Sensitivity Analysis

Sensitivity analysis in data mining and statistical model building/fitting generally refers to the assessment of the importance of predictors in the respective (fitted) models. In short, given a fitted model with certain model parameters for each predictor, what the effect would be of varying the parameters of the model (for each variable) on the overall model fit.

In Statistica Data Miner, sensitivity analysis is available via several options; the particular statistics and measures that will be reported will dependent on the statistical or data mining method for which the sensitivity analysis is requested; for example:

1. In Statistica Automated Neural Networks, the program will compute the Sums of Squares residuals or misclassification rates for the model when the respective predictor is eliminated from the neural net; ratios (of the reduced model vs. the full model) are also reported, and the predictors (in the results table) can be sorted by their importance or relevance for the particular neural net. See also SANN regression example.

2. In all tree models (Classification and Regression Trees and Boosted Trees), sensitivity and predictor importance is computed from the average importance of each predictor at each split point (split node) in the final tree model (for additional details, see also Predictor Importance in Statistica GC&RT, Interactive Trees, and Boosted Trees).

3. In general linear models and generalized linear/nonlinear models (GLM and GLZ), various predictor statistics are computed so that analysts can gage the contribution of the respective predictors to the overall fit, and hence the importance of those predictors.