It sounds as if you are planning on using some p-value cut-off in univariable models to determine which variables to include in a multivariable model. This is a recipe for over-fitting. See Mike Babyak's nice (readable) article for more details. HTH.

https://www.cs.vu.nl/~eliens/sg/local/theory/overfitting.pdf

Thank you Bruce Weaver! The article doesn't mention anything about covariates. Should we care for covariate significance or its just the variable significance in univariate model for selection of variables into multivariable model?

Perhaps I misunderstand your question. I thought the "Univariate Pretesting or Screening" section starting on p. 417 was relevant. HTH.

Thank you Bruce Weaver

Hi If you are still interested in this topic then I think Bruce Weaver understood what you wanted to do, but you are using the wrong method to select variables. Since logistic regression is mentioned above I'll give you an example of a much better method using logistic regression. Pay particular attention to the references. I will include a second paper in Cox regression because that stresses another important advantage of this approach, called adaptive lasso. Best wishes, David Booth

I suppose to use a Wald test for regression to select those variables with a p-value

when we use 0.25 as a cut of value

when we use 0.25 as a cut of value