I want to do a linear regression on the response Y (attached)with a predictor vector X. Therefore, I used Matlab polyfit function: [p,s,mu]=polyfit(X,Y,1) but it returns p=[NaN,Inf]
I have Inf values in vector Y. Is it the cause?
Best Answer
Yes. infs in your data will surely cause NaNs, and essentially garbage results. Drop them out first if you expect something meaningful.
y = [rand(10,1);inf];
x = rand(11,1);
p = polyfit(x,y,1)
p =
NaN Inf
Surely you don't expect anything reasonable to come out of such a computation with INFs in it. What would the slope of such a line be with some data points at infinity? Of course, any such point at infinity would have infinite leverage.
It's not a bug. What's happening is you're multiplying Inf * 0, which is undefined (it's equivalent to trying to calculate 0/0). Try something like this, it will just set the diagonal elements to Infinity without multiplying the off-diagonal elements by Infinity as well:
Like that, each value of "a" is correlated to each value of "b", but applying the formula of the correlation, the correlation of two single numbers is NaN. To compute the correlation correctly, traspose the input vectors
Best Answer