[Nick Cox]

In principle using all the information you have should be a good idea, especially as reducing education to two levels seems problematic. The graph in #11 doesn't surprise that much in a world in which many criminals, sports people, rock stars and celebrities earn far more than academics.

[Pita Fouta]

Thank you very much Fei Wang


After cleaning the outliers data the coefficient of education is still negative but much less than before. It was -0.90 and it is now -0.030.
Also, the R-squared has increased (it was 0.32 and has increased to 0.31).

I don't know if I should continue to clean the data to force my coefficient back to positive. Please find attached these 3 pictures following your explanations Fei Wang ? Please, what do you think about it ?

Thanks in advance.

[Pita Fouta]

Thanks, Nick Cox , Yes I think the same as you, it would be better if I remove the outliers as they are isolated cases but keep the university and technical school to take into account all my data.

Yes, it's true that other people earn more than academics (universities). But in my case, the data you see is for farmers so a more physical job.
But in several other studies, I have seen that the level of education still has a significant impact on their income. The higher it is, the higher their income is.

Please, what do you think?

Thanks in advance.

[Nick Cox]

It's your project. But

* I never advocate removing outliers unless it's clear that they are incorrect values and you can't work out correct values.

* Removing data points to fit your prior expectations is a first-order statistical sin and I can only hope that your mentors,supervisors, advisors, examiners, reviewers don't endorse it.

Otherwise I can only circle back to #2 and my broad suggestions for why coefficients can be puzzling.

[Carlo Lazzaro]

Pita:
what your model seems to oberlook is the risk of endogeneity. It is well known (

https://www.stata.com/bookstore/microeconometrics-stata: pages 177-180;

https://www.stata.com/meeting/german...621.beamer.pdf) that regressing income on education level moves individual ability to residuals. However individual ability is a latent variable taht influences both the regressand (other things being equal, on average smarter people negotiate higher wages) and the oredictor you're interested in (other things being equal, on average smarter people reach higher educational level).

[Fei Wang]

The raw relations seem ok to me -- It's possible that income doesn't shoot up until education reaches some high levels. I wouldn't tailor the data to some "desired" results, and would maintain the four-category education variable and see what happens if covariates are included.

[Pita Fouta]

Thank you all for your messages. Fei Wang Carlo Lazzaro Nick Cox

Of course, forcing statistics to look like our expectations is a big mistake in statistics. I wasn't going to do it for my work.

I will take your advice and hope that I will get a robust and well-fitted regression.

[Carlo Lazzaro]

Pita:
I did not mean that your intent was to torture data and statistics to obtain what you've already decided to get from them (usually, those who follow this approach do not seek for help).
In addition to what others have wisely advice, in my opinion the main issue with your regression is the risk of endogeneity.
I would recommend to discuss all the critical topics you came across with your supervisor/teacher/more experienced colleagues.