Table 2 Performance of seven approaches to genetic association analyses for trait change in simulated and empirical longitudinal data

From: Analyzing longitudinal trait trajectories using GWAS identifies genetic variants for kidney function decline

 

Simulated data

Empirical data

 

UKB scenario for eGFR-trajectories

UKB eGFR-trajectories

Approaches

T1E [%] (CI)

Power [%] (CI)

Bias [%]

T1E [%] (CI)

“Power” in 9

“Bias” in 9

Identified in 595

Without singletons

Difference model

5.1 (4.7, 5.5)

12.0 (11.4, 12.7)

0.6

4.9 (4.5; 5.3)

4/9

0.0

2 (2)

Time model RI&RS

4.7 (4.2, 5.1)

25.6 (24.7, 26.4)

0.0

4.8 (4.4; 5.2)

7/9

8.2

7 (2)

Age model RI&RS

4.8 (4.4, 5.3)

31.4 (30.5, 32.4)

0.7

5.0 (4.5; 5.4)

8/9

Reference

9 (2)

Age model RI&RS uncorr.

7.5 (7.0, 8.1)

38.8 (37.8, 39.8)

0.7

8.8 (8.2; 9.3)

8/9

0.7

13 (3)

Age model RI-only

32.4 (31.4, 33.3)

59.0 (58.1, 60.0)

0.6

43.1 (42.2; 44.1)

9/9

16.9

98 (41)

BLUPs&LinReg

5.3 (4.9, 5.8)

44.8 (43.9, 45.8)

−37.7

5.0 (4.6; 5.5)

8/9

−38.5

13 (3)

Including singletons

Age model RI&RS

5.2 (4.8, 5.7)

44.1 (43.1, 45.1)

0.5

4.8 (4.3; 5.2)

8/9

0.2

12 (6)

  1. We compared seven approaches (“Methods” section and Supplementary Table 2) regarding type I error, power, and bias: six approaches analyze individuals with ≥2 trait assessments over age/time (no singletons, i.e., individuals with =1 trait assessment), the 7th approach repeats age model RI&RS including singletons. Simulations were based on distributions of age, global/random trait effects, and random error as in UKB 350K for eGFR and simulated genotypes (EAF = 30%; 10,000 simulation runs; “Methods” section and Supplementary Table 3). This scenario covers a setting as in UKB (~50% singletons) for trajectories of a trait like eGFR with prounouced age effect on trait. We show estimates of type I error (T1E), power, and bias from 10,000 simulation runs. In empirical analyses using UKB 150K (no singletons) or 350K (including singletons), we show permutation-based type I error, proxies of power and bias (based on 9 SNPs known for eGFR-decline10), and number of SNPs identified with Pdecline < 0.05/595 (Pdecline< 5 × 10−8) among the 595 SNPs known for association with cross-sectional eGFR8.
  2. Simulated data: T1E = proportion of SNPs with Pdecline < 0.05 across 10,000 simulated SNPs given zero true effect on decline, βdecline = 0 (95% CI using SEs from exact binomial test); Power = proportion of SNPs with Pdecline < 0.05 across 10,000 simulated SNPs given true effect on decline, βdecline = −0.025 (95% CIs derived from SEs using exact binomial test); Bias = relative bias of effect estimates given true effect on decline, βdecline = −0.025, derived as average (across 10,000 simulation runs) of (\({\hat{{{{\rm{\beta }}}}}}_{{{{\rm{decline}}}}}\)−βdecline)/βdecline; Empirical data: T1E = proportion of SNPs with Pdecline < 0.05 among 10,000 permutation-based “null-SNPs” using eGFR-trajectories of UKB individuals (95% CIs using SEs from exact binomial test); “Power” in 9 = proportion of SNPs directionally consistent with Pdecline < 0.05 in UKB 150K (for age model RI&RS: additionally in UKB 350K) among the 9 SNPs known for eGFR-decline10; “Bias” in 9 = relative deviation of effect estimate from reference among the 9 SNPs known for eGFR-decline10 derived as average across the 9 SNPs of (\({\hat{{{{\rm{\beta }}}}}}_{{{{\rm{decline}}}}}-{\hat{{{{\rm{\beta }}}}}}_{{{{\rm{decline}}}}({{{\rm{reference}}}})}\))/\(\,{\hat{{{{\rm{\beta }}}}}}_{{{{\rm{decline}}}}({{{\rm{reference}}}})}\); Identified in 595 = number of SNPs with Pdecline < 0.05/595 (in parentheses: with Pdecline < 5 × 10−8) among 595 SNPs tested.
Back to article page