Updating linear regression efficiently when adding observations and/or predictors in R I would be interested in finding ways in R for efficiently updating a linear model when an observation or a predictor is added. biglm has an updating capability when adding observations, but my data are small enough to reside in memory (although I do have a large number of instances to update). There are ways to do this with bare hands, e.g., to update the QR factorization (see "Updating the QR Factorization and the Least Squares Problem", by Hammarling and Lucas from 2008), but I am hoping for an existing implementation.
 A: If the algorithm you are looking for is indeed something like Applied Statistics 274, 1992, Vol 41(2) then you could just use biglm as it does not require you to keep your data in a file.
A: There is rank one QR update function in matlab here
that saves you a factor $p$ in the complexity of updating the coefficients of a p-variate linear regression.
Despite searching for days a couple of months ago, i've not been able to find an equivalent in R (beware there are many qr.update functions in cran but when you look under the hood they're just fake --i.e. they call lm.update all the same).
Update: try in the source of the package 'leaps'. In the R-source, you will find a function 'leaps.forward', which calls a FORTRAN routine 'forwrd', located inthe /src of the package which seems to implement rank 1 QR update. 
A: Why don't you try the update capability of the linear model object
update.lm( lm.obj, formula, data, weights, subset, na.action)

Take a look at this links


*

*For a general explanation of the update function:


http://stat.ethz.ch/R-manual/R-devel/library/stats/html/update.html 


*

*For a particular explanation about update.lm:


http://www.science.oregonstate.edu/~shenr/Rhelp/update.lm.html
A: I've been also looking since long time for an equivalent to the matlab qr update, leaps seems a nice way!
In R, you could look at the recresid() function in package strucchange, that will give recursive residuals when you add an observation (not variable!). My guess is that this will require little modification to obtain recursive betas (the betar in the code?). 
