Model selection for component network meta-analysis in connected and disconnected networks: a simulation study

Table 3 Selected CNMA models in 1000 simulations of connected networks

Scenario	\(n_{diff}\)	Additive CNMA	CNMA with one 2-way interaction			CNMA with two 2-way interactions
Scenario	\(n_{diff}\)	Additive CNMA	\(A\!*\!B\)	\(A\!*\!C\)	\(C\!*\!D\)	\(A\!\!B\)+\(C\!\!D\)	\(A\!\!B\)+\(A\!\!C\)	\(A\!\!C\)+\(C\!\!D\)
No heterogeneity (\(\tau ^2=0.00\))
A	35	693	119	94	84	3	2	5
B1	257	324	436	148	61	16^a	12^a	3
B2	298	325	45	154	436	17^a	8	15^a
C1	687	61	746	99	17	34^a	28^a	15
C2	732	51	13	104	742	38^a	7	45^a
Low heterogeneity (\(\tau ^2=0.01\))
A	52	652	136	98	108	3	1	2
B1	281	337	426	136	59	16^a	18^a	8
B2	284	317	49	133	450	25^a	10	16^a
C1	676	75	704	109	24	29^a	47^a	12
C2	738	70	13	98	716	45^a	7	51^a
Moderate heterogeneity (\(\tau ^2=0.10\))
A	188	458	188	142	179	14	13	6
B1	392	277	346	178	132	34^a	17^a	16
B2	420	262	122	195	356	28^a	13	24^a
C1	656	117	550	135	53	65^a	64^a	16
C2	702	93	48	161	546	65^a	18	69^a

The correctly chosen model is printed in bold
^aCombination of two 2-way interactions includes the correct interaction
\(n_{diff}\): number of networks rejecting the additivity assumption with significant p-value for \(Q_{diff}\) (\(p < 0.05\)), where \(Q_{diff}\) is the Q statistic for the difference between additive CNMA and standard NMA model

ISSN: 1471-2288