Replicate structure - technical, biological, and what the model assumes

1 · Learn the move · Flipped interaction

Hand a model 24 rows and it assumes 24 independent observations - silently, instantly, and wrongly if those rows are three mice measured eight times each. Technical replicates measure your pipetting; biological replicates measure the world; and n is the number of independent draws from the thing you want to generalize about, which is usually the smaller number. Nothing downstream survives getting this wrong, so it cannot be left to assumption - and the flipped interaction is the fix: the model interviews you, one question at a time, about what varies between which rows - same animal? same passage? same vial split in two? same day? - and it does not proceed until it can restate your unit of replication back to you and you have confirmed it. Two minutes of interview beats a beautiful analysis of an n that does not exist.

Before analyzing my data, interview me about its replicate structure - one question at a time, no analysis until done. Establish: what the rows are; which rows came from the same animal / donor / culture / vial; what was independently prepared vs split from one preparation; what shares a day, batch, or passage. Then restate: the unit of replication for my question [question], the true n per group, and which rows are technical replicates of the same unit. Wait for my confirmation. If my design gives n below what my planned comparison needs, say so plainly.