PPML RESET Test results

[Tina Samuel]

Dear all,
I was running a PPML model as follows: bilateral exports = Env. Provisions + Std_ depth + country pair FE + importer FE + Exporter FE.
when I ran the reset test for this model. I got Prob > chi2 = 0.0009, so I thought of adding another control, which is language. The control reduced the significance a bit but it was still significant. I retested and got Prob > chi2 = 0.0010. So I thought that by adding another control variable that the reset test would adjust. However, whenever I added another control the significance would increase but I will get very small probability 0.0000. The Prob > chi2 = 0.0010 is the closest thing I got to a less than 0.05.
Any suggestions what might be wrong with the model.
Just noting that the literature haven't indicated any controls outside the initial the model "bilateral exports = Env. Provisions + Std_ depth + country pair FE + importer FE + Exporter FE." and all the other variables I am using as a control are a matter of trial and error, but it seems like I am not getting closer.
I would very much appreciate your expert opinion on this.
Thanks in advance.

[Maxence Morlet]

As I see it (surely experts on this forum will find better solutions, but just in the meantime) you have two solutions:

- Ignore it or admit misspecification (that might be my strategy on a paper I'm elaborating at the minute, so this post is relevant for me as well )

- Include plenty of interactions and higher order polynomials.

Also note the following recommendation from Santos Silva and Tenreyro (2022): when performing a RESET test, do not include - to obtain fitted values - fixed effects that are numerous and that do not have a lot of observations available to estimate them. These estimates will be very imprecise, noisy and will affect the results of the RESET test.

[Tina Samuel]

Dear Maxence,
Thank you for your advice. I did include the fixed effects when preforming the RESET test which might have affected my results.
I also like the interactions idea, I haven't attempted it but I could try.
Thank you.

[George Ford]

RESET is most sensitive to non-linearity in the Xs (and the FE). It's also sensitive to how many quadratic terms you include in the regression.