Removing Bias, Uncovering Bias: Classification Trees and Federal Sentencing Data

• In Event: Poster Session: Presidents and Executive Politics & Public Policy
  In Poster Session: Poster Session: Public Policy

Abstract

Recent developments in statistical modeling have produced sophisticated techniques to isolate the true effects of an indicator on a dependent variable. We apply this approach to federal sentencing data. In this paper, we utilize classification trees to approximate a random trial using observational data. We are able, as a consequence, to identify a more accurate measure of the effects of felon characteristics, including gender, race, number of children, and severity of previous offense, on the length of prisoner sentences. For example, this approach permits us to take every African-American member in the data set and match on the output node of a classification tree to find a Caucasian member that was demographically similar. Prior to these operations, we find that the average African American offender receives 32 more months in prison than a similar white offender. After using the matching procedures above, we find the average African American sentence is still 6 to 8 months longer than that of a white feeling, despite all other factors of the cases being virtually identical. All told, the use of classification trees allows for a more refined and accurate description of why certain offender receive harsher sentences than others.

Author

• Howard Bartlett Sanborn, Virginia Military Institute