The accuracy of imputation was high for all scenarios and for all chromosomes, although the low-density panel had only a density equivalent to 384 markers across the whole genome (Table 1), which is approximately one SNP every 8 to 9 centimorgans. The accuracy of imputation was slightly lower for chromosome Z than for the two autosomes. Both the accuracy and the differences in accuracy between chromosome Z and the two autosomes were affected by the genotyping status of the immediate ancestors of the test individuals.
Scenario SC1, which had all ancestors genotyped at high-density, had a higher accuracy of imputation than SC2, which had only male ancestors at high-density and female ancestors at low-density, and than SC3, which had only great-grandparents genotyped at high-density and all other ancestors at low-density. Scenario SC4 was a more extreme case of SC2, in which the test individuals were three additional generations removed from their female ancestors that were genotyped at high-density. Despite this, the accuracy of imputation did not appear to be worse for the autosomes in SC4 compared to SC2, but it was slightly lower in SC4 for chromosome Z (still within the bounds of sampling error due to SC4 having a large sampling variance). The genotyping status of the immediate ancestors of the testing individuals has been shown to be an important factor in determining imputation accuracy for autosomal chromosomes in other species, e.g. [1, 8]. In this study, this trend was also observed for chromosome Z.
The accuracy of imputation on chromosome Z was much more variable across individuals than it was for the two autosomes. With the exception of SC3, for which it was 0.08, the standard deviation of accuracy was at most 0.02 for the autosomes. For chromosome Z, the variability was large and increased with the increasing difficulty of the imputation scenario. For SC3 and SC4, the standard deviations of accuracy were 0.10 and 0.22 respectively. Thus, although the mean accuracy was lower for chromosome Z than for the autosomes, some individuals had high accuracy, while others had low accuracy. The low accuracy in certain individuals for chromosome Z was due to the higher rate of Mendelian errors for chromosome Z in comparison to the autosomes, which in turn may be caused by lower reliability of genotyping platforms for markers on sex chromosomes than for autosomes. AlphaImpute checks for consistency between the genotype information and the pedigree. Individual SNP genotypes are set to missing in both the parent and the offspring if they conflict. This results in removal of SNP that exceed a threshold for the proportion of individuals having that SNP missing from the full imputation involving the use of haplotype information. For autosomes, these SNP are imputed using single-locus segregation analysis  but for sex chromosomes they are naively imputed as the parent average genotype. For chromosome Z, particularly for SC4, a greater proportion of SNP were excluded from the analysis than for the autosomes (Table 1).
The good performance of imputation of genotypes on chromosome Z for some individuals can be explained by the fact that imputation of markers on sex chromosomes is less challenging than on autosomes for a number of reasons. Heterogametic individuals are phased de-facto, thus avoiding the possibility of phasing errors for these individuals, other than due to genotyping errors. The highly accurate phasing of heterogametic individuals helps in surrogate definition and partitioning in the long-range phasing step, and in the haplotype library phasing step of AlphaImpute for homogametic individuals. Imputation of the gamete received from the heterogametic parent by a homogametic individual is also de-facto. Computation time for imputation for all chromosomes was of the order of minutes for this dataset but was faster for chromosome Z than for the autosomes, because the phasing was computationally less demanding and genotype probabilities were not calculated for the reasons aforementioned.
Using the imputation approach outlined in this paper, which was specifically designed to impute genotypes on sex chromosomes, did not always outperform the standard autosomal imputation procedure of AlphaImpute. Using the autosomal approach yielded imputation accuracies of 0.97 ± 0.04, 0.92 ± 0.08, 0.84 ± 0.11, and 0.89 ± 0.04 for SC1, SC2, SC3, and SC4, respectively. The autosomal approach was better than the specifically designed approach for SC1 but worse for the three remaining scenarios. Good performance of the autosomal approach for imputation of sex chromosomes may be due to the pedigree haplotype library imputation step, which is independent of the mode of inheritance. However, in the presence of highly accurate genotyping of sex chromosome markers and high-density genotypes on close ancestors of the individuals to be imputed, the imputation approach outlined in this paper is expected to be more optimal than the standard autosomal imputation approach implemented in AlphaImpute.
The pseudo-autosomal region of chromosome Z and chromosome W was ignored in this study due to the difficulty in both identifying and sequencing SNP in this region. If these can be reliably identified, they can be treated as an artificial autosomal chromosome in AlphaImpute. Compared to chromosome Z, chromosome W is very small, contains only a handful of known genes  and has very few known SNP reported in Assembly 2 of the chicken genome.