How do I interpret the coefficient values and make inferences from regression based on a Box-Cox- transformed dep. variable?

[Hank]

Actually, I did a regression just recently from another, similar, dataset. Both the response and the predictors are almost the same as in the one we have discussed above. I increased the R2 from 0.18 (using Box-Cox transformation of the response (lambda -1)), to 0,51.

In the Box-Cox regression, all diagnostics has been done (Reset-test, outliers-robustness/proc robustreg, multicollinearity, normal residuals, robust s.e. etc.). I only have in mind to look for endogeneity problems. So I feel OK with the Box-Cox model, that is,

I feel it meets the assumptions of linear regression OK. However, if using mspline/spline can increase the explanatory power on, already tested, variables, it is tempting to present those results as well.

But I am not sure how good it is to use this new information about monotone spline/spline, because I am afraid I am overfitting the model. What do you think?

I must add, I am quite new to regression modelling on this level, and unfortunately, is the only one on my job that know how to do it. So I really appreciate all your feed-back and help!

Have a nice weekend

Best regards,

Hank

The code is shown below:

proc transreg data=hem_reg2_2 solve test nomiss plots=all;

      ods exclude where=(_path_ ? 'MV');

      model mspline(response/ nknots=2) = identity(x1_dummy) spline(x2 x3 x4 x5 / nknots=2);

run;

I have reduced the nknots=2, just to reduce the possibility of overfitting the model.

[Ksharp]

For each unit change in the x variable, the transformed Y variable decreases by -0.01068. (  Assuming other independent variable not change )

OP,You can refer to Logic model 's Odds Ratio

Message was edited by: xia keshan

[Hank]

Even though the answer is already posted, thanks a lot. Do you have any ideas to my questions in the second post in this thread?

/Hank