Variation in sentencing outcomes represents the actions of a number of members of the criminal justice system. To isolate the part of the variation that is due to the discretion of the judge (or other sentencing agent, such as a prosecutor), one can model the sentencing guidelines themselves. Such a model captures any non-linearity in the sentencing grid. In practice, modeling the guidelines rather than legal factor scores (as is common in the literature) means that more of the variation that race and legal factors share in common will be attributed to the racial status of the offender. Using data from Maryland, we find that African Americans have 20% longer sentences than whites, on average, holding constant age, gender, and recommended sentence length from the guidelines. We find more judicial discretion and greater racial disparity than is generally found in the literature. Moreover, when we begin to try to explain this discretion, we find that judges tended to give longer sentences (relative to those recommended by the guidelines) to people in the part of the guidelines grid with longer recommended sentences (who are disproportionately African American) than they gave to people in the part of the grid with lower recommended sentences.
There is a large literature on racial discrimination in sentencing outcomes that begins with the disproportionate representation of African Americans and other minorities in prison.1 IMAGE FORMULA10But black offenders tend to have other characteristics, such as longer criminal histories, that are considered by most observers of the sentencing process to be legitimate reasons for harsher outcomes. Researchers have dealt with this correlation between race and legally relevant factors by dividing racial disparity into two parts: warranted and unwarranted disparity. Warranted disparity is the variation in sentence outcomes due to legally relevant factors, such as criminal history, crime type, and crime severity. Unwarranted disparity (Stolzenberg & D'Alessio 1994) is the variation in sentencing outcomes that can be reasonably identified as being the sole result of race or other extra-legal factors (e.g., gender), after all legally mandated sentencing factors are taken into account.2 Partialing out warranted and unwarranted disparity is generally done using standard regression techniques.
In general, the body of research on unwarranted disparity in sentencing outcomes has shown that legal factors have large effects on sentencing outcomes. Furthermore, there is little evidence of direct racial discrimination once these legal factors are included in the statistical models. This evidence led Sampson and Lauritsen (1997:362) to conclude that "there is little evidence that racial disparities reflect systematic, overt bias on the part of criminal justice decision makers (as a whole)." In response, some researchers have argued for an interactive, rather than additive, model to examine whether racial discrimination occurs indirectly through certain court contexts (e.g., plea vs. jury trial) or individual characteristics (e.g., employed vs. unemployed). This can be done by including interaction terms in multivariate models (Miethe & Moore 1986), testing for differences among age and gender subgroups of each race group (Steffensmeier et al. 1998), or estimating separate models for each racial group (Albonetti 1997).3
This type of research has emphasized greater statistical rigor and model specification over the development of a theoretical framework for understanding the sentencing process (Albonetti 1991). But the statistical parsing of unwarranted and warranted disparity may conflict with the desire to understand the key factors driving the sentencing process. For example, Dixon (1995) tested her organizational context perspective on sentencing using the same basic "racial disparity" statistical model (described previously) as Albonetti (1991) used to test her bounded rationality model of judicial decisionmaking. …