Composite endpoints: what improved, exactly?

Start with the event definition

A composite endpoint combines specified outcomes into one measure. In an event-based composite, a participant commonly reaches the endpoint when any listed event first occurs. Combining related events may make a trial feasible when individual events are uncommon. It does not mean that the treatment has demonstrated a benefit for each component.

The protocol should tell you exactly what qualifies, how events are confirmed and whether the analysis counts people with a first event, time until that event, or recurrent events. Those are different questions.

Work the example from the inside out

Imagine two groups of 1,000 people followed for a full year. Their invented outcome is the first urgent visit or hospital admission. A person is assigned to just one first-event category. Nobody is lost to follow-up; these simplifying assumptions are part of the example, not expectations about real research.

In the standard-care group, 20 people have an admission as their first event and 80 have an urgent visit first. The total is 100 people. In the new-strategy group, the corresponding counts are 20 and 40, giving 60 people. The combined risk falls from 10% to 6%: a 4 percentage-point difference, or 40 fewer people with a first event per 1,000. Relative to the starting 10%, that is a 40% reduction.

Nothing in this arithmetic shows fewer first admissions: both groups have 20. All 40 fewer composite events come from the urgent-visit category. The example supplies no death data, no disability measurements and no confidence intervals. A claim of improved survival would have no basis here.

Follow one person past the first event

Consider an invented participant, Person A, who visits an urgent clinic in February and is admitted to hospital in June. In our first-event table, Person A appears under urgent visits. The June admission does not move that person into the admission row. Person B, admitted in March without an earlier urgent visit, appears under admissions.

Now imagine that several people in each urgent-visit row later entered hospital. Our table would still look exactly the same. To answer “Were there fewer people hospitalised at any time?”, we would need another analysis that includes those later admissions. The first-event table cannot supply that answer. This is why the heading above a component table matters as much as the figures inside it.

Ask what each component means to the reader

Three features deserve attention: how important each event is to patients, how often it happens and whether effects on the components plausibly differ. A frequent, less consequential event can dominate a combined result. Wide component confidence intervals may leave substantial uncertainty even when the overall estimate is precise.

For the invented example, write separate questions beside the table: How burdensome were the urgent visits? Were admissions shorter or longer? Did symptoms improve? Were there additional adverse effects? Those answers are absent. Naming them prevents the attractive total from filling gaps it cannot fill.

Rewrite the headline without losing the finding

An accurate description of our arithmetic is: “In an invented one-year example, 60 rather than 100 people per 1,000 experienced a first urgent visit or admission; the difference was in urgent visits.” That sentence is longer than “major events cut by 40%”, but it tells the reader what changed.

A shorter version could say: “Fewer first urgent visits drove the combined result.” It would still need the absolute figures and follow-up nearby. We cannot write “40 fewer admissions”, because the admission counts did not fall. We cannot write “40 lives saved”, because survival was not measured in the example.

Nor should we quietly drop the urgent-visit finding. Avoiding an urgent visit could matter. The useful task is to describe that benefit accurately and investigate its importance, rather than promote it into a different outcome.

Connect the result to another study

Build a small comparison card before putting two headlines side by side. For our example, the card would read: urgent visit or admission; first occurrence; complete one-year follow-up; people affected per 1,000. Leave the population and treatment fields explicitly “invented”.

If a second report counts every admission over three years, its number answers another question. A result for a different composite cannot be treated as a replication simply because both papers use a phrase such as major events. Seek matching component definitions, follow-up and counting rules before discussing agreement. A survival result, if available, deserves its own line.

What a more useful next report would show

For a follow-up to this invented exercise, the most useful additions would be a record of every admission, the duration of visits and admissions, symptoms, adverse effects and uncertainty around the comparisons. A patient could then see both the frequency and burden of events. This is a reporting wish list, not a prediction of benefit.

For real trial reports, separate component results remain important. A significant composite does not automatically license claims of significant benefit for every component; formal claims require the relevant testing framework. Keep descriptive component patterns separate from confirmed component effects.

Open the bundle: which first events changed?

Hypothetical one-year counts. Rows for the two components are mutually exclusive and sum to the total because each person is classified by the first event only. Later admissions are not shown. No significance test is supplied.