A smoothed outcome series can help a reader see past annual fluctuations. It can also delay the appearance of a recent change. When a college's three-year average rises, the newest annual result may still have fallen. Understanding which observation enters the window and which leaves it is essential before calling the latest movement an improvement.

NIST's time-series guidance explains averaging as a smoothing method. The illustration below applies a simple trailing three-year arithmetic average to fictional college rates. It uses equal cohort sizes so that averaging the annual rates also matches pooling their numerators and denominators. This equality should not be assumed when real cohort sizes differ.

Reconstruct the two windows

Assume a consistent outcome definition and one hundred eligible students in each of four years. The fictional annual outcome rates are forty, eighty, ninety and seventy percent. Therefore the numbers meeting the outcome are forty, eighty, ninety and seventy respectively.

Year Annual rate Trailing three-year average
1 40% Not yet defined
2 80% Not yet defined
3 90% 70%
4 70% 80%

The latest annual rate falls twenty percentage points, from ninety to seventy. Yet the rolling average rises ten points, from seventy to eighty. Both statements are correct. The second window replaces Year 1's forty with Year 4's seventy while retaining Years 2 and 3. The net addition of thirty points divided across three years raises the average by ten.

The identity is useful: the change in a three-year moving average equals the entering value minus the departing value, divided by three. It does not depend only on whether the entering value exceeds the immediately previous year. That distinction explains the apparently contradictory directions.

Do not treat overlapping windows as new independent samples

Two adjacent windows share most of their observations. The averages above both contain Years 2 and 3. Presenting the two averages as though they came from separate sets of three years overstates the amount of fresh information. A smooth sequence of increases can partly reflect how old observations move through the window.

Before comparing institutions, record whether the average is trailing or centred, the window length and the treatment of missing years. A centred average may use later observations unavailable at the time a historical decision was made. A trailing average has a different timing interpretation. The label “three-year” is insufficient without the endpoints.

If cohort sizes vary, establish whether the publisher averages rates equally or pools counts. These methods answer different questions and can differ materially. Do not reconstruct a pooled average from rounded rates when the underlying counts are unavailable. Keep an unresolved calculation unresolved rather than substituting equal weights silently.

Read annual and smoothed evidence together

A recent fall need not prove a lasting reversal, just as a rising smoothed average need not prove current improvement. Inspect definition changes, coverage, uncertainty and the broader annual pattern. The four invented observations are too few to identify a durable trend or its cause.

For a shortlist worksheet, retain the latest annual observation alongside the smoothed statistic, with dates and cohort coverage. Ask what decision each is intended to inform. A longer-term description may be useful for broad context, while an operational question about recent conditions may require more current evidence.

The correct summary of this example is explicit: “The three-year average rises to eighty percent because a low old observation leaves the window, while the newest annual rate falls to seventy percent.” That sentence preserves the value of smoothing without allowing a reassuring average to erase an important recent movement.