New JMP 19 Graph Builder features in action

[#1]

In my first Raw Data Studies post since the JMP 19 release, I was surprised to find myself already relying on several new JMP 19 features. Since none of these made the main New in JMP 19 article, I thought it would be good to recap them as a meta-blog post. In the original post, I was exploring data on high school and college American football players. At one point, I made a dot plot with two groups on the Y axis indicating whether a player was drafted by the NFL or not. Each dot is a player and the X axis shows their high school rating.

This chart uses several JMP 19 features:

• the "ordinal" jitter direction which moves the dots in the two groups both toward the center. I take responsibility for the cryptic name, which is a reference for the application of this style in an ordinal or nominal logistic fit chart.
• the hexagonal grid dot plot for tighter packing.
• the jitter smoothing, where one stack can take dots from its neighbor based on an optimization criterium that balanced density smoothness and error in the X direction.

The next graph shows the percent of players from selected colleges that were drafted, separating the players by their high school star rating.

The graph is using the Line of Fit element, but the response variable is actually binomial (drafted or not). In that case the confidence interval is binomial, which means it can be imbalanced and it can never go beyond the 0 to 1 interval. Notice South Carolina had only a few five-star players, so the confidence interval is very wide, and all of them were drafted, so the average is 100%, but the confidence interval is also capped at 100%. Also new with JMP 19, the footer text is part of the output when you use right-click > Copy Graph.

Finally, the original post has a smoother showing the trend curve of draft percentage versus player high school rating. Here is a breakdown of that relationship by school.

It makes use of the new p-spline smoother and a couple features it enables:

• The trend curve is constrained to be non-negative while still being smooth.
• The trend curve is constrained to be non-descending. The idea is that a higher player rating should never result in a lower draft likelihood prediction. For example, Florida has only a few players rated at the low end (around 0.8), but two of them got drafted. If unconstrained the trend curve would start much higher and then go down.

For a little more context, here's that same constrained curve for Florida overlaid with the confidence interval for the unconstrained curve.

It's not guaranteed but somehow comforting that the constrained fit is still within the confidence interval of the unconstrained fit.