0

I have a threshold VAR. Since nonlinear irf's are not possible in R, I want to get my way around it with IRF's.

Following are the codes for my multivariate time series and then the TVAR

complete=ts.intersect(GDPdiff2,Inflationdiff2,Euribordiff2,CISSdiff2)

tvarcomplete= TVAR(complete, lag=4, nthresh=1, thDelay=1, thVar=complete[,1], trim=0.15)

I want to divide the existing time series with the 4 variables into two (one VAR with the values of all variables which have real GDP growth > 0 and one with < 0 ) and the calculate the corresponding IRFs.

   SPLITGDPup <- complete[(GDPdiff2>(0))]

did not work.

Heike
  • 107
  • 2
  • 11
  • 1
    Make it easier for others to help you: Please provide a minimal reproducible example; read [here](https://stackoverflow.com/questions/5963269/how-to-make-a-great-r-reproducible-example) for more details – duckmayr Nov 10 '17 at 22:15

1 Answers1

0

I have a two part answer to your question: (1) I'll show you how to subset complete correctly, (2) I'll point you to a resource for impulse response functions of nonlinear VARs.

Subsetting Multivariate Time Series

There are a couple of issues with the way you attempted to subset complete: (1) you were essentially subsetting a matrix like it was a vector, and (2) you were using a logical vector of length length(GDPdiff2) when you wanted a logical vector of length nrow(complete). To illustrate, I first make some example data since you didn't provide yours:

# set the seed for reproducibility
set.seed(123)
# make example data
ts1 <- ts(rnorm(10))
ts2 <- ts(rnorm(8))
complete <- ts.intersect(ts1, ts2)
complete

Time Series:
Start = 1 
End = 8 
Frequency = 1 
          ts1        ts2
1 -0.56047565  1.2240818
2 -0.23017749  0.3598138
3  1.55870831  0.4007715
4  0.07050839  0.1106827
5  0.12928774 -0.5558411
6  1.71506499  1.7869131
7  0.46091621  0.4978505
8 -1.26506123 -1.9666172

Now we'll try subsetting it like you did:

# attempt to subset like yours
complete[ts1 > 0]

[1]  1.55870831  0.07050839  0.12928774  1.71506499  0.46091621 -0.55584113
[7]  1.78691314  0.49785048 -1.96661716

That resulted in a vector instead of a matrix. Why? From Hadley Wickham's Advanced R:

Because matrices and arrays are implemented as vectors with special attributes, you can subset them with a single vector. In that case, they will behave like a vector.

Moreover, the elements it pulled from the second column of complete don't line up with the positive elements of the first column. Why? Let's look at the logical vector you'd use to subset:

ts1 > 0
[1] FALSE FALSE  TRUE  TRUE  TRUE  TRUE  TRUE FALSE FALSE FALSE

There are 10 elements, while complete only has 8 rows because you've used ts.intersect which only gives complete cases (ts2 has less observations than ts1). The combination of these two issues are why your subsetting strategy didn't work. Here's how to do it right:

complete[complete[, 'ts1'] > 0, ]

            ts1        ts2
[1,] 1.55870831  0.4007715
[2,] 0.07050839  0.1106827
[3,] 0.12928774 -0.5558411
[4,] 1.71506499  1.7869131
[5,] 0.46091621  0.4978505

Alternatively, if you know and would prefer to use the column number:

complete[complete[, 1] > 0, ]

            ts1        ts2
[1,] 1.55870831  0.4007715
[2,] 0.07050839  0.1106827
[3,] 0.12928774 -0.5558411
[4,] 1.71506499  1.7869131
[5,] 0.46091621  0.4978505

Impulse Response Function for TVAR

I would caution you to do more Googling before you ever make a statement like "Since nonlinear irf's are not possible in R". Almost anything is possible in R, and because of the great R community, so many of the things you could want to do (regarding extant statistical methods) have already been implemented by someone if you know where to look.

In your case, if you had looked at the GitHub repository for the package you're using you'd see that while a generalized impulse response function hasn't been implemented in the tsDyn package, one of the authors has made code for it that you can find here that does this for TVAR results.

duckmayr
  • 16,303
  • 3
  • 35
  • 53