Table 3.

Power (percent) for detecting a treatment effect based on 4000 simulations in a two-arm (placebo/treatment) parallel design study, by μ and the standard deviation of the values after treatment (${\sigma }_{Trt}$) Results are for at least 99% of the simulations, except as indicated.

	μ (64 per arm, σTrt = 1)				μ (90 per arm, σTrt = 1.4)
	2.0	2.5	3.0	4.0	2.0	2.5	3.0	4.0
Model
Nonparametric analyses
Wilcoxon test (NQ only)	60.3	74.0	75.8	76.8	38.9	60.5	70.1	77.0
Wilcoxon test (both NQ and ND)	67.9	79.0	76.9	76.8	53.2	68.2	72.7	77.1
Gehan test (NQ only)	61.0	74.9	76.8	78.0	36.7	60.3	71.5	77.1
Mixed models using both baseline and follow-up data
Equal variances	82.3	90.5	92.1	93.2	37.7	63.2	81.6	93.8
Unequal variances	77.0	89.9	92.4	93.1	64.2	81.4	90.8	94.5
Mixture models using likelihood L2 and follow-up data only
Equal variances	51.6	63.5	67.1	56.1d	NS	NS	NS	NS
Unequal variances	51.4	61.5	63.6	51.7d	NS	NS	NS	NS
Mixture models using likelihood L2 and both baseline and follow-up data
Equal variances	19.8c	25.1b	26.6a	30.6a	NS	NS	NS	NS
Unequal variances	28.2c	33.2b	30.7a	25.1a	NS	NS	NS	NS
Generalized estimating equation logistic regression models for a significant decrease
NQ and ND combined	36.2	52.7	67.1	68.3	17.7	41.0	65.2	83.3
ND separate from NQ	52.5	64.5	71.4	68.8	33.5	55.2	72.0	84.0

Assumptions: treatment reduces the underlying mean μ by 0.5, the standard deviation of the untreated mean is 1, and Type I error is 0.05. Quantification limit is log₁₀(80) = 1.90. NQ: non-quantifiable; ND: not detectable.

NQ only: NQ and ND were combined.

^aResults for 90% to 98.9% of all simulations.

^bResults for 80% to 89.9% of all simulations.

^cResults for 70% to 79.9% of all simulations.

^dResults for 60% to 69.9% of all simulations.

NS: not shown, excessive Type I error.