Table 1.

Resolution comparison between single-trait fine-mapping and flashfm.

	Trait 1 (A+D); Trait 2 (A+C); Vary sample size			Trait 1 (A+D; Trait 2 (A+C); Vary proportion missing trait 1data
N	Median percentage size reduction	Single-trait group A coverage	Multi-trait group A coverage	Trait 1 proportion missing, p1	Median percentage size reduction
1000	0	1	0.986	0	28.5
2000	10.5	0.997	0.99	0.1	31.4
3000	28.5	1	1	0.2	27.8
4000	32.5	1	0.997	0.3	18.8
5000	33.3	1	1	0.4	16.7
				0.5	10.5

	Trait 1 (A+D); Trait 2 (C); Vary sample size			Trait 1 (A+D); Trait 2 (A+C); Vary trait correlation
N	Median percentage size reduction	Single-trait group A coverage	Multi-trait group A coverage	Cor(Y1, Y2)	Median percentage size reduction
1000	0	0.944	0.944	0	33.3
2000	0	0.98	0.97	0.2	33.3
3000	0	1	0.99	0.4	28.5
4000	0	0.988	0.984	0.6	21.1
5000	0	0.996	0.996	0.8	14.3

When traits share a causal variant, flashfm tends to yield smaller SNP groups than those from single-trait fine-mapping, regardless of amount of missing data and trait correlation; both methods have similar resolution and accuracy when there are no shared causal variants. In simulations with a shared causal variant A, (trait 1 is A+D, trait 2 is A+C), β_A = log(1.4) for both traits 1 and 2; trait 1 has a second causal variant D and trait 2 has second causal variant C, both with β = log(1.25). In the non-shared causal variant setting (A+D, C), all causal variants have β = log(1.25). Traits 1 and 2 have correlation 0.4 and were both measured on all individuals, unless otherwise specified. When proportion missing data and trait correlation vary, sample size is 3000. The region has 345 SNPS and was simulated to mimic the LD structure of the IL2RA region, 10p-6030000-6220000 (GRCh37/hg19). Results are based on 300 replications.