Going along with what Phil is saying, perhaps you are taking a log of a very very small number, resulting in NaNs. Or, alternatively, columns have a very small norm, so when you normalize, it is effectively a division by zero?
On Tue, Aug 2, 2011 at 2:57 PM, Phil Steitz <[email protected]> wrote: > On 8/2/11 12:24 PM, dan wrote: > > I have been using the OLSMultipleLinearRegression class successfully > > for a while now, but I am having trouble in my current application. > > > > The code is very simple, and looks like this: > > > > OLSMultipleLinearRegression regression = new > OLSMultipleLinearRegression(); > > regression.setNoIntercept(true); > > regression.newSampleData(ys, z_bars); > > double [] new_eta = regression.estimateRegressionParameters(); > > > > When I run this code with my current data, all of the regression > > coefficients come back as NaNs. > > > > In the input data, the z_bars are vectors that have been normalized to > > sum to 1, and the ys are the logs of the "true" response variables (I > > am trying to reproduce the results from a research paper, in which it > > was claimed that logging the response variables made them more > > normally distributed, resulting in a better fit). Is there something > > wrong with my setup? It seems like, even if the logged data is not > > very linear, that it should still be possible to obtain some OLS fit, > > even if it is a poor one. Any help would be appreciated. > > Are you sure there are no NaNs in your input data? > > Phil > > > > Thank you, > > Dan > > > > --------------------------------------------------------------------- > > To unsubscribe, e-mail: [email protected] > > For additional commands, e-mail: [email protected] > > > > > > > --------------------------------------------------------------------- > To unsubscribe, e-mail: [email protected] > For additional commands, e-mail: [email protected] > >
