Sources verified: October 28, 2026
In the previous article, When a New Drug “Works,” What Has Actually Been Proved?, we moved from the streptomycin trial to tumor shrinkage, survival time, independent review, and the hazard ratio, or HR. We examined what studies observe, how outcomes are assessed, and how results in two groups are compared.
We also left a question open: if the final results take years, can patients gain access to treatment sooner?
Tests can now look for tiny numbers of residual tumor cells in bone marrow or faint signals of tumor DNA in blood. These advances are changing clinical research. Yet detecting a change earlier still leaves an evidentiary distance to confirming a real benefit to patients.
In this article, we follow that line further, examining how research makes judgments amid time constraints, uncertainty, and the costs patients bear.
1. Why approve first and continue verification afterward?
In 1992, the FDA established regulations for accelerated approval. These allow some medicines for serious diseases that address unmet medical needs to receive earlier approval on evidence such as a surrogate endpoint reasonably likely to predict clinical benefit. Studies must still be completed after approval to verify the anticipated benefit.[1]
A surrogate endpoint uses a change measurable earlier to help predict a later benefit to patients. It requires a scientific basis. Evidence from disease mechanisms, previous follow-up, and clinical trials can help explain why the measure may predict clinical benefit. The FDA must also consider the particular disease, population, and treatment setting to assess whether it is appropriate for that study.[2]
Our earlier article, When Urine Protein Falls, Are the Kidneys Protected?, addressed this very question.
In IgA nephropathy, a reduction in urine protein can be measured directly. Existing follow-up studies and pooled analyses of randomized trials support its use as a surrogate endpoint reasonably likely to predict long-term kidney benefit. For an individual medicine, however, subsequent research must still verify whether reduced urine protein translates into slower loss of kidney function.[3]
If later studies fail to confirm the anticipated benefit, or if the benefit is insufficient to outweigh the risks, the FDA can begin procedures such as withdrawing approval of the relevant indication. Accelerated approval offers an earlier treatment opportunity while leaving questions that still need answers.[1]
2. Making patients' lives part of the evidence
The previous article noted that a CT scan can show stable disease while a patient continues to experience treatment-related nausea and fatigue. Imaging cannot answer for patients about these experiences.
Patients also care about whether they can eat, sleep, work, and go out with their family, as well as how much discomfort, monitoring, and travel for care treatment brings.
In 2024, the FDA issued final guidance on core patient-reported outcomes in cancer clinical trials. It recommends systematically collecting information on disease-related symptoms, symptomatic adverse events, the overall impact of side effects, physical function, and role function in daily life.[4]
For example, a patient who previously shopped and cooked independently may need family members to provide transport and care after starting a medicine. Even if tumor measures improve, these changes should be recorded and included in the benefit–risk assessment.
Questionnaires also require sound methods. Whether the instrument is suitable, when questions are asked, and whether sicker patients are more likely to leave them unanswered can all affect results. We need to know whether the people who provide answers adequately represent the patients who were supposed to be observed.
3. After patients stop or switch treatment, are we asking the same question?
Start with one particular patient. Suppose a lung cancer trial compares a new drug with standard treatment and measures how long patients live. This patient is initially randomized to the control group and receives standard treatment. Later, the disease progresses. The patient stops the original treatment and switches to the new drug being studied.
We can ask two different questions about this experience.
First: starting from the initial treatment choice, what actually happens along the way?
Since the patient has switched treatment, researchers continue following them and record how long they live afterward. When comparing the groups at the end, they still count this patient in the original control group, rather than moving them into the new-drug group because they later received it.
This question does not require patients to remain on their original medicine. It includes the discontinuations and switches that actually occur. It asks: starting with a choice of the new drug or standard treatment, and allowing for subsequent treatment changes, how do the groups' eventual survival outcomes differ?
Second: what would have happened if this patient had not switched to the new drug?
We are still concerned with the same patient, but now want to know what would have happened after progression if they had received subsequent usual care without this new drug. Extended to the trial's control group, the question becomes: how different would survival in the two groups have been if no control-group patients had switched to the new drug?
But the patient has actually switched. We cannot make the same person live through a second “no-switch” course, so we cannot directly observe their outcome along that other treatment path.
Could we simply remove patients who switched and examine those who did not?
That simple approach cannot answer the second question. The patients left behind may be precisely those whose disease is better controlled and who do not need to switch. Their outcomes cannot directly represent what would have happened without switching to a patient whose disease has progressed and who needs another treatment. The ability to switch may also depend on physical condition, treatment access, and other factors. Removing those who switch can change the composition of the groups, undermine the comparability established by randomization, and bias the estimated treatment effect.
