Research questions look very well behaved in proposals.
They sit inside neat boxes. They begin with verbs like “investigate” and “estimate.” They rarely interrupt. Mine, I assumed, would remain where I had put it: somewhere near heat, cardiovascular disease, and perhaps acute myocardial infarction—AMI, the clinical name for a heart attack.
I entered my first week expecting a clean story. Hong Kong has become hotter; cardiovascular disease is consequential; therefore, perhaps the recent data would reveal a shift in the city’s temperature–health pattern. I could almost see the arc before the analysis began. In my head it already had a title.
This is generally a warning sign.
On 17 July, our lab meeting took that tidy arc and held it up to the light. The general Hospital Authority file available to the project does not contain reasons for admission. That is not a small footnote. Without knowing why someone was admitted, I cannot honestly construct an AMI outcome from that file.
So AMI left the project.
Not because heart attacks became uninteresting during the meeting. Not because a model failed. The question disappeared because the data could not answer it. There was something bracing about this. Research is often described as following the evidence, but sometimes the first evidence is simply a locked door with an unusually informative sign.
What remained was a different and more defensible path: governed aggregates of stroke events, organised by month. The exact outcome file had not yet arrived, so there were no stroke coefficients to celebrate, mourn, or place in a colourful graph. The immediate task was to prepare for monthly stroke aggregates without imagining what the undelivered fields might contain.
“Monthly aggregates” may not sound like the opening of a thriller. They cannot tell us whether a particular hot Tuesday triggered a particular admission. They do not recreate the daily, delayed temperature–risk models used in other studies. They ask a quieter question: how does thermal exposure accumulated across a month relate to stroke burden recorded for that month?
Quieter is not the same as trivial.
The meeting also changed my idea of what an analysis should look like. I had imagined finding one glamorous model: the statistical equivalent of a perfectly framed photograph. Instead, Professor Bishai pushed us toward a panel of roughly ten or more methods.
One model might use monthly mean temperature. Another might look at maximum or minimum temperature. Others could count official very hot days, hot nights, or cold days; preserve sequences of extreme weather; or examine exposure from the previous month. The point is not to manufacture ten discoveries. It is to ask whether one scientific picture survives several reasonable ways of looking.
This was my introduction to the strange democracy of methods. A continuous temperature value and a five-night heat spell can describe the same summer while noticing different things. One sees level. One sees persistence. Neither automatically wins because its name looks more sophisticated in a table.
Then the heatwave world widened.
Work linked to Chao Ren and Wang treats heat not only as a high number but as a spell: very hot days, hot nights, and combined day–night events. The combined definitions are especially intuitive. A city may endure the afternoon and still struggle if the night offers no relief. These daily events can be translated into monthly burden measures, but the translation has to remain visible. A daily heatwave study, a monthly stroke analysis, and an excess-mortality calculation are related scientific relatives—not identical twins borrowing one another’s passports.
By the end of the meeting, my original question had not been answered. It had been dismantled and rebuilt.
I had arrived wanting to test a dramatic possibility: perhaps the old cold-dominant Hong Kong story had yielded to a new heat regime. That hypothesis is still a question, not a conclusion. We do not yet have real stroke estimates. More importantly, I now understand that “prove a regime shift” is a dangerous instruction to give yourself before opening the data. It quietly turns analysis into casting: every number auditions for a role in a story already written.
The new instruction is less cinematic and more useful: build something honest enough to survive contact with real data.
That means keeping the time scale straight. It means labelling each method before comparing results. It means refusing to turn synthetic code checks into findings. It means waiting for the outcome dictionary rather than filling blank spaces with confidence.
It also means learning that collaboration is not a decorative ring around the science. Hogan’s weather questions determine what “heat” becomes in the dataset. Roro’s outcome knowledge determines what a stroke month actually means. Professor Bishai’s insistence on multiple methods changes what counts as convincing. My job is not to make those different forms of expertise disappear into one smooth paragraph. It is to make them work together without sanding off their edges.
Being twenty-something and early in a research life, I had thought anticipation meant waiting to discover whether the answer was yes or no. Week 1 taught me a different wait: discovering what the question can honestly be. There is disappointment in giving up the clean story. There is also relief. The project no longer needs reality to cooperate with my first draft.
It only needs us to listen carefully when reality edits it.
Week 1 ended with less certainty than it began.
Oddly, it also ended with a better project.