How to create an econometric model?

**krishnarohan** · 06-06-2008, 11:31 AM

I am doing reserach on a topic which requires me to create my own econometric model (time sreies analysis). I have a list of variables that I must include. I am not sure how to decide what would be a bit-fit equation.

In other words, if i want an equation of the form y = a + b + c + d + error term, how do i know if i have to include "b"and not "b(square)" or log(b) or....? How can I test for possible misspecifications in my model?

Do softwares such as Eviews, etc. have an option which suggests which specification would be the best?

**asquare** · 06-06-2008, 11:49 AM

In general, your econometric specification should be driven by your economic model. In other words, the regression you run should be derived from your optimization problem (it's often a first order condition or transformation of a first order condition).

There are a lot of ways to test for model fit, depending on what your options are. A "minimum chi squared" test is one good alternative for testing linear vs. higher power models. Do NOT simply rely on the r-squared of the regression to tell you which model fits better. Model selection is one of the fundamental parts of econometric analysis, and it's not something you want to rely on your software to do for you. It should be based on the economic principles of the question, and tested to the extent possible.

**guan** · 06-06-2008, 01:04 PM

That being said, there is a lot of (somewhat controversial) literature on automatic model selection. Autometrics (PcGive, formerly PcGets) can do this.

**486hunter** · 06-06-2008, 02:03 PM

There is no consensus on how to specify a model. Regression modeling is probably mroe art than science in many ways. In general, I would say that your model ought to be guided by relevant theory. That said, there are ways to test for which model fits the data better, provided all models have the same dependent variable. AIC/BIC/chi squared test, etc.

**GymShorts** · 06-06-2008, 04:23 PM

Look into the following:

regression specification error test (RESET): tests for nonlinear effects of the independent variables, for example, whether x1 squared or x1 cubed should be included.

Davidson-MacKinnon test: test whether x1 or log(x1) should be included.

But as the others said, don't let these tests do the thinking for you.

See chapter 9 in Wooldridge's Introductory Econometrics.

**bibas** · 06-06-2008, 04:28 PM

I think you should be less concerned about the ways to fit your model than to actually define the variables of interest/relevance. Time series should give you a high R-squared just by its nature. Generally, people use the log of variables. Squares and other specifications might (at times) get you a better fit but will make it difficult for you to explain. Also, try lagging the variables. Use the differences instead of levels (again, by doing so, you lose the explanatory information available in levels).

There are several techniques which you can use but, recently there has been a growing tendency to use cointegration techniques. So, if that is something you are familiar with, go for it.
Is this a time series or a panel? I know you have written time series but just wanted to confirm. There are some advanced techniques like System GMM which address many of the problems encountered in simple OLS.

**LouisBD** · 06-11-2008, 10:54 AM

I take it that you are not a time series expert (no offense, I am also not an expert). If I am correct, then I can recommend you the following book:
"Applied econometric time series (W. Enders)".

Quite thin, easy to understand, covers a broad range of material and will probably adress most of your questions (just like the one you're posing now).

If you are using Eviews, check the users guide, which gives also a lot of info.

personal statement: For a more indepth explanation: Time series analysis (hamilton), but this book is much thicker and more difficult than Enders.

**Smileysquared** · 06-11-2008, 02:26 PM

You cannot just use the log of variables when you feel to. One only uses log to smooth out the variability seen in data. Be very cautious when using the log of any variable. OLS estimation is only justified if all your variables are I(0) otherwise you will end up with the spurious regression problem. Think about the reasons you are creating this model-is it for short term forecasting, long term forecasting etc. This will help you narrow which types of econometric modelling you may want to use. Each has its own advantages and disadvantages. It is alot of fun. Take care.