Multicollinearity in prognostic factor analyses using the EORTC QLQ-C30: identification and impact on model...

Area of Science:

• Oncology
• Biostatistics
• Health Outcomes Research

Background:

• Advanced breast cancer treatment relies on prognostic factor analysis for survival and chemotherapy response.
• Quality of Life (QL) data, measured by the EORTC QLQ-C30, is often included in these analyses.
• High correlations between QL subscales may lead to multicollinearity and model instability.

Purpose of the Study:

• To investigate the impact of multicollinearity, specifically from global QL, on prognostic factor analysis in advanced breast cancer.
• To assess the stability of multivariate models used for predicting chemotherapy response.
• To provide recommendations for optimizing the use of QL data in prognostic models.

Main Methods:

• Utilized clinical and QL variables from an EORTC first-line chemotherapy trial in advanced breast cancer.
• Employed forward and backward selection methods for multivariate model building.
• Performed correlation analysis, regression modeling, principal component analysis, factor analysis, and bootstrap techniques to assess multicollinearity and model stability.

Main Results:

• Multivariate models for chemotherapy response showed instability with different variable selections.
• Global QL was highly correlated with multiple other QL subscales, indicating potential multicollinearity.
• Factor analysis suggested global QL was redundant and its inclusion minimally impacted component loadings.
• Bootstrap analysis confirmed that excluding global QL enhanced model stability.

Conclusions:

• Global QL exacerbates multicollinearity issues in prognostic factor analyses using the EORTC QLQ-C30.
• Excluding global QL from prognostic models improves the stability and reliability of findings.
• Recommends omitting global QL in future prognostic factor analyses with the QLQ-C30 to ensure robust results.