Is it appropriate to consider regressors in multiple regression model that have no impact when taken individually?

[Zubai Khan]

Hi
I am using stata to analyse the impact of different regressors on private equity fundraising using one step GMM. I am new to stata. The following command that i use:
xtabond2 fr_gdp l.fr_gdp boone_indicator depmoney_ratio disc_index ,gmm( fr_gdp ,lag(2 2) collapse) iv( boone_indicator depmoney_ratio disc_index yr*) small orthogonal robust

However, the problem is that when I use the three regressors boone_indicator depmoney_ratio disc_index, they are not significant at all. When I put market capitalisation as the fourth variable, all three become significant. There is no high correlation between the regressors. Market capitalisation is highly correlated with dependent varialbe while rest three have insignificant correlation with dependent variable. When I use all the four regressors in different models, all of most of them show significant impact.
P-Values of the three are as under when they are taken without market capitalisation:
booneindicator: 0.477
depmoney_ration 0.350
disc_index 0.587
cons 0.508
------------------------------------------------------------------------------


When they are taken with marketcapitalisation, they output is as under:

marketcap_gdp 0.000
booneindicator: 0.007
depmoney_ration 0.026
disc_index 0.065
cons 0.022


My question is that:
Do the three variables qualify to be included in models when they have no impact when taken individually in simple regression model or three of them combined in multiple regression model (only depmoney_ratio is significant at 10% level when taken alone in simple reg model). Should we consider regressors that are not significant when taken alone but gets significance in multiple regression?

[Richard Williams]

Welcome to Statalist.

Note that, when posting code and output, you should use code tags. See the section in the Statalist FAQ about asking questions effectively.

To answer your Q: Did the N drop when you added the last variable? If missing data causes you to lose cases, cross-model comparisons can be problematic.

Suppressor effects could cause what you are observing, and might be very theoretically interesting if they were there. For a discussion, see

https://www3.nd.edu/~rwilliam/stats2/l35.pdf

If you post your output for both models (again, using code tags) we might be better able to advise you.

[Nick Cox]

Richard Williams gave excellent advice. I rather fear that you are in a situation where skewness and outliers are messing things up too. If your panels are different countries then I'd expect clearer results on logarithmic scale. I see GDP everywhere, which can mean Giant Data Problems,

[Zubai Khan]

Welcome to Statalist.

Note that, when posting code and output, you should use code tags. See the section in the Statalist FAQ about asking questions effectively.

To answer your Q: Did the N drop when you added the last variable? If missing data causes you to lose cases, cross-model comparisons can be problematic.

Suppressor effects could cause what you are observing, and might be very theoretically interesting if they were there. For a discussion, see

https://www3.nd.edu/~rwilliam/stats2/l35.pdf

If you post your output for both models (again, using code tags) we might be better able to advise you.

Thank you for you advice. I dont understand hoe to use code tag (going to check it on the forum).
N is 31 (out of all 32 countries used)when three of them are used; adding marke capitalisation reduce it to 28.
I checked the interaction between boone indicator and market capitalisation is significant and it might have created difference? However additing interaction term makes the boone indicator strongly negative from strongly positive. Is there possibility of interactions?