Thanks to Kit Baum, a new command called rangerun (with Nick Cox) is now available on SSC. To use it, you need to install rangestat as well. To install both, type in Stata's Command window:
rangerun runs a user-supplied Stata program for each observation in the sample. At each pass, the data in memory is cleared and replaced with observations that fall within the interval bounds specified for the current observation. The user's program is run and results are collected from new variables left in memory when the program terminates without error. rangerun provides the same functionality as rangestat except that, instead of using Mata functions, statistics are calculated by running a user-written Stata program that may contain any number of Stata statistical commands.
rangerun provides a general solution to any type of rolling window problem. To contrast the mechanics of rolling compared to rangerun, here's the last example from rolling's help file:
With rolling, execution times increase exponentially as the data size increases. With rangerun (and rangestat), execution times increase linearly with the number of observations in the sample. If you double the number of observations, the run time for rangerun will be twice as long. For large problems, rangerun will be orders of magnitude faster than rolling.
Everything that can be done with statsby can also be done with rangerun. Since statsby's execution times also increase exponentially as the data size increases, rangerun will be orders of magnitude faster for large data problems. See the help file for a fully spelled out example that shows how to specify the interval bounds so that rangerun only runs the user's program once per group.
Because rangerun manages all the data in Mata and runs the user's program from Mata, there is very little overhead. The commands in the user's program do not require in or if qualifiers since the data in memory is only populated with observations in range for the current observation. So even with the most efficient Stata loop, rangerun will be significantly faster (and more convenient). Here's an example that calculates a regression on a rolling window of 7 years (including the current observation) and stores the constant term. A minimum of 4 observations is required:
The help file for rangerun includes the above examples as well as many others. Each example is completely self-contained and can be run easily via a click to run link.
Code:
ssc install rangerun ssc install rangestat
rangerun provides a general solution to any type of rolling window problem. To contrast the mechanics of rolling compared to rangerun, here's the last example from rolling's help file:
Code:
webuse lutkepohl2, clear tsset qtr rolling ratio=(r(mean)/r(p50)), window(10): summarize inc, detail list in 1/10 clear all program myprog if _N < 10 exit summarize inc, detail gen ratio = r(mean)/r(p50) end webuse lutkepohl2, clear rangerun myprog, interval(qtr -9 0) use(inc) list qtr inc ratio in 1/19, sep(0)
Everything that can be done with statsby can also be done with rangerun. Since statsby's execution times also increase exponentially as the data size increases, rangerun will be orders of magnitude faster for large data problems. See the help file for a fully spelled out example that shows how to specify the interval bounds so that rangerun only runs the user's program once per group.
Because rangerun manages all the data in Mata and runs the user's program from Mata, there is very little overhead. The commands in the user's program do not require in or if qualifiers since the data in memory is only populated with observations in range for the current observation. So even with the most efficient Stata loop, rangerun will be significantly faster (and more convenient). Here's an example that calculates a regression on a rolling window of 7 years (including the current observation) and stores the constant term. A minimum of 4 observations is required:
Code:
clear all
webuse grunfeld
local nobs = _N
gen alpha = .
quietly forvalues i = 1/`nobs' {
capture regress invest kstock if company == company[`i'] &inrange(year, year[`i']-6, year[`i'])
if _rc == 0 & e(N) >= 4 replace alpha = _b[_cons] in `i'
}
program myprog
if _N < 4 exit
regress invest kstock
gen alpha_rr = _b[_cons]
end
rangerun myprog, interval(year -6 0) by(company)
assert alpha == alpha_rr
Comment