I am a student, new to research and have chosen a challenging field to perform my research in. While I have been reading about similar research's, as part of my literature review, I observed various researchers using various models in their study. I am having a tough time understanding why researchers use different models/tests and how they choose it, although its the same kind of research being performed.

My research is about the impact that macroeconomic variables have on Stock indexes. I have 6 indexes and 5 variables x 2 countries (Quarterly data). Each of the index/variables have about 60-70 observations.

  • What I am planning to find -
    1. Which of the MV's could potentially explain the movements of the index prices. Which of the MV's impact the Stock prices.
    2. Is there a variable/combination of variables that specifically create a downward/upward movement.

Should you have any suggestions of statistical tests/models for the above, I'd be pleased to learn and know more about them.

Thanks in advance.


1 Answer 1


One of the reasons why there are so many different models is that "having impact" is not that easy to define. It is not to test whether a variable has "any" effect on another variable, so researchers first have to come up with a model (based on theoretical considerations) that they can estimate and test. And of course the tests are different depending on the model they assume.

Basic model

One of the most basic (but powerful and widely used) concepts for these type of hypotheses is Granger-causality. Simply put, variable $A_t$ is said to Granger-cause variable $B_t$, if $A_t$ is useful in forecasting $B_t$. So you might try to find out which macro variables Granger-cause your indices. However to do formal tests you need to restrict the previous definition a bit.

Let us say, that variable $A_t$ Granger-causes variable $B_t$ if $$E[B_t|B_{t-1}, B_{t-2}, ...] \neq [B_t|A_{t-1}, B_{t-1}, A_{t-2}, B_{t-2}, ...]$$ or in words, knowing the past values of $A_t$ helps us get a better prediction of the current value of $B_t$. The nice thing is, that the above hypothesis is easy to test: you just have to fit a linear model.

So if your index variable is $Y_i$, and your explanatory variables are $\{X_{1t}, \dots X_{mt}\}$, then you just have to estimate the simple linear model $$ Y_t = \alpha + \sum_{s=1}^S \gamma_s Y_{t-s} \sum_{k=1}^m \beta_{ks} X_{k, t-s} + \epsilon_t $$ by OLS. Then $X_k$ Granger-causes $Y$ (in the expectation sense), if $\beta_{k1}, \dots \beta_{kS}$ are jointly significant.

Practically to test which explanatory variables Granger-cause $Y_t$ you have to do the following:

  • Run a linear regression with $Y_t$ as the dependent variable and some lags of $Y_t$ and the explanatory variables at the right hand side
  • Test whether all the lags of a given explanatory variable are jointly significant
  • Repeat it for each of your indices

By estimating these models you also get an answer to your second question (you just have to look at the coefficients).

Another way to go would be to estimate a VAR model where you include all of your indices and explanatory variables, and then see whether the lags of a given explanatory variable for a given index are jointly significant. This will have a slightly different interpretation: you would test whether your explanatory variables Granger-cause your indices even if you condition on all the indices. In your case I would be careful with it due to the relatively low number of observations.

Practical concerns

  • Deciding how many lags to include in the regression might be tricky. It depends on the amount of data you have and the model you have in mind. For quarterly data, you should probably include at least 4 lags.

  • As have panel data you could be inclined to include a country dummy to get fixed effects estimates. However because you have lagged dependent variables on the right hand side, your estimates would be biased if you did this. There are ways around it, but they are quite advanced.

If you have questions how to implement it in practice feel free to ask for details.

  • $\begingroup$ Thanks for your response Martin. Like you mentioned, I am not certain if my data is good enough to employ these models. Although reading about them in a couple of the research papers motivated me to find out more about them and see if I could use them. I am currently using SPSS and found no option to use the Granger Causality model. Any suggestions about a tool where I can run the model without having the need to script? Also, do you happen to know of any alternative tests I can run on SPSS that can resemble the results like Granger causlity or be as robust as the model? $\endgroup$
    – Shiva
    May 29, 2016 at 21:53
  • $\begingroup$ I am not really familiar with SPPP, but should be able to test Granger-causality even if it does not have a menu item for it. First you have to run a simple OLS regression with lagged variables. "SPSS LAG Example - Days Between Orders" here may help if you have trouble creating the lagged variables. The second step is a simple F test, which SPSS should also be able to do. That said, I often hear that SPSS is not really suited for time serias analysis, so maybe that is the reason why it is not straightforward. $\endgroup$ May 30, 2016 at 7:42
  • $\begingroup$ Thank you again for breaking it down for me. I am asked to analyze my work using EVIEWS. I read somewhere that EVIEWS offers Granger causality. $\endgroup$
    – Shiva
    May 31, 2016 at 14:19
  • $\begingroup$ You are welcome. Yes, you are right, Eviews is very well suited for time series analysis in general, and there are built-in Granger-causality tests. You can also consider accepting this answer if you are satisfied with it :) $\endgroup$ May 31, 2016 at 18:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.