This does not make the second question impossible to study. Using information on disease, treatment, and follow-up before and after switching, researchers can apply prespecified methods and explicit assumptions to estimate the between-group treatment effect under a “no switch to the new drug” scenario. But that estimate cannot be treated as a directly observed fact or an exact reconstruction of each patient's alternative life. Researchers also need to test whether the conclusion holds under other reasonable assumptions.[5]
The ICH E9(R1) guideline, finalized in 2019, emphasizes precisely this sequence: clarify which treatment question is being asked, then arrange data collection and analysis accordingly. Including treatment switches that actually occur and estimating what would happen without switching answer different questions and must not be conflated.[5]
4. Failing to detect a difference does not mean ruling out risk
Suppose a cancer-drug trial has 200 patients in each group. Follow-up has reached only three months, with three deaths in the control group and four in the new-drug group. Does the one additional death mean the new drug increases the risk of death? These numbers alone cannot establish that.
Even if the two treatments have identical true effects, the death counts are unlikely to be exactly equal every time. Differences in patients' conditions, the timing of deaths, and random variation can produce a difference of one or two cases. The additional death must be recorded carefully and investigated alongside its cause, timing, and other safety information, but cannot be attributed directly to the new drug.
The reverse is also true. If there are three deaths in the control group and only one with the new drug, we still cannot conclude that the drug is necessarily safer or prolongs life. With only a handful of deaths overall, the apparent difference may be unstable; some treatment-related harms may also emerge later. A statistical analysis that “finds no significant difference” therefore establishes neither that the treatments are identical nor that a detrimental effect on survival has been excluded.
How many months of follow-up are enough? There is no single number for every cancer. In rapidly progressing advanced disease, many deaths may accumulate over a short period. After surgery for a disease with a lower recurrence risk and longer survival, years of observation may be needed. Three months is not worthless in every study, but information is clearly limited in this example with only a few deaths. Longer follow-up is not the only consideration: enrollment, accumulated deaths, completeness of follow-up, and whether observation covers the period in which delayed harm could emerge all affect judgment.
The key is how much these data narrow the uncertainty. A study reports not only a risk estimate but also a range expressing its precision, usually called a confidence interval. When events are few, that range is often wide. The data may allow both “the new drug is beneficial” and “the new drug causes substantial harm.” Failure to distinguish a difference then mainly tells us that we cannot yet separate these possibilities.
Consider a separate teaching example. Suppose a study specifies in advance that it will assess whether the new drug increases the proportion dying within two years by five percentage points—approximately five additional deaths per 100 patients. This number illustrates the method; it is not a universal safety threshold for cancer drugs. If observed death proportions are similar but the confidence interval's upper bound still permits an increase of eight percentage points, the harm of concern has not been excluded. Only if analysis using the prespecified method places that upper bound below five percentage points is there a basis to say that, within the specified statistical uncertainty, the data exclude an increase of that magnitude. This still does not mean there is no risk at all, or exclude a smaller increase or effects beyond two years.
“Enough information” should therefore be planned around a clear question established beforehand: what magnitude of survival harm must be detected or excluded, and how many patients, deaths, months of follow-up, and how precise an estimate will that require? These standards should reflect the disease, treatment setting, and expected benefit; they cannot be chosen after seeing the results. An undersized trial may still fail to exclude important risks even after follow-up ends. If concerning safety signals emerge during the study, researchers cannot wait merely because the planned analysis time has not arrived.[6]
The FDA's 2025 draft guidance on overall survival, or OS, specifically discusses planning survival safety assessments when OS is not the primary endpoint, including information requirements and precision. At this section's current verification date, the document remained a draft.[6]
Keeping disease controlled for longer is an important outcome. For example, after treatment a tumor may shrink or remain stable, then start growing again or develop new metastases later. If the new medicine delays the point at which patients meet the study's progression criteria compared with control treatment, it has extended this period of disease control. This does not require a tumor to keep growing slowly throughout; it may first shrink, remain stable for an extended period, and only later progress again.
If progression eventually occurs, does that erase the value of the earlier treatment? It does not. For some cancers that are currently difficult to cure, an important aim is to control disease for as long as possible. In a simplified example, neither of two treatments ultimately prevents progression, but one controls disease for about six months and the other for about a year. Although both eventually see progression, patients' experiences are different: the second provides additional time before disease progresses. Later progression does not erase that earlier period.
The value of that time, however, also depends on how patients spend it. If disease control comes with less pain, preserved daily activity, or a later start to another burdensome treatment, those may be tangible benefits. But if patients endure severe side effects throughout, overall benefit may be limited despite delayed progression on imaging. Symptoms, quality of life, and time to the next treatment require their own supporting data; “delayed progression” alone does not establish that all have improved.
We should therefore distinguish whether treatment extends disease control, how patients live during that time, and whether they ultimately live longer. Delaying progression can have value, but extra disease-control time cannot be converted directly into the same amount of added life.
