Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
From | ymarchenko@stata.com (Yulia Marchenko, StataCorp) |
To | statalist@hsphsun2.harvard.edu |
Subject | st: Re: xtmixed slope cross validation |
Date | Thu, 01 Dec 2011 10:39:49 -0600 |
Petra <petroula.proitsi@kcl.ac.uk> uses -predict, reffects- after -xtmixed- on a test dataset which does not include the original data used in estimation and receives a "no observations" error: > I have predicted the slopes for memory decline in a sample of demented > patients - in order to validate my model I split my sample into a "training" > and "test" set with the idea to fit a model on the training set and then > predict the slopes of the "test"set and compare these to the slopes of the > whole sample. > > in the help for the predict command says that you can run a model using one > dataset (I used the "train"), use a different dataset (then used the "test") > and then used predict (in my case I typed: predict r_slope1 r_inter1, > reffects)- However, I get that message: no observations. Does anyone have > any idea of how to predict the slopes for the "test" set, using the slopes > of the "train" set? Unlike linear prediction, fitted values, etc., the computation of the estimates (BLUPs) of random effects requires the original data used in estimation be in memory. We will improve the error message to clarify this. To obtain BLUPs of random effects for the test data, Petra can do the following: . use <entire_data> // load training+test data . generate byte training = ... // create indicator for the training sample . xtmixed ... if training // run -xtmixed- on the training sample . predict ... if !training, reffects // predict REs for test sample -- Yulia ymarchenko@stata.com * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/