Build a model based on your understanding. Include the variables you deem relevant predictors and don't include those you deem irrelevant. If you don't know anything, it's impossible to build a meaningful model. Eventually, the model is your decision - it is not dicated by the data. There can be different models representing different aspects of reality, all equally valid. You must know what aspects should be represented.

As David said, using p-values is a (very!) poor way to do variable selection. It is generally not good to search for a "good model" with only the data used to fit the model. You can only test the model performance (e.g. the positive and negative predictive values) on a new data set that was not used to fit the model (if you have enough data, split it into a "training set" [to fit] and a "test set" [to check]). If you do this it should become obvious that p-values or alike from the fitted model are irrelevant for model selection, and procedures based on such statistics usually fail, as they give vastly over-confident results, often including a suboptimal or wrong set of variables. They also end up with models of wich you have no good idea what aspect they should represent and what alternative aspects they don't represent. Without knowing such things, it's impossible to get a meaningful interpretation. And if you knew such things, you would not need such kind of variable selection.

Depends on your research question but p-value is a very poor way to do variable selection. see the attached

Build a model based on your understanding. Include the variables you deem relevant predictors and don't include those you deem irrelevant. If you don't know anything, it's impossible to build a meaningful model. Eventually, the model is your decision - it is not dicated by the data. There can be different models representing different aspects of reality, all equally valid. You must know what aspects should be represented.

As David said, using p-values is a (very!) poor way to do variable selection. It is generally not good to search for a "good model" with only the data used to fit the model. You can only test the model performance (e.g. the positive and negative predictive values) on a new data set that was not used to fit the model (if you have enough data, split it into a "training set" [to fit] and a "test set" [to check]). If you do this it should become obvious that p-values or alike from the fitted model are irrelevant for model selection, and procedures based on such statistics usually fail, as they give vastly over-confident results, often including a suboptimal or wrong set of variables. They also end up with models of wich you have no good idea what aspect they should represent and what alternative aspects they don't represent. Without knowing such things, it's impossible to get a meaningful interpretation. And if you knew such things, you would not need such kind of variable selection.

The standard approach of throwing all your predictors into a model ignores, as Jochen Wilhelm says, using your prior knowledge. I recommend that you map the relationships between your predictors and both the outcome and each other.

Have a look at directed acyclic graphs, which are a wonderful tool for making these ideas explicit using a simple graphical method. http://www.dagitty.net

Thank you so much (@ Ronán Michael Conroy)for sharing the wonderful tool.

Thank you so much (@ Jochen Wilhelm) for the clear explanation, it is very helpful.

Thank you so much ( @ David Eugene Booth) for your message and attachment.

In this case, it is better to use the five variables in the final model

Hey,

I would to add something here, if I'm not wrong, then sometimes we may find that some variables are not significant in univariate analysis but they might turn up to be significant in binary logistic regression or multivariate analyses.