Step 1 - Derivation of the discriminant function

• Variable selection
  
  
  • Dependent variable
    
    
    • Groups containing the observations of the dependent variable must be both mutually exclusive and exhaustive
    • Artificially constructed groups may be created from both order and interval data
      
      
  • Independent variables
    
    
    • Theoretical selection
    • Intuitive selection
      
      
• Sample division - Split sample (cross-validation ) approach
  
  
  • Analysis sample
  • Hold-out (validation) sample
    

A proportionately stratified sampling procedure - the number of elements contained in each group of the hold-out sample should reflect proportionately the number of observations in each group of the analysis sample.


The number of elements in the analysis and hold-out samples need not be equal.

For greater accuracy the sample can be split several times.


Not appropriate for small samples.

• Computational method
  
  
  • Simultaneous
  • Step-wise - This procedure is especially useful when interpreting the relative importance of the discriminating (independent) variables.
    
    
• Statistical significance
  
  
  • Acceptable level of significance is usually 0.05 or better. Beyond this level better classification is unlikely.
  • The standard statistic for measuring a discriminat functions significance is the Chi-square statistic.
    

Go to Step 2