Table 2  Power of parametric and estimating equations approaches under three error distributions: Gaussian, autoregressive (GA); t, 4 d.f., autoregressive (tA); and Gaussian, Matérn (GM)
Neighbor correlation
0.610.830.94
Error model
MethodGAtAGMGAtAGMGAtAGM
Lik0.960.960.600.900.920.610.960.950.58
EES0.920.900.740.760.810.750.890.900.72
EE0.900.900.700.750.800.710.860.880.69
• Shown is the proportion of times the P-values under the alternative were smaller than the 5th percentile of the null. The parametric approach (Lik) has the highest power when the covariance structure of the error is correctly specified. This is true for both Gaussian and t-distributed errors. The estimating equations approaches have slightly less power than the likelihood method with correctly specified correlation. When the covariance structure is incorrectly specified, the estimating equations approaches have greater power. Estimating equations with shrinkage (EES) appears to have slightly higher power than estimating equations without shrinkage (EES).