Changes to Least squares solve function in JMP14

yvesprairie

Hi all,

 

I scratched my head trying to find out why a script I posted here was not functioning on someone else's computer when it was working flawlessly on mine. It turns out that the order of the X and Y variables in the Least Squares Solve function in JSL has been inverted in JMP 14 relative to JMP 13. Naturally, this renders my initial script unusable for anyone using a JMP version <14, which is a bit annoying. Are these subtle changes documented somewhere?

 

Best regards, Yves

Milo

That would be very frustrating. I'm sorry to hear that's how you discovered the change in the Least Squares Solve() JSL function.

 

Swapping the order of parameters in a JSL function like this is extraordinarily rare. When a change disrupts compatibility with previous versions, we list it in the Release Notes.

 

We're in an active development environment at JMP, but do our best to also maintain backward compatibility. You're correct that the order of parameters changed from JMP 13.0 (when the function was introduced) to 14.0. Not only that, but by default, Least Squares Solve() in 14.0 includes an intercept term. 

 

There were a number of reasons motivating the change. First, Least Square Solve() was introduced (by me) in JMP 13.0, so it was not in circulation for long. Second, I switched the order and the default intercept term in tandem so that in nearly all use cases, Least Squares Solve() should result in a script not running rather than in incorrect numbers, the latter of which would be particularly hard to discover. Third, in 14.0, I introduced the JSL function Linear Regression(), which essentially solves the same problem as Least Squares Solve(), but with a more user-friendly access to standard output from a linear regression (e.g. standard errors, t-tests, R^2, etc.). With this new function, it made more sense to order the parameters as y, X, since it solves a y=Xb problem. Because of the similarities between these two functions, we elected to have them behave in the same way, with the parameters in the same order.

 

I went back and forth on the decision to switch the ordering for Least Squares Solve() so that it matched Linear Regression(), but I ended up making the call on switching it, since for JMP 14 and on, it seemed cleaner and more intuitive. 

 

If you really want a script to run in either JMP 13 or JMP 14 using Least Squares Solve(), you could do something like:

Names Default To Here(1);
// generate data<BR />random reset(123);
x = j(10,2,random normal()); 
beta = [1, 2];
y = x * beta + j(10,1,random normal(0,.2) );
<BR />// check version
jmpver = num(left(jmpversion(),2));
if(jmpver >= 14,
	Least Squares Solve(y, X, <<nointercept), <BR />  // where in latest versions, the parameter order is switched <BR />  // and an intercept is added by default
	/*else if*/ jmpver == 13,
	 Least Squares Solve(X, y)  <BR />  // this function was introduced in JMP 13.0 with limited functionality
);