Depletion of Susceptibles

In plain language

Depletion of susceptibles happens when the patients most likely to be harmed by a drug have the problem early, stop the drug, switch away, die, or disappear from follow-up. The patients still exposed later are the ones who tolerated the early period, so late exposed person-time can look safer than the starting population really was. Analysts should look for it whenever a safety result is based on current users, long-term users, or a hazard ratio pooled across early and late follow-up.

Depletion of susceptibles

is the pharmacoepidemiologic version of survivorship selection. After a drug, vaccine, device, or treatment strategy begins, people differ in unmeasured susceptibility: some are biologically or clinically prone to the event, others are not. If the susceptible patients experience the event, stop treatment, switch, die, or are removed from follow-up early, the treated population that remains on day 90, day 180, or year 2 is no longer comparable with the population that started. It is a selected pool of tolerators. Analyses that compare late exposed person-time with early exposed person-time, or that enroll prevalent users who already survived early exposure, will therefore underestimate early harms and can make a hazardous exposure look safer over time.

Core conceptual distinction

Depletion of susceptibles is not just "risk changes with duration." It is selection of the risk set by prior exposure and outcome history. A true pharmacologic hazard can be highest early because biology changes after initiation, but depletion adds a second mechanism: the people who would have had early events are removed from the later risk set. That means a declining observed hazard is not automatically evidence of adaptation, tolerance, or waning causal effect. It may be the arithmetic consequence of losing the most susceptible patients first. This is why prevalent-user cohorts are dangerous for safety questions: they condition on surviving and remaining observable after the early high-risk window.

The estimand problem follows directly. A new-user study estimates the risk faced by patients at initiation. A prevalent- user study estimates the risk among survivors who are still exposed after some unobserved duration. A duration-response curve that does not separate biologic latency from depletion describes the changing composition of the exposed risk set, not necessarily the changing effect of the exposure.

Pros, cons, and trade-offs

This entry names a bias mechanism, so the trade-offs are between design responses: - New-user restriction vs prevalent-user inclusion: New-user restriction starts follow-up at initiation and observes the early risk period before susceptible patients are depleted. Cost: fewer eligible patients and less information on long-term stable users. Prefer it for initiation safety, effectiveness, and early adverse-event questions. - Duration-stratified analysis vs pooled current-use analysis: Stratifying by time since initiation can reveal a steep early hazard and later attenuation. Cost: duration strata are themselves selected over time; late strata should be interpreted as survivor/tolerator strata, not as the original population aging on treatment. Prefer duration-stratified reporting whenever biology or clinical practice suggests early susceptibility. - Marginal risk contrasts vs conditional hazard ratios: A hazard ratio can drift toward the null as susceptible individuals are removed from one or both arms, even when a subgroup-level effect persists. Risk differences at fixed horizons, cumulative incidence curves, and explicit susceptible-subgroup diagnostics are often more interpretable than a single pooled Cox hazard ratio. - As-treated censoring vs intention-to-treat follow-up: As-treated analyses censor exactly when intolerance or early toxicity may be informative. They answer an on-treatment question but can amplify depletion if discontinuation is driven by prodromal symptoms or early adverse effects. Intention-to-treat-like follow-up keeps initiators in the risk set for a policy question, but dilutes risks after discontinuation. Report both when the safety signal is plausible.

When to use

Diagnose this mechanism whenever exposure duration matters: early drug toxicities, vaccines during waves of infection, chronic therapies with intolerance-driven stopping, pregnancy or oncology regimens where frail patients stop early, and comparative safety studies where one arm has more early discontinuation. It is especially important when a protocol proposes "current users" or "ever users," when the first observable fill may not be the true first fill, or when a hazard ratio is summarized across a long follow-up horizon despite visibly non-proportional hazards.

