Why student analytics fail: the missing denominator

Most institutional dashboards count events accurately and cohorts carelessly. The numerator is the easy half. Almost every misleading figure in education analytics is a denominator problem.

A principal is shown that attendance is 91 per cent. It is a real number, computed from real registers. It is also useless, because nobody in the room can say whether the denominator was students enrolled today, students enrolled at the start of term, or students who appeared in at least one register that month. Those three produce figures several points apart, and the gap is largest exactly where attention is needed: in the sections with the most churn.

This is not a reporting bug. It is what happens when a dashboard is built from event tables. Events are easy to count and they carry no notion of who should have been there.

Three denominators, one label

Take a single section of forty-two students in a term where two left in week three, one joined in week six, and one was on approved medical leave for a month.

All three are defensible. None is wrong. But a dashboard that does not say which one it used cannot be compared across sections, across terms or against a target, and every meeting that cites it will re-litigate the same question.

Six familiar figures and the base they need

In every case the numerator is uncontroversial and the base is where the argument lives.

Absence of data is not data about absence

This is the most dangerous habit in institutional dashboards, and it is nearly invisible. A section where the teacher stopped marking attendance produces no absent rows. Counted from events, it reports perfect attendance. The worst-managed section in the school appears at the top of the table.

The same pattern runs through everything. A department that never records grievances looks harmonious. A programme that stops entering internal marks looks like it has no failures. A hostel that abandons its incident register becomes the safest building on campus.

What makes an early-warning signal usable

Every institution wants to identify students at risk before results. Most attempts fail for reasons that have nothing to do with the model and everything to do with what happens after the flag.

A simple signal that a mentor acts on beats a sophisticated one that nobody owns. Start with two inputs, attendance trend and internal marks trend, and make the follow-up visible in the record. Sophistication can come after the loop exists.

Five rules for a dashboard people believe

None of this is analytics in the fashionable sense. It is arithmetic with its assumptions written down. An institution that gets the denominators right has better information than one running predictive models on cohorts nobody has defined.