You should not select variables by their p-value. You should build a model based on theory. If you consider a variable interesting or important enough to include it in the model, then include it. Otherwise don't.

Hello Noshin,

In addition to Jochen's response with which I agree, let me add one point. If your proximate goal is to somehow arrive at some subset of your candidate IVs that would serve as an "optimal" model (given your data set), please note that there are much better methods available than by sifting through the bivariate relationships of each IV with the DV. Following your path would assure, for example, that all instances of suppressor variables would most likely be missed.

If "optimal" subsets are your aim, then you should have a look at methods such as those outlined in machine learning texts (including lasso/adaptive lasso). However, do note that this isn't hypothesis testing as you likely know it. As well, you're not following theory to arrive at a model.

Good luck with your work.

In general I agree with Dr Wilhelm. I note in some cases where prediction is the objective optimal predictive models can be obtained by the addition of modern methods. I attached an example. Best wishes David Booth

Dear Jochen Wilhelm,

Most of the reviewers want to see "statistically significant" results. One of my manuscript had 6 insignificant IVs out of 9 IVs. Reviewer objected it. I am struggling to justify my results based on statistical significance. Please show me some direction in this regard.

Laxman Singh Bisht find a better journal to submit

or educate reviewers that it is a bad practice

(theres decades of papers around on the topic)

the significant/insignificant game makes absolutely no sense, especially to select variables to include in a model (without mentioning the problem of multiple testing)

As mentioned above select your variables according to previous theory or using your knowledge/arguments as the subject matter expert.

Stefano Nembrini thanks