Minimac Diagnostics

minimac is a tool for imputation of missing genotypes into phased haplotypes. At the end of each run, minimac generates summaries of imputation quality and stores those in a .info file.

Basic Descriptors

Marker and Allele Labels

The first three columns in the .info file list marker name and alleles for each marker. Typically, the most common allele will be listed first, but this is not guaranteed.

Estimated Allele Frequency

The next column in the .info file lists the estimated frequency of allele 1 - this corresponds to the average number of imputed copies of allele 1 for each individual, divided by two.

Estimated Imputation Accuracy (ImpRsq)

Frequency information is followed by an estimate of the squared correlation between imputed genotypes and true, unobserved genotypes. Since true genotypes are not available, this calculation is based on the idea that poorly imputed genotype counts will shrink towards their expectations based on population allele frequencies alone; specifically $2p$ where $p$ is the frequency of the allele being imputed.

Currently, minimac uses the following definition:

${\hat {r}}^{2}={{Var({\mbox{Estimated Counts}})} \over {{\hat {p}}(1-{\hat {p}})}}$

Leave One Out Statistics

To evaluate imputation quality, Minimac hides data for each genotyped SNP in turn and calculates 3 statistics, described below.

looRsq : Estimated R-squared in Leave-One-Out Analysis

This first statistic is calculated by hiding all known genotypes for the SNP, imputing it and then estimating imputation accuracy. It doesn't use the known genotypes for the SNP at all.

empR : Correlation Between Imputed and True Genotypes

Whereas looRsq statistic completely ignores experimental genotypes, this one is based on a comparison of imputed and experimental genotypes. A negative correlation between imputed and experimental genotypes can indicate allele flips.

empRsq : Squared Correlation Between Imputed and True Genotypes

Whereas looRsq statistic reports the estimated imputation accuracy, this one reports the actual imputation accuracy - as estimated by comparing genotypes generated using imputation (after hiding any known genotypes for the marker) and the previously hidden known genotypes. By comparing empRsq and looRsq it should be possible to tell whether estimates of imputation accuracy are well calibrated.