I have no direct experience with such horrific models, but your formula is
a mess and Google suggests the biglm package with ffdf.
Specifically, you should convert your discrete variables to factors before
you build the model, particularly since you want to use predict after the
fact, for which you will need a new data set with the exact same levels in
the factors.
Also, your use of I() is broken and redundant. I think formulas
lny ~ id + year + x1 + I(x1^2) + x2 + I(x2^2)
or
lny ~ id + year + x1^2 + x2^2
would obtain the intended prediction results.
--
Sent from my phone. Please excuse my brevity.
On June 17, 2017 11:24:05 AM PDT, Miluji Sb <milujisb at gmail.com> wrote:
Dear all,
I am running a panel regression with time and location fixed effects:
###
reg1 <- lm(lny ~ factor(id) + factor(year) + x1+ I(x1)^2 + x2+ I(x2)^2
,
data=mydata, na.action="na.omit")
###
My goal is to use the estimation for prediction. However, I have 8,500
IDs,
which is resulting in very slow computation. Ideally, I would like to
do
the following:
###
reg2 <- felm(lny ~ x1+ I(x1)^2 + x2+ I(x2)^2 | id + year , data=mydata,
na.action="na.omit")
###
However, predict does not work with felm. Is there a way to either make
lm
faster or use predict with felm? Is parallelizing an option?
Any help will be appreciated. Thank you!
Sincerely,
Milu
[[alternative HTML version deleted]]