The Category Difference

The NCCN breast cancer guidelines, version 6.2026, cover first-line treatment of hormone-receptor-positive, HER2-negative advanced breast cancer. For an aromatase inhibitor combined with a CDK4/6 inhibitor, the guideline assigns ribociclib category 1. The same combination with abemaciclib or palbociclib is category 2A.

Category 1 means high-level evidence and uniform panel consensus. Category 2A means lower-level evidence with uniform consensus. The guideline states the basis for the difference: the overall survival benefit seen with ribociclib.

That is one of two systems NCCN runs, and it is not the one that says which drug to prefer. The Categories of Evidence and Consensus — the 1 and the 2A above — record how strong the evidence is that an intervention is appropriate. A separate set, the Categories of Preference, records the panel’s assessment of a regimen’s value — Preferred being defined as based on superior efficacy, safety, and evidence taken together. It is not an ordinal ranking and the guideline does not present it as one. On that second system the guideline lists all three aromatase-inhibitor-plus-CDK4/6-inhibitor combinations as preferred first-line options — for postmenopausal patients and for premenopausal patients on ovarian suppression alike — and explains the evidence grade elsewhere, in the categories-of-evidence discussion rather than beside the preference listing: Due to the clear OS benefit seen with ribociclib in combination with AI, it is a category 1 recommendation. Abemaciclib and palbociclib are category 2A. The preference designation does not differ across the three. Only the category of evidence does.

And there the guideline stops. We asked whether it anywhere tells a reader how to read a category 1 sitting beside a category 2A when all three regimens are equally preferred — whether it cautions, in the discussion, the principles section, a footnote or a panel comment, against taking the higher evidence grade for the better drug. It does not. The tables print the preference tier and the evidence category side by side and offer no interpretive guidance about what the difference between them means. That is not a criticism of the guideline, which is a reference document and not a reading primer. It is the answer to why this misreading is so easy: the document that draws the distinction carefully is silent on the one question a reader arrives with, and the silence is filled by whoever is reading.

The version of this assessment published on 28 August did not say that. It reported the category difference, asked what a reader should make of it, and never mentioned that the guideline itself had already answered the question elsewhere in the same document. That was a selective omission: the one fact most likely to change a reader’s conclusion, left out of a page built to criticise how the rest of the document reads. It is corrected here and recorded below. It is also, exactly, the failure this assessment is about — an evidence label read as a ranking of drugs, this time by us, after eighteen automated checks and two reviews on the question of whether it was one.

The distinction is specific to that setting, and it is worth being precise about. In combination with fulvestrant rather than an aromatase inhibitor, both ribociclib and abemaciclib are category 1, and palbociclib is not. So the claim is narrow: first-line, with an aromatase inhibitor, ribociclib alone carries category 1.

That second setting is worth pausing on, because the guideline gives the same reason for it: ribociclib and abemaciclib with fulvestrant are category 1 because those combinations showed an overall survival benefit, in MONALEESA-3 and MONARCH 2. Which means abemaciclib is a category 1 drug in one setting and a category 2A drug in another — on the same evidence standard, consistently applied. The grade is not a verdict on the molecule. It is a record of which of its trials produced a significant survival result, and the guideline says so plainly.

The guideline also records that the three drugs have not been directly compared in clinical trials.

On the endpoint every one of these trials was powered to measure, nothing separates them. Among the three postmenopausal aromatase-inhibitor trials — the three the guideline lists side by side — the progression-free hazard ratios are 0.54, 0.57 and 0.58. MONALEESA-7, in a different population, reports 0.55. Nor does the choice of readout rescue a difference. Across the investigator-assessed progression-free analyses these four trials have published between them — primary, updated and final, eight readouts, all named in the sources below — every hazard ratio falls between 0.52 and 0.58. That is a claim about the eight we checked, not about every analysis ever run on these datasets.

That comparison is across trials, and a comparison across trials is weak evidence whichever endpoint it uses. It is offered here for exactly what it is worth, which is the same as what the overall-survival comparison below is worth: these trials were never designed to be read against each other.

The guideline does not make that mistake, and it is worth being exact about what it does instead. It grades each regimen on its own trial, against that trial’s own significance threshold — a per-trial test, not a comparison of effect sizes. Then it prints the resulting grades in one table, in one column, beside each other. Nobody performed a cross-trial comparison; the format performs one on the reader’s behalf, and a category 1 sitting next to a category 2A is read as a ranking of the drugs whether or not it was meant as one.

The grade turns instead on overall survival, and the three trials did not design that analysis the same way. PALOMA-2 powered for it: its final survival analysis was planned after at least 390 events, with 80% power to detect a hazard ratio of 0.74 or better at a one-sided 0.025. MONARCH 3 did not: survival there was a gated secondary endpoint, its final analysis planned at about 315 events with alpha split between the whole population and a visceral-disease subgroup. So palbociclib's null result comes from an analysis built to find an effect of that size, and abemaciclib's near-miss comes from one that was not. An earlier version of this page said none of these trials was powered for survival, which is false and flattened exactly this difference; an outside reviewer caught it.

