In the GWAS by subtraction lecture, the COG latent factor was described ‘causing’ the CP observed variable. Given that the CP observed variable is not a direct measure of cognitive performance but rather the summary statistics of a GWAS of cognitive performance, how do we conceptualise this ‘causal’ relationship?
Replying to “In the GWAS by subtraction lecture, the COG latent…”:
It’s part of the theory of latent variables, and why the arrow goes from the circle to the square — the big latent variable is theorized to cause variation in the smaller, more specific indicators (so “cognitive ability” the latent unmeasurable variable that causes test scores, cognitive performance outcomes, etc). As opposed to something like SES, which is caused BY its indicators (like education, income, etc) — in this case, the arrows go from the indicators to the latent variable.
Replying to “In the GWAS by subtraction lecture, the COG latent…”:
Thanks
I have a few questions about the GWAS-by-subtraction model in the cog and noncog paper - can you explain how the latent cog factor is quite genetically similar to CP? In the genomicSEM google group there is a reply to a GenomicSEM user stated that “latent variable Cog is equivalent to the Cognitive Performance… these GWAS are highly correlated with a LDSC genetic correlation of 1 (p=.00).” With this in mind, does the path loading from cog to EA showing a genetic causal effect also represent shared genetic covariance (implying Cog shares genetic variance with EA but not noncog (fixed to zero))? If the path loading represents genetic covariance, can you explain if the strong rg of 1 between CP and Cog also includes the covariance between Cog to EA represented by the path loading?
Replying to “I have a few questions about the GWAS-by-subtracti…”:
The latent Cog variable is exactly the same as the Cognitive Performance GWAS because it was modeled to have all of the variance in cognitive performance explained by this latent factor (with no residual variance).
The rg=1 between CP and Cog should not be thought of as a genetic correlation, but rather a clever way of just ‘moving’ the variance in cognitive performance from one part of the model to another
Replying to “I have a few questions about the GWAS-by-subtracti…”:
Yes, the loading from cog to EA captures its shared genetic covariance.
I did a gwas-by-subtraction. Can you explain why there is chi-square p-value of 1 in latent trait sumstats after doing a GWAS-by-subtraction? Is this because the model is perfectly identified and the model implied matrix does not deviate from the observed matrix?
Replying to “I did a gwas-by-subtraction. Can you explain why t…”:
Yep! technically, you’re just reproducing the observed correlations, and so you’re not “testing a hypothesis” (like Sarah said)