If B predicts A, can A also predicts B?

[statsnewbie28]

Hi everyone, I'm new to multiple regression analysis. I have a question and I hope someone can help answer it. Thank you.

In a multiple regression analysis, B, along with several other predictors, was found to be a predictor of A. Literature also says B is a predictor of A.

Since A and B are significantly correlated, I am wondering: can A also be one of the predictors of B? When running a multiple regression analysis for B, should A be added to the model as one of the potential predictors?
(In literature, no one has tested whether A is a predictor of B.)

 

[hlsmith]

Given your description - I have many questions. Can you provide a context for what you are writing about?


B and A are not transformations of the same thing are they?


Based on temporary of events, does B happen before A. So if we said good grades are predict graduating, we can say graduating predicts good grades.


I guess we just need to know more about A and B as well as their relationship.

 

[statsnewbie28]

Thanks hlsmith for your reply.

A is science knowledge; B is scientific misconceptions.

A doesn't come before B, or vice versa. Data collected using two sets of questions presented in a survey.

A is the sum score of a set of questions assessing science knowledge; B is the sum score of a set of questions assessing scientific misconceptions.

 

[statsnewbie28]

I asked this question because I read this: independent variables predict the dependent variable, but the reverse is not true; the dependent variable cannot predict the values of the independent variables.
http://www-users.cs.umn.edu/~ludford/Stat_Guide/Multiple_Regression.htm

 

[hlsmith]

Well in simple logistic regression you can run it both ways and you will get the exact same answer. Doesn't mean it is right. However if you have other "independent" variables in the model, you may not get the same answers by flipping the groups, due to their affects in the model.

So you can do it, but what the link you referenced is saying is one should deemed the predictor and dependent variable, which is what logistic is set up for.

Is your conundrum that you don't know which variable to list as dependent?

 

[noetsi]

One minor point. It is not general usage to say multiple regression, multiple logistic regression etc. When you say linear or logistic regression everyone knows that means it can contain more than one predictor. Its less wordy that way.

Substantively, in the real world, A can cause B and B cause A (simultaneously). However, regression assumes this does not occur, technically when it does it is called a feedback loop. When you have a feedback loop then you are violating the assumptions of linear and logistic regression. In theory the type of regression used in Structural Equation models can deal with this, although there are differences between SEM models that have feedback loops and those that do not.

 

[statsnewbie28]

Thanks, hlsmith.
I know which variable to list as dependent.
Since literature says science knowledge predicts scientific misconceptions, I now want to find the predictors of science knowledge. And it seems reasonable to think that scientific misconceptions can be a predictor of science knowledge. However, it now seems that I shouldn't put science knowledge in the regression model. Right?
Btw, I am using the stepwise method.

 

[statsnewbie28]

Thanks noetsi for introducing the SEM; however, it's too advanced for me. The only method I can handle at my level is linear regression.

 

[noetsi]

Its too advanced for me right now and I had a year of it in graduate school But it is a useful technique for many reasons and if you continue to work in statistics I strongly recommend it. I think you will find the literature remains overwhelmingly tied to regression as compared to SEM.

Incidentally theorist, as compared to practitioners, strongly reject stepwise and for good reason. It capitalizes heavily on chance so if you got a different sample you might well get completely different results. If variables you include in the model are correlated strongly with ones left out of the model this will bias your results. Since I read some of the critiques of stepwise I stopped using it. This is especially problematic when for the last variable you add two variables are close to be chosen.

I once read an article entitled "Death to Stepwise: Think for yourself" a pretty strong statement for an academic Its much better to use theory to generate your model.

 

[hlsmith]

I agree with noetsi on the stepwise. It may take more time to insert a variable then another - and look at all of the combinations, but it is far superior than letting a machine automate it. Would you let an online data site have the final decision in who you are dating. It would probably do a reasonable job based on the available info, but it would not have the contextual information that is in your head.