Question about multivariate regression analysis

[Daniel Wing JC]

Dear colleagues, Good morning. I would like to ask about multivariate regression analysis.
I researched the OR between risk factors and outcome diseases.
In the a univariate logistic regression analysis, I got such risk factors which would be able to into the multivariable

logistic disease ib1.gender // reference group is male
logistic disease ib2.dry_skin// reference group is no
logistic disease ib2.wheezing_noise// reference group is no
logistic disease ib2.sever_symdrome// reference group is no I found those four groups P<0.2, so I would like to run multivariate analysis, however, the groups (e.g. dry_skin, wheezing_noise, sever_symdrome) have 0.5-1% missing data.

Therefore in the multivariate regression analysis,
Is it possible to run

logistic disease ib1.gender ib2.dry_skin ib2.wheezing_noise ib2.sever_symdrome

Looking forward to hearing from you.
Thank you very much.

[Marcos Almeida]

If I understood the query correctly, this is all about multivariate regression and, yes, you can add several predictors. Shall you have missing values, the (estimated) sample size will decrease. You may try it out and see the results.

[Daniel Wing JC]

Thanks Marcos. The missing data didn't have a significant influence the results as I have tried.
Best regards, Daniel

[Carlo Lazzaro]

Daniel:
as an aside to Marcos' helpful advice, you're actually planning a multiple (not a multivariate) logistic regression, as you have one regressand only and different predictors.
While is true that such a scant percentage of missing values is probably immaterial, it would be advisable (especially if you're planning to submit your research to a technical journal) to check whether their missing mechanism is ignorabe or not.

[Bruce Weaver]

Hello Daniel. Carlo has already pointed out the distinction between multivariate (i.e., multiple outcome variables) and multivariable (multiple explanatory variables). I'll add that univariable pre-screening of candidate explanatory variables, like many other algorithmic approaches to variable selection, is ill-advised because it increases the likelihood of overfitting. See Mike Babyak's (2004) article for a nice overview of the issues. See also the Stata FAQ on "stepwise" variable selection. HTH.

• https://people.duke.edu/~mababyak/pa...regression.pdf
• https://www.stata.com/support/faqs/s...sion-problems/

[Daniel Wing JC]

Hello Carlo and Bruce, Thank you very much for your suggestions. Sorry for such delay to reply. Yes, I only looked at one outcome corresponding to several risk factors. I will correct the wordings in the text.
As Bruce raised an important information. When I ran univariable analysis, I found 11 factors with statistical significant. However, when doing multivariable analysis, I only found 3 factors reporting statistical significant. I don't know whether the missing data played some sort of effects to the multivariable analysis.
Many thanks again. Jiancong

[Carlo Lazzaro]

Daniel:
as Bruce wisely highlighted, there's no gain in running univariate pre-regression screening, as you actually regresses the regressand on one predictor at time. Hence, no wonder that you got different results with different regression models.This is probably not what you want, as regression is basically aimed at investigating the contribution of each predictor (when adjusted for the other ones) to variation of the regressand.

[Daniel Wing JC]

Hello Carlo, Thank you very much for your great inputs. In my paper, I will make the explanation of the contribution of of predictor after adjustment.
Could I please ask another question? As I post it for several days without any helps. Please refer to below link. Many thanks again.
https://www.statalist.org/forums/for...tive-incidence

Best regards, Jiancong

[Carlo Lazzaro]

Jiancong:
sorry, but I've never used -stcrreg-.

[Daniel Wing JC]

Thanks Carlo. No problems. Jiancong