Could you speak statistician?

If you build software for statisticians, epidemiologists or health economists, the requests arrive in their language: intervals, confounders, denominators. Some of it you follow. The rest you guess, and a wrong guess costs a week.

Eight questions place you on the ladder, from building what is asked to reviewing the analysis alongside the people who asked. Answer for how you actually work. Nothing is recorded until you choose to send it.

1. Q1An analyst asks for "the rate, with a 95% interval". What do you build?

   The rate, shown clearly. The interval goes in a tooltip.Both, from whatever columns the analyst's script gives me.Both, kept together wherever the number appears, exports included.Both together, and I ask what the interval should tell the reader.
2. Q2"Admissions rose by a fifth at one hospital." Your first thought?

   Worth a headline tile on the dashboard.Rose from what? I check last year's figure.Per how many people? The catchment may have grown too.Per how many, counted how, and did the counting rule change?
3. Q3People who use your health app have better outcomes. What does that show?

   That the app works.Something promising. More data would settle it.Maybe only that healthier people sign up.Probably selection. I would ask what comparison could answer it.
4. Q4A figure in your output is challenged. Can you say where it came from?

   Eventually. Give me an afternoon with the logs.Yes, I can point to the query that made it.Yes: data version, code commit and query, all recorded.Yes, and an analyst could rerun it and get the same number.
5. Q5An analyst needs last March's numbers exactly as published. Can you produce them?

   Not easily. The data has been updated since.Roughly, by filtering today's data to March.Yes, from the snapshot we keep of each release.Yes: snapshot, code version, and a note of what changed since.
6. Q6One record in eight has no outcome recorded. What do you do?

   Drop them. There are plenty left.Drop them, and note the count in the output.Check whether the missing ones differ from the rest.Check, then ask the analyst: why they are missing changes the answer.
7. Q7The meeting turns to "intention to treat". Where are you?

   Nodding. I will look it up afterwards.Roughly there. It is something about trials.With them: people are analysed in the group they were assigned.With them, and I know what it changes in the data I build.
8. Q8When do you want a statistician involved?

   When they send me the spec.When the results look odd.Before I build anything that turns data into a claim.Early, and I review their pipeline as they review my code.

The button opens an email to us, copied to Rabbie, our robot, who keeps the diary. Your result travels in it so we can prepare. You can change or delete it before you send, and nothing leaves your browser until you do. The developer door, adding analysis to tech, is on the two doors. If you work with data and evidence yourself, try chatbot or do-bot? instead.

Back to the two doors →