Applies to: NeuralTools, all releases

 

What is Variable Impact Analysis?

NeuralTools measures how sensitive the net’s predictions are to each independent variable. This analysis is done only on the training data. Each independent variable is assigned a Relative Variable Impact value, expressed as a percent; the values always add up to 100%.

• A lower percent indicates the variable has less effect on predictions.

• These results are relative: a variable with 1% impact is less important than one with 99%, but may still be useful if prediction accuracy is critical.

Purpose

Variable Impact Analysis helps you identify which independent variables are likely contributing most to accurate predictions. Low-impact variables might be removed or replaced with other variables to improve predictive power.

How the calculation works

1. Categorical variables: For each case, the analysis steps through all valid categories, measuring changes to the predicted value (using raw numeric net outputs for category predictions).

2. Numeric variables: For each case, the analysis steps through the range of values from minimum to maximum in the training data, recording changes in predicted values.

3. Delta values: For each case and variable, the difference between the maximum and minimum predicted value is calculated.

4. Mean delta: Average the delta across all cases for each variable: MeanDelta1, MeanDelta2, ..., MeanDeltan.

5. Relative impact: Calculate each variable’s impact as

Caveats

• Only the training set is used; Auto-Testing or Auto-Prediction data are excluded.

• Impacts are relative to the trained net; different networks or training runs may produce different impacts.

• Results are intended to guide variable selection, not provide definitive proof that a variable is irrelevant.

Using the results

• Variables with higher impacts are generally more influential in predictions.

• Low-impact variables may be removed or replaced during experimentation.

• Use the analysis to refine the set of independent variables for improved prediction accuracy.

 

 

Related Articles:

Interpreting the Percentages in Variable Impact Analysis

How to eliminate Variables Based on Impact Analysis?