# Growth theory / Initial GDP: positive coefficient sign

I estimated growth regressions for several EU countries. In each of them, the sign of "initial GDP"'s coefficient is statistically significant but positive, which contradicts growth theory as well as most of other published papers on this topic. I tried to change the variables, the period, etc. but it remains positive.

I can't find any explanation for this result in the literature, which would imply that there is no growth convergence among European countries.

Did anybody already face this problem? How did you solve this? I can't find the source of this problem. Personally, I replaced "initial GDP" by "lagged (one period) GDP" but it would be better if I could keep "initial GDP" among the explanatory variables.

• I think that the lagged period is too short. So, your model possibly had serial-correlation, that means convergence effect can not express. – pirapat Dec 6 '15 at 4:51
• As pirapat mentioned, have you checked for serial correlation? Can you try using other, similar variables? Have you made sure that things are all in the correct units? Just some simple suggestions. – majmun Jan 5 '16 at 23:06
• What is the period between initial and actual GDP? How many control variables did you use? Sample size? Give some more information. – Fitzroy Hogsflesh Aug 4 '16 at 6:14
• What are your other variables ? – Yann Sep 2 '16 at 14:26
• probably you are not interested in a solution anymore. In case you are, you need to tell about the data and time periods you used. If you regressed gdp in 2010 on gdp in 2000, then you would certainly find no convergence. If you instead used 1960 as initial GDP (usually the first year were meaningful GDP numbers are available), you certainly will find convergence. – Fitzroy Hogsflesh Dec 1 '16 at 9:51

It's really difficult to give a satisfactory answer without having the data at hand. I would do two things:

1. Check carefully the data and computation of growth rates. By experience, it happens that some "particular" result is the consequence of data or computation issues. You may for example observe any issue by plotting your growth rates and initial GPDs with the name of the countries.

2. Use your data but exactly the same period, same countries and same methodology than the literature and check if you still find a contradictory result. Then, you can mix: use the same source of data, same countries, same period but your methodology, etc...

This would be something best handled by closely looking at your residuals.

Your unobserved variation is the single most important statistical insight when it comes to accurate empirical examination. I could almost guarantee you're dealing with serial correlation, some simultaneity issues, and possibly multicollinearity. These issues are generally best tackled through autoregressive vector specifications to avoid the issues of causality.