Scalable methods for estimating marginal coalescent trees across the genome present new opportunities for studying evolution and have generated considerable excitement, with new methods extending scalability to thousands of samples. Benchmarking of the available methods has revealed general tradeoffs between accuracy and scalability, but performance in downstream applications has not always been easily predictable from general performance measures, suggesting that specific features of the ARG may be important for specific downstream applications of estimated ARGs. To exemplify this point, we benchmark ARG estimation methods with respect to a specific set of methods for estimating the historical time course of a population-mean polygenic score (PGS) using the marginal coalescent trees encoded by the ancestral recombination graph (ARG). Here we examine the performance in simulation of seven ARG estimation methods: ARGweaver, RENT+, Relate, tsinfer+tsdate, ARG-Needle, ASMC-clust, and SINGER, using their estimated coalescent trees and examining bias, mean squared error (MSE), confidence interval coverage, and Type I and II error rates of the downstream methods. Although it does not scale to the sample sizes attainable by other new methods, SINGER produced the most accurate estimated PGS histories in many instances, even when Relate, tsinfer+tsdate, ARG-Needle and ASMC-clust used samples ten or more times as large as those used by SINGER. In general, the best choice of method depends on the number of samples available and the historical time period of interest. In particular, the unprecedented sample sizes allowed by Relate, tsinfer+tsdate, ARG-Needle, and ASMC-clust are of greatest importance when the recent past is of interest-further back in time, most of the tree has coalesced, and differences in contemporary sample size are less salient.