What type of model should be used?

[Davidkp]

Ladies and Gentlemen, brace yourselves for the newbiest post you have ever seen on this forum (probably).

I desperately need some help with selecting the correct model for my economic dissertation.

My study is on the determinants of intra-EU FDI, so my variables are as follows:

Dependant variable:

- FDI IN (Inward FDI from the other countries in the sample)

Explanatory variables:

- Monetary Union Dummy (to show the effect of the introduction of the Euro in 1999)
- Number of Patents
- Relative Labour cost
- Tax rate
- Growth rate
- Value added output

My sample covers the 11 initial Euro countries and covers 25 years (85-09).

As you would expect, the data is not stationary.

I selected the aforementioned variables by selecting the least correlated variables from a correlation matrix in order to avoid problems associated with excess multicollinearity.

I then constructed my model as follows:

log(FDI) = EMUDUMMY + log(PATENTS) + log(OUTPUT) + log(LABOURCOST(-1)) + GROWTH + TAX

I used a fixed effect panel data analysis.

The coefficients had the expected significance however the Durbin-Watson statistic was too low.

Now as you've probably gathered so far, my econometric proficiency is not great. I want to build a specific type of model such as an AR(1) model, or one using the LDV, which is relevant, and that I'll be able to interpret, (and preferably simple!). However I do not know which model to use.

Using an LDV or using AR(1) in my fixed effects model generates a better DW stat, however am I correct in assuming I cannot use these in a fixed effects model?

Given my data, could you recommend what kind of model I should pursue? What transformations (if any) should I apply to the data - e.g.: Should I take differences?


I'd really appreciate your help, I'm desperate! (I know I probably deserve it, but please go easy on the slating)

Cheers!

[wjgatt]

Some questions you might find useful from my end:

1. If you suspect your variables are non-stationary, did you test for the order of integration, so that you would know how many times you need to take differences to make them stationary?

2. Did you construct the correlation matrix on the level or on the first differences (assuming they are I(1)) of your non-stationary variables?

3. As for the model, you might want to consult the literature for suggestions, but given that you have non-stationary series with a long span, why not consider an error-correction model in the panel data approach (to benefit from the information contained about long run relationships-if any) ? I've never done this but I guess the model used depends on what you're really after in your dissertation.

Hope this is useful
W

[Davidkp]

Some questions you might find useful from my end:

1. If you suspect your variables are non-stationary, did you test for the order of integration, so that you would know how many times you need to take differences to make them stationary?

2. Did you construct the correlation matrix on the level or on the first differences (assuming they are I(1)) of your non-stationary variables?

3. As for the model, you might want to consult the literature for suggestions, but given that you have non-stationary series with a long span, why not consider an error-correction model in the panel data approach (to benefit from the information contained about long run relationships-if any) ? I've never done this but I guess the model used depends on what you're really after in your dissertation.

Hope this is useful
W

Thanks so much for your reply!

I have not tested the order of integration, as I can't seem to be able to do it in my panel (When I go to 'Proc' the option 'Make Residual Series' is not there) - any ideas where I could be going wrong?

I constructed the correlation matrix without making the differences, and the correlations were all reasonably low.

Thanks for the suggestion of pursuing an error correction model, however I can't find any decent explanations online on how to carry one out, and the Hill Griffiths & Lim textbooks don't seem to cover it :(

Would it be incorrect if I used the LDV or AR(1) in my fixed effects panel without taking the differences (OLS)? Or would it be best just to make each variable stationary and carry out the fixed effects panel (OLS)?

Many Thanks

David

[wjgatt]

I have not tested the order of integration, as I can't seem to be able to do it in my panel (When I go to 'Proc' the option 'Make Residual Series' is not there) - any ideas where I could be going wrong?

You need to test each series individually for stationarity, not the residual. Stationarity in the residuals from a regression involving non-stationary variables is something you would want as it would hint at cointegration (if you're interested in the long run relationship). First test for trend stationarity (i.e ADF test with intercept and trend) and if the trend is not s.significant, test using only an intercept.


I constructed the correlation matrix without making the differences, and the correlations were all reasonably low.

Low correlations do not mean that the fact that your variables are not stationary does not matter, at least to my knowledge.

Would it be incorrect if I used the LDV or AR(1) in my fixed effects panel without taking the differences (OLS)? Or would it be best just to make each variable stationary and carry out the fixed effects panel (OLS)?

I'm not sure what you mean by LDV; it could be lagged dependant variable or limited dependent variable, which are different. As for the second question, again this depends on your intial results, what is typically suggested in the literature and what you're after. I think it's best if you discuss this with your dissertation supervisor.

