LECTURE HUB

🔋Lecture 10 - Power, sample size and stopping early

🚦 Status🟣 Formatted

🧠 Mastery🐥 Apprentice

📅 Last Review8 Jun 2025

🔤 Statistical power is the chance of seeing an effect that is really there. Low power means a real improvement will usually look like nothing.
Sample size depends on three things: how variable the outcome is, how small an effect you care about, and how sure you want to be. Halve the effect, quadruple the sample.
Stopping early because it looks good inflates false wins. Stopping early because it looks bad wastes true ones.
Last new material. From here it is revision, and the assignment I handed in a day late tonight.

handed assignment 2 in at the door. a day late. he took it without comment
two ways to be wrong: cry wolf / sleep through the wolf
power = 1 − sleeping. 80 % is the usual ask and almost nobody checks
n ≈ 16 p(1−p) / δ². per GROUP
halve the effect you want to see → four times the sample. ouch
day vs night shifts: needed ~4,000 each, had 2,000. so lecture 7 was a coin toss all along
peeking: only if you planned the peeks
no class for me next week (Lisbon). this is the last new material

Write 🔤 Statistical power
Add the rule of thumb to 📐 Sample size for comparing two proportions with the triage example
Go back to lecture 7 and work out how many packets the shift comparison needed
Start revision for the final: one lecture a day from the 27th

📝LECTURE NOTES