Week 2 Session: Taking It Apart

The one week you build a number out of nothing — and then discover that your number and the official one were counting different things

1 March 2027

Phones away for this one. You will want to look something up about twenty minutes in, and not being able to is the exercise.

The target#

How many bus stops are there in the ACT network?

That is the whole brief, and it is deliberately underspecified. You will not be told what counts as a stop. Notice, later, when that starts to matter.

This is the only week of the semester whose central activity is building a number from scratch. Every week after this one works on estimates that already exist — yours, somebody else’s, or an institution’s. So take the chance to find out what you actually do when nobody hands you a starting figure.

Three passes#

Pass one — build the chain. In groups, get from something you know to the thing you don’t, in as few steps as you can manage. There is no prescribed route. Suburbs, routes per suburb, stops per route is one chain; kilometres of road and stops per kilometre is another; buses in the fleet and stops per bus’s run is a third. Whichever you pick, every link is a number and a reason, written where a stranger can read both.

The smallest set of sub-quantities that gets you there is usually smaller than the first one you write. Try to delete a link.

Pass two — audit somebody else’s units. Swap chains with another group. Your job is not to argue with their numbers; it is to check that each multiplication produces the unit it claims to. Stops-per-route times routes gives stops. Stops-per-route times suburbs gives nothing at all — and a chain that quietly produces “stops per suburb per route” has a bug in it that no amount of better guessing will fix.

This is the cheapest structural check in the course and almost nobody runs it unprompted. It is also only a necessary condition, not a sufficient one: stops-per-route times routes is dimensionally impeccable and still counts a stop served by four routes four times. Units passing tells you the chain is not nonsense. It does not tell you the chain is right.

Pass three — mark the weak link. Circle the one number in your own chain you would least like to defend if someone pushed. Don’t fix it. Just mark it. We come back to exactly this instinct next week, and sharpen it into something more useful than a hunch.

What the record actually says#

The ACT Government publishes a bus stop dataset on its open data portal. The counts below were not read off the portal page; they come from querying the linked file directly while this session was being prepared. That is all the record says:

• 2,581 rows.
• 2,581 distinct stop_id values — so every row carries its own identifier.
• 2,570 distinct stop_name values — eleven fewer than there are rows, which tells you some names repeat, not how many names or how often.
• Each row carries a name, an id, and a latitude and longitude.

Two things that figure is not.

It is not a 2027 count. The dataset is a July 2017 snapshot, and it is the most recent bus stop extract the portal publishes openly; the current stop list lives in Transport Canberra’s GTFS feed, which needs an access key. You estimated the network you can see out the window. The record describes the network of a decade ago.

And it is not, on its own evidence, a count of physical stops. It is a count of stop records. Whether two records at the same intersection are two stops or one stop with two sides, whether a shelter and an unsheltered pole both count, whether a stop served by four routes is one thing or four — the dataset does not say. That shortfall of eleven is a hint that something is going on, not an answer.

Comparing definitions, not just numbers#

So don’t ask yet whether you were close. Ask, in your group:

• What were you counting? Write it as a sentence someone could disagree with. “Places a bus stops” and “poles you could stand at” are different sentences and different numbers.
• What is the dataset counting? Say what you can support from what is above, and say where you would have to guess.
• If your number and 2,581 differ, how much of the gap is ten years of network change, how much is the two of you counting different things, and how much is your chain being wrong?

That third question usually cannot be settled, and saying so is the correct answer rather than a failure. Two careful people can produce different honest numbers for the same question because they were answering different questions. An estimate whose target is undefined is not precise enough to be wrong.

What you hand in#

One page from each group: the chain, the units audit somebody else wrote on it, the circled weak link, and the sentence saying what your number was a number of. Bring it back in week 3 — we will be reading someone else’s estimate then, and it helps to remember what your own looked like from the inside.