There's a discord between the variant calls released as part of Phase 1 and Phase 3 by about 2 million variants. Most recently the 1000 genomes consortium released the following information to account for this. So its entirely your preference on how you'd want to treat a high MAF variant that was may be called in Phase 1 but deleted in Phase 3.

sample_dropout - sites missing because the samples carry the ALT alleles were dropped in phase3;
not_called_in_p3 - sites missing because they were not called initially by any phase3 caller;
failed_svm_filter - sites missing because they failed phase3 SVM filter;
ovl_p3_larger_event - sites missing because they overlap with phase3 released indels or SVs; or they overlap with unfiltered mvncall set;
not_in_p3_shapeit2_GL_files - sites missing because they are not included in the phase3 shapeit2 genotype likelihood files;
possible_patchup_candidate - missing SNP sites that cannot be well explained. They may be possible patch up candidates;
indel_unknown - missing indel sites that cannot be explained by sample dropout;
sv_unknown - missing structural variation sites that cannot be explained by sample dropout;