First difference or seasonal difference in VAR/VECM I have monthly data on house price, rental price, wage index and interest rates. I want to use VAR to produce impulse response function.
Is there any reason why I should use first difference, x(t)/x(t-1) - 1 instead of e.g. x(t)/x(t-12) - 1 for montly time series?
 A: The question is quite specific, but an answer can be based on general principles. If the data generating process happens to be better approximated* by an integrated VAR model than by a seasonally-integrated VAR model, your may be better of with the former when forecasting. 
When could this be the case? Whenever your time series are not seasonally integrated but vanilla-integrated. (Note that seasonal integration is just a special case of seasonality.) Tests of seasonal integration and vanilla-integration as well as macroeconomic theory can help you decide on that.
You may also need to consider (seasonal or vanilla) cointegration to account for possible long-term equilibrium relationships. VECM could be used instead of VAR then. Again, statistical tests and macroeconomic theory can be of help.

*Considering (1) the true vs. modeled functional form and (2) estimation accuracy. According to Shmueli "To Explain or to Predict", (2010)), when interested in explanation, you may care more about getting the functional form right and less about estimation accuracy. Meanwhile, if your interest is in forecasting (I see you have added the forecasting tag), both functional form and estimation accuracy matter as per bias-variance decomposition.
