Model-Based Clustering with Measurement or Estimation Errors.

Area of Science:

• Statistics
• Bioinformatics
• Computational Biology

Background:

• Model-based clustering using finite mixture models is a prevalent technique.
• Existing methods like MCLUST do not inherently account for estimation errors in clustered summary statistics.
• Summary statistics, such as regression coefficients, often possess associated estimation errors.

Purpose of the Study:

• To propose MCLUST-ME, an extension of Gaussian finite mixture modeling that incorporates estimation errors.
• To investigate the impact of explicitly modeling estimation errors on clustering performance.
• To provide new insights into data by accounting for uncertainty in summary statistics.

Main Methods:

• Developed MCLUST-ME, a Gaussian finite mixture model extension.
• Assumed each observation comprises a true component distribution and an independent measurement error distribution.
• Evaluated performance through simulations and application to RNA-Seq data.

Main Results:

• MCLUST-ME generates different groupings than standard MCLUST by considering unique estimation error covariances.
• Explicitly modeling estimation errors can enhance clustering performance under specific conditions.
• The degree of improvement is contingent upon factors like the distribution of error covariance matrices.

Conclusions:

• MCLUST-ME offers an improved approach to clustering when dealing with summary statistics containing estimation errors.
• Accounting for estimation errors provides valuable insights and potentially better data groupings.
• The effectiveness of MCLUST-ME is influenced by the characteristics of the error covariance matrices.