We consider the problem of estimating the latency of a feasible but unused Autonomous System-level path on the Internet. This problem arises in evaluating the overhead incurred by censorship and surveillance circumvention schemes that alter the Internet routing infrastructure, and the cost of attacks against such schemes. Since these paths are not advertised by the current routing infrastructure, they cannot be directly measured by end hosts, leading researchers to estimate the costs indirectly. Using traceroute measurements of observed Internet paths, we measure the accuracy of the two methods used in the literature to date, finding that these methods have poor accuracy and correlation, explaining as low as 3% of the variation in observed AS path latencies, and at most 42%. We also describe an improved method that can balance accuracy and path coverage. At the high end our estimator can explain up to 83% of variation in observed AS path latencies, while still being able to achieve 56% when maximizing the number of paths able to be estimated.