How to analyses sensitivity for understanding which variables are the most effect on the predictive model I have a dataset with 150 observations. The dataset has 9 input parameters and 1 output parameter. I have built a predictive model (Random Forest) using the dataset. And now, I want to know that which variables are the most effect on the constructed predictive model. How can I do?
P/s: Please help me with an example because I am a newbie in R.
Thanks in advance!
 A: As you used RF to build you model, you can use a sensitivity measure (also called importance measure) which comes (almost) for free when you build the Random Forest: mean decrease in accuracy (also called permutation accuracy importance). As @Frank Harrell said, it is better to estimate a few times the random forest training to have multiple importance measure and have a better idea of their true importance. 
See the paper of Strobl et al. (2007) for more details about the method.
Dependending on the library you used for your RF estimation, you could have it already computed or you may have to recompute it yourself.
In R, RandomForest and cforest packages provide it. In Python, scikit-learn does it too (feature_importances_ parameter). Same in Mllib.
If using R, use cforest without bootstrap, as advised in Strobl et al.
Note also that you can still apply any classical Sensitivity Analysis tool provided your problem is a regression (and not a classification). See HSIC indices, for example: https://hal.archives-ouvertes.fr/hal-00903283/document and https://www.rdocumentation.org/packages/sensitivity/versions/1.14.0/topics/sensiHSIC 
A: I am not convinced that with random forest the importance measures are stable.  Use the bootstrap to repeat the entire process and compute the importance measures a couple of hundred times, and show the variability in importance measures for each specific predictor.  You can use this also to compute 0.95 confidence limits on the importance ranking of each predictor.  If those confidence intervals are wide, you really can't know much about predictor importance.
