Table 5B.

Post hoc comparisons results.

Diagnostic tasks	Tractography algorithm	(I) PC number	(J) PC number	Mean difference (I-J)	Sig.	95% confidence interval
						Lower bound	Upper bound
AD vs. NC	Tensor-RK2	15	75	0.16667	0.003	0.0293	0.3040
			150	0.16759	0.003	0.0302	0.3050
		20	75	0.15370	0.011	0.0163	0.2911
			150	0.15463	0.010	0.0173	0.2920
AD vs. MCI	Tensor-FACT	10	150	0.14378	0.022	0.0092	0.2784
	Tensor-RK2	10	150	0.17437	0.016	0.0150	0.3338
	Tensor-SL	10	40	0.12890	0.017	0.0105	0.2473
			100	0.12099	0.038	0.0026	0.2394
			150	0.19536	0.000	0.0770	0.3137
		15	150	0.17532	0.000	0.0569	0.2937
		20	150	0.12342	0.030	0.0050	0.2418
	Tensor-TL	10	150	0.17099	0.001	0.0383	0.3037
		15	150	0.14747	0.013	0.0148	0.2802
	ODF-FACT	10	150	0.11424	0.019	0.0084	0.2200
	Probtrackx	10	100	0.14219	0.000	0.0519	0.2325
		15	100	0.12236	0.000	0.0320	0.2127
		20	100	0.10876	0.004	0.0184	0.1991
		25	100	0.12500	0.000	00347	0.2153
		30	100	0.13544	0.000	0.0451	0.2258
		35	100	0.13397	0.000	0.0436	0.2243
		40	100	0.10506	0.006	0.0147	0.1954
		45	100	0.09726	0.019	0.0069	0.1876
		50	100	0.09515	0.026	0.0048	0.1855
	Hough	10	150	0.12416	0.008	0.0154	0.2329
		15	150	0.11994	0.014	0.0112	0.2287
MCI vs. NC	ODF-RK2	40	150	0.11920	0.012	0.0125	0.2259
	Probtrackx	10	100	0.12047	0.002	0.0247	0.2162
		15	100	0.09728	0.041	0.0016	0.1930

Only tests that passed Bonferroni correction are shown here. Using PCA as a feature extraction method, the AUCs for some tractography algorithms are statistically affected by the number of PCs for specific diagnostic tasks. Moreover, a smaller number of PCs tends to give better performance (higher AUC) than higher numbers of PCs for these tractography algorithms when using PCA.