Background Population genetics and association studies usually rely on a set of known variable sites that are then genotyped in subsequent samples, because it is easier to genotype than to discover the variation. This is also true for structural variation detected from sequence data. However, the genotypes at known variable sites can only be inferred with uncertainty from low coverage data. Thus, statistical approaches that infer genotype likelihoods, test hypotheses, and estimate population parameters without requiring accurate genotypes are becoming popular. Unfortunately, the current implementations of these methods are intended to analyse only single nucleotide and short indel variation, and they usually assume that the two alleles in a heterozygous individual are sampled with equal probability. This is generally false for structural variants detected with paired ends or split reads. Therefore, the population genetics of structural variants cannot be studied, unless a painstaking and potentially biased genotyping is performed first. Results We present svgem, an expectation-maximization implementation to estimate allele and genotype frequencies, calculate genotype posterior probabilities, and test for Hardy-Weinberg equilibrium and for population differences, from the numbers of times the alleles are observed in each individual. Although applicable to single nucleotide variation, it aims at bi-allelic structural variation of any type, observed by either split reads or paired ends, with arbitrarily high allele sampling bias. We test svgem with simulated and real data from the 1000 Genomes Project. Conclusions svgem makes it possible to use low-coverage sequencing data to study the population distribution of structural variants without having to know their genotypes. Furthermore, this advance allows the combined analysis of structural and nucleotide variation within the same genotype-free statistical framework, thus preventing biases introduced by genotype imputation.