3

I performed a correlation analysis on my IVs to see which are related. As this is data from an experiment, I also have variables that are in general not so easy to capture from people in real life without a questionnaire. So I was wondering if there's a chance that I can combine a set of (highly correlated) variables to replace another variable.

For example, variable A is correlated with variables B, C, D, E. Could I somehow create a replacement for variable A by combining variables B,C, D, and E?

I know that a factor analysis is able to do this, but I believe this wouldn't work in my case. Performing a factor analysis on my the example that I provided, would need variable A to be included, while I want to find a way to exclude it. Or am I wrong here and is factor analysis the right approach.

user40037
  • 97
  • 2
  • 6
  • 1
    Donot omit any variable but rather go for multiplying or dividing the correlated to get a new variable. This is called 'feature engineering'. A ton of methods are available in google scholar research papers – Saumyajit Parida Jun 09 '20 at 15:51

1 Answers1

4

You can try regressing A in terms of B, C, D and E and then use that equation rather than the exact 'A' value

You might want to give Structural Equation Modelling a go. This Introduction to structural equation modeling gives you a lot of references and links to get started on.

Community
  • 1
Abhilash
  • 105
  • 8
  • Thanks! I believe you mean performing a regression with A as my DV and B, C, D, and E as my IV? I guess the regression equation would be fine eventually, but for now I do want to test whether the "replacement A" is a good enough predictor. SEM model sounds like a good idea, I already had that in mind but I never performed such test (nor worked with the software) before. Think it's a little tress hold that I need to overcome. Do you have any recommendations for the software to use? I can get a hold on AMOS, MPLUS, and of course R – user40037 Sep 26 '14 at 17:55
  • Yes, I was implying that you use A as your DV and rest of them as your IVs. Even I have just started exploring SEM, but your case seemed like it could use it. I predominantly use R , so my answer may be biased. I have found 'lavaan' package to be really helpful. It even has its own tutorial which explains all the things needed to work with the package – Abhilash Sep 27 '14 at 14:54
  • This is the one i use : [Lavaan tutorial](http://lavaan.ugent.be/tutorial/index.html) . Hope this helps ! – Abhilash Sep 27 '14 at 14:58