolating an effect there depends on unsupported modeling. Check whether outcome and death dates are captured, and whether loss of observable follow-up is related to measured health history.

For sustained strategies, inspect adherence patterns and reasons for switching or stopping. If one strategy is rarely followed in certain clinical histories, the data may not support that comparison. A naïve analysis restricted to people who remain adherent can be biased because adherence may depend on evolving health status. Similarly, a treatment-policy analysis should not silently censor at switching: doing so changes the question and can introduce selection bias.

A randomized trial is the main alternative when feasible; random assignment can balance both measured and unmeasured baseline causes in expectation. An observational new-user active-comparator design can improve comparability by focusing on patients at a similar treatment decision, but it cannot guarantee exchangeability. A prevalent-user comparison includes people who have already tolerated and persisted with medication, so it targets a different population and can create selection from prior survival and treatment history. Finally, an initial-treatment effect and a sustained-strategy effect are alternatives in the sense that they answer different questions, not interchangeable analysis options.

The main limitations are residual confounding, incomplete treatment or outcome records, uncertain washout history, and limited overlap in treatment choices. Long-term follow-up magnifies these concerns because changes in health can influence subsequent treatment. The protocol makes the estimand clear and exposes the data requirements; it does not turn observational data into randomization.