Prediction with one very strong and many weaker variables I want to create an RF model, with about 100 weak variables and one very strong variable.
The strong variable is a probability score, I do not have visibility on how it was derived. It may be using many of the variables I already have, and also some I do not have.
If I create Random Forest, wouldn't it mask the strong predictor with the weaker ones, since it randomly selects the variables at each stage?
What is the best way to deal with a very strong variable and many weaker ones - with the idea being to add value over the stronger variable?
 A: Since each variable has an equal chance of getting selected with each tree in the RF, and some trees in the RF will not include this strong variable, you don't need to worry about the strong "masking" the weak.  The RF's prediction is based on a consensus of all the trees in the forest regardless of whether or not they contain the strong variable.  You could examine the amount that each variable contributes across the entire forest using importance(), assuming you're using R's randomForest.  RF should be a good way to deal with this situation, but really you should hold back a random sample and use it to test several methods.  Also, other than importance() an RF is somewhat "black-boxy" - if all your after is pure predictive power then that's ok, but if you're trying to understand all the interactions then other techniques may be better.  
A: Multivariable linear regression may be an option for you; however, you need to ensure that the appropriate assumptions and adequate sample size are met. For feature selection, I recommend first assessing multicollinearity of potential predictor variables (without regards to the outcome) and the R package, glmnet, to make use of lasso and ridge regression techniques.