If a study reports progression-free survival, death must also be considered. Only time during which a patient has neither met progression criteria nor died counts as “progression-free survival.” Even if this measure improves, whether overall survival is extended and whether other treatment-related harms exist still require assessment alongside subsequent evidence.
A safety conclusion should therefore specify the population, the duration of observation, and the magnitude of risk excluded, rather than broadly declaring that a treatment has been “proved safe.”
5. Should eligibility criteria be stricter or broader?
Patients in a study also need to resemble, as closely as possible, the people who will actually use the medicine.
Suppose a lung cancer trial recruits only patients in relatively good physical condition who can perform most daily activities. There will then be limited direct evidence about efficacy and side effects in patients with the same cancer who can look after themselves but can no longer work and have markedly reduced activity.
During early exploration, when safety and appropriate dosing remain unclear, scientifically justified eligibility restrictions can help manage risk. As safety data accumulate, however, those restrictions should be reassessed, especially in later trials intended to verify efficacy.
Appropriately broader criteria allow more patients to participate, may accelerate recruitment, and may make results more applicable to the eventual treatment population. The cost is that some patients may find treatment or follow-up harder to tolerate or complete. Monitoring, follow-up support, and analyses of different populations must be planned in advance. Dedicated study cohorts can be prespecified where necessary to accumulate evidence gradually.[7]
Eligibility should therefore evolve with the evidence. Within the limits of safety and the study's purpose, research should include people likely to receive the medicine in the future; studying these populations need not always wait until after approval.
The FDA's related final guidance, issued in July 2026, likewise recommends adjusting performance-status eligibility based on existing safety evidence so that later trials better reflect the intended treatment population.[7]
Another distinction matters: broadening study eligibility does not automatically broaden an indication. Adding data on less physically fit patients with the same lung cancer may improve information on the treatment population, safety, or dosing. Expanding the disease, treatment stage, or population covered by an existing approval, however, requires appropriate evidence and regulatory review. Whether an indication actually changes depends on the original approval's scope.[8]
6. Why does finding no residual cells in bone marrow matter?
More sensitive tests are enabling researchers to see changes that were previously difficult to distinguish.
Multiple myeloma is a malignancy of plasma cells, with tumor cells commonly found in bone marrow. Even after a complete response by conventional criteria, small numbers of tumor cells may remain. These cells can be a source of later disease progression.[9]
Consider a hypothetical example: two patients both reach a complete response. More sensitive bone-marrow testing still detects residual tumor cells in one. In the other, a test capable of identifying one tumor cell among 100,000 normal cells finds no residual disease. Under the relevant criteria, the latter can be classified as negative for minimal residual disease, or MRD-negative.[9]
The conventional results appear similar, but the depth of response differs.
Previous studies have found that patients who become MRD-negative tend, overall, to have longer progression-free survival. Pooled analyses of randomized trials have also found an association between a treatment's improvement in the MRD-negative rate and its improvement in progression-free survival. This raises the possibility of using MRD as an earlier signal to help predict later disease control.[10]
As more patients achieve conventional responses, more sensitive assessment can also help distinguish treatment differences. In draft guidance issued in January 2026, the FDA discussed using the proportion of patients who achieve both a complete response and bone-marrow MRD negativity during a study to support accelerated approval in specified populations and study conditions. At this article's verification date, the document remained a draft and did not constitute blanket acceptance across all myeloma treatment settings.[9]
A negative result still has sampling and detection limits and cannot be equated with a cure. Subsequent progression, survival, and safety require continued observation. MRD is also studied and used in other diseases, but the relationship between test results and patient outcomes needs separate validation in each setting.[9,11]
7. What can tumor DNA in blood tell us?
Beyond bone-marrow testing, DNA in blood is another approach attracting attention.
Cells release DNA fragments into the bloodstream, forming cell-free DNA, or cfDNA. The portion originating from tumors is called circulating tumor DNA, or ctDNA. Normal cells also contribute cell-free DNA, so the total amount of cfDNA in blood cannot be treated directly as the amount of tumor.[12,13]
This also shows why MRD and ctDNA are not competing names. MRD describes residual disease; ctDNA can be a tool for detecting it at the molecular level. Bone-marrow testing is commonly used in myeloma, whereas tumor-DNA signals in blood can be investigated in some solid tumors.
For a hypothetical example, consider a patient with colon cancer whose scans show no recurrence after surgery. If blood still contains DNA features matching the original tumor, that may suggest residual disease not yet visible on imaging and help researchers identify a group at higher risk of recurrence. A negative test must also be interpreted in light of how much DNA the tumor releases, sampling time, and test sensitivity.[13]
In clinical trials, ctDNA can serve different purposes: identifying people more likely to need treatment, observing how signals change after therapy, or studying whether a reduction or clearance of the signal predicts long-term outcomes. Each use requires its own validation; they cannot substitute for one another.[14]
An actual example makes this distinction clearer.
