IRF with I(1) variables? VECM needed?

Here is a question related to IRF, impulse response functions, in the vars package.

My data has following attributes: I do have two log-transformed variables that are tested as I(1). I.e. the two variables are stationary after first differencing and not stationary at level, allowing for "constant and trend". The I(1) variables are found to be cointegrated.

• What do I have to consider when drawing an impulse response function (irf())?
• Can I use my variables at level (which are not stationary) or do I have to use the first-differenced data?
• In case: When do I have to use a VECM?
• VECM stands for vector error correction model, so you do not need to repeat model after VECM. – Richard Hardy Oct 25 '17 at 13:41
• Okay. Will keep it in mind. Let me assure on the procedure: Case 1: Variables are both I(1) and cointegrated: Estimate IRF with VECM. Case 2: Variables are both I(1) but not cointegrated: Estimate IRF with first-differenced variables. Correct? – WolfgangRtry Oct 25 '17 at 14:09
• I think that's right. – Richard Hardy Oct 25 '17 at 14:37