W

[Davidkp]

I have not tested the order of integration, as I can't seem to be able to do it in my panel (When I go to 'Proc' the option 'Make Residual Series' is not there) - any ideas where I could be going wrong?

You need to test each series individually for stationarity, not the residual. Stationarity in the residuals from a regression involving non-stationary variables is something you would want as it would hint at cointegration (if you're interested in the long run relationship). First test for trend stationarity (i.e ADF test with intercept and trend) and if the trend is not s.significant, test using only an intercept.


I constructed the correlation matrix without making the differences, and the correlations were all reasonably low.

Low correlations do not mean that the fact that your variables are not stationary does not matter, at least to my knowledge.

Would it be incorrect if I used the LDV or AR(1) in my fixed effects panel without taking the differences (OLS)? Or would it be best just to make each variable stationary and carry out the fixed effects panel (OLS)?

I'm not sure what you mean by LDV; it could be lagged dependant variable or limited dependent variable, which are different. As for the second question, again this depends on your intial results, what is typically suggested in the literature and what you're after. I think it's best if you discuss this with your dissertation supervisor.

W

Thanks again for your help - I really appreciate it!

I've checked the stationarity of the variables and none of them are stationary, however they are all stationary after testing for the unit root in the first difference (therefore they are I(1) variables).

As a result, would it be best just to use a fixed effects panel and OLS using the first differences of the variables?

By LDV I meant lagged dependant variable, I've been advised by some people (notably my supervisor) to include it, however others (lecturers) have told me not to. I'm worried it will be difficult to interpret as there is no economic reasoning behind it. E.g.: you would include a LDV when modelling growth as it is consistent with the Solow model.

I also am unsure about the issues associated to using an LDV in a fixed effects panel.

I'm not overly concerned with the results, I just want a model that is consistent and that I am able to interpret.

Thanks again for all your help so far!

[wjgatt]

As a result, would it be best just to use a fixed effects panel and OLS using the first differences of the variables?

By LDV I meant lagged dependant variable, I've been advised by some people (notably my supervisor) to include it, however others (lecturers) have told me not to. I'm worried it will be difficult to interpret as there is no economic reasoning behind it. E.g.: you would include a LDV when modelling growth as it is consistent with the Solow model.

In that case yes you can estimate your model in first differences, with the lagged dependant variable. You can then interpret the coefficients on the explanatory variables as the short run elasticities. If say your model is d(y) = c + e.d(x) + f.d(y(-1)), then e is the short run elasticity.

You can then interpret the long run elasticity as e/(1-f) (i.e you solve for d(y), dropping time subscripts). This depends on whether it makes sense, but I guess it's a good way to incorporate all of your requirements.

W

[Davidkp]

As a result, would it be best just to use a fixed effects panel and OLS using the first differences of the variables?

By LDV I meant lagged dependant variable, I've been advised by some people (notably my supervisor) to include it, however others (lecturers) have told me not to. I'm worried it will be difficult to interpret as there is no economic reasoning behind it. E.g.: you would include a LDV when modelling growth as it is consistent with the Solow model.

In that case yes you can estimate your model in first differences, with the lagged dependant variable. You can then interpret the coefficients on the explanatory variables as the short run elasticities. If say your model is d(y) = c + e.d(x) + f.d(y(-1)), then e is the short run elasticity.

You can then interpret the long run elasticity as e/(1-f) (i.e you solve for d(y), dropping time subscripts). This depends on whether it makes sense, but I guess it's a good way to incorporate all of your requirements.

W

Thanks again for your post, you're a real help!

Interpreting the coefs as short run elasticities is great. Can you construct the equation in eviews to find the LR elasticities? Or do you just have to work them out outside of eviews?

Also, do you know how one could test the stationarity of the residual of a panel? As now that I've determined all my variables are I(1), I would like to see if the panel is co-integrated.

Many thanks for all your help so far!

[wjgatt]

You're welcome.

I'm not aware of any specification that can be written so that you have both S and L run elasticities readily available in the context that you're working with.

As regards panel data, sorry I can't be of help here. Refer to the manual.

W

[Davidkp]

You're welcome.

I'm not aware of any specification that can be written so that you have both S and L run elasticities readily available in the context that you're working with.

As regards panel data, sorry I can't be of help here. Refer to the manual.

W

Thanks again for your advice.

I've since been advised that using a LDV in a fixed effects model will give a biassed result. My Dissertation supervisor and econometrics teacher are just contradicting themselves! :(

No idea how to proceed now!