When NOT to use - and when it is actively misleading

- Do not label every time-varying hazard as depletion. Some hazards decline because induction ends, dose is reduced, or clinical monitoring improves. Depletion is specifically selection of who remains at risk. - Do not "fix" depletion by adjusting for post-initiation covariates such as early lab response, blood pressure control, or first-month tolerability. Those variables often lie on the pathway through which susceptible patients are selected. - Do not treat a late low rate among prevalent users as reassurance about initiation safety. The late rate excludes the patients who had early events, stopped, or died. - Do not reset time zero at a later refill and call it a fresh start unless the new episode satisfies a real washout and observability rule. Otherwise the patient had to survive to that refill, creating immortal time and survivor selection.

Data-source operational depth

- Claims: The central diagnostic is an exposure-duration table. For each initiator, derive first qualifying fill after a fully observed washout, construct days-supply episodes, and tabulate outcome, discontinuation, switch, death, and disenrollment by early duration bands such as 0-30, 31-90, 91-180, and 181-365 days. Require continuous medical and pharmacy enrollment across the washout; exclude unobservable Medicare Advantage-only spans when fee-for-service claims are needed. If early discontinuation is much more common in one arm, an as-treated hazard ratio after censoring is partly a comparison of selected survivors. Sensitivity analyses should vary washout, grace period, and lag assumptions. - EHR: Susceptibility can be hidden in labs, severity scores, adverse-event notes, and dose holds that never become claims. Use EHR data to distinguish biologic adaptation from stopping due to early toxicity, but remember that patients who stop returning to the system may look event-free. Link dispensing and mortality when possible. - Registry: Registries often enroll patients after treatment start, which can make the entire registry cohort a prevalent survivor cohort. Use explicit treatment-start dates, adverse-event stop reasons, and vital status to separate true long-term tolerators from patients whose early susceptible period is simply unobserved. - Linked data: Linked claims-EHR-vital records are the best substrate because pharmacy fills define initiation, clinical data describe susceptibility and reasons for stopping, and mortality captures the competing removal process. Reconcile order, fill, administration, and event dates before assigning duration bands.

Worked example

A claims analyst studies 180-day risk of serious infection after initiating Drug A versus Drug B. Both arms have 5,000 new users after a 365-day washout. In the first 30 days, Drug A has 80 infections and 600 discontinuations; Drug B has 40 infections and 200 discontinuations. From day 31 to day 180, Drug A has 60 infections among the remaining on-treatment patients and Drug B has 70. A pooled as-treated Cox model after censoring at discontinuation may suggest little difference, because Drug A's most susceptible patients either had the event or stopped early. The correct diagnostic report shows the early event rate, the early discontinuation rate, and the cumulative incidence from initiation under an intention-to-treat-like contrast. If Drug A's early discontinuers have high rates of hospitalization or death immediately after stopping, the "on-treatment" analysis is not a neutral safety estimate; it is conditioned on tolerating Drug A.

Worked example

Scenario

Two new-user drug cohorts are followed for serious infection. Drug A has more early infections and many more early discontinuations than Drug B. A pooled as-treated analysis hides the early Drug A excess because susceptible Drug A patients leave the on-treatment risk set quickly.

Dataset

Simplified 180-day new-user follow-up by treatment arm and duration band.

Steps

• Start both arms at first qualifying fill after a fully observed washout; do not enroll prevalent users.

• Report the first 30 days separately because early biologic susceptibility and intolerance are plausible.

• Notice that Drug A loses 680 patients in the first 30 days through infection or discontinuation, compared with 240 for Drug B.

• The lower later rate in Drug A is conditional on surviving and tolerating the early period; it is not reassurance about initiation safety.

• Compare cumulative incidence from initiation and run sensitivity analyses that keep follow-up after discontinuation, vary the grace period, and examine post-discontinuation events.

Result

Drug A's apparent late safety is compatible with depletion of susceptibles. The report should foreground early risk and early discontinuation, not summarize the whole study with a pooled as-treated hazard ratio.