Reliability estimation and optimization are computationally prohibitive, especially in higher dimensions. Often metamodels are used to address the computational cost involved in repeated expensive function evaluations in optimization. However, metmodels demand more samples in higher dimensions and performance also deteriorates in higher dimensions with limited samples. An attractive alternate option is to use dimension reduction techniques such as active subspace which helps to alleviate the computational cost and also permit reduced sampling. Usually metamodels are built and optimization is performed in the reduced dimension. Reliability estimation requires propagating uncertainties through the metamodels which is not cumbersome if the uncertainty information is known. If the uncertainty is not known, the underlying distribution is assumed which could lead to flawed decisions if the assumption is wrong. We propose an approach to propagate the uncertainties in the reduced dimension. Though large samples are required to estimate reliability, the number of times reliability is estimated is less in the reduced dimension. Demonstration of the proposed approach on numerical examples will be discussed in the presentation.