The aim of this study was to compare threshold sire model TS), threshold sire-maternal grandsire model TS-MGS) and linear sire-maternal grandsire model L) for genetic analysis of dystocia. Threshold models were based on Bayesian approach. In the study, a total of 19439 dystocia records from Holsteins in USA were used. The effects of calving year-season, sex of calf, parity of dam, sire of calf and herd effects were included in all models and also maternal grandsire effect of calf was included in only sire-maternal grandsire models. Variance-covariance estimates were greater in threshold models than in linear model. Estimates of heritability (+/- SE) of dystocia based on direct genetic effects h(D)(2)) and maternal genetic effects h(M)(2)) were 0.18 +/- 0.004 and 0.14 +/- 0.004 from TS-MGS and 0.12 +/- 0.003 and 0.09 +/- 0.003 from L, respectively. Heritability estimates based on direct genetic effects from TS was 0.20 +/- 0.009. Genetic correlation between direct and maternal genetic effect were -0.087 +/- 0.006 from the TS-MGS and -0.253 +/- 0.010 from L. It was concluded that the threshold models were better than the linear model in the analysis of dystocia. The higher heritability estimates on the underlying scale from threshold models should allow greater genetic improvement than those using linear model estimations.