When Tumor Shrinkage Doesn't Mean Living Longer

In 2019, FDA granted accelerated approval to voxelotor for sickle cell disease based on a surrogate endpoint: hemoglobin increase. The mechanism was sound. Voxelotor increased hemoglobin oxygen affinity, which should reduce sickling and improve outcomes. The biology made sense. The FDA agreed. Patients got access.

In September 2024, the drug was withdrawn globally after post-marketing data showed higher rates of death and serious complications in treated patients compared to standard care. Hemoglobin went up. Patients did worse.

I've written about the surrogate trap before: why biologically plausible mechanisms fail to translate into clinical benefit. Voxelotor fits the pattern exactly. But the broader question is not whether surrogates sometimes fail. Of course they do. The question is whether the accelerated approval pathway, which has now been used for roughly 170 oncology indications since 1992, handles those failures well enough to justify the access it provides.

The answer depends on which cases you look at.

The pathway working as designed

Tarlatamab for extensive-stage small cell lung cancer is the best recent example of accelerated approval doing exactly what it was designed to do. FDA granted approval in 2024 based on objective response rate. Twelve months later, the approval converted to traditional based on confirmatory data showing overall survival benefit: 13.6 months with tarlatamab versus 8.3 months with chemotherapy.

Patients who received tarlatamab under accelerated approval gained access roughly a year before survival data matured. The surrogate predicted correctly. Tumor shrinkage translated to living longer.

This pattern repeats. Epcoritamab converted in 2025 after confirmatory trials verified benefit. The median time from accelerated to traditional approval has dropped from 4.3 years before 2014 to 2.3 years recently, partly because FDA now expects confirmatory trials to be underway at the time of approval, not merely planned.

Across the full record, about 61% of accelerated approvals have converted to traditional approval. Another 19% have been withdrawn after confirmatory trials failed or revealed safety concerns. The remaining 20% are still under review. A 61% conversion rate is substantial for a pathway explicitly designed to trade early uncertainty for speed. But the 19% withdrawal rate is not a rounding error. It represents patients who received drugs that did not ultimately demonstrate clinical benefit, and in some cases, like voxelotor, caused harm.

(The underlying data comes from recent systematic analyses by Tibau et al., Liu et al., and Jenei et al..)

The cases that don't fit the binary

The interesting regulatory question is not what happens when surrogates clearly succeed or clearly fail. It is what happens in between.

Lurbinectedin for refractory metastatic small cell lung cancer received accelerated approval in 2020 based on objective response rate. The confirmatory trial, ATLANTIS, did not meet its primary endpoint of overall survival. In a traditional framework, that would end the conversation. The trial failed. Withdraw the drug.

FDA maintained the approval. The trial showed no survival detriment, and few treatment options existed for this population. Alternative confirmatory trials were allowed to proceed. In a November 2024 commentary, FDA's Gautam Mehta and Richard Pazdur articulated the logic: "A failed confirmatory trial after accelerated approval in oncology does not always equate to a failed drug."

That statement deserves more attention than it has received. It reframes accelerated approval as a provisional state rather than a binary switch. And it acknowledges something that the approval/withdrawal framework obscures: trials can fail for reasons that have nothing to do with whether the drug works.

Gefitinib is the clearest example. It received accelerated approval for non-small cell lung cancer in 2003. Five confirmatory trials in unselected populations failed. The approval was withdrawn in 2011. Then, in 2015, gefitinib was granted traditional approval for patients with EGFR mutations. The drug worked. The original trials enrolled the wrong patients.

Accelerated approval did not identify an ineffective drug. It exposed flawed early development assumptions. The surrogate was not wrong. The population was.

Where the design decisions matter

The factors that predict whether an accelerated approval will convert are visible before the first patient is enrolled. This is what makes the pathway a design problem, not just a regulatory one.

Surrogate choice matters enormously. Objective response rate has a reasonable track record across many tumor types. Novel surrogates, like hemoglobin increase in sickle cell disease, require extraordinarily strong mechanistic justification, and voxelotor shows that even apparently strong justification can be wrong. The question is not whether the surrogate moves. It is whether its movement predicts something patients care about.

Effect size matters. Marginal response rates in the 15 to 20% range rarely sustain accelerated approval in competitive landscapes. Response rates above 40% strengthen the case considerably, not because a threshold exists, but because larger effects are harder to explain away when confirmatory data arrive.

Confirmatory trial readiness matters. If you approach five years without confirmatory data, expect scrutiny. The FDORA provisions now give FDA explicit authority to require that confirmatory trials be underway at the time of accelerated approval and to initiate withdrawal proceedings if sponsors fail to conduct them with due diligence.

The hardest design decision is what to do when you are in the complicated middle: the drug has activity, the confirmatory trial did not meet its endpoint, and the question is whether the trial failed or the drug failed. Gefitinib's decade-long arc from accelerated approval to withdrawal to traditional approval in a biomarker-selected population is a reminder that these questions sometimes take longer to resolve than anyone wants.

The cost of getting it wrong in both directions

Accelerated approval encodes a specific asymmetry: the cost of withholding a potentially effective therapy from seriously ill patients is judged to be higher than the cost of provisional access to a drug that might not ultimately verify. That asymmetry is defensible in oncology, where many patients have no alternatives and limited time. It is less defensible when the confirmatory evidence is slow to arrive, the withdrawal process is slow to engage, or the surrogate is weaker than anyone acknowledged at the time of approval.

Voxelotor is the cautionary case. Tarlatamab is the success case. Gefitinib is the case that resists easy categorization. The accelerated approval pathway will be judged not by how it handles the clear wins and clear failures, but by how it handles the drugs that sit between them.

The record so far suggests the pathway works when surrogates are validated, effects are substantial, and sponsors complete verification promptly. When those conditions are not met, it becomes a gamble that can take years to resolve, during which patients are exposed to drugs whose clinical benefit remains unproven.

Acknowledging that trade-off clearly, before the first patient enrolls, is the most important design decision a sponsor can make.

This analysis builds on the error asymmetry framework I explored in a previous post. For a deeper look at how surrogate failures trace back to mechanistic overconfidence, see The Surrogate Trap.

📬 For more essays on experimental design, regulatory evidence, and statistical decision-making across domains, subscribe to the Evidence in the Wild newsletter or browse the archive.