Statistical significance for hierarchical clustering in genetic association and microarray expression studies

Area of Science:

• Genetics
• Bioinformatics
• Statistical Analysis

Background:

• Vast datasets in molecular genetics (gene expression, haplotypes) necessitate data reduction for interpretation.
• Hierarchical clustering is commonly used to group similar observations into fewer classes.
• Researchers often select the most significant result at a specific clustering step, overlooking the process itself.

Purpose of the Study:

• To develop a method for assessing the overall statistical significance of hierarchically clustered data.
• To address the issue of inflated significance when focusing on a single optimal clustering outcome.

Main Methods:

• Propose using the strongest result (smallest p-value) as an experiment-wise statistic.
• Evaluate the significance level of this statistic for a global assessment.
• Applied the approach to haplotype association and microarray expression datasets.

Main Results:

• Selecting a single clustering step often yields significance levels that are too small.
• The proposed method provides a more accurate global assessment of statistical significance.
• Over-reliance on one clustering outcome can lead to formally significant results for non-significant overall experiments.

Conclusions:

• The clustering process itself impacts statistical significance and must be incorporated.
• A single, optimal clustering result may not reflect true experimental significance.
• Accurate interpretation of genetic data requires methods that account for the entire clustering procedure.