Estimating effective population size from linkage disequilibrium: severe bias in small samples