[R] panel data unit root tests

jukka ruohonen jukka.ruohonen at helsinki.fi
Fri Jan 6 22:56:08 CET 2006

When finally got some time to do some coding, I started and stopped right 
after. The stationary test is a good starting point because it demonstrates 
how we should be able to move the very basic R matrices. I have a real-
world small N data set with 


Thus, a good test case. For first id I was considering something like this:

lag <- as.integer(lags)
lags.p <- lags + 1
id <- unique(group)
id.l <- length(id)
y.l <- length(y)
yid.l <- length(y)/id.l
  if (lag > yid.l -2) 
        stop("\nlag too long for defined cross-sections.\n")

#for (i in id) {
  lagy <- y[2:yid.l]
      lagy.em <- embed(lagy, lags)
  id.l <- length(id)
  dy <- diff(y)[1:yid.l-1]
      dy.em <- embed(dy, lags)
#     }
print(levinlin(ws, year, id, lags = 3))

Couldn't figure the loop over units out but with N = 1 the data 
transformation seemed to work just fine. Now we should pool the new 
variables within the panel and regress y over yt-1 and dy-t1 +....+ dy-t-j 
with, say, BIC doing the job for d's (H0: y-1 ~ 0) for each in the panel. 

Now the above example puts the right-hand on columns, and if we are dealing 
with panel models in general, we should store the new variables together 
with dX's, which should then give clues to IV estimator with e.g. 
orthogonal deviations, e.g. k <- y ~ yt-1 + x + as.factor(id)). So one 
confusing part is the requirement of some big storage base for different 
matrices doing the IV business with lags/levels - the amount of instruments 
can be enormous with possibly calculation problems in a GMM dynamic panel 
estimator a la Arellano & Bond. Therefore one should code the theoretically 
relevant instruments beforehand with various transformation matrices. Thus, 
should I start to study something that can be done with the newly added 
SparseM package? 


Jukka Ruohonen.

More information about the R-help mailing list