Choosing variables that affect the dependent variable most? I would like some help in choosing best methods of choosing variables that affects the dependent variable the most.
My data consists of gathered survey responses where all answers have scoring between 1 and 7.
The data is then calculated into fractions(0%-100%) of people answering 6 and 7 for each independent variable.
The dependent variable is customer satisfaction(0%-100%), while there is 16 independent variables giving information about the customers experience of service quality at a store.
Since the employees at the store are not able to focus on improving all the factors at the same time I want to understand which independent variables that should be focused on and that improves customer satisfaction the most.
The method used today is stepwise regression. But as i have read in numerous articles and forums this is an "outdated" method yielding incorrect answers.
I have looked at lasso regression and relative importance as methods that could work, but would love some guidance in the right direction on this and tips on methods/models that can be useful!
 A: There are no importance measures that replicate unless the sample size is enormous.  The data do not possess the needed information to reliably measure the contribution of each predictor, unless you have a tightly controlled experiment with orthogonal factors.  This is true whether you use ordinary regression or penalized regression such as lasso. If you use the bootstrap to get confidence intervals for the importance measures you'll get the picture.  An example is in Chapter 4 of RMS.
A: I suggest finding non linear relationships between a feature and the response by obtaining the Maximal Information Coefficient between both. While correlation may not show relationships (values near 0), there may be strong non linear relationships that suggest strong ability to predict.
In R you can use for this the package minerva.
A: If you are using PLS-SEM for the model estimates, you could consider using the Importance-Performance Map Analysis (IPMA) in SmartPLS. The analysis can help with assessing the importance of an independent variable for the dependent variable. A good example of the use of IMPA analysis is available in Rahman et al (2022).
