Average ACF and PACF with panel data

[Luca Gagliardone]

Hi,

is there a command that extends corrgram to panel data?

If not, do you have suggestions on how to compute the autocorrelation and partial autocorrelation plots and statistics in a panel setting?

Thanks in advance.

[Clyde Schechter]

I'm not aware of any. I think you have to iterate corrgram over the values of the panel identifier. -corrgram- leaves the autocorrelations behind in r() both as scalars for each lag and as a row matrix. So it goes like this:

clear*
webuse grunfeld


capture program drop one_company
program define one_company
    tsset year
    corrgram invest
    matrix AC = r(AC)
    matrix PAC = r(PAC)
    local company = company[1]
    clear
    matrix combo = AC \ PAC
    svmat combo, names(col)
    gen statistic = cond(_n == 1, "AC", "PAC")
    gen company = `company'
    exit
end

runby one_company, by(company) status
order company statistic, first

Note: -runby- is by Robert Picard and me, and is available from SSC.

[Luca Gagliardone]

Thanks Clyde.

Do you know if there is a sensible way to "average" the ID-level ACF and PACF to obtain aggregate statistics? Assuming all the IDs follow the same data-generating process.

Perhaps bootstrap?

[Clyde Schechter]

Sorry, I don't know. Perhaps somebody else who does will jump in.

[Luca Gagliardone]

Thanks Clyde!