Consider the situation an oncologist is actually in. A patient with newly diagnosed metastatic non-small-cell lung cancer is sitting in front of them, symptomatic, losing weight, frightened, and asking when treatment starts. Comprehensive molecular profiling has been sent. It will take two to three weeks.
Everyone in the room knows the argument for waiting. Everyone in the room is also looking at a patient who may not have two to three weeks of tolerance for waiting — and the data justify that fear. In a cohort of 4,389 patients with advanced non-squamous NSCLC, 17% died before receiving any first-line therapy at all.
So treatment starts. That decision is defensible, made in good faith, and — as we will see — not actually the problem.
The problem is what happens when the report comes back.
I.The information arrives, and the plan does not change
In a retrospective study of 964 patients with stage IV NSCLC in a community-based integrated health system, 52% initiated first-line treatment before NGS results were available. Among those who started early and were subsequently found to have a targetable alteration, 59% went on receiving chemotherapy and/or immunotherapy rather than switching to targeted therapy. The paper is titled, with unusual bluntness, Half of oncologists fail to use ordered NGS results to guide their first-line treatment decision.
The larger cohort tells the same story. Among biomarker-positive patients, 49.5% received non-targeted therapy despite a positive result. Only 31.4% received the appropriate targeted therapy.
The obvious defence is that switching mid-treatment is risky, or that these patients were too unwell. The evidence does not support it. In a separate analysis of 4,244 biomarker-positive patients, 73.2% started on a targeted agent and 8.2% started chemotherapy and switched within 56 days of the result. Their outcomes were indistinguishable.
- Duration of first-line therapy: 13.1 months (immediate targeted) vs 14.2 months (early switch)
- Time to treatment failure: 14.9 vs 15.6 months
- Overall survival: 35.2 vs 36.8 months
None of those differences was significant. The safe path exists, it is documented, and it costs the patient nothing. Roughly half of oncologists do not take it.
And there is a mortality signal attached. Among patients with targetable alterations, those who received guideline-concordant targeted therapy after their results lived a median of 24.6 months versus 14.7 months for those who received discordant non-targeted therapy. The adjusted hazard ratio for starting non-targeted therapy before results came back, rather than waiting, was 1.58.
II.The result that settles the argument
You could still argue that none of this is really about anchoring — that it reflects sicker patients, pragmatic judgement, unmeasured confounding. So consider a case where the sequence itself, and nothing else, is the variable.
In 126 patients with EGFR-mutant NSCLC treated with both PD-(L)1 blockade and an EGFR tyrosine kinase inhibitor, the order in which they were given determined whether patients were hospitalised.
This is the hardest available evidence that the first decision is not merely first. Give immunotherapy to a patient whose EGFR status you have not yet confirmed, and you have not simply missed an opportunity — you may have made the correct drug dangerous.
III.And then the door closes
The second-line plan assumes there will be a second line. Often there is not.
In real-world US data on KRAS G12C-mutant advanced NSCLC, fewer than 40% of patients received any second-line therapy — 39.4% in one database, 37.4% in another. Of those who did, roughly 30% received a KRAS G12C inhibitor, which made it the most commonly used second-line option.
IV.Why the first impression holds
The cognitive account is not mysterious, and it is not unique to oncology.
Start with how the patient is framed at the outset. In a prospective study of 468 terminally ill patients and 343 doctors, prognoses were accurate 20% of the time. 63% were overestimates, and physicians predicted survival 5.3 times longer than patients actually achieved. The finding that should give everyone pause is the modifier: the better the doctor knew the patient, the more likely they were to be wrong. This is not ignorance. It is attachment, and attachment is an anchor.
Aviation named the downstream effect decades ago and built its safety system around it. Plan continuation bias is the continuation of an original plan even when information suggests it should be abandoned — and it strengthens the closer you get to completing the plan. The initial motivation to finish is strong; mid-course signals to change are individually weak and get rationalised away, even as they accumulate into something serious.
Read the oncology version back: the patient started treatment, is tolerating it, has scans in six weeks. The report arrives. Each individual reason to switch is weaker than the momentum of the plan already running.
Aviation's response was not to teach pilots harder about weather. It was to define decision points in advance — stabilised approach criteria, mandatory callouts, go-around gates — so that continuing requires an active decision rather than passive drift. The bright line is set before the pressure arrives, at a moment when nobody is under it.
V.What this means for education
If the second-line decision is foreclosed in week one, then educating oncologists about second-line evidence is an intervention delivered after the decision has already been made. It is aimed at the right topic and the wrong moment.
Three things follow.
- The high-leverage target is the first decision, not the later one. Not which agent to give at progression, but what the clinician commits to doing when the report lands — decided before the report lands.
- The teachable moment is the return of the result, not the point of progression. That is the moment the plan should be reopened, and it is precisely the moment nothing currently prompts anyone to reopen it.
- Sequence deserves the emphasis that agent selection gets. The EGFR data show the order is not a detail downstream of the choice. In some cases the order is the choice.
And there is a fourth, uncomfortable implication for how these programmes are designed. Almost every later-line educational initiative is justified on the grounds that later-line decisions are difficult and the evidence is evolving. Both true. But if 60% of patients never reach that decision, and the ones who do arrive there carrying constraints set months earlier, then the later-line programme is teaching a decision that has already been substantially determined — to an audience that will apply it to a minority of their patients.
VI.The honest caveats
- Attrition is partly biology, not only decision-making
- Patients who deteriorate fastest both start treatment soonest and are least likely to reach second line. Confounding by indication runs through all of this, and it means the attrition figures cannot be read purely as missed opportunity.
- The switching evidence shows equivalence, not licence
- The 56-day analysis says an early switch does not appear to cost outcomes. It does not say switching is always right, or that it works at any interval.
- The sequence finding rests on small numbers
- Six severe events in 41 patients, and five in 21 for the short-interval group. The signal is striking and biologically coherent, and it is still a small retrospective series. Treat it as a strong warning rather than a settled rate.
- Uptake data reflect a moving target
- KRAS G12C inhibitors were newly available across much of the period these cohorts cover. Current practice may look better — though the attrition that limits access will not have changed.
- And we are not neutral
- We design education delivered at the point of care, so we came to this question already holding a view. A reason to check our reasoning rather than dismiss it.
VII.The short version
An oncologist under time pressure starts treatment before the molecular report returns. That is reasonable, and the evidence says an early switch would cost the patient nothing. Then the report arrives, and about half the time the plan does not change — not because switching is dangerous, but because plans in motion resist being reopened.
By the time the later-line decision arrives, a substantial share of patients are not there to receive it, and some of those who are have been given a first-line therapy that makes the right next drug more hazardous.
So the leverage is not where the field keeps putting it. If we want better later-line decisions, the intervention has to reach the clinician in the first week — at the moment the report lands, with a reason to stop and reopen a plan that is already running.