Treatment failure sounds like it should be easy to find in real-world data. A patient starts a therapy. The therapy does not work. The patient changes treatment. The endpoint fires.

That is the story people tell when the protocol is still a slide. It is not what the data actually contain.

Claims data can show fills, administrations, procedures, rescue medications, hospitalizations, and treatment changes. EHR data can show labs, scores, imaging, medication orders, notes, and sometimes clinician-assessed response. Registries can add structured disease activity and reasons for switch. None of those sources, by themselves, contain a universal variable called "treatment failed."

Treatment failure is a constructed endpoint. If the construction is not explicit, the study silently mixes biology, adherence, toxicity, access, clinician preference, formulary pressure, death, and missing observability into one number.

The first question is not "did they fail?"

The first question is: what kind of failure are we measuring?

Clinical non-response is closest to the disease state. It asks whether the patient improved enough after starting therapy, using a disease-specific assessment such as a lab value, imaging result, disease activity score, patient-reported outcome, or clinician judgment.

Proxy treatment failure is different. It asks whether the care system behaved as if the therapy was not adequate. Signals might include dose escalation, interval shortening, add-on therapy, rescue medication, hospitalization, discontinuation, switch, or next line of therapy.

Strategy failure is different again. It treats the initiated regimen as a real-world strategy, where non-adherence, intolerance, affordability, and access barriers may be part of the endpoint because they are part of routine treatment performance.

Those are all legitimate constructs. They are not interchangeable.

The dangerous version is the claims-only study that calls a switch "non-response" without chart, lab, imaging, registry, or clinician-assessment validation. A switch may mean inadequate efficacy. It may also mean toxicity, pregnancy planning, insurance change, step therapy, cost, remission, patient preference, or a planned sequence of care. If the source cannot distinguish those mechanisms, the endpoint label has to say so.

The adequate trial window is the gate

No patient should be allowed to fail a treatment before the treatment had a fair chance to work.

That sounds obvious, but it is a common design failure. A switch three days after index is often an access correction, intolerance signal, or administrative artifact. A refill gap before the drug could plausibly affect the disease is not evidence of pharmacologic failure. A missing post-index lab is not proof of response or non-response.

The protocol needs an adequate-trial rule before outcome classification starts:

  • minimum exposure or persistence
  • first eligible assessment date
  • allowed grace periods and inpatient bridging rules
  • whether early events are excluded, classified separately, or counted as immediate strategy failure
  • which data source can trigger each component

Once that gate is set, the analyst can define the first qualifying failure component. Without it, the endpoint becomes a detector for treatment churn rather than treatment failure.

Intercurrent events are not cleanup details

Treatment failure endpoints are full of intercurrent events: discontinuation, rescue therapy, switch, add-on therapy, dose escalation, death, loss to follow-up, and treatment interruption.

ICH E9(R1) gives trialists a vocabulary for this problem: intercurrent events affect either the interpretation or the existence of the measurement. RWE studies need the same discipline. A rescue medication can be the endpoint under a composite strategy. It can be ignored under a treatment-policy strategy. It can censor follow-up under a while-on-treatment strategy. It can support a hypothetical estimand if the question is what would have happened without rescue.

The database should not decide that for you. The protocol should.

Death is the cleanest stress test. For a non-fatal failure endpoint, death before switch prevents later observation of switch. Ordinary censoring can make a high-mortality group look artificially "failure-free." Depending on the question, death may need to be a competing event or an unfavorable component of a net clinical failure endpoint. It should not be handled by default.

Composites need component accountability

Composite treatment failure endpoints are useful because routine care is multidimensional. A patient can fail through inadequate disease control, rescue therapy, escalation, acute care, discontinuation, or switch. Combining those signals can match the decision problem better than any single component.

The tradeoff is interpretability.

If one frequent, low-specificity component dominates the endpoint, the composite stops meaning what readers think it means. Discontinuation is the usual suspect. It is observable and common, but it carries many meanings. Rescue therapy or hospitalization may be more clinically specific but less frequent. A lab-defined non-response may be clinically closer to the target construct but missing or irregularly measured.

The fix is not to avoid composites. The fix is to store and report the trigger:

  • failure date
  • failure component
  • failure source
  • adequate-trial status
  • later treatment-pattern events
  • component-specific counts
  • sensitivity analyses that remove weak or ambiguous components

The first-failure date matters for time-to-event analysis. The component history matters for interpretation.

The operating checklist

Before using a treatment-failure or non-response endpoint, the protocol should be able to answer these questions:

  1. Is the endpoint clinical non-response, proxy treatment failure, strategy failure, loss of response, or a composite?
  2. What minimum exposure or persistence is required before a patient can fail?
  3. What is the first eligible response assessment date?
  4. Which components can trigger failure?
  5. Which data source is authoritative for each component?
  6. How are switch, rescue therapy, discontinuation, dose escalation, death, and loss of observability handled?
  7. Are claims-only treatment changes labeled as proxies unless validated against clinical sources?
  8. Will the manuscript report component counts and sensitivity analyses?

If those answers are not written before the analysis, the endpoint is not ready.

What this looks like in practice

Imagine a patient starts biologic A for inflammatory bowel disease on January 5. The protocol defines an adequate trial as at least 90 days of persistent therapy. After that window, any of the following can trigger proxy treatment failure: steroid rescue, dose escalation, switch to another advanced therapy, IBD hospitalization, or EHR-documented non-response.

The patient reaches the adequate-trial mark on April 5. Steroid rescue occurs on April 20. Dose escalation occurs on May 18. Switch to biologic B occurs on July 1.

For a time-to-first-failure endpoint, failure occurs on April 20. The component is steroid rescue. The later dose escalation and switch remain important, but they do not move the first-failure date.

That is a proxy treatment-failure endpoint unless clinical source data confirm non-response. If the question is pharmacologic non-response, the analyst still needs evidence that the patient received an adequate trial and that rescue reflects uncontrolled disease rather than another mechanism. If the question is real-world strategy failure, the rescue event may be enough.

Same patient. Same data. Different estimand. Different interpretation.

The point

Treatment failure is not a code to be discovered. It is an endpoint contract to be specified.

The contract has to say what failure means, when a patient becomes eligible to fail, which source can prove each component, and how intercurrent events change the meaning of the endpoint. If those decisions are visible, the evidence can be challenged, refined, and reused. If they are hidden, the study may still produce a clean estimate, but no one should trust what the estimate represents.

For a practical implementation layer, see the RWEdnesdays companion concept on treatment failure and non-response in RWE and the treatment failure endpoint checklist.

References

  1. ICH. E9(R1) Addendum on Estimands and Sensitivity Analysis in Clinical Trials. 2019. ich.org
  2. US Food and Drug Administration. Real-World Data: Assessing Electronic Health Records and Medical Claims Data to Support Regulatory Decision-Making for Drug and Biological Products. 2024. fda.gov
  3. Andrade SE, Kahler KH, Frech F, Chan KA. Methods for evaluation of medication adherence and persistence using automated databases. Pharmacoepidemiol Drug Saf. 2006;15(8):565-574. doi.org/10.1002/pds.1230
  4. Cramer JA, Roy A, Burrell A, et al. Medication compliance and persistence: terminology and definitions. Value Health. 2008;11(1):44-47. doi.org/10.1111/j.1524-4733.2007.00213.x
  5. RWEdnesdays. Treatment Failure and Non-Response in RWE. rwednesdays.com
  6. RWEdnesdays. Treatment Failure Endpoint Checklist. rwednesdays.com