The effective population size (Ne) is a key parameter in evolutionary and population genetics. Single-sample Ne estimation provides an alternative to traditional approaches requiring two or more samples. Single-sample methods assume that the study population has no genetic sub-structure, which is unlikely to be true in wild populations. Here we empirically investigated two single-sample estimators (ONeSAMP and LDNE) in replicated and controlled genetically structured populations of Drosophila melanogaster. Using experimentally controlled population parameters, we calculated the Wright–Fisher expected Ne for the structured population (TotalNe) and demonstrated that the loss of heterozygosity did not significantly differ from Wright’s model. We found that disregarding the population substructure resulted in TotalNe estimates with a low coefficient of variation but these estimates were systematically lower than the expected values, whereas hierarchical estimates accounting for population structure were closer to the expected values but had a higher coefficient of variation. Analysis of simulated populations demonstrated that incomplete sampling, initial allelic diversity and balancing selection may have contributed to deviations from the Wright–Fisher model. Overall the approximate-Bayesian ONeSAMP method performed better than LDNE (with appropriate priors). Both methods performed best when dispersal rates were high and the population structure was approaching panmixia.