Thanks Bryan, this is very helpful. The main impetus for using SNP count rather than region size, for me, was that the sample I will be trying to impute is ~20,000 subjects and using a default region size of 5mb leads to a reasonable number of regions to impute that won't be handled by my computers capabilities.

If you would like to explain further why setting chunk boundaries based on reference variants is preferable I would be very interested (for example, what would be the extremes of this? how few study variants would indicate possible problems?)...Therefore controlling region interval by reference variant count might lead to a finer control on running time and memory usage.

A related question to my specific imputation is whether imputing all subjects together produces just a helpful results as splitting the sample by subject, imputing and then combining the subjects back into one dataset for association testing. Really I would just like the option of using the default IMPUTE2 settings of 5Mb to impute and I believe splitting the sample by subjects would allow this as running times and memory tends to increase with sample size.  Can you provide any further guidance on that approach. My main two worries with it are a) the accuracy of the imputation will be decreased due to a smaller sample sample size and b) when combining the subjects from two imputations runs things might get sticky with SNPs that are difficult to impute having genotypes that correlate with which imputation.

I also notice that sometimes the buffer has less input SNPs, so perhaps I could control this as well.

Just to finally check, when you suggest controlling reference variant count, the benefits of this are mainly for memory usage and running times compared to controlling input SNPs count, i.e. there is no benefit to imputation accuracy,