There is a graph in the link https://odelaneau.github.io/GLIMPSE/rsquare_eur.html. This graph shows the imputation accuracy of certain MAF bins. Now I understand that $r^2$ is the correlation between imputed genotype and true genotype. What I don't understand is how they calculate this $r^2$ correlation. Can anyone suggest to me how I can define the r^2 correlation between true genotype and imputed genotype? I suppose it depends on the individual who defines his own $r^2$. So, it may differ for the different algorithms. However, different aspects or different $r^2$ definitions will help me understand and represent this graph slightly differently.
This is done by downsampling. Take the 1000 genomes, set some genotypes as missing
./., impute them using GLIMPSE, then measure correlation between the genotype dosages in the imputed dataset and the same sample at high coverage - in this case, the 1000 genomes sequenced to 30x coverage by the NY Genome Centre.