Genotype imputation takes advantage of shared haplotypes between a target individual and haplotypes within a reference panel. The process of imputation itself is the exercise of estimating partially unobserved haplotypes in the target individual and as such, the results of imputation are based not only on local evidence but also on non-local evidence, such as reads on linked sites.
For a more detailed description of the statistical model underlying imputation from low-pass sequencing, please see the Supplementary Note in Wasik et al. 2021 or Rubinacci et al. 2021.