As explained in Method Overview, the LDAK weightings are intended to correct for short-range SNP-SNP correlations caused by LD. When the dataset contains strong population structure or closely related samples, extensive long-range correlations may also be present. Ideally the weightings should not correct for these correlations. Therefore, when calculating weightings using a highly-related and/or structured dataset, we recommend adding --decay YES and providing a value for the halflife (in kb) using the option --half-life.
The impact of this is that, having calculated the squared-correlation between two SNPs, LDAK will then multiply this value by exp(-a d), where d is the distance in kb between the two SNPs and a=log(2)/halflife. When two SNPs are close to each other, exp(-a d) is close to 1, reflecting that we believe most of their observed correlation is due to LD; but when two SNPs are distant, exp(-a d) is close to 0, reflelcting that most of their observed correlation is due to the structure and/or relatedness. Note that by default, d is measured in kb, but if window-cm is used, d will be measured in centiMorgans (so halflife should also be specified in centiMorgans).
Finally, we emphasize that for most human association studies, there is no need to model the decay of LD; we think it is only necessary when there is considerable structure and/or relatedness, such as that observed in animal or plant studies, or possibly very isolated human populations (and even then, the impact on estimates is usually fairly slight).
