The issue addressed in this paper is how sentencing disparity should be assessed when sentencing decisions are made under sentencing guidelines systems. Much of the debate on this issue has focused on using specific values from within the sentencing grid (e.g., minimum recommended sentence) or on using interaction terms in regression models in order to capture the nonadditive effects of offense severity and prior record on length of sentence. The proposed alternative of using quantile regression models offers several advantages over these traditional methods of analyses of sentence length. Quantile regression models allow for an examination of the effects of case and offender characteristics across the full distribution of sentence lengths for a given sample of offenders. This enables researchers to address such issues as whether offender characteristics, such as race or offense severity, have the same effect on sentence length for the 10 percent of offenders who receive the longest sentences. This paper illustrates the application and interpretation of quantile regression models by using 1998 sentencing data from Pennsylvania. The findings show that the effects of case and offender characteristics vary across the distribution of sentence lengths. This means that traditional linear models that assume a constant effect fail to capture significant differences in how case and offender characteristics influence sentencing decisions. This paper also briefly discusses the application of quantile regression models to jurisdictions without sentencing guidelines, to other issues in criminology and criminal justice that rely on the analysis of highly skewed distributions, and to the testing of theories that either directly or indirectly imply nonconstant relationships. 4 tables, 9 figures, and 47 references