Correct me if I misunderstood your question, but it seems like you are splitting or combining samples into various sized groups and then running those through GATK.

From the supplemental material of GATK on haplotype calling:

Whereas the initial phase of the algorithm is run per sample, the second stage combines the genotype likelihoods over all samples in order to determine the most likely alternate allele frequency in the cohort. The likelihood for a given set of genotype assignments at a given frequency is simply the product of the genotype likelihoods for each sample given that sample’s assigned genotype (Box 1). We then apply a population genetic prior to the allele frequency likelihoods based on θ, the population specific heterozygosity, and choose the most likely allele frequency and associated genotype assignments. The variant quality score for a polymorphic call is given as -10*log10(probability that the site is actually monomorphic).

Most variant callers use some kind of Bayesian modeling for within-group variation, since being able to model within-group variation significantly increases the likelihood that the remaining variation is due to inter-group variation (usually treatment vs control, tumor vs normal in most experimental designs). Splitting the samples into artificial cohorts of sizes that are not true to the actual data will affect the results of this calculation substantially.