My original model is: $Y=\alpha(\delta*X_1^{-\rho}+(1-\delta)*X_2^{-\rho})^{-\eta/\rho}exp(e_t)$

Which I partially linearize to: $ln(Y)=ln(\alpha)-(\eta/\rho)*ln(\delta*X_1^{-\rho}+(1-\delta)*X_2^{-\rho})+e_t$

I've used a nonlinear least squares algorithm to estimate the parameters.

Now, I am trying to determine if the "intercept" in this partly linearized model (in this case $ln(\alpha)$) is different for the different years of data that I have. I have 3 years of data, with about 24 observations in each year.

My first thought is to just add a dummy variable for each year as $ln(\beta*Year)$ and just test if the two coefficients that would result (three years minus one comparison years) are jointly statistically different from zero using an F-test. But I doubt that this would be legit since I am trying to test if one parameter ($ln(\alpha)$) varies over years rather than if all the parameters vary over the years.

My second thought is to just split up the sample by year, run three different regressions, and somehow test if all the $ln(\alpha)$ parameters are different. I don't what test I would use or how I would use it, though, to do this. I'm thinking maybe a t-test with unequal variance on each pair of the coefficients, but I feel like pairwise tests on each parameter does not get me a legitimate "joint" test. I have also read about the Chow test, but again I want to test if just one parameter is different across the subpopulations, not all the parameters, which is what the Chow test does.

Any thoughts on what the right procedure is?


  • $\begingroup$ One approach to this kind of problem is pretty much covered - with a partly worked example - in my answer here: nls curve fitting of nested/shared parameters. The alternative - if the data are all completely independent - is to compute the standard error of the difference in the estimates from the separate regressions and do a form of t-test based on that. $\endgroup$
    – Glen_b
    Commented Mar 11, 2014 at 1:13

1 Answer 1


I would fit 2 models, the 1st you have above (reduced model) and another with dummy variable for year effectively giving 3 intercepts (either 3 different intercepts, or get the 2nd and 3rd by adding coef to the baseline slope) (full model), then do a full and reduced model F test to compare. The null is that the reduced model fits just as well as the full model.

The other parameters may change between models, but this is just fitting the best models under the 2 different conditions.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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