Use of the bootstrap in analysing cost data from cluster randomised trials: some simulation results

Table 4 Control ICC = 0.01, 24 clusters of size 25 per arm.

How does the ICC change as a result of the intervention?		Between-cluster distribution combination¹	Rejection Rate (%)²
			Huber-White		Single Bootstrap		Double Bootstrap
			Lower	Upper	Lower	Upper	Lower	Upper
No change	-	N, N	2.65	2.45	3.55	3.46	3.11	3.08
		LN, LN	2.57	2.64	3.80	3.68	3.25	3.27
		N, LN	2.42	2.76	3.64	3.67	3.21	3.39
Double		N, N	3.08	2.35	3.65	3.57	3.13	3.37
		LN, LN	2.88	2.37	4.00	3.75	3.36	3.40
		N, LN	2.84	2.46	3.93	3.45	3.29	3.27
		N, N	2.55	2.59	3.38	3.40	3.26	2.99
		LN, LN	2.44	2.82	3.91	4.05	3.53	3.55
		N, LN	2.07	3.00	3.50	3.77	3.04	3.41
Halve		N, N	1.97	3.27	3.23	3.97	2.79	3.30
		LN, LN	2.10	3.03	3.49	3.83	2.93	3.04
		N, LN	1.92	3.20	3.24	3.77	2.87	2.95
		N, N	2.66	2.59	3.66	3.58	3.06	3.06
		LN, LN	2.54	2.46	3.63	3.71	3.00	3.14
		N, LN	2.38	2.57	3.43	3.51	2.87	3.05

¹ Control arm is first, followed by intervention arm. N, N denotes normal distribution in both arms; LN, LN denotes lognormal distribution in both arms; N, LN denotes normal distribution in control arm and lognormal distribution in intervention arm
² Entries in bold indicate where double bootstrap method achieved rejection rate closer to nominal 2.5% than Huber-White method.

ISSN: 1472-6963