Probability binning comparison: a metric for quantitating univariate distribution differences

Area of Science:

• Biostatistics
• Data Analysis
• Computational Biology

Background:

• Comparing data distributions is crucial for detecting statistical significance between samples.
• Applications include identifying responses in test samples and monitoring instrument stability over time.

Purpose of the Study:

• To introduce and validate a novel statistical method called Probability Binning.
• To provide a robust metric for comparing univariate data distributions.

Main Methods:

• A variant of the chi-squared statistic is applied to univariate distributions.
• A control distribution is divided into equal-event bins, minimizing maximum expected variance.
• These bins are then applied to test distributions to compute a normalized chi-squared value.

Main Results:

• Monte-Carlo simulations determined the distribution of chi-squared values.
• A metric analogous to a t-score was derived, estimating the probability of distribution differences.
• The metric effectively ranks samples by similarity to a control and was applied to immunophenotyping data.

Conclusions:

• Probability Binning offers a valuable metric for assessing differences between flow cytometric data distributions.
• The method can rank distributions by similarity and quantify contamination in overlapping populations.
• It can also be used to gate significantly different subsets from control samples.