Dichotomisation using a distributional approach when the outcome is skewed

Area of Science:

• Biostatistics
• Health Research Methodology
• Statistical Modeling

Background:

• Dichotomizing continuous outcomes risks information loss, yet is often needed for risk communication.
• Existing distributional methods for dichotomization apply to normally distributed data, not skewed health outcomes.
• Skewed data is common in health research, necessitating alternative dichotomization approaches.

Purpose of the Study:

• To present a robust methodology for dichotomizing skewed continuous outcomes.
• To evaluate the method's validity and robustness using simulation studies.
• To enable the presentation of dichotomized outcomes with precision, reflecting mean comparisons.

Main Methods:

• Developed a distributional methodology for dichotomizing skewed variables.
• Utilized data from multiple observational studies for illustration.
• Conducted a simulation study to assess method robustness to non-normality.

Main Results:

• The skew-normal method is reliable for skewed variables like gestational age where transformations fail.
• Normal distributional methods are reliable for outcomes transformable to normal (e.g., birthweight, blood pressure, BMI) or with minor deviations from normality.
• The distributional approach effectively handles skewed data, allowing for precise proportion comparisons.

Conclusions:

• The developed distributional method is applicable to common skewed health data.
• Researchers can provide both continuous and dichotomized estimates without information or precision loss.
• This facilitates a practical understanding of mean differences in terms of proportions.