Estimation of type I error probability from experimental Dixon's "Q" parameter on testing for outliers within small...

Area of Science:

• Statistical analysis
• Data science
• Outlier detection methods

Background:

• Significance tests in statistical software typically provide the probability (p) of type I error.
• Dixon's Q-test is widely used for outlier detection in small datasets (N=3-12).
• Current Q-test application relies on comparing experimental Q-values with tabulated critical values, yielding only a range for p.

Purpose of the Study:

• To develop a method for estimating the precise probability (p) of type I error for a given experimental Q-value and dataset size (N).
• To overcome the limitation of Q-test yielding only a range for p.
• To provide explicit expressions for p in terms of Q and N for small N.

Main Methods:

• A stochastic (Monte Carlo) approach was employed to estimate p for given Q and N.
• Explicit mathematical expressions for p were derived based on Dixon's equations for N=3 and N=4.

Main Results:

• A Monte Carlo simulation method was successfully developed to estimate the type I error probability (p) for Dixon's Q-test.
• Explicit formulas for p were derived and presented for datasets with N=3 and N=4 observations.
• The new method allows for a more precise determination of statistical significance when identifying outliers.

Conclusions:

• The developed Monte Carlo approach provides a valuable tool for accurately determining the probability of type I error in Dixon's Q-test.
• Explicit formulas for small N enhance the practical application of the Q-test by enabling precise p-value calculation.
• This work addresses a significant limitation in outlier detection methodology, improving statistical inference accuracy.