Hi Eli,

The Johansen test will in any case tell you whether you can set up a CVAR (Cointegrted VAR model) or not, and also the rank value. This is not the same as setting up a VAR model.

Depending on the set up of your model equations and the form of your series it may be required to take logs, which is usually the case with the variables that you are using. With the ADF or augmented ADF (or other unit root tests) you are trying to check whether your series are stationary or not. Having unit roots is a requirement for cointegration.

What I would do in your case is to take logs, then check for stationarity of the series with ADF or other unit root test. Then verify if cointegration exists by means of the trace or eigenvalue Johansen test, and if rank 0 can be rejected then set up a VECM model.

Thank you, Hector. 

You need to tell us your sample size. Second, why Johansen estimator? Is it because your ADF test suggests I(1) however uncertainty exists in the power of the unit root tests. Do some preliminary experimentation using other estimators CCR, FMLS and EG -- they all assume I(1)ness. If they are all in annual growth rates, you are saying that no linear combination of the variables exists. You should be careful about for focusing only on the short-run relationship.

Hi Chuck,

The sample size of my data is 116, spanning from 1985 to 2014. I have quarterly frequency data, which are seasonally adjusted using the X-12 method. All the variables are in real log form. The ADF test on all log variables at level fails to reject the null hypothesis that variable has unit roots. The ADF test was conducted in all three cases; 1) with constant only; 2) with both constant and trend; 3) without constant and trend. Therefore, I did the Johansen test to examine the long run relationship between the variables. 

The objective of my research is to investigate the effects of fiscal shock, which is usually a short term phenomenon. Thus, I am using the quarterly data instead of annual data to preclude the impact of systematic discretionary spending, which responses after a year or two to shocks. Consistent with most of the literature on fiscal shock, I wish to use the SVAR model. We know that a main prerequisite of  VAR is that all the variables must be I(0). According to the ADF test, all my variables in logs are I(1). I decided not to use SVAR in difference because I do not want to lose the information on the long run relationship. Instead, I plan to use it for the robustness check.

The ADF test on the same variables expressed in annual growth rates found all series to be I(0), which suits the VAR requirement. Nevertheless, my concern is whether it can still be used to examine the short run dynamic relationships of the variables. 

As you suggested, I will try to use the other estimators on the log variables in lieu of the ADF test.

Thank you, for your helpful contribution.

If you are concerned on short-run relationship, using seasonally adjusted series maybe a mistake.

Bare in mind that your series have 3 modifications: (a) they were transformed in Log (therefore you are measuring their elasticity, and not brute impact);  (b) they were smoothed by X-12, therefore the short-run variations were also smoothed; (c) you used real-value, so any monetary impact (that is significant in short-run, even if not in long-run) were ignored. So, basically, you just neutralized part of the information that could be found on short-run impact. The main problem is you are using a method that is adequate to study long-run variations, but is not complete adequate to short-run.

Also, you have just a bunch of variables and are testing a know theory. Although people tend to use VAR for everything now a day, VAR is not necessary (and is not the most adequate) for everything. If you know the theory you want to test, and there are only a few variables involved, a 2SLR or a 3SLR model may be more accurate, since they will give you more information.

Using 3SLR you may find not only a long-run relationship but also you can estimate conjunctural impact. If you model is well built, you will be able to detect changes in elasticity, punctual impacts, mid-term changes on behavior, etc.

Thank you Mr. Apoena. Are methodologies you suggested implementable in the E-views?

Yes, They are implementable on E-views. But it has been a long time since I don't use that software (I'm corrently using R), I'll leave a link for an youtube video that explain how to use a 3 stage regression on eviews.: https://www.youtube.com/watch?v=qYC9g3cFkiw

I hope it helps.

If you need futher help, I'll be glad to give any assistance I can.

Thank you Apoena. I really much appreciate your prompt response. I will keep you posted if I encounter difficulties. May I have your email address to ask any specific question in regards to those methods.

Best regards,

Eli

Wontok Eli:

Mi hamamas that yupela doing research on PNG economy. YOu may be aware the first macroeconometric model of PNG was published by me and I have the copy with me.

I also have some econometric models( Money supply, inflation, exports, imports functions) relating to PNG. YOu can googlise my Name to find most of these papers  Kannapiran C Arjunan

1. I have the quarterly data on GDP and more than 48 variables from  1978 to 1994 with me.

2. Interest rate is already in % so its is naturally stationary.

3. If you have quarterly data (as you indicated then) then you must use X -4  i.e X1 -X5 to difference them (X -12 is useful with monthly data only.

4. I am sure cointegration must be tested at levels only not with differenced variable to capture the long-run relationship.

5. I totally agree with Dr Aponea on causal models 2SLS or 3SLS if you have simultaneous sets. If  single equation, you can use OLS or CORC_OLS or Error Correction models depending upon the behavior of the variables and functional specifications.

Please feel free to contact me if you need any assistance.

https://www.abebooks.co.uk/9789980750921/Macroeconomic-Model-Papua-New-Guinea-9980750928/plp

Thankutru

Kannan Australia

5. If you choose log function then its ideal to have all variables in log except Interest and inflation rates (expressed in %).

6. I suggest you use the stationary first difference of fourth order (for quarterly data).

Thank you, Dr. Kannapiran. I greatly appreciate your excellent contributions in developing macro-econometric models relating to macro economic issues of PNG. I am doing this project for my master's thesis and very humbled by your sincerity to help.

I will revise the seasonality adjustment procedure and carry out the cointegration tests as recommended.

My main intention is to use the SVAR to estimate the effects of fiscal policy shock on the economy following the similar approach by Blanchard and Perotti (2002) and Perotii (2004). The noble feature of this approach is that I can be able to estimate: a) the automatic reaction of government spending and net taxes to innovations in the GDP and other sets of macro variables, b) "systematic" discretionary response of fiscal policy, c) "random" discretionary fiscal policy shocks. Using the quarterly data essentially eliminates the channel (b) and concentrates my analysis to the short run impact, which is the main objective of the study. 

Below are three issues I have in mind for your assistance.

1). If my variables are not stationary in levels but exhibit cointegration relationship. Shall I still proceed with the SVAR analysis in levels in lieu of running the model in first difference? I want to preserve the long run relationship between the variables. My baseline variables are; government spending (Gt), net tax (Tt), GDP (Yt), GDP deflator (Pt) and 6-months Treasury bill rate (Rt), all in real terms. The Gt, Tt, Yt and Pt are in log values except for the real interest rate. 

2). According to the body of literature that adopted the approach used by Blanchard and Perotti (2002), they calculated the elasticity coefficients of responses of Gt and Tt to changes in output and other sets of variables in the system to derive the cyclically adjusted reduced form spending and tax residuals. Due to data limitation, I cannot calculate the respective elasticities. Notwithstanding this, I plan to use the short-run tax buoyancy coefficient calculated by Paolo and Joao (2017) as a close proxy for the tax elasticity and set the restriction for the elasticity of spending to zero. The rationale for having the spending elasticity to zero is based on the assumption that government outlay is determined independent of changes in output and hence, does not respond immediately to shock in output. I am not sure if this is acceptable.

3). Will the methods suggested by Apoena generate the outcomes similar to what I aim to find?

I really need your expert advice on the aforementioned queries.

Bikpla Tank yu, true.

Mi amamas true lo you tok pisin lo me. 

Wontok...

Please send email to [email protected] so that I can reply to your email. That's easy to communicate  Thankutru Gutpela