Aggregation is the combining of many values into one, either across time or across a hierarchy. It is where a surprising number of measurement errors originate, because the arithmetic always succeeds and only sometimes means anything.
Each metric needs an aggregation rule stated in its definition, because there is no universal correct answer.
Sum works for counts of discrete events. Incidents in a quarter is the sum of incidents per month.
Average works for rates and durations, with a caveat below.
Maximum works for anything where the worst reading in the window is what matters, such as peak backlog.
Last value works for point-in-time state. Agent coverage at quarter end is the last daily reading, not the sum of daily readings.
Applying sum to a point-in-time metric is the classic failure. Thirty daily coverage readings of 94 percent do not aggregate to 2,820 percent, and yet a naive rollup will produce exactly that.
Averaging ratios is wrong more often than it is right.
Three business units report remediation rates of 90, 80 and 50 percent. The average is 73 percent. If those units have 10,000, 5,000 and 200 findings respectively, the actual organization-wide rate is close to 86 percent.
Ratios aggregate by summing numerators and summing denominators, then dividing. The simple average weights a 200-finding unit equally with a 10,000-finding one, which is almost never what the reader assumes.
This one shows up in real board packs regularly.
The same discipline applies going up an org chart. A regional figure should be computed from the underlying records, not averaged from the country figures beneath it.
Where the underlying records are not available at the parent level, state that the rollup is an average of averages, because the number will be quietly wrong and someone should know.
For duration metrics, medians and percentiles usually beat means. A mean time to respond is dragged around by a handful of long-running cases and can improve while the typical case degrades. Publishing the median alongside the mean, and a high percentile alongside both, costs nothing and prevents a recurring argument.
From the blog