Parametric methods outperformed non-parametric methods in comparisons of discrete numerical variables.

Area of Science:

• Biostatistics
• Medical Research Methodology

Background:

• Discrete numerical variables, representing counts per individual, are common in medical research.
• Current methods for analyzing and presenting these variables lack standardization.
• There is a need for clear recommendations for discrete numerical data analysis.

Purpose of the Study:

• To provide recommendations for the analysis and presentation of discrete numerical variables.
• To evaluate statistical methods for comparing discrete numerical data.

Main Methods:

• Two simulation studies were conducted.
• Investigated hypothesis tests and confidence interval methods for discrete variables with varying outcome ranges (e.g., 0-5).
• Utilized the difference between means as the primary effect measure.

Main Results:

• The Welch U test and its confidence intervals demonstrated strong performance across most simulated scenarios.
• The Brunner-Munzel test performed well, except with small sample sizes (n=10 per group).
• The ordinary T test, Wilcoxon-Mann-Whitney test, and bootstrap methods showed unsatisfactory performance.

Conclusions:

• The difference between means is a suitable effect measure for comparing independent discrete numerical variables.
• Parametric hypothesis tests and confidence intervals are encouraged for analyzing discrete numerical data.
• The Welch U test is recommended for its robust performance in comparing discrete numerical variables.