|
Post by Jian Yang on Sept 21, 2015 1:46:31 GMT
For a case-control study, V(G), V(e), Vp, V(G)/Vp are all on the observed scale. V(G)/Vp_L is the estimate of variance explained on the underlying liability scale under a threshold model. On the observed scale (0-1 disease status), the genetic variance can be greater Vp per definition, i.e. if the heritability on the underlying scale (h2L) is high and the disease prevalence is low, it is possible that the heritability on the observed scale (h2O) can be greater than 1. By default, GCTA does not allow any estimate of variance component to be negative. In this case, Ve is constrained at 10-6, so that the estimate of V(G)/Vp is constrained at 0.9999. You could specify the option --reml-no-constrain to allow V(G)/Vp to be greater than 1. However, you need to be cautious that any artefacts between cases and control will be estimated as 'genetic' variance, especially when cases and controls were genotyped separately (e.g. on different plate or at different labs). When using GCTA to analysis a case-control study, very stringent QC on SNPs are required. Please refer to Lee et al (2011 AJHG) for the QC steps and some other technical details of applying the method in case-control studies. For a quantitative trait (which is relatively robust to the artefacts in SNP data as compared to a case-control study), it is likely that your sample size is small so that the estimate varies within a great range (i.e. large standard error). It may also suggest that the true parameter (i.e. variance explained by all SNPs) is relatively large.
|
|