Null alleles are alleles that for various reasons fail to amplify in a PCR assay. The presence of null alleles in microsatellite data is known to bias the genetic parameter estimates. Thus, efficient detection of null alleles is crucial, but the methods available for indirect null allele detection return inconsistent results. Here, our aim was to compare different methods for null allele detection, to explain their respective performance and to provide improvements. We applied several approaches to identify the 'true' null alleles based on the predictions made by five different methods, used either individually or in combination. First, we introduced simulated 'true' null alleles into 240 population data sets and applied the methods to measure their success in detecting the simulated null alleles. The single best-performing method was ML-NullFreq_frequency. Furthermore, we applied different noise reduction approaches to improve the results. For instance, by combining the results of several methods, we obtained more reliable results than using a single one. Rule-based classification was applied to identify population properties linked to the false discovery rate. Rules obtained from the classifier described which population genetic estimates and loci characteristics were linked to the success of each method. We have shown that by simulating 'true' null alleles into a population data set, we may define a null allele frequency threshold, related to a desired true or false discovery rate. Moreover, using such simulated data sets, the expected null allele homozygote frequency may be estimated independently of the equilibrium state of the population.