In biology you often count things: cells of type A out of total cells of type B, mutant flies out of total flies, etc. The most common move in papers is to compute a ratio per animal and run a t-test on the ratios. This throws away how many cells you actually counted: "5/100" and "50/1000” becomes same, and feeds strictly [0,1] bound data to t-test. The principled alternative is count regression with offset(log(N)): model the raw count directly, bring the total in as a statistical weight, respect the non-Gaussian nature of count data. This week I decided to test this assumption in practice: Setup. Four methods across two pipelines: Animal-level: Welch's t-test on ratios vs CMP GLM (glmmTMB(..., family = compois())) Field-level: LMM with (1 | EmbryoID) vs CMP GLMM with the same RE Three metrics: Type-I error, size-adjusted power (Lloyd correction), median 95% CI width. The interesting bit.…