In association mapping, haplotype-based methods are generally regarded to provide higher power and increased precision than methods based on single markers. For haplotype-based association mapping most studies use a fixed haplotype effect in the model. However, an increase in haplotype length raises the number of parameters in the model, resulting in low accuracy of the estimates especially for the low-frequency haplotypes. Modeling of haplotype effects can be improved if they are assumed to be random effects, as only one parameter, i.e. haplotype variance, needs to be estimated compared to estimating the effects of all different haplotypes in a fixed haplotype model. Using simulated data, we investigated statistical models where haplotypes were fitted either as a fixed or random effect and we compared them for the power, precision, and type I error. We investigated five haplotype lengths of 2, 4, 6, 10 and 20. The simulated data resembled the Danish Holstein cattle pedigree representing a complex relationship structure and QTL effects of different sizes were simulated. We observed that the random haplotype models had high power and very low type I error rates (after the Bonferroni correction), while the fixed haplotype models had lower power and excessively high type I errors. Haplotype length of 4 to 6 gave the best results for random model in the present study. Though the present study was conducted on data structure more frequent in livestock, our findings on random vs. fixed haplotype effects in association mapping models are applicable to data from other species with a similar pedigree structure.