[R] Using predict on a biglm object returns NA

[Thomas Lumley]

On Fri, 20 Mar 2009, Pradheep K E wrote:

> When I run predict on the model output on a new data (for testing) or on the
> same data, I get only NA's. I'm able to run predict with some other models
> constructed with biglm. One reason I suspect is that the model itself has a
> few undefined terms (NA's). I'm wondering if there's any way to ignore the
> undefined terms (treat them as 0).

No, there isn't.  Well, you can easily get extract the coefficients and do the same calculations as predict.biglm does -- they aren't that hard. But the NA means that it wasn't possible to work out from the training data what the coefficient was, so setting it to zero (rather than, say, 42) doesn't seem like a reliable strategy. You'd probably be better off refitting a model that can actually be estimated from the data.


     -thomas

Thomas Lumley			Assoc. Professor, Biostatistics
tlumley at u.washington.edu	University of Washington, Seattle