st: Constructing confidence intervals for a sum of forecasts

[Ian Sue Wing <isw@bu.edu>]

Dear StataListers,

I am generating a dynamic forecast of two variables, x and y, after 
running a vector autoregression, and I want to construct the forecast 
confidence interval for the sum of their first differences, D.x + D.y. 
My question is what is the correct way to do this.

I can easily generate the se's of the first differences of individual 
variables using -fcast compute, diff-. Let's call these se_Dx_hat and 
se_Dy_hat. It is simple to compute the se of their sum if I assume 
independence:

se(Dx_hat+Dy_hat) = sqrt( se_Dx_hat^2 + se_Dy_hat^2 )

However, I am not clear about how to correctly handle potential 
correlation in the first-differences, which this expression omits.

From first principles,

D.x + D.y = x(t) - x(t-1) + y(t) - y(t-1)

var(D.x + D.y) = var(D.x) + var(D.y) + 2 cov(D.x, D.y)

Now,

var(D.x) = var(x(t)) + var(x(t-1)) - 2 cov(x(t), x(t-1)) = se_Dx_hat^2,

with a similar expression for y. What I don't have is the covariance term:

2 * [ cov(x(t),y(t)) - cov(x(t),y(t-1)) - cov(x(t-1),y(t)) + 
cov(x(t-1),y(t-1)) ] = 2 * [ 2 * cov(x(t),y(t)) - cov(x(t),y(t-1)) - 
cov(x(t-1),y(t)) ]

Should I just assume that each of the terms in square brackets is a 
constant, given by, first, the data, and second, the appropriate 
elements in e(V) generated by my var? If not, can anyone recommend an 
alternative way of doing this calculation?

Thanks,

-i

--
Ian Sue Wing                      675 Commonwealth Ave., Boston MA 02215
Associate Professor               Tel: (617) 353-5741
Dept. of Geography&  Environment  Fax: (617) 353-5986
Boston University                 Web: http://people.bu.edu/isw

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/