The RATS Software Forum

Thanks, Mr.Doan. Right. I know that F values are important to decide causality direction. I meant that Granger causality test does not tell us what are the numbers and signals of 'b' and 'a' in the model of 'Y=b*X + a'. So, I said that we need to do subsequent tests such VAR, cointegration test, and so on.

I am considering about your last sentence "a 10% reduction in forecast error in a small data set may be statistically insignificant while a 2% reduction in a large one may statistically significant."

bok1234 wrote:Thanks, Mr.Doan. Right. I know that F values are important to decide causality direction. I meant that Granger causality test does not tell us what are the numbers and signals of 'b' and 'a' in the model of 'Y=b*X + a'. So, I said that we need to do subsequent tests such VAR, cointegration test, and so on.

I'm really puzzled about what you're trying to do. If you're talking about a static y(t)=bX(t)+noise model, the Granger causality test tells you effectively nothing about that.

Dear Mr.Doan,

I think that there might be some communication error. I just referred to general process of time series analysis. Let me explain more concretely.
First, confirm the cause and effect relationship through Granger causality test.
Second, based on GCT and if there exists such a relationship, we can identify dependent varible (effect) and independent variable (cause).
Third, based on 2nd step, we can do ECM, Johansen cointegration test, or VAR.
I did not mean static model as a following test after GCT.

Let me ask you one thing about '2 variable Granger causality test (GCT)' vs. '2 variable VAR model' (VAR) . VAR produces impulse response functions and many people analyze and interprete IRF not only as 'impact size and direction' but also as 'causal relationship'. If we do VAR test with IRF, then GCT can be skipped?

bok1234 wrote:Dear Mr.Doan,

I think that there might be some communication error. I just referred to general process of time series analysis. Let me explain more concretely.
First, confirm the cause and effect relationship through Granger causality test.
Second, based on GCT and if there exists such a relationship, we can identify dependent varible (effect) and independent variable (cause).
Third, based on 2nd step, we can do ECM, Johansen cointegration test, or VAR.
I did not mean static model as a following test after GCT.

That's completely wrong. Granger causality tests are not and never had been designed to show cause and effect. For the types of models you cited, typically all variables of interest are treated as endogenous. Granger causality tests if employed (often, they aren't, in fact, most of the time, at least in good empirical work, they aren't) are to look at very specific dynamic exogeneity hypotheses (such as real-nominal variables).

bok1234 wrote: Let me ask you one thing about '2 variable Granger causality test (GCT)' vs. '2 variable VAR model' (VAR) . VAR produces impulse response functions and many people analyze and interprete IRF not only as 'impact size and direction' but also as 'causal relationship'. If we do VAR test with IRF, then GCT can be skipped?

Again, if you look at good empirical work with VAR's, you will rarely see Granger causality tests, or if you do it's for very specific reasons. Pro forma GCT's are included by writers who are just doing a software output dump rather than careful empirical work.

Dear Mr.Doan,

Thank you for your explanation, Mr.Doan, and I am sorry if I made you angry with my own, or wrong concept about GCT. Let me ask you something about your yesterday teaching.

1. In spite of your explanation, can I apply GCT to test exogeneity hypotheses - not to test cause and effect hypotheses - between 'actual inflation rate' and 'inflation expectation' ? You said GCT could be used for such a limited case "to look at very specific dynamic exogeneity hypotheses (such as real-nominal variables)". So, I am wondering if my case, the exogeneity test between 'actual inflation rate' and 'inflation expectation' might be THAT CASE.

2. In the aspect of statistical concept what is the difference between exogeneity and causality? Which one is the broader or more comprehensive concept? Or are these concepts are independent?

3. What is the best or proper test to analyze the cause-and-effect relationship between variables? VAR model test with IRF is enough to see the causal relationship? Or should we give up any empirical tools in finding such a relationship and shift the resposibility to economic theories? Several years ago you taught me multivariate granger causality test. Does multivariate GCT also have the same sense, that is to say, 'very very specific or extremely rare or hardley useful' as you referred to the usefulness of GCT yesterday?

bok1234 wrote:Dear Mr.Doan,

Thank you for your explanation, Mr.Doan, and I am sorry if I made you angry with my own, or wrong concept about GCT. Let me ask you something about your yesterday teaching.

