Best Practices for Binary and Ordinal Data Analyses.

Area of Science:

• Psychometrics
• Biostatistics
• Behavioral Genetics

Background:

• Human trait and disorder measurement often uses binary or ordinal questionnaire data.
• Clinical assessments frequently exhibit non-normal distributions in the general population.
• Recoding data into binary or ordinal categories can lead to loss of statistical power and violate analysis assumptions.

Purpose of the Study:

• To investigate the impact of treating ordinal data as continuous in statistical analyses.
• To evaluate biases in correlation and genetic variance component estimation when using inappropriate methods.
• To highlight the importance of appropriate statistical modeling for ordinal and binary data.

Main Methods:

• Simulation studies were conducted to compare different correlation estimation methods (Pearson vs. polychoric).
• Analysis extended to the classical twin model to assess genetic and environmental variance components.
• Odds ratios were examined in relation to population prevalence.

Main Results:

• Treating ordinal data as continuous, especially with skewed distributions, biases correlations towards zero.
• Pearson correlations are biased, while maximum likelihood estimates of polychoric correlations are unbiased but have larger standard errors.
• Incorrectly treating binary data as continuous underestimates genetic variance and overestimates unique environmental variance in twin models, with biases increasing as prevalence declines.

Conclusions:

• Appropriate statistical modeling of ordinal and binary data is essential to avoid biased correlations and parameter estimates.
• Failure to use methods like maximum likelihood estimation for polychoric correlations can lead to significant analytical errors.
• Accurate analysis of psychological and clinical data requires careful consideration of data distribution and measurement properties.