Two plants share each pot: Yh,i = Pi + δ + εh,i and Ya,i = Pi + εa,i, with pot effects Pi ~ N(0, σP2) and noise ε ~ N(0, σ2). The within-pair differences remove the pot effect, so the paired t-test sees only the noise 2σ2; the unpaired (Welch) test estimates its standard error from the total variance σP2 + σ2 and becomes conservative. The gain is governed by the within-pair correlation ρ = σP2/(σP2 + σ2): the variance of the mean difference is 2τ2(1 − ρ)/n. Darwin's real maize data have ρ < 0, and there pairing loses.
code/ch13_experimental_design.R (paired_sim, Figure 13.10).