Much has been written about extra-legal factors producing biases in criminal sentencing, such as "are black men sentenced to longer prison terms?" Exploring this question, we must engage in formal causal reasoning, and rephrase it as "does ethnicity cause the length of the sentence?" But, we know that establishing a causal effect without randomized experiments is difficult. And, for the judicial system, all we have is observational data. In this webinar, we explain how Bayesian networks can help formalize causal assumptions, which, however, must be justified purely on theoretical grounds. Given such causal assumptions, a Bayesian network can then facilitate causal inference from observational data.
Koon, Booker, Gall, and Entropy?
Another important angle in this discussion is the fact that criminal justice system is a probabilistic domain, with many elements of uncertainty. Formally quantifying uncertainty (or entropy) is the second major focus of our webinar. Bayesian networks are well-suited for this task, as entropy is a key measure in machine learning networks from data. This can help us understand to what extent characteristics of the crime and the defendant reduce the uncertainty regarding the expected sentence. Of particular interest is how varying degrees of judiciary discretion — as a result of legislative action (e.g., PROTECT Act) and Supreme Court rulings (e.g., Koon, Booker, and Gall) — can be quantified by computing the entropy of sentences.
In this webinar, we examine the data of federal criminal cases related to drug-trafficking received by the United States Sentencing Commission that had sentencing dates between 1996 and 2011.