Posted by BigML’s predictive models are great for helping us discover when “something happens” in a dataset. As an example, let’s do a bit of digging in the legal world.
Recent books on the supreme court, such as Jeffery Rosen’s The Supreme Court and Jeffery Toobin’s The Nine, have discussed how the U. S. Supreme Court is becoming increasingly polarized and less willing than courts from decades previous to find consensus among its justices.
Assuming for a moment that this is true, one interesting question is, “What are the types of decisions on which there is the most disagreement?” Luckily, we have the data to answer it:
http://scdb.wustl.edu/data.php
Harold J. Spaeth has compiled a list of post-war U. S. Supreme Court decisions, over 8300 in total. With each decision, Spaeth has associated a number of attributes: The type of decision that was made, the type of law involved, the issue involved, the lower court decision, whether the decision was conservative or liberal in nature, and so on. He also gives the size of the minority vote in each case. We can make a simple classification problem out of this by labeling each decision with a “contentious = yes” if the decision had a minority vote size of three or more and “contentious = no” if the minority vote size was 2, 1, or zero.
We can use BigML’s predictive models to help us identify particular types of cases that tend to have contentious votes. Generating a tree from this data, we can mouse over various nodes at higher levels of the tree. If the node “expects” a contentious vote, it means that a majority (but not all) of the cases in the decision tree that reach this node are contentious, and this might be an interesting node to examine further.
At the second “level” of the tree, we see only a few nodes that expect a contentious vote. One of these is cases where the issue is “First Amendment” in which the decision was to “Affirm” the decision of the lower court:
First amendment cases where the decision is to reverse the lower court are not as controversial:
But why stop there? The beauty of decision trees is the ability to dig deeper and find insights of greater specificity. Expanding the contentious path above, we can see that first amendment cases where the decision was to affirm, and the lower court’s decision was to reverse are classified positively only when the basis for the decision was Judicial Review. Certainly, one can imagine judicial review of laws governing first amendment freedoms being a very contentious topic on the court.
Some of the other places where votes tend to be contentious are in the areas of Due Process and Civil Rights, both very hot-button, and the disagreements tend to happen most when the court invokes judicial review of state law. Why might this be?
One possibility is that during the middle and middle-late part of last century, the court began to use due process to protect rights not explicitly protected in the constitution, often protecting practices (the distribution of contraception, for example) prohibited by state law. Again, it is easy to imagine the justices at odds about this new use of due process.
By contrast, we see some of the least disagreement among justices in areas like Unions, Economic Activity, and (perhaps unsurprisingly) Federal Taxation. All very important, but generally not as ideologically charged as the First Amendment.
And so we are able, using this data, to get a fairly rich picture of when supreme court justices are disagreeing with one another. Whether your own intuition is confirmed or invalidated by this tree, it’s nice to have data to point us in the right direction.
The Official Blog of BigML.com 