# goodness of fit in regression/ models include a significant number of insignificant variables

I have performed analyses using two different regression models, multi-nominal regression and Tobit regression.

In the first model I measured the impact of 9 variables on choosing travel mode while in the second model I addressed the impact of these variables on trip time duration. After submitting to a peer review journal I am asked to answer some comments of reviewers. One comment is:

(Multi-nominal regression model) "models include a significant number of insignificant variables. This approach misrepresents the magnitude, and possibly in some cases the sign of other coefficients. This needs to be addressed."

I showed the significant codes in my paper by these signs: Signif. codes: *α ≤ 0.10, **α ≤ 0.05, ***α ≤ 0.01

What does he/she mean by this comment?

I think I was questioned because I have some variables that are not significant but one category of them is significant. For example, the chi-square of income and education is not significant but one category of them (low level income), or (high level education) is significant. However, the sig is under 0.05. How I should I convince the reviewer?

The second question that I should to answer is about the Tobit regression which I performed in Stata. I addressed the effect of several variables on time duration of work, shopping and leisure trips to city centre. I am asked to answer this comment:

(Tobit reg) "it is found that the goodness-of-fit of the leisure model is about 6 times for shopping and 4 times for working purpose travels. What are significant variables that could describe the travel time of work and shopping trips and overlooked? Why are the same group of variables that describe the travel time of leisure trips significantly better than work and shopping trips? Clarification is needed". Pseudo R2 for work, leisure and shopping trips are 0.082 0.32 0.052 respectively. the numbers of each kind of trip are 1375, 756, and 545 respectively.

To tell the truth, I do not have enough knowledge of regression models; I just run the models and I find it very difficult to answer these comments. I would be grateful if anyone could help me answer these comments.

• Thanks for your consideration to my questions. I added some more details about my models and also uploaded two pics of my results. Hope it makes my questions more clarified. – Awan Aug 24 '17 at 6:52
• There are various minor confusions on both sides here. Trivially, the term multi-nominal is not a standard substitute for multinomial in my view (although the term does make sense) (fixed in edits). Also minor: in text visible on editing you mention 1 independent variable and 9 dependent variables. These terms are the wrong way round. Also, the reviewer's wording "a significant number of insignificant variables" is clumsy as the first significant just means "a lot" while insignificant does carry its technical sense. – Nick Cox Aug 24 '17 at 9:50
• On bigger issues I agree with @MarquisdeCarabas. It's standard to include a bunch of indicators (dummies) that collectively capture a categorical variable even if some do not show up as significant at conventional levels. You can find explanations of this in good modelling texts, e.g. IIRC stat.columbia.edu/~gelman/arm Otherwise answer comments about too many insignificant predictors by leaving out predictors that don't have a good rationale. My guess is that the reviewer may overstate the risks of a few useless predictors but the only way to be sure is to try it. – Nick Cox Aug 24 '17 at 9:55
• It is harder to comment on your use of Tobit models. The statement "goodness-of-fit of the leisure model is about 6 times for shopping and 4 times for working purpose travels" makes no sense to me. The reviewer seems to be asking for substantive comments on which predictors explain what, which seems fair to me, and beyond our scope here. But why did you use Tobit regression at all? If it's because durations are necessarily positive, it seems to me that there are better functional forms out there, especially generalized linear models with logarithmic link. – Nick Cox Aug 24 '17 at 10:01
• As a different kind of comment: I am intermittently an author, a reviewer and a journal Editor and have experience of all these levels. It's fine to re-submit to an Editor and say that some comments by reviewers are not clear to you. Then an Editor has choices, e.g. to ask the original reviewer to clarify. But inevitably your chance of a better decision does depend on how far you can go to meet their suggestions (unless you can show that those suggestions are wrong or irrelevant). – Nick Cox Aug 24 '17 at 10:04