1. In spite of your explanation, can I apply GCT to test exogeneity hypotheses - not to test cause and effect hypotheses - between 'actual inflation rate' and 'inflation expectation' ? You said GCT could be used for such a limited case "to look at very specific dynamic exogeneity hypotheses (such as real-nominal variables)". So, I am wondering if my case, the exogeneity test between 'actual inflation rate' and 'inflation expectation' might be THAT CASE.

I find it hard to imagine that a test for whether inflationary expectations are exogenous would be at all interesting, so the question would be the other direction. In certain small economic models inflationary expectations are created as some type of weighted average of their own past and past actual inflation. That would mean that inflationary expectations would be a moving average of past inflation, so inflationary expectations would not Granger cause inflation. I assume your inflationary expectations series is generated some other way (surveys??) so the GC test would be testing whether there was some independent information in that that wasn't just extrapolated from past inflation.

The result in Sims' Money, Income and Causality is that in a bivariate system, exogeneity and lack of Granger causality are equivalent---in some cases, such as yours, the hypothesis would be naturally described as testing for Granger causality, while in others (the money-income relationship in that paper), it would be testing for exogeneity.

Thank you for fast reply. I need to say my test purpose. In the aspect of anchoring of inflation expectation, some economists say that monetary policy anchor inflation expectation very well if actual inflation DO NOT or RARELY affect long run inflation expectation. Of course there are many definitions or empirical tests for anchoring, this kind of tests are considered as not bad. So, in this point of view, I hope to analyze many countries, especially inflation targeting countries anchoing situation with 2 variable. Therefore, I am asking again, for this test reason, can I use GCT for verifying non-endogeneity of inflation expectation? Or 2-variable VAR with IRF is more appropriate test for it?
And I am expecting your answer for question 3 and a little more about concet question 2.
I appreciate again your helpful advice.

If there is no GC from inflation to inflation expectations, I would probably be more likely to interpret that as meaning that the inflation measure (almost any standard inflation measure) was too noisy a proxy to be useful. But the story about anchoring inflation expectations would be exactly the type of thing that would be testable using a GC test. If you reject (i.e. the lagged inflation is significant), then that hypothesis is rejected.

Regarding your other questions, Hamilton has a section about GC tests and how they relate to more general notions of causality.

Thank you, Mr.Doan. May I expect your answer about the 3rd question? It was about your recommendation for proper cause-and-effect test. I rewrite it again. "3. What is the best or proper test to analyze the cause-and-effect relationship between variables? VAR model test with IRF is enough to see the causal relationship? Or should we give up any empirical tools in finding such a relationship and shift the resposibility to economic theories? Several years ago you taught me multivariate granger causality test. Does multivariate GCT also have the same sense, that is to say, 'very very specific or extremely rare or hardley useful' as you referred to the usefulness of GCT yesterday?"

Macroeconomics isn't an experimental social science. There is no test for "cause and effect". That section in Hamilton largely addresses that.

Some of your questions are getting a quite a bit away from support for the RATS software.

Dear Mr.Doan,

1. Thank you for your opinion on macroeconomics and I agree with you. My question intention was to know empirical tools which were close to or mimicking cause-and-effect experiment if we suppose that kind of tool exists.

2. I apology that some of my questions go too far beyond your usual service. I think that Q&A about rolling regression drove me to go deeper and deeper, and made me ask you fundamental, cronic to me, problems about econometrics. Not to make this sort of accident happen again, please let me know your warning by saying 'you are crossing the line' at any time.

Hi Tom!

I am trying to create a rolling sample Granger causality graph, as produced by the dummy code, but in a multivariate set up. In particular, I have a bond dataset for different ratings and I want to replicate the time-varying causalities in the system.

Can you please help me on how to amend the rats syntax for that?

Thank you in advance

That seems to be combining two rather flawed ideas---the pairwise causality tests in a multivariate system and rolling causality tests. Note that the whole point of this example is that the "interesting" graph that it produces is entirely sampling variations---there is no actual relationship between the series.