I've been running rolling regressions using C.F. Baum's -rollreg-
routine. The routine is great, but it does require that the data
contains no gaps. Unfortunately my data (grouped by companies) has quite
a few gaps. Annoyingly, I may have many years of data for a company with
only 1 missing observation. However, as far as I can see I have to
delete all the data from this company. e.g.
. tsset id year
panel variable: id, 1 to 3
time variable: year, 1980 to 1997
. rollreg y x1 x2, move(4) stub(rr)
Number of gaps in sample: 2
Observations with preceding time gaps
----------------------------------
Record | id year
----------+-----------------------
4 | 1 1983
9 | 1 1988
----------------------------------
sample may not contain gaps
r(198);
So to run this file I need to remove id == 1 from the sample. Note that
when id == 1 & year == 1986 I actually have enough obs to estimate a
regression with 4 years data. In practice a relatively small number of
gaps are resulting in a drastic loss of data.
Can anyone suggest how I can overcome (or minimize) this problem and
still use -rollreg-?
One solution I tried is to write my own version of -rollreg- that can
deal with gaps. Based on help I've received from people on this list I
was able to come up with a program (shown below) that worked on a small
"testing" dataset. e.g.
Seems to work. But when I use -rollreg2- on a larger data set, problems
emerge. The program runs for a while producing results and then the
internal call to -regress- starts to repeatedly return an error code
(_rc = 902). I can't find a list of the meaning of these codes, but
experimentation seems to suggest it is a memory problem. But I don't
understand why memory should be a problem as prior to running the
program, -describe- reveals 95% memory free. I'm at a loss to explain
this. Any help would be greatly appreciated.
Here is a copy of -rollreg2-
capture program drop rollreg2
program rollreg2 , byable(onecall)
version 8.2
syntax varlist(min=2), MOVE(integer) STUB(string)
qui generate rc = .
local by "`_byvars'"
tempvar gr
if _by() == 1 {
qui egen `gr' = group(`by')
}
else {
qui gen `gr' = 1
}
qui levels `gr', local(groups)
local k: word count `varlist'
local depvar: word 1 of `varlist'
qui forvalues i = 2/`k' {
local v: word `i' of `varlist'
local vr: word `i' of `varlist'
confirm new variable `stub'_`v'
confirm new variable `stub'_se_`v'
generate `stub'_`v' = .
generate `stub'_se_`v' = .
local reglist "`reglist' `v'"
}
qui {
confirm new variable `stub'_cons
confirm new variable `stub'_se_cons
confirm new variable `stub'_r2
confirm new variable `stub'_RMSE
confirm new variable `stub'_sd_residual
confirm new variable `stub'_N
generate `stub'_cons = .
generate `stub'_se_cons = .
generate `stub'_r2 = .
generate `stub'_RMSE = .
generate `stub'_sd_residual = .
generate `stub'_N = .
}
local max = 0
local min = 1
qui count
local total = r(N)
quietly {
foreach g of local groups {
count if `gr' == `g'
local max = r(N) + `max'
if `=`max' - `move' + 1' < 0 | `=`max' - `min'' < `move' - 1 {
local min = `max' + 1
continue
}
forvalues i = `min'/`=`max' - `move' + 1' {
local j = `i' + `move' - 1
if `j' <= `total' {
capture regress `depvar' `reglist' in `i'/`j'
replace rc = _rc in `j'
if _rc == 0 & e(N) == `move' {
tempvar res
predict `res' if e(sample), res
summarize `res'
replace `stub'_sd_residual = `res' in `j'
replace `stub'_r2 = e(r2_a) in `j'
replace `stub'_RMSE = e(rmse) in `j'
replace `stub'_N = e(N) in `j'
replace `stub'_cons = _b[_cons] in `j'
replace `stub'_se_cons = _se[_cons] in `j'
forvalues l = 2/`k' {
local v: word `l' of `varlist'
replace `stub'_`v' = _b[`v'] in `j'
replace `stub'_se_`v' = _se[`v'] in `j'
