5
$\begingroup$

Let's say I have to make a multiple regression like:

$ Y_i = \beta_0 + \beta_1 x_i + \beta_2 w_i + ... +\beta_3 z_i + \epsilon_i $

Then I run a Ramsey RESET test upon it and discover that my linear specification is not good. What is the best way to cope with non-linearity? I know that I could specify a log-log model, a log-lin model, or add some powers on variables, or try interaction effects.

What I don't understand by reading Verbeek and Stock - Watson is: how to choose the best non-linear specification? Should I try all of them and then take a look at Akaike's Index (or Bayesian or Hannan Quinn)? Or is there a way to understand which specification is the best?

Sorry if I wasn't clear, English is not my native language.

Thank you in advance!

$\endgroup$
2
  • $\begingroup$ Definitely do NOT use AIC to select an optimal transformation of the dependent variable (which I think you were suggesting in the log-linear model). Likelihoods are not comparable between two models where the response has been transformed in different ways, so the AIC is not meaningful. $\endgroup$
    – Macro
    Feb 7 '12 at 0:44
  • $\begingroup$ Of course you're right! My big mistake! So, would $ R^2 $ be a better indicator? $\endgroup$
    – Luigi
    Feb 7 '12 at 10:27
4
$\begingroup$

I've become rather enamoured of late with generalized additive modelling to handle non-linearity. The gam() function from the mgcv package for R makes things very easy as it incorporates automated generalized cross-validation to avoid overfitting.

$\endgroup$
2
  • $\begingroup$ Hi Mike. Thank you for your answer. Unfortunately (since my econometrics course is for beginners), I can use only one of the methods I've wrote about in my first question and only the program called gretl! Thank you again! $\endgroup$
    – Luigi
    Feb 5 '12 at 18:12
  • $\begingroup$ +1; @Luigi, that's a shame, because this approach really is the answer to your concern. I believe gretl and R are somewhat similar, and R is also free, but much more extensive than gretl. There are also a great deal of online tutorials, etc.; R may well be worth your while to learn. $\endgroup$ Feb 5 '12 at 20:15
4
$\begingroup$

I've never heard of gretl but a parametric version of the excellent gam suggestion by Mike is to use additive regressions such as restricted cubic splines (natural splines). R and Stata make this easy to do. With regression splines (piecewise polynomials) you can model almost any relationship that is fairly smooth, and you still get all the advantages of ordinary models (confidence limits, predictions, formulas, etc.). A good default strategy is to figure what complexity the sample size and signal:noise ratio will support, translate that to the number of knots (join points) in the spline functions, and to fit those functions without later trying to simplify the model. In R that would like like

require(rms)
f <- ols(y ~ rcs(x1,4) + rcs(x2,4)) # 4 knots for x1,x2 at default locations

Very few relationships in nature are linear, so it is good to learn about flexible nonlinear modeling.

$\endgroup$

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.