4
$\begingroup$

I have asked several related questions (1, 2, 3), but now I would like to ask the most basic questions and hope to get a very solid answer.

I have 40 treatment variables, and I am interested to find out which ones are related to my dependent variable. I want to do this in an entirely data-driven way. I also have two variables that I would like to control for. One of these control variables is significantly correlated with several of my predictors.

My approach at the moment is to run an adaptive LASSO, forcing in the two control variables (by setting lambda to 0 at both steps of the adaptive LASSO).

  1. Does using Adaptive LASSO make sense? If not, what approach would be better?
  2. Does my way of dealing with the control variables make sense? If not, how should I do it?
Improve this question
$\endgroup$

2 Answers 2

Reset to default
0
$\begingroup$

To find out which variables have the most correlation with a dependent variable, the simplest way I can think of is to perform a Parial Least Squares (PLS) analysis in that data with your dependent variable defined as output. Then, you can obtain the VIP (Variable Importance in Projection ref, ref) values, oe even simpler to analyze that, the VIP plot, and immediately you get a visually representation of the variables with the biggest influence for the prediction of your dependent variable. USually, the variables with a VIP>1 (this treshold varies between 0.8 and 1.2) are the ones with the most impact.

You can also yield a plot representative of how each variable correlates negatively or positively with your dependent variable (of course you can also assess this through the loadings plot), so a PLS could really be a very easy way to solve your problem.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thank you! Is there a special way to deal with control variables? Or are they simply included in the model. $\endgroup$
    – Dave
    Commented Sep 16, 2020 at 12:11
  • $\begingroup$ Are the control variables used for prediction purposes? If not, I would say not to include them, but I really am not sure how to deal with that specific issue and I don't want to lead you to a wrong answer on that $\endgroup$
    – Johanna
    Commented Sep 16, 2020 at 14:04
0
$\begingroup$

I want to do this in an entirely data-driven way

If you are interested solely in prediction then there are many different approaches including partial least squares and regularization.

However,

I have 40 treatment variables, and I am interested to find out which ones are related to my dependent variable

implies that you are thinking causally. If so, you cannot do this "in an entirely data-driven way" and hope to get sensible results. There is no way to know which variables are potential confounders or competing exposures, and should be included, and which are mediators and should be excluded.

See this answer for the kind of things that can go wrong with a data-driven procedure:
How do DAGs help to reduce bias in causal inference?

Improve this answer
$\endgroup$
4
  • $\begingroup$ Thanks for the reply. Are you saying that I need to instead create a model based on knowledge of the subject area, and then test it? With regards to data-driven approaches, is there on that is "less bad"? $\endgroup$
    – Dave
    Commented Sep 16, 2020 at 12:35
  • $\begingroup$ Yes, if you are thinking causally that's exactly what I mean. Take a look at the linked answer to see what can go wrong. That should convince you that anything else is bad, and I don't think "less bad" is a good way to be thinking. Why do you seek after a driven approach? $\endgroup$
    – Robert Long
    Commented Sep 16, 2020 at 12:43
  • $\begingroup$ I should say, in my paper, I have an analysis that tests a theoretically derived hypothesis directly. I wanted to compliment that with an entirely undirected exploratory analysis. $\endgroup$
    – Dave
    Commented Sep 16, 2020 at 12:58
  • $\begingroup$ Hmmm. I really don't think that would be useful for your readers. It's good that you are testing the theoretically derived hypothesis. Comparing it to a data driven approach is meaningless in my opinion. You might consider other theoretically derived hypotheses instead. $\endgroup$
    – Robert Long
    Commented Sep 16, 2020 at 13:02

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.