VAR and preliminary steps

[Pablo123]

Hi,

I’m trying to do a VAR that combines macroeconomic variables and Google search indices. I have read somewhat confusing instructions or opinions regarding some of the preliminary steps prior to estimating a VAR model.
My plan is to
1. Test stationarity with ADF
2. Test cointegration with Johansen method (only for the endogenous variables)
3. Estimate a VAR and study the Granger causality of the variables

Does this sound like a good way to proceed? I’m working with quarterly data from Q12004 to Q22016 (ie. 50 observations). The results from the ADF tests show that some of the variables are I(1) and I took the first difference to make them I(0). Some of the variables are i(0) so I left them untouched. I used the AIC criteria to choose the lag length and basically used the trend and intercept for all the variables as the difference was not that clear to me, but I also checked the results for using only the constant and the results are the same.

I have a couple of things I’ve tried to search an answer for without luck, as it seems these are dividing the opinions of many researchers..
1. Is it necessary to use logged variables? I read somewhere that the log is mainly used for variables that are right-skew. This is only the case for one of my variables, so is there any reason why I should take the log for only that variable?
2. Adjusting for seasonality. Some of the variables have clear patterns, but not all of them. Should I adjust only these variables or would it be better to add a dummy for each quarter? Or is this necessary at all?
Any help would be highly appreciated.

[jmagomez]

Dear Pablo123,
I will try to give you my opinion about your questions:
Does this sound like a good way to proceed? A.: Yes, it looks like a good way to proceed.
I’m working with quarterly data from Q12004 to Q22016 (ie. 50 observations). The results from the ADF tests show that some of the variables are I(1) and I took the first difference to make them I(0). Some of the variables are i(0) so I left them untouched. I used the AIC criteria to choose the lag length and basically used the trend and intercept for all the variables as the difference was not that clear to me, but I also checked the results for using only the constant and the results are the same. A.: You could use BIC (Schwarz Criterion) instead of AIC.

I have a couple of things I’ve tried to search an answer for without luck, as it seems these are dividing the opinions of many researchers..
1. Is it necessary to use logged variables? I read somewhere that the log is mainly used for variables that are right-skew. This is only the case for one of my variables, so is there any reason why I should take the log for only that variable? A.: I think log variables are easier to understand and reduce a little bit the volatility.

2. Adjusting for seasonality. Some of the variables have clear patterns, but not all of them. Should I adjust only these variables or would it be better to add a dummy for each quarter? Or is this necessary at all? A.: I think you should put dummies that will help you understand a little bit the behavior of your dependent variable, without taking so much information from your independent variables.
Best Regards.

[Pablo123]

Hi,

Thanks so much for your help and opinions!

Best Regards,
Pablo

[jmagomez]

Dear Pablo123,

You're welcome!

Regards.

[Pablo123]

Hi,

I ran into a problem with my model. The Granger causality tests suggest the variables are not granger caused in the order assumed. The next step would've been to create a simple forecasting or nowcasting model using the information from the granger causality tests, but now it's not possible. Does the bad test results from granger causality tests mean that I shouldn't try to create a model with these variables or can the test results be ignored?

-Pablo

[jmagomez]

Dear Pablo123,

It would be better if you achieve a good Granger Causality Test. Do you run other tests?

What is the p-value, Jarque-Bera, Colinearity, Cointegration Tests..?

Do you think that your explaining variables are good to understand your explained variable?

Regards.

[Pablo123]

Hi,

Thanks for the response. I'm studying if Google search indices could be used to predict shifts in the housing market, so I'm not certain if there exists a relationship between these two. However, I tried adding some macroeconomic data to the model and handling the Google search indices as contemparenous exogenous dummies and the model reached an adjusted R^2 of 0,579, which I think is acceptable. My plan is to continue by testing the MSE of the model.

Thanks again for your help,
Pablo