Lag is not severity

Fast and trivial

The fastest event was a request for printed copies of the weekly rota, one week after they stopped. Somebody wanted paper. The cost of the removal was approximately zero and the cost of reversing it was the price of printing.

Second fastest: chasing unpaid invoices weekly, which crept back at two weeks because the person who did it simply carried on doing it. Nothing had gone wrong. Nothing would have.

Third: the Monday planning hour, argued about at the time, asked back after two weeks, and reinstated at under half its old length. The argument had been vigorous. The consequence was a shorter meeting.

Three of the five fastest events on the record were about preferences, habits and comfort. None of them revealed anything.

Slow and serious

The three slowest were all problems: thirty-one weeks, forty-seven weeks, sixty-four weeks. Between them they produced a scramble through an audit, a probable contribution to three departures, and the same piece of analytical work done four times over.

So the top of the distribution is trivial and the bottom is serious, which looks exactly like a relationship.

Why there is no relationship

The middle of the distribution breaks it, and the middle is where most of the record lives.

Two events at fourteen weeks, one trivial and one costly. Two at twenty-two weeks, same split. The lag is doing no work as a predictor.

Where the appearance comes from

The correlation at the extremes is real and it has an explanation that is not about severity.

Fast events are fast because somebody's own work was made harder immediately. Work being made harder immediately is, by construction, a small and recoverable problem — if it were large, it would not be recoverable by asking for the thing back.

Slow events are slow because nobody's own work was affected. Problems that nobody's daily work surfaces are problems that accumulate, and accumulation is what makes a cost large.

So both ends are explained by the same variable — whether anybody is inconvenienced today — and the lag is a symptom of that variable rather than a cause of anything.

Lag measures how long until somebody is standing in the right place. Severity measures what has piled up while nobody was. They are driven by the same thing and they are not the same thing.

The error this prevents

The practical consequence is a rule about triage during a transition.

The natural approach is to deal with what comes in. Requests arrive, they get handled, the queue empties, and the emptying of the queue feels like completion. Attention follows the signal, which is exactly what attention is supposed to do.

But the signal is systematically composed of the cheap problems. The queue of incoming requests is a queue of things somebody noticed within a fortnight, and a thing somebody noticed within a fortnight is a thing that was inconveniencing a person who could fix it.

Managing the transition by managing the queue therefore means spending the available attention on precisely the subset that needed the least. On this record, the eight asked-back items consumed most of the discussion during the first quarter. All eight were resolved. All eight were cheap. The five that mattered produced no queue entries at all.

What to do instead

Keep handling the queue — it is cheap and the items are real. But do not treat it as the measure of how the transition is going, and do not let an empty queue end the attention.

The two activities are separate:

1. Handle what arrives. Fast, cheap, mostly preferences. Done by week fourteen.
2. Go and look at what has not arrived. Slow, uncomfortable, requires a list and a calendar. Not done by anybody on this record until a problem forced it.

The second is not a more diligent version of the first. It is a different activity aimed at a different population, and nothing in the first will ever prompt it.

The one case where they do line up

There is a single genuine relationship in the record and it is worth stating because it is narrower than the one people assume.

Among the removals that were reversed, the lag predicts the weight of what came back. The eight fastest reversals came back at a median of about sixty per cent; the eight slowest at about forty-five.

That looks backwards until the mechanism is named. A fast reversal is a request, and a request is negotiated — the hours are visibly not there and everybody knows it, so what goes back is trimmed. A slow reversal is a response to something that happened, and what goes back is sized by the incident.

But the eight slowest reversals here are mostly quiet returns rather than incidents, and a quiet return is trimmed hardest of all, because nobody rebuilds more of a procedure than they personally need.

So the relationship is real, it runs through two different mechanisms, and it still says nothing about severity.