On May 15, 2026, the FDA approved atezolizumab and its subcutaneous formulation as adjuvant treatment for specified adults with muscle-invasive bladder cancer: after cystectomy, these patients had ctDNA molecular residual disease detected by an FDA-authorized test. The FDA also approved the Signatera CDx companion diagnostic to identify such patients.[15]
The supporting IMvigor011 trial randomized eligible ctDNA-positive patients to treatment or placebo. Its primary outcome was disease-free survival—the time to recurrence or death—and it also assessed overall survival. The study reported benefits on both.[15,16]
In this example, ctDNA helped determine whom to study. The evidence establishing treatment value still included whether patients later experienced recurrence and how long they lived. “Using ctDNA to select patients” cannot be taken to mean “proving efficacy solely because ctDNA becomes negative.”
Researchers are also exploring combinations of DNA mutations, chemical markers such as methylation, and DNA-fragment lengths and breakage patterns to identify fainter tumor signals. As methods become more refined, it becomes all the more necessary to establish whether they identify the target consistently and accurately, and what those signals mean in a particular cancer and treatment setting.[14]
The FDA's related 2024 final guidance discusses several uses of ctDNA in drug development for early-stage solid tumors. It notes that using ctDNA changes as an earlier efficacy endpoint requires further randomized-trial evidence and evidence on long-term outcomes.[14]
New tools thus let us see changes sooner while moving the evidence question forward: once people at higher risk have been identified, can treatment guided by that signal actually improve their outcomes?
Returning to what patients really care about
Across these two articles, we have traveled from a randomized controlled trial in 1948 to today's residual-disease testing and circulating tumor DNA. Different tools solve different problems, and each has its limits.
The next time we see “clinical trial success,” we can keep asking: do the study participants resemble the people who will later use the medicine? Has the study established a change in a measure, delayed progression, longer life, or improvement in daily living? How large is the benefit, how long does it last, and what do patients bear in return?
Earlier signals need later verification. Patients' experiences need systematic recording. Design choices need to be made explicit before results are seen.
How clinical trials are designed determines which benefits can be seen, which costs are recorded, and what doctors and patients can rely on when making decisions later.
Before the Conclusion: there are still questions worth pursuing.
Medicine · Science · Ideas and Reading
Follow Before the Conclusion to stay curious and think independently together.
This article discusses medical and scientific knowledge and does not replace individual medical advice.
Principal sources
- FDA. Accelerated Approval. The 1992 background, postapproval verification, withdrawal of the relevant approval, or indication changes.
- FDA. Surrogate Endpoint Resources for Drug and Biologic Development.
- Thompson et al. Proteinuria Reduction as a Surrogate End Point in Trials of IgA Nephropathy, CJASN, 2019.
- FDA. Core Patient-Reported Outcomes in Cancer Clinical Trials, final guidance, 2024.
- ICH. E9(R1): Estimands and Sensitivity Analysis in Clinical Trials, final version, 2019.
- FDA. Approaches to Assessment of Overall Survival in Oncology Clinical Trials, draft guidance, August 2025.
- FDA. Cancer Clinical Trial Eligibility Criteria: Performance Status, final guidance, July 2026; Enhancing Participation in Clinical Trials, final guidance, 2025.
- FDA. Frequently Asked Questions about Labeling for Prescription Medicines.
- FDA. Minimal Residual Disease and Complete Response in Multiple Myeloma: Use as Endpoints to Support Accelerated Approval, draft guidance, January 2026; document-status page.
- Landgren et al. EVIDENCE meta-analysis, Blood, 2024. Distinguishes patient-level prognostic association from trial-level association between treatment effects.
- FDA. Hematologic Malignancies: Regulatory Considerations for Use of Minimal Residual Disease, final guidance, 2020.
- NCI. Circulating tumor DNA.
- FDA. Integrative liquid biopsy approaches to capture minimal residual disease. Used for the cfDNA/ctDNA distinction, test performance, and the research background on residual disease; a research project is not treated as completed clinical validation.
- FDA. Use of Circulating Tumor DNA for Curative-Intent Solid Tumor Drug Development, final guidance, November 2024. Particular attention was paid to the different roles of patient selection, signals of drug activity, earlier endpoints, and randomized-trial validation. Methylation and fragmentomics appear in the section on assay methods.
- FDA. May 15, 2026 announcement approving atezolizumab for a postoperative ctDNA-defined bladder cancer population.
- Powles et al. ctDNA-Guided Adjuvant Atezolizumab in Muscle-Invasive Bladder Cancer, NEJM. Original IMvigor011 study.
← Return to Part 1: When a New Drug “Works,” What Has Actually Been Proved?