How Not to Do WLS Fitting in Calibration with Heteroscedastic Data

Area of Science:

• Data analysis
• Calibration modeling
• Statistical methods

Background:

• Least-squares fitting requires accurate weights for optimal results, typically derived from inverse variances.
• Traditional weighted least-squares (WLS) uses replicate measurements for variance estimation, which can be imprecise.
• Ordinary least squares (OLS) sometimes outperforms WLS when variance estimates are poor.

Purpose of the Study:

• To investigate the effectiveness of variance-function estimation for determining optimal weights in least-squares fitting.
• To compare the performance of variance-function weighted least-squares against ordinary least squares and traditional WLS.
• To identify the best weighting strategy within variance-function fitting.

Main Methods:

• Utilized variance-function estimation to model data variances as functions of the response variable (y).
• Compared variance-function weighting with ordinary least squares (OLS) and traditional weighted least-squares (WLS) using Monte Carlo simulations.
• Evaluated four weighting methods within variance-function fitting, including iterative reweighting.

Main Results:

• Variance-function estimation provides precise calibration weights, significantly improving WLS performance.
• Variance-function weighted least-squares consistently outperformed OLS across various numbers of x values and replicates.
• Iterative reweighting within the variance-function fitting method yielded the best results.

Conclusions:

• Variance-function estimation is a superior method for obtaining weights in least-squares fitting compared to traditional WLS approaches.
• This method enhances the accuracy and reliability of calibration models, especially when data precision varies.
• Iterative reweighting is recommended for maximizing the benefits of variance-function weighting.