Assumptions - normality, independence, variance, and what breaks
Lesson 3 of 6 · Statistics you can defend
1 · Learn the move · Fact-check list
Every test is a contract: valid if its assumptions hold, undefined if they do not. The model states the assumptions when asked and stays silent otherwise, so the move is a standing fact-check list. For the test you chose: name each assumption, name how it is checked in this dataset, record the status, name the fallback if it fails. Two honesty rules make the list worth keeping. Independence cannot be checked from the data - it comes from the design, and cells from the same mouse or wells from the same plate are not independent no matter what a plot says. And at small n, normality checks pass almost anything; a passing Shapiro test at n=4 is absence of evidence. Write DOUBTFUL when doubtful. The list goes in your analysis log, next to the result it protects.
I am running [test] on [data description]. Build the assumption fact-check list: one row per assumption with columns ASSUMPTION / HOW CHECKED IN THIS DATASET / STATUS (HOLDS, DOUBTFUL, FAILS, CANNOT CHECK FROM DATA) / FALLBACK IF IT FAILS. Independence must be justified from the design I describe, not from the data. Where my description leaves a check impossible, write [NEEDS: ...]. Then draft the check code, but do not interpret its output for me.
2 · Your turn. You write the prompt
You settled on a two-sample comparison of protein abundance between treated and control lysates, n of 5 per group, one prep day each. Before running it you want the assumption list on paper - especially because both groups were processed on different days, which a colleague waved off as fine. Build the list.
Remember: the AI sees only your prompt, not this page. If the situation isn't in your prompt, it doesn't exist.
Optional. These shape the output when you run your prompt below, not your score.