variance-stabilizing transformations

Many real measurements are noisier when they are larger: count data scatters more when counts are high, and a proportion near 1/2 wobbles more than one near 0 or 1. That changing spread (heteroscedasticity) breaks tools that assume constant noise. A variance-stabilizing transformation reshapes the data with a clever function so the spread becomes roughly the same across the whole range.

The idea rests on the delta method. If a variable has variance that depends on its mean, say Var roughly equal to v(mu), then transforming by a function h spreads or compresses the noise: the transformed variable has variance approximately (h'(mu))^2 times v(mu). To make that constant, choose h whose derivative cancels the growth of v — formally h'(mu) proportional to 1/sqrt(v(mu)), so h is the integral of 1/sqrt(v(mu)). Work that integral out for common cases and famous transforms drop out: the square root for Poisson-like counts (whose variance equals the mean), the arcsine of the square root for proportions, and the logarithm when the standard deviation is proportional to the mean.

These transforms are practical staples: they make residuals more nearly equal in spread, pull skewed data closer to symmetric, and let ordinary methods (which assume constant variance) work better. The honest caveat: the stabilization is only approximate, it leans on the delta-method linearization, and the transformation changes the scale you are working on — a mean on the transformed scale is not the transform of the original mean, so interpret results with that in mind.