For multivariate failure time data, estimators of parameters in marginal hazard regression models can be proposed under working independence assumption. However these estimators are often low efficient. To improve efficiency of estimators, it should be considered to utilize the potential correlation between failure types, and we can use the weighted partial likelihood to obtain the weighted estimators. It is difficult to get an optimal weight for estimators of parameters because of "curse of dimension". Fan, Zhou, Cai and Chen proposed three different criteria to choose the weights based on the variance of component estimators of parameters, then it can give out the combination weighted partial likelihood estimators. In this paper, we establish asymptotic normality of the combination weighted partial likelihood estimators and obtain their asymptotic covariances. Some simulations are conducted to perform the combination weighted partial likelihood estimators and working independence estimators. A HIV-1 RNA data set from an AIDS clinical trial study for comparing treatments is analyzed using the proposed estimators.