We want the slope of Y on X, but we only see W = X + U. Regressing Y on W shrinks the slope by the factor λ = σx2/(σx2 + σu2). SIMEX turns this around: add extra noise of variance ρσu2 to W, record the average slope Ω̂(ρ) for several ρ ≥ 0, fit a curve and read it off at ρ = −1, where the total measurement error variance would be zero.
Model: X ~ N(0, 1), Y = X + ε with ε ~ N(0, 0.52), so the true slope is 1; U ~ N(0, σu2) with σu known. SIMEX uses ρ = 0, 0.25, ..., 2. The quadratic extrapolant is fitted by least squares; the rational one, γ1 + γ2/(γ3 + ρ) (10.7), by least squares in (γ1, γ2) for each γ3 > 1 and a one-dimensional search over γ3. The rational form is exactly the shape of the large-sample curve Ω(ρ), so it removes the bias; the quadratic leaves some, more as σu grows. The method of moments divides the naive slope by λ̂ = (sW2 − σu2)/sW2.
code/ch10_other_topics.py (simex_all, fit_rational, fig_10_2).