why regstats give an error when I try to build a quadratic model

Hello

I am trying a very simple model for a parabola with noise

>> Y=sin([0:0.01:pi])+rand(1,315)/1; >> X=[1:315]+rand(1,315); >> % X is my data and Y is considered to be the response >> regstats(Y,X,'quadratic')

returns

Warning: The exact method is impossible. Will be forced to use approximate method. > In stats/private/pvaluedw at 46 In dwtest at 74 In regstats at 352 In regstats at 155

If I inverse

>> s=regstats(X,Y,'quadratic')

>> s.rsquare

ans =

   1.3575e-04

it works but then the rsquare is close to zero while it is a simple parabola and the model should be efficient. I expect a rsquare close to 1 ?

Any help for the two different issues is welcome

Vincent

Thanks in advance

Vincent

Answer 1

[Tom Lane]

The warning is just an indication that the significance of the Durbin-Watson statistic is outside the range of a table of exact values, so the quantity is being approximated. That should not be a problem for you.

When I try your example I get rsquare around 0.5 each time. If you plot X vs. Y you will see there is substantial scatter, so I would not expect a value near 1. I can't reproduce your result showing a value near 0.