What does mapping quality mean, and why do variant callers filter on it?

Why

The SAM specification defines MAPQ as -10 log10 of the probability that the mapping position is wrong, rounded to an integer, with 255 meaning not available. It is separate from base quality. A read can have perfect base calls and still be ambiguous about where it belongs.

If a read fits two places equally well, the chance its chosen position is wrong is one half, which corresponds to a MAPQ of about 3. Such reads cluster in segmental duplications, pseudogenes and repeats, the regions the Genome in a Bottle stratifications flag as difficult. Small differences between paralogous copies then look like heterozygous variants when reads from both copies pile onto one.

Filtering on MAPQ removes those reads from the evidence. The scale differs between aligners, so a threshold that works for one tool does not carry over directly. Filtering has a cost too.

In a duplicated gene, removing low-MAPQ reads can leave no coverage at all, turning a false positive into a silent false negative, so those regions should be reported as not assessable rather than as variant-free.