ab99
New Member
Posts: 8
|
Post by ab99 on Sept 10, 2014 14:57:39 GMT
Dear All,
I am using the "--reml-pred-rand" option to predict the random effects by the BLUP method. I am not sure if I completely understand this method correctly. I have the following two questions:
1) Based on the manual, the output file "*.indi.blp" will output two columns for each GRM matrix: an intermediate variable, and the total genetic effect. Can somebody explain what this "intermediate variable" is? I am guessing what we care about is the "total genetic effect" column, so can I completely ignore the "intermediate variable" column?
2) Is it meaningful to calculate the correlation (or more precisely, the correlation square) between the observed phenotype and the "total genetic effect" column? I am thinking this would be the R2 between the observed phenotype and the predicted genetic effect, so it should be <= the total variance explained by all the SNPs (i.e., V_G/V_p), which is the main result from GCTA. Is this right?
Thank you very much! ab99
|
|
|
Post by Zhihong Zhu on Sept 11, 2014 6:41:27 GMT
Re 1) Yes. Re 2) Yes, the R^2 will be similar to hg^2 in .hsq file. But SE, which is important, is unable to obtain from that.
|
|
ab99
New Member
Posts: 8
|
Post by ab99 on Sept 11, 2014 15:30:05 GMT
Re 1) Yes. Re 2) Yes, the R^2 will be similar to hg^2 in .hsq file. But SE, which is important, is unable to obtain from that. Hi Zhihong, Thank you for your reply! Just a follow-up with 2): what you said was what I originally thought. However, we've looked at 5 quantitative traits, among which 4 has R^2 (calculated as I mentioned in my original thread) significantly larger than hg^2 (i.e., V_G/V_p) in .hsq files. This does not make sense to us. Do you have any thoughts on this? Also, I appreciate if you can explain a bit what the "intermediate variable" is in the .indi.blp files. Thanks, ab99
|
|
|
Post by Jian Yang on Sept 12, 2014 9:52:46 GMT
Re 1) The prediction analysis needs to be done in an independent sample, i.e. using the SNP BLUP effects to compute a predictor of in a new sample.
Re 2) "intermediate variable". You can ignore this. It is only used to compute SNP BLUP.
|
|
ab99
New Member
Posts: 8
|
Post by ab99 on Sept 16, 2014 2:52:21 GMT
Re 1) The prediction analysis needs to be done in an independent sample, i.e. using the SNP BLUP effects to compute a predictor of in a new sample. Re 2) "intermediate variable". You can ignore this. It is only used to compute SNP BLUP. Hi Jian, Thank you very much for your reply! I understand that the prediction analysis is better to be done in an independent sample. But my question is if I do want to get a predicted phenotype in the original data set (i.e. get the y_hat), can I do that? Specifically, can I use 1) the "total genetic effect", or 2) the original phenotype minus "residual effect" as the predicted phenotype? I may be completely wrong about this, but I do hope to get an answer from you. Thank you again! ab99
|
|
|
Post by NR on Sept 30, 2014 9:53:21 GMT
Hi Jian and ab99,
I did the same observation: when I calculated the BLUP in the GREML analysis sample, the explained variance is much higher than the SNP-heritability. I don't understand this, because it should be the same? (if I understand Chapter 26 of Lynch & Walsh correctly). It does not happen with all phenotypes. For height I get more or less the same amount of explained variance.
Thanks!
|
|