a high correlation coefficient between the dependent variable and a control variable

[David Lu]

Dear all,

I have an control variable which has a high correlation coefficient of 0.6985 with the dependent variable.It's cross-sectional data, what things should I concern about the high correlation coefficient ? Can I still get it into the model (it's a very important control variable)?

pwcorr dep ctrl1 ctrl2 iv1 iv2,sig

Best,
David

Cross-post at :
http://stats.stackexchange.com/quest...a-control-vari

[Carlo Lazzaro]

David.
yes, you can.
However, taking a look at the -pwcorr- outcome table, your (regression?) model will probably suffer from multicollinearity.

[David Lu]

David.
yes, you can.
However, taking a look at the -pwcorr- outcome table, your (regression?) model will probably suffer from multicollinearity.

Dear Carlo,

Yes, you've already forseen the pitfall. I got an insignificant regression with incorporating this variable into the model. After droping it, the model got significant coefficients. In that case, what should I do?

Plan 1: find alternative variable to replace the existing one. Yes, it has one, but if I use it, my sample size will be much smaller. Then, the issue of biasness emerges.
Plan 2: drop it directly. If I do that, how can I control the effect of that variable?

Thanks,
David

[Carlo Lazzaro]

David:
- do not put statistical significance as your totem; instead, consider what Others did in the past when presented with the same topic in your research field;
- go for a more parsimoniuos/different model (your Plan 1). The biasedness issue will emerge if your missing values are informative (again, p<<whatever> is not the threshold that split good and bad);
- multicollinearity includes more than one variable. It's up to you what to drop. If you drop, you cannot control for,
- as a final remark, I would also check whether your regression model suffers from endogeneity (you may end up with a variable included within residuals which is correlated with -depvar- and one -indepvar-).

[David Lu]

David:
- do not put statistical significance as your totem; instead, consider what Others did in the past when presented with the same topic in your research field;
- go for a more parsimoniuos/different model (your Plan 1). The biasedness issue will emerge if your missing values are informative (again, p<<whatever> is not the threshold that split good and bad);
- multicollinearity includes more than one variable. It's up to you what to drop. If you drop, you cannot control for,
- as a final remark, I would also check whether your regression model suffers from endogeneity (you may end up with a variable included within residuals which is correlated with -depvar- and one -indepvar-).

Dear Carlo,

Yes, you're right. Endogeneity is the most frequently mentioned issue that I cannot escape, and what I read most is based on the OLS context. Since now I use glm with a log link, how can I check or do something to alleviate the problem of endogeneity? My data is cross-sectional data. Here is the command what I use:

glm dep c.iv1t##c.iv2 c.iv1t#c.iv1t cv1 i.cv3 ,family(poisson) link(log) vce(robust)

Thanks again,
David

[Nick Cox]

In the natural sciences a strong predictor could be the sign of a good theory and grounds for rejoicing.

In the social sciences, it seems, such a variable is treated with suspicion as likely to be just another version of the response and on the grounds that no theory could be that good.

[Carlo Lazzaro]

David:
set aside some famous class-room drills (i.e., ability embedded within residuals, that affects both educational attainments (-indepvar-) and wage (-depvar-)), the usual source of endogeneity examples is the literature of your research field.
Endogeneity is a strange (and dangerous) beast: sometimes it's apparent, sometimes is reported in literature and sometimes again seems to be in the beholder's eyes (unfortunately, the beholder in point is often a reviewer/discussant).
As per your code, you seem to have a count regression model; I'm not clear with what you're going to test with the second interaction, that does not seem to include the conditional main effect of the predictors: I do not see any detail dealing with endogeneity issue, though.