An empirical Bayes (EB) pairwise FST estimator was previously introduced and evaluated for its performance by numerical simulation. In this study, we conducted coalescent simulations and generated genetic population structure mechanistically, and compared the performance of the EBFST with Nei's GST, Nei and Chesser's bias‐corrected GST (GST_NC), Weir and Cockerham's θ (θWC) and θ with finite sample correction (θWC_F). We also introduced EB estimators for Hedrick’ G’ST and Jost’ D. We applied these estimators to publicly available SNP genotypes of Atlantic herring. We also examined the power to detect the environmental factors causing the population structure. Our coalescent simulations revealed that the finite sample correction of θWC is necessary to assess population structure using pairwise FST values. For microsatellite markers, EBFST performed the best among the present estimators regarding both bias and precision under high gene flow scenarios (FST≤0.032). For 300 SNPs, EBFST had the highest precision in all cases, but the bias was negative and greater than those for GST_NC and θWC_F in all cases. GST_NC and θWC_F performed very similarly at all levels of FST. As the number of loci increased up to 10 000, the precision of GST_NC and θWC_F became slightly better than for EBFST for cases with FST≥0.004, even though the size of the bias remained constant. The EB estimators described the fine‐scale population structure of the herring and revealed that ~56% of the genetic differentiation was caused by sea surface temperature and salinity. The R package finepop for implementing all estimators used here is available on CRAN.