Creating IRF plots after ARIMA

[Mia Giuriato]

Hello,

For reasons I cannot figure out, I am able to create IRF plots after an ARIMA regression using the varbasic command, but not using the arima and irf commands. My data and code is as follows.

Data:

* Example generated by -dataex-. To install: ssc install dataex
clear
input double(year injuries_million_hrs)
1983 0
1984 0
1985 22.963167080003675
1986 20.50819302311273
1987 81.9957772174733
1988 18.352664806929965
1989 17.903500134276253
1990 35.12654337249943
1991 21.644084671659236
1992 23.458759500797598
1993 62.786463238525776
1994 42.98394549635711
1995 0
1996 0
1997 17.462673535318256
1998 0
1999 15.519515791107317
2000 12.670256572695598
2001 0
2002 12.90955565309442
2003 0
2004 0
2005 0
2006 11.513176830882946
2007 0
2008 0
2009 17.809439002671414
2010 15.291922806373673
2011 16.838418535731126
2012 15.557180416621293
2013 26.243275160740062
2014 0
2015 0
2016 48.19044865307696
2017 0
end

Code:
tsset year
varbasic injuries_million_hrs, lags(1/3)
^this creates an IRF plot without issue

tsset year
arima injuries_million_hrs, ar(1/3)
irf create naaame, set(naaame) replace
irf graph irf
^this returns the error code "nothing to graph r(198)"

The reason I would like the arima + irf combination to work is because my full dataset is panel data with N=25102 and T between 1 and 35 years. I am able to run the following arima model on the full panel dataset:
arima injuries_million_hrs, arima(0,1,3) condition
but I get the same "nothing to graph" issue when I try and create an irf plot after the arima regression.

Thank you in advance for any help you are able to provide.

[Scott Merryman]

For reasons that are not clear to me, it appears that the long variable name is the problem. You should pass this along to Technical Support (https://www.stata.com/support/faqs/t...-tech-support/) The code below works:

rename injur hrs
tsset year
arima hrs, ar(1/3)
irf create arima, replace  set(myirf1)
irf graph irf

[Mia Giuriato]

Thank you so much Scott! I will let Technical Support know.

I have another question. I am hoping to look at the effect of a particular exogenous shock on hrs. When I create the plots, it seems like I am looking at the effect of an unspecified shock to hrs on hrs. How do I specify that a particular shock I have data for should be used?

Here is the data including the shock:

clear
input double(year hrs shock)
1983                  0 0
1984                  0 0
1985 22.963167080003675 0
1986  20.50819302311273 0
1987   81.9957772174733 0
1988 18.352664806929965 0
1989 17.903500134276253 0
1990  35.12654337249943 0
1991 21.644084671659236 0
1992 23.458759500797598 0
1993 62.786463238525776 0
1994  42.98394549635711 1
1995                  0 0
1996                  0 0
1997 17.462673535318256 0
1998                  0 0
1999 15.519515791107317 0
2000 12.670256572695598 0
2001                  0 0
2002  12.90955565309442 0
2003                  0 0
2004                  0 0
2005                  0 0
2006 11.513176830882946 0
2007                  0 0
2008                  0 0
2009 17.809439002671414 0
2010 15.291922806373673 0
2011 16.838418535731126 0
2012 15.557180416621293 0
2013 26.243275160740062 0
2014                  0 0
2015                  0 0
2016  48.19044865307696 0
2017                  0 0
end

If I run:

tsset year
arima hrs shock, ar(1/3)
irf create arima, replace set(myirf1)
irf graph irf

The plot says that both the impulse and the response are hrs. Am I simply controlling for "shock" here, rather than seeing the effect of this particular shock on hrs?

[Scott Merryman]

The shock is a one unit change. You could rescale it by just multiplying the one unit change by the scalar factor.

With the AR(1/3) it is pretty easy to calculate the irf by hand. The general form is irf(j) = ar1*irf(j-1) + ar2*irf(j-2) + ar3*irf(j-3)

arima hrs, ar(1/3)

irf create arima, replace  set(myirf1)
irf table irf

local IRF0 = 1
local IRF1 = _b[ARMA:L.ar]*`IRF0'
local IRF2 =_b[ARMA:L.ar]*(`IRF1') +_b[ARMA:L2.ar]*`IRF0'
local IRF3 =_b[ARMA:L.ar]*(`IRF2') +_b[ARMA:L2.ar]*`IRF1' +_b[ARMA:L3.ar]*`IRF0'
disp `IRF1'
disp `IRF2'
disp `IRF3'

[Jean-Claude Arbaut]

The problem in your first question seems to come from the following lines in arfimairf_create.ado, which is called by irf_create.ado in the case of an ARIMA or ARFIMA model.

qui gen str13 response = "`e(depvar)'"
qui gen str13 impulse = "`e(depvar)'"

A quick test shows that a variable name (in -arima-) longer that 13 characters leads to the error.
If I change these lines to create str25 variables instead, the code runs without error, and apparently correctly.
However, I am not sure why they were str13 in the first place and whether or not it can break further code.

HTH

Jean-Claude Arbaut