Reads from paired-end and mate-pair libraries are often utilized to find structural variation in genomes, and one common approach is to use their fragment length for detection. After aligning read pairs to the reference, read pair distances are analyzed for statistically significant deviations. However, previously proposed methods are based on a simplified model of observed fragment lengths that does not agree with data. We show how this model limits statistical analysis of identifying variants and propose a new model by adapting a model we have previously introduced for contig scaffolding, which agrees with data. From this model, we derive an improved null hypothesis that when applied in the variant caller CLEVER, reduces the number of false positives and corrects a bias that contributes to more deletion calls than insertion calls. We advise developers of variant callers with statistical fragment length-based methods to adapt the concepts in our proposed model and null hypothesis.