Survival was a secondary endpoint in all three — which, as the paragraph above shows, covers both a powered analysis and an unpowered one. That is not a comment on how much survival matters — it is the outcome that matters most, and this page is not going to imply otherwise while criticising others for the reverse. It is a statement about statistical power. A trial sized to detect a difference in progression is not necessarily sized to detect one in survival, so a survival result from such a trial can fail to reach significance while the underlying benefit is real, and can reach it while the benefit is smaller than the point estimate suggests. That is the machinery the grade is resting on.

Palbociclib is a different case, and the difference is real. A hazard ratio of 0.96 on an interval from 0.78 to 1.18 is not a near miss. It is a result compatible with a meaningful benefit at one end and with harm at the other, centred almost exactly on no effect — which is not the same as proof that the drug does nothing, and this page has just spent a paragraph saying so about the other two. What PALOMA-2 established is that it did not demonstrate a survival benefit. That is all it established, and it is less than “palbociclib does not work”. PALOMA-2 is also the trial whose investigators recorded a large and disproportionate imbalance in missing survival data between its arms — survival status was unknown for 13.3% of the palbociclib arm and 21.2% of the placebo arm — which they said limits interpretation. They then did something about it. A sensitivity analysis using recovered data cut the unknowns to 9.2% and 11.7% and moved the hazard ratio from 0.96 to 0.92 (0.76–1.12). Four hundredths — and this piece has spent several hundred words on what four hundredths does and does not establish, so it will not now call that no movement. What did not change is the conclusion: the interval still runs through 1.0 and the investigators still reported no survival benefit. The defect is real, and their own remedy points where their headline number already pointed. A different number, from a trial with a defect its own authors flagged, is still not a measurement of a different drug — but this piece is not claiming palbociclib’s result sits where abemaciclib’s does. It does not.

A hazard ratio of 0.80 does not mean 20% fewer deaths. It does not mean a 20% better chance of anything, and it is not a probability that applies to a person. It is a ratio of rates: over the course of follow-up, among patients still alive at each moment, deaths occurred in the treated group at about four-fifths the rate of the comparison group. It is an average of that ratio across years of follow-up, and if the ratio changed over time — which it usually does — the single number conceals the change.

What a person wants to know is how much longer. The trials answer that too, in months, and the answer is worth setting beside the grade.

Two warnings about that table, both of which apply to us and not only to other people. These are medians, not lifespans: a median difference of 13.1 months is the gap between the times by which half of each group had died, and no individual is promised anything by it. And comparing medians across trials is a cross-trial comparison, with everything that costs — different patients, different follow-up, different trial sizes. That is not what this page criticises the guideline for; the guideline grades each trial on its own threshold. It is a weakness of this table, owned here, and the table is offered as an illustration of scale rather than a comparison of drugs.

Palbociclib is the case where the absolute and relative pictures agree. 2.7 months of median difference, a hazard ratio of 0.96, an interval running through the no-effect line, and a sensitivity analysis that moved the estimate to 0.92 without changing the conclusion. Whatever is true of the other two, PALOMA-2 did not show a survival benefit.

No randomised trial has tested one of these drugs against another. Several research programmes have compared them indirectly or observationally. This page examines four of them. It does not claim to have the whole literature — an outside reviewer found two more that we had missed, one of which is now the fourth below.

A network meta-analysis (Scientific Reports, February 2024) pooled seven phase III randomised trials, 4,415 patients, at a median follow-up of 73.3 months. It found no statistically significant difference in overall survival between any pair of the three. Its MONARCH 3 input is the final overall-survival result — HR 0.804 (0.637–1.015), at 97.2 months, the figure given above — not the earlier interim one.

P-VERIFY took 9,146 patients from US oncology records and weighted them by inverse probability of treatment. On overall survival it found nothing anywhere: ribociclib versus palbociclib 0.98 (0.87–1.10, P = 0.75), abemaciclib versus palbociclib 0.95 (0.84–1.08, P = 0.43), abemaciclib versus ribociclib 0.97 (0.82–1.14, P = 0.70). A second paper from the same cohort found nothing on progression either. Its arms are very uneven — 6,831 patients on palbociclib, 1,279 on ribociclib, 1,036 on abemaciclib — which is what prescribing actually looked like and is also a limit on what the smaller arms can show. For the comparison this piece turns on, abemaciclib against ribociclib, the two arms are of similar size and the imbalance is not the issue; it is their size against palbociclib’s that limits the other two comparisons.

One thing about it a reader should have. Its full name is the Palbociclib Verifying Evidence of Real-world Impact study, and its own poster states that it was funded by Pfizer Inc., which makes palbociclib. Industry-funded real-world studies are ordinary, they are held to the same methodological standards as any other, and this does not make its findings wrong. The fact is recorded here and no inference is drawn from it.

And the same disclosure has to run the other way, or it is not a disclosure but a point-scoring device. We could not open PALMARES-2’s declaration of interests. That is a statement about our access and not about its investigators, of whom we know nothing and about whom this page will not speculate. What follows from it is only this: we have read one study’s funding and not the other’s, so neither is offered here as the disinterested one.

