*Thread Reply:* that being said, it can be interesting to compare the factor structure in phenotype and genetic space when possible. For example, if you are examining a set of traits that are all within UK Biobank it should be relatively straight forward to fit the model for the genetic correlation matrix using Genomic SEM and to fit a model using standard approaches to the phenotypic correlation matrix. You can then do something like eyeball the differences between standardized factor loadings to get a sense of if/where the factor structure shifts. A great example of this is from the Genomic SEM g-factor paper (de la fuente et al., [2021]) where they plot the two sets of loadings against each other in the attached figure
*Thread Reply:* @Michael Neale thanks. The pheWAS analogue might have been confusing, let's forget about the SNPs.
Long story short - i want to compare the personality profiles for phenotypes like it is done in LDSC. I did a crude comparison in this paper , but I'd like to incorporate the SEs somehow. Right now, the p values assume N=30 (30 personality facets) https://www.nature.com/articles/s41562-019-0752-x
Video instruction on how to copy the files and get set for the practical
Hi everyone this is the practical worksheet in PDF format, copy the files using: "cp -r /faculty/andrew/GenomicSEM_practical/ ./" alternatively download the files from box: https://utexas.app.box.com/s/sounavy84gwygj0j2askcyaoo2ostbgu its a light set of files so shouldnt crash the system if you all copy at the same time on Monday but feel free to download earlier.
*Thread Reply:* so we (almost exclusively) use the DWLS estimator after having evaluated both in the initial genomicSEM paper (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6520146/). It should not matter too much if all your GWAS are similarly powered, but I realise thats not always the case. IFfyour GWASes do vary in sample size it might be worth it to try both and see whether your conclussions remain unchanged regardless.
*Thread Reply:* it's a great question! since QSNP captures deviations from the factor model, it can be informative to do something like plot the unstandardized factor loadings from the common factor model that does not include the individual SNPs against the SNP-phenotype betas from the GWAS sumstats for the particular SNP that was significant for Q (i.e., not going through the work of pulling the independent pathways estimates from a Genomic SEM model, though they should be similar). within the scatterplot, if you then draw a line of best fit through the estimates a QSNP significant SNP will typically show one or two univariate GWAS estimates that strongly deviate from the line indicating that this trait is driving the QSNP effect. For example, in the attached scatter plot from the g-factor paper from de la Fuente et al. (2020) they see that reaction time (RT) strongly deviates from the line; this reflects a SNP that has a particularly strong association with RT, but RT does not load strongly on the g-factor, so it does not fit the model and is identified as significant for QSNP. if you have a lot of QSNP effects, what I have done is create a table that lists the univariate GWAS betas and Z-statistics for those SNPs, in which case there tends to be some effects that stand out as the outlier in the bunch that are clearly driving the QSNP effect. I've then created a scatterplot for just the top 4 or 5 most significant QSNP effects.
just putting at the front of the messages here for the Session B crew that the files for the Genomic SEM tutorial can be copied using: "cp -r /faculty/andrew/GenomicSEM_practical/ ./" the files can also be download from box: https://utexas.app.box.com/s/sounavy84gwygj0j2askcyaoo2ostbgu. the pdf for the tutorial is also attached. see everyone in a few hours!
*Thread Reply:* We updated the model as it did not converged
*Thread Reply:* no it wasn't covered so I'm glad you asked! That warning reflects the fact that the genetic covariance (S) matrix is non-positive definite for this set of traits, so it was smoothed to positive definite prior to running the model. In this case, the largest difference in the genetic covariance matrix pre and post-smoothing is very small (.0001) so it's not something to be overly concerned about. We write on the github towards the bottom of one of the wiki pages (https://github.com/GenomicSEM/GenomicSEM/wiki/3.-Models-without-Individual-SNP-effects) that if the difference is > .025 this would be an interpretive concern and since smoothing at that level is going to oftentimes be caused by less well-powered GWAS being included in the model you might consider excluding those to see how estimates change.
*Thread Reply:* Hi Alex! I would suggest saving the full model output for a handful of SNPs that are producing that warning and comparing to the model output for the model without individual SNP effects. If it doesn't look like it's producing estimates that are any different than the model without individual SNPs, then that is some indication that this may be a relatively benign warning since the model without SNPs is not producing that warning about the non-positive definite theta matrix. The most common culprit for a theta (i.e., residual) matrix type warning or error is a negative residual variance, which you can check when looking at the runs for the individual SNPs. This can also occur in the specific context of estimating residual covariances. This reflects a somewhat less common situation where you are estimating residual covariances between some variables that have very little residual variance to begin with after accounting for the variance explained by the factors. My sense is the userGWAS model is then throwing this error when the variance explained by the SNP via the factor just pushes that theta matrix to be non-positive definite. In the absence of negative residuals, a misspecified model, or really bizzare model estimates I would consider this a safe warning to ignore (particularly because you do not see the warning in the model without SNPs.
For your question about the p-values > .05 for the GWAS-by-subtraction model, someone else ran into that same issue (https://github.com/GenomicSEM/GenomicSEM/issues/38) and found that the culprit was misspecified arguments for the sumstats function. You might check that the p-values from your sumstats output are not also attenuated first, which would be some indication that the arguments are off there.