An Empirical Investigation Of Step-Down Canonical Correlation With Cross-Validation.

Area of Science:

• Multivariate statistics
• Psychometrics

Background:

• Canonical analysis is a statistical method for examining relationships between two sets of variables.
• Stepwise procedures offer a method for variable selection within statistical models.

Purpose of the Study:

• To explore applications of stepwise procedures in canonical analysis.
• To evaluate alternative stepping decision rules.
• To assess the impact of variable reduction on canonical analysis results.

Main Methods:

• A stepdown procedure utilizing smallest interest multiple correlation was employed.
• Data from the Minnesota Importance Questionnaire and Minnesota Vocational Interest Inventory were used.
• A double cross-validation design with two random halves of 500 male participants was implemented.

Main Results:

• Significant variable reduction (up to 75%) was achievable with minimal loss in canonical R (R[SUBc]).
• Cross-validation coefficients were often higher after variable reduction compared to using the full variable sets.
• The applied stepwise method demonstrated robustness in maintaining predictive accuracy.

Conclusions:

• Stepwise procedures are valuable for simplifying canonical analysis models.
• Variable selection via stepwise methods can enhance the generalizability of findings.
• The smallest interest multiple correlation criterion is an effective stepping rule for canonical analysis.