Small-sample estimation of negative binomial dispersion, with applications to SAGE data.

Area of Science:

• Bioinformatics
• Statistical Genetics
• Computational Biology

Background:

• Accurate estimation of the dispersion parameter is crucial for negative binomial models.
• Small sample sizes, common in gene expression studies, pose challenges for traditional methods.
• Existing methods for dispersion estimation can introduce bias, impacting downstream analyses.

Purpose of the Study:

• To develop a novel quantile-adjusted conditional maximum likelihood estimator for the negative binomial dispersion parameter.
• To evaluate the performance of the new estimator against existing methods, focusing on bias in small samples.
• To investigate the influence of dispersion estimation on hypothesis testing and develop an improved statistical test.

Main Methods:

• Derivation of a quantile-adjusted conditional maximum likelihood (CACML) estimator.
• Comparative analysis of the CACML estimator's bias against other methods using small sample datasets.
• Development of an 'exact' hypothesis test incorporating the improved dispersion estimation.

Main Results:

• The CACML estimator demonstrated superior performance, exhibiting less bias than all compared methods in very small samples.
• The motivating data from serial analysis of gene expression studies confirmed the practical relevance of small sample performance.
• The derived 'exact' test significantly outperformed standard approximate asymptotic tests in hypothesis testing scenarios.

Conclusions:

• The CACML estimator provides a more accurate and less biased estimation of the negative binomial dispersion parameter, especially for small sample sizes.
• This improved estimation directly benefits statistical analyses in fields like gene expression studies.
• The developed 'exact' test offers enhanced power and reliability for hypothesis testing in these contexts.