Week 5 began the way Week 4 had ended: with a Results heading and nowhere to put a number.
Then a file arrived.
I had been practising a particular welcome. I would open monthly stroke aggregates. I would check that the months lined up. I would finally run the panel we had been specifying since the question changed its mind. The sentence I had ready was about stroke and temperature, organised by calendar month, at last in contact with real counts.
The file was not stroke.
It was monthly counts of the first recorded hospitalisation after a first diagnosis of coronary heart disease, or of heart failure, among people already living with type 2 diabetes and/or hypertension.
Coronary heart disease, CHD, concerns the heart’s blood vessels. Heart failure is when the heart cannot pump well enough. Both are cardiovascular outcomes. Neither is a stroke. Admission cause was still not recorded, so I still could not say why each person had been admitted. Stroke had been mentioned. It had not been attached.
The data did not refuse to come. They refused to be the series I had been addressing in my head.
A first hospitalisation after a first diagnosis is a narrower object than “heart problems this month.” Imagine a pool of people who can still contribute a first event. When someone in that pool is hospitalised after their first CHD or HF diagnosis, they leave the pool. They do not sit there waiting to be counted again as a first event. Over a decade, the pool shrinks. The yearly totals fall. That fall is not, by itself, evidence that Hong Kong became healthier. It is what this kind of count often does.
We still cannot treat a monthly total as a daily trigger. We still cannot turn a city-wide population number into the number of cohort members still at risk. The honest question is quieter than incidence. Given the month of the year and the slow drift of time, how do this month’s ways of writing heat and cold sit beside this month’s first-event counts?
Then, at last, there was something to run.
We ran the labelled panel we had written down while the Results section was blank. Mean temperature. Maximum and minimum. Official hot nights, very hot days, cold days. Heat and cold in the same design, because the local mortality literature had already warned us not to delete winter just because summer feels like the emergency.
Age bands and sex splits were not in the file, so those models did not run. A stroke subtype split did not run, because there is still no stroke file.
Getting to work, it turns out, is mostly this: matching months, refusing to fill gaps with invented fields, and letting some of the clever recipes stay on the counter.
Stroke is still a named absence. Hogan’s weather writing is still his. Roro still owns the health-data methods, including whatever this file’s columns actually mean. Professor Bishai still owns whether a complete exploratory panel is allowed to remain exploratory.
My job, now that there is something to run, is not to rush a headline into the blank that Week 4 left for Results. It is to describe the counts we actually have, including the ways they fail to be the counts I expected.
Week 5 ended with less rehearsal and more work.
The Results heading finally had somewhere to put a sentence. The sentence was not about stroke.