PALMARES-2 (Annals of Oncology, April 2025) followed 1,982 patients across eighteen Italian centres and weighted them the same way. On progression-free survival it found abemaciclib ahead of palbociclib at an adjusted hazard ratio of 0.76 (0.63–0.92), p = 0.004, and ribociclib ahead of palbociclib at 0.83 (0.73–0.95), p = 0.007. It also reports overall survival, as an exploratory endpoint, pointing the same way. This page does not print those survival figures. The secondary summaries we could reach disagree with each other about which numbers are the survival ones and which are the progression ones, the paper itself is behind a wall we could not open, and a figure we cannot attribute to a primary source is a figure this page does not carry. What the authors do say about that endpoint is that the data are immature — 464 events at a median 31.3 months — and that follow-up is badly uneven between the arms: 45.7 months on palbociclib against 25.2 on ribociclib and 22.4 on abemaciclib. That imbalance cuts against their own finding, and it is their caveat, not ours.

A fourth, an indirect comparison from reconstructed patient data (Cancers, 2023), rebuilt patient-level survival curves from PALOMA-2, MONALEESA-2 and MONARCH 3 — 1,827 patients — and estimated every pairwise difference directly. On overall survival it separates nobody: abemaciclib versus ribociclib 0.933 (0.753–1.157), p = 0.528. On progression-free survival its one-stage model puts abemaciclib against ribociclib at 0.722 (0.520–1.002), p = 0.051 — the closest any study has come to separating those two, and it does not, while its two-stage model on the same data gives 0.921 (0.597–1.420), p = 0.710. That borderline figure belongs in this piece precisely because it is the one number that cuts against its conclusion.

None of these is a randomised head-to-head trial and none should be read as one. Together they are the best comparative evidence in existence, and on palbociclib they contradict each other. The larger cohort, on both survival and progression, finds nothing. The smaller one finds palbociclib behind on progression, on intervals that exclude the no-effect line — abemaciclib’s by 0.08 and ribociclib’s by 0.05, which is significant but not comfortably so, and reports an exploratory survival result pointing the same way on data its own authors call immature. An earlier version of this page reconciled them by saying they had measured different endpoints. That was wrong — both programmes reported both endpoints — and it was the tidier story rather than the true one. Whether palbociclib is behind the other two is unsettled, and this page does not settle it.

They do not contradict the guideline’s own statement, and it is worth saying so plainly. In guideline language, “directly compared in clinical trials” means a randomised head-to-head trial, and none exists. The Flatiron study says as much itself: it opens by describing its own purpose as working in the absence of randomised trials that directly compare the three. The guideline is accurate as written. What these comparisons add is not a correction to it, but the only evidence there is on the question it records as untested.

An earlier version of this page said cardiac monitoring was “the one clear difference” between the drugs, and that it ran the other way from the grade. That was selective. Each of the three carries its own labelled warning, and reading the labels side by side is the only fair way to do this.

Ribociclib is the only one requiring cardiac monitoring: an ECG before treatment and again at about day 14 of the first cycle. Its label also requires serum electrolytes “prior to the initiation” and at the beginning of the first six cycles, liver function tests every two weeks for the first two cycles, and blood counts on the same schedule.

Abemaciclib carries a labelled warning for venous thromboembolism — thrombotic events in 2% to 5% of patients treated across the trials the label reports, which include the early-breast-cancer trial as well as the two metastatic ones, with monitoring for signs and symptoms — and diarrhoea in 81% to 90% of patients across four trials, grade 3 in 8% to 20%, with patients instructed to begin antidiarrhoeal treatment at the first loose stool.

Palbociclib carries neutropenia: 80% of patients in PALOMA-2, grade 3 or worse in 66%, with blood counts required on day 15 of each of the first two cycles as well as at the start of every cycle.

These are three different experiences of taking a drug, and they do not line up with the grade in any direction. The category 1 drug is the one with the cardiac requirement. One category 2A drug carries a thrombosis warning and the other a two-in-three rate of grade 3 neutropenia. The point is not that one of them is worse. It is that the category records which trial produced a significant survival result, and the burden a patient actually carries sits in a different column entirely — one the grade does not look at.

This piece is not arguing that the three drugs are identical. They are not. It is arguing that the evidence does not establish that one of them works better than another.

If you are taking one of these drugs, or choosing between them, the question underneath all of this is whether a category 2A drug is a worse drug. There are two answers and they point the same way. The guideline’s own answer is no: it lists all three as preferred first-line options, and the category records how strong the evidence behind each one is, not which drug works better. The evidence’s answer is that nothing has been shown to separate them: no randomised trial has compared them, and none of the four comparative studies on this page can tell abemaciclib and ribociclib apart.

Which leaves the question of why a reader would have thought otherwise. Not because the guideline says so — it does not. Because a category 1 and a category 2A printed in one column look like a ranking, and the preference designation that would correct the impression sits somewhere else in the document. This page made that mistake in its first published version, which is the most direct evidence it can offer that the mistake is easy to make.

That is a statement about evidence and not a recommendation. This assessment does not say what anyone should take. Prescribing decisions turn on toxicity, monitoring burden, comorbidities, interactions and cost, none of which a page like this can see, and all of which your oncologist can.