Different fields of research have different approaches here. In the social sciences (e.g. sociology, economics, political science), we NEVER let bivariate analyses dictate what variables to enter in a multiple analyses like, say, regression. Theory, prior research and common sense dictate which variables to enter in a multiple (regression) analysis. But for pure exploratory work, cutoffs might come in handy vor large data sets.

Hello Wubet,

The practice of requiring a bivariate relationship exist between a possible IV with the chosen DV before using the IV in a multivariable model is a "quick and dirty" approach to pruning a large set of potential IVs to a more manageable size.

There are three serious problems with this "method."

1. You are not guaranteed to find the "optimal" ensemble of IVs for explaining differences in DV scores. There are many other methods available (often used in data mining applications) which fare far better towards this goal.

2. You are running a risk of mistakenly incorporating at least one IV into the model, due to the sheer number of bivariate relationships evaluated. This risk increases as the number of candidate IVs increases.

3. You'll be guaranteed to miss any instance of suppressor effects among IVs.

So, I too would be skeptical of this approach.

Good luck with your work.

Dear Wubet Worku Takele,

You can think in a simple way if you do research in social science.

In social science, no research has validity if there is no significant relationship between an independent variable (IV) and the dependent variable (DV).

Hence, the bivariate analysis, especially correlation analysis, between IV and DV is a basic idea and a tool for validating the research with the IV and DV. No social science research can be meaningful if the study variables have no relationship.

Also, if there are two or more IVs, the bivariate analysis can help you identify the IVs having similar information in the model (in the context of the level of multicollinearity). For example, if there is a correlation coefficient >=0.9 or =0.9 or

Dear Wubet, here is my suggestion.

There are two types of stepwise regression in multivariable regression i.e., backward and forward methods. When we use forward method stepwise regression, the relationship of outcome variable with each predictor is analysed and added to the multivariable model with regard to your cut point (0.25, 0.2).

When we use backward method stepwise regression, all of the probable predictors variable are introduced in to out multivariable model and remove those predictors with no significant association one by one. One senior professor in my university recommend us to use backwards stepwise regression. I am also to the motion of this idea. Forward method stepwise regression is time taking and produces suppressor effect so backward stepwise regression is the preferred method of multivariable analysis.

Best regards,