The differences between maximum likelihood (ML) and regression (REG) interval mapping in the analysis of quantitative trait loci (QTL) controlling endosperm traits of F3 seeds were investigated analytically and numerically by simulation. Analytically, the similarity between the conditional probabilities in REG interval mapping and the conditional posterior probabilities in ML interval mapping determines the approximation of REG to ML interval mapping. Therefore, investigating the factors affecting the similarity between the two kinds of probabilities can distinguish differences between the two methods. Simulations are performed to compare the two methods in detecting QTL under different sample sizes, heritability levels, interval sizes and dominance effects. The results show that the approximation of REG to ML nterval mapping might not be good for QTL with large first and second dominance effects. The REG method tended to give estimates with larger mean square error (MSE) and smaller likelihood-ratio test (LRT) statistics. It implies that the ML method tended to be more accurate, precise, and powerful as compared to the REG method, especially, with smaller sample size, lower heritability level and wider marker interval. Another advantage of the ML method over the REG method is that it provided a better estimate of the residual variance. The REG method tended to overestimate the residual variance. Furthermore, the REG method had the problem to estimate the two dominance effects, and the ML method had the ability to estimate these two dominance effects separately.