ATOMIC: hazard ratio and absolute benefit at three years

Why this news deserves a closer look

On 8 October 2026, the FDA approved atezolizumab with chemotherapy for adjuvant treatment of stage III dMMR colon cancer. The current news is regulatory; the results discussed here come from the ATOMIC paper published in The New England Journal of Medicine in March 2026. Approval and scientific publication are separate milestones.

Adjuvant treatment is given after surgery with the aim of reducing the chance that disease will return. Time to recurrence therefore plays a central role. When treating people after tumour resection, we need to understand both the expected benefit and the adverse effects and treatment burden involved in achieving it.

The population is equally important. dMMR means that the DNA mismatch repair system is deficient. ATOMIC studied this subgroup of stage III colon cancer after resection. Its findings should not automatically be extended to all colorectal cancers, other stages or mismatch repair-proficient tumours. In precision medicine, explaining whom the trial represents is part of explaining its results.

For a research team, this kind of news connects three tasks: understanding the evidence behind a regulatory decision, identifying the population it concerns and communicating the size of the effect. A hazard ratio alone addresses only part of that assessment. The practical question is how much the course of disease changes, over what period and at what cost in treatment burden.

What ATOMIC studied and which numbers belong together

ATOMIC was a phase III, randomised, open-label trial involving 712 participants. It compared atezolizumab plus mFOLFOX6 for six months, followed by atezolizumab to complete twelve months of treatment, with mFOLFOX6 for six months. The primary endpoint was disease-free survival; overall survival and adverse events were secondary endpoints. Median follow-up in the published analysis was 40.9 months. Trial design and results in the original article.

One way to express the difference is that approximately ten additional people per hundred would remain event-free at three years, according to these estimates and in a population comparable to the trial population. This translates estimated probabilities; it is neither an observed count in two groups of one hundred people nor an individual guarantee.

A median follow-up of 40.9 months also does not mean that every participant was observed for exactly that length of time. Recruitment dates and observation times vary. That variation is one reason survival methods are needed.

What a hazard ratio of 0.50 means

The hazard is the instantaneous event rate among people who have not yet experienced the event. For recurrence or death, the relevant group at each instant consists of people who remain alive and recurrence-free. The HR compares this rate between the two groups:

HR(t) = h_treatment(t) / h_comparator(t)

An HR of 0.50 expresses a lower instantaneous event rate in the treatment group: approximately half that in the comparator in the summary provided by the analysis. Interpreting this as a constant relationship throughout follow-up requires the proportional hazards assumption to be appropriate. The number alone does not mean that each person's probability of experiencing the event over the next three years is halved.

The distinction becomes clearer when we separate two questions. The hazard concerns event occurrence at that instant, among those still event-free. Cumulative probability concerns how many people will have experienced the event by a specified time. That probability depends on the entire preceding course and on the comparator group's baseline risk.

ATOMIC's estimates illustrate this. The complements of three-year disease-free survival are 100 − 86.3 = 13.7% and 100 − 76.2 = 23.8%. Their ratio is approximately 0.58; the relative reduction in that cumulative probability is approximately 42%. These are descriptive calculations from rounded percentages, without CIs calculated here. They differ from the 50% obtained by expressing 1 − HR because they refer to a different measure.

The absolute difference has the same magnitude either way: 23.8 − 13.7 = 10.1 percentage points lower event probability, or 86.3 − 76.2 = 10.1 points higher probability of remaining event-free. Keeping “at three years” in either sentence is essential.

The HR's 95% CI, from 0.35 to 0.73, describes uncertainty in that estimate under the analysis used. It is not an interval for three-year survival or for the 10.1-point difference. Nor can the CI for a difference be obtained by subtracting the endpoints of the two group-specific intervals: the difference needs its own uncertainty estimate.

There is a further nuance. The groups of people remaining event-free change over follow-up. Even in a randomised trial, an HR should therefore not be interpreted as a constant individual causal effect. This issue is discussed by Hernán in The Hazards of Hazard Ratios. Probabilities at relevant times help complete the interpretation.

What a Kaplan–Meier curve adds

Kaplan–Meier estimates the probability of remaining event-free over time. The curve steps down when events are observed; the size of each step depends on how many people were at risk immediately beforehand. A horizontal segment indicates that no new events were recorded during that interval, although follow-up may still have been lost.

A censored participant contributes information up to the last time they were known to be event-free. Their subsequent course is no longer observed. A censoring mark is not an event and does not establish that the person remains disease-free in the future. Removing all censored participants from the outset would also discard the follow-up information they did contribute.

The usual interpretation requires censoring not to systematically select people whose subsequent course would differ from that of those still observed, within the groups or conditions considered. If people with a worse prognosis leave follow-up earlier, the curve may be overly optimistic. The method does not automatically correct informative loss to follow-up.

The numbers-at-risk table is therefore part of the result. A tail supported by very few people warrants more caution than a period with substantial follow-up. Read it alongside confidence bands, censoring marks and the rules defining the event and the end of follow-up. The BMJ introduction to survival analysis explains this construction.

Kaplan–Meier and the HR serve different functions: the curve describes probabilities over time; the HR usually comes from a model or comparison of hazards. If an effect emerges late or changes during follow-up, a single HR can hide that pattern. Assessing proportional hazards may require model diagnostics such as Schoenfeld residuals and interactions with time. Complementary measures can include survival differences at prespecified horizons or restricted mean event-free time, which summarises the area under the curve up to a common limit. We have not performed these analyses on ATOMIC.

From statistics to useful clinical interpretation

First, preserve the endpoint. Disease-free survival and overall survival are not interchangeable. ATOMIC's highlighted result concerns recurrence or death as a composite event. That HR does not establish a 50% reduction in mortality or a particular gain in years of life. Overall survival results and their maturity must be assessed separately.

This also matters for competing risks. If the question concerned recurrence alone, death beforehand would prevent recurrence from being observed and would need to be handled as a competing event according to the estimand. Censoring those deaths and calculating 1 − Kaplan–Meier does not generally estimate the real-world probability of recurrence. For the composite of recurrence or death, both are part of the event. Changing the definition changes the question. BMJ: bias from censoring competing events.

Second, consider toxicity alongside benefit. The paper reported grade 3 or 4 adverse events in 84.1% of the atezolizumab + mFOLFOX6 group and 71.9% of the mFOLFOX6 group. These are adverse-event proportions; they do not imply that every event was attributable to the added drug. Assessment requires examining the nature, severity and duration of events and treatment discontinuations. It is also not a matter of mechanically subtracting the excess adverse-event percentage from the benefit: these are different outcomes.

Third, retain population, horizon and uncertainty. Similar relative effects can correspond to very different absolute differences when baseline risk changes. Applying a result to another population requires assessing clinical similarities, concomitant treatments and follow-up. A number needed to treat also cannot be obtained from 1/(1 − HR): it requires an absolute risk difference, an explicit horizon and uncertainty.

Useful scientific communication brings together five elements: the population studied, event, time, effect size and uncertainty. ATOMIC matters for its clinical result and for the opportunity to practise that complete reading before turning it into a headline.