The Biggest Number Is the Weakest

A prespecified analysis of the SELECT trial asked whether semaglutide still cut cardiovascular events in people whose blood tests suggested liver fibrosis. It reports three answers, for three definitions of high risk: 26%, 21% and 34%.^[1] The 34% is the biggest, it is the one that will be quoted, and it is the only one that was not statistically significant. Understanding why is more useful than any of the three figures.

What the analysis did

SELECT enrolled people who carried both obesity and diagnosed heart disease, none of whom had diabetes, and gave half of them weekly semaglutide. The drug cut the trial’s combined endpoint — death from cardiac causes, heart attack short of death, or stroke short of death — by a fifth. This analysis takes that population and splits out those at high risk of significant liver fibrosis, using the FIB-4 score.

FIB-4 is a screening calculation from age, platelet count and two liver enzymes. It is not a biopsy and not a scan. “High risk of fibrosis” throughout this analysis means a blood-test score above a threshold, which is a reasonable way to find people worth investigating and is not a diagnosis.

The three answers

Look at the third row’s interval. It runs from 0.39 to 1.10 — and 1.10 is above 1, which means the data are compatible with semaglutide reducing events by 61%, and also compatible with it increasing them by 10%. When an interval crosses 1, the result has not distinguished benefit from no effect.

Why the biggest number is the weakest one

Which is why the ordering of these three rows by headline size inverts their ordering by reliability. The 26% figure, with the narrowest interval and a P of 0.0004, is the most trustworthy and the least dramatic. The 34% is the reverse.

A confidence interval that crosses 1 has not shown an effect, however impressive the number in front of it.

None of which means the 34% is wrong. The true effect in that group may well be large — the point estimate is the best single guess available. It means the analysis cannot establish it, and anyone quoting “34% reduction” without the interval is transmitting confidence the data do not contain.

What the analysis does establish

The two significant rows are the substance, and they say something worth knowing: semaglutide’s cardiovascular benefit in SELECT held up in people whose blood work suggested liver fibrosis, at roughly the magnitude seen in the trial overall. Liver risk did not appear to blunt the cardiovascular effect.

That is a modest, useful finding. It is also a subgroup analysis of a trial designed to answer a different question, which is the standing caveat on all such work — prespecified subgrouping, as here, is considerably better than choosing the groups after seeing the data, and it does not turn a subgroup into a trial.

The habit worth taking away

• Read the interval before the headline. If it crosses 1 for a ratio, or 0 for a difference, the finding is not established.
• Be most suspicious of the largest number in a set of subgroups. It is usually the smallest group.
• Check whether subgroups were prespecified. Chosen in advance is a different thing from chosen after looking.
• Ask what the risk score actually is. Here, “high fibrosis risk” is a blood-test threshold, not a diagnosis.

SELECT turns up across this register in other guises: what the trial says about hospital days, in days spent in hospital, and how long the evidence base actually runs, in how long is the evidence.