[open] JASP: Some questions about multiple linear regression
I have some questions about multiple linear regressions in JASP.
My first question is about the difference between the Forward and the Stepwise method for entering predictors in a multiple linear regression. I found the following in the Help documentation:
- Forward: Predictors are entered sequentially based on the criterion specified in "Stepping method criteria".
- Stepwise: Predictors are entered sequentially based on the criterion specified in "Stepping method criteria"; after each step, the least useful predictor is removed.
Does this mean that in a forward model it is theoretically possible to end up with a model containing all variables that were entered in the box
Covariates, whereas with a stepwise model the maximum number of predictors that can be included in the model is the total number of Covariates - 1?
I was also wondering whether someone has advice about which of the four methods (Enter, Forward, Backward or Stepwise) is 'best' for students that are using multiple linear regression for the first time. In JASP, the default is "Enter", does this mean that this is the safest/easiest option?
Also, I found the following discussion on github:
Do I understand correctly that theory-driven entering of predictors is not yet possible, but might be added in the future?
My final question is about interactions. There is no option to include interactions in multiple linear regression analyses yet, is there? As a work-around, if it turns out that some of my students want to test for interactions, could I advise them to manually add a column containing the multiplication of the values from two predictors? Or is there something wrong with this method?
Any help would be very appreciated!!
Sorry for the delayed response.
With respect to your question about forward/stepwise regression:
Yes, when you apply the forward method, it is theoretically possible to end up with a model that contains all variables. However, I think, that this is also possible when using the stepwise method.
For the stepwise procedure, each time a variable is entered into the model, a removal test is made for the least useful predictor. If this test indicates that the predictor is not useful anymore, it is removed.
However, it might be the case that this removal test never suggests to remove a predictor (of course, this depends on the removal criterion you specify which in JASP can be done under "Options->Stepping Method Criteria").
With respect to the question about which of the four methods (Enter, Forward, Backward or Stepwise) is "best" for students that are using multiple regression for the first time:
In my opinion, it is easiest and safest to use the "Enter" method, because it is immediately clear which variables will be in the model, namely the ones that the student selects.
Furthermore, as Andy Field writes in "Discovering Statistics Using R": "Stepwise regressions are generally frowned upon by statisticians" ("stepwise" includes Forward, Backward, and
Stepwise). There are a number of articles about issues with these stepwise methods, for a short list of issues see: stata.com/support/faqs/statistics/stepwise-regression-problems/
With respect to your question about theory-driven entering of predictors:
Theory-driven entering of predictors is at the moment possible in the sense that you can select the variables that you think should be in the model and then enter them to obtain the results for that
model. What is not possible at the moment is to enter the variables in blocks (which some people call hierarchical regression).
With respect to your question about interactions:
It is possible to include them in a linear regression model in JASP already. The way to do this is to go to the "Model" tab and then select all variables that should be part of the interaction term by pressing "control" (on a Mac, I think, it is "command") and then clicking the relevant variables. Finally, you need to press on the arrow pointing to the right (i.e., pointing to "Model Terms"). Note that, at the moment, this only works when the Enter method is used.
Also your suggested work-around (i.e., multiplying the values from the two predictors to form a new column) will work.