LD Decay

The LDAK weightings are intended to correct for short-range SNP-SNP correlations caused by LD. When the dataset contains related individuals or strong population structure, extensive long-range correlations may also be present. Ideally the weightings should keep these correlations intact. To achieve this, when analysing a highly-related / structured dataset, we advise adding –decay YES when calculating weightings to turn on LD decay modelling. (By default, LD Decay is turned off, which is appropriate for most human association studies; LD Decay is normally only required for animal/plant studies.)

When calculating weightings, LDAK first constructs a SNP-SNP similarity matrix C. With LD decay turned off, values of this matrix simply equal the observed correlation squared between pairs of SNPs. When LD decay is activated, LDAK instead multiplies each correlation squared value by a decay function. The idea is that when two SNPs are proximal, this function takes value close to one, so correlation squared values are barely affected. But when two SNPs are distant, the value is close (or equal) to zero, meaning that longer-range correlations are essentially ignored when calculating weightings. For two SNPs distance basepairs apart, this function is
exp(-a distance),       if (distance < maxdistance)
0,                                 otherwise.
The value of a is set by specifying the functions halflife, such that exp(-a halflife)=0.5.
The default value for halflife is 1Mbp, and for maxdistance is 3xhalflife.

Note that regardless of whether LD Decay is on or off, when populating C,  correlation squared values <mincor are automatically set to zero, to appreciate that even predictors in linkage equilibrium will have a non-zero correlation. mincor has default value 0.01.

In summary:
–halflife <base_pairs> (default 1Mbp) – set halflife.
–maxlife <base_pairs> (default 3x halflife) – sets maximum range of LD decay.
–mincor <float> (default 0.01) – sets squared correlation cutoff.