Dear SAS Community,
I apologize for the lengthy message.
As you know, the VARMAX procedure in SAS allows for the computation of the accumulated impulse response in the transfer function and provides standard errors for the long-run response elements.
I am wondering if similar functionality exists in PROC ARIMA (or in HPF procedures). While I understand that VARMAX can be used to model a univariate series, my understanding is that the mathematical formulation of the transfer function in a VARMAX model differs from that in an ARIMA model. Specifically, in an ARIMA model, the transfer function has a specific rational form, where the numerator and denominator polynomials define the distributed lag structure of the input series.
Consider the following scenario: I have a transfer function model in PROC ARIMA where the input variable St is a standard step function, taking the value 0 before a certain time point and 1 thereafter.
The code is:
proc arima data=A;
identify var=Y crosscorr=St;
estimate p=(1)(12) q=1 input=( / (1) st ) method=ml outmodel=Results;
run;
The estimated transfer function weights are given by ω0 (SAS estimates it as ‘NUM1’) and δ1 (SAS estimates it as ‘DEN1,1’), where the accumulated response evolves as follows:
ω0 (initial period response),
ω0*(1+δ1) (next period response),
ω0*(1+δ1+δ1^2) (following period response), etc.
While I can manually compute the accumulated response since PROC ARIMA estimates ω0 and δ1, I am looking for a more efficient ("automatic?") way to obtain these values directly in SAS.
Additionally, PROC ARIMA provides standard errors (and thus, confidence intervals) only for the estimated parameters ω0 and δ1, but in reality, each accumulated response term has its own confidence interval. I am unsure how to obtain these confidence bands in SAS for all these separate added responses. Is there a formula for it?
Any insights on this would be greatly appreciated. Thank you again.