Effects of categorization method, regression type, and variable distribution on the inflation of Type-I error rate...

Area of Science:

• Epidemiology
• Biostatistics
• Statistical modeling

Background:

• Categorizing continuous variables leads to signal loss.
• This signal loss can inflate Type-I errors when the variable is a confounder in regression analyses.
• Previous research has not fully explored how Type-I error inflation varies across different regression models, confounder distributions, and categorization methods.

Purpose of the Study:

• To analytically quantify the impact of categorizing continuous variables on Type-I error rates.
• To estimate Type-I error inflation in various regression scenarios (logistic vs. linear), confounder distributions, and categorization techniques.
• To identify factors influencing the severity of Type-I error inflation.

Main Methods:

• Analytical quantification of categorization effects.
• Conducted 9600 Monte Carlo simulations to assess Type-I error inflation.
• Evaluated scenarios involving logistic and linear regression, diverse confounder distributions, and multiple categorization methods.

Main Results:

• Type-I error inflation was unacceptably high in most tested scenarios (often >10% and sometimes 100%).
• Error inflation increased with larger sample sizes, fewer categories, and stronger confounder-exposure/outcome associations.
• An exception was observed when categorizing a proxy for a dichotomous latent variable.

Conclusions:

• Categorizing continuous confounders significantly inflates Type-I errors in regression analyses, posing a substantial risk to research validity.
• Researchers must be aware of this inflation, particularly in scenarios with large sample sizes, few categories, or strong confounding.
• Online tools are provided to help researchers assess and understand potential Type-I error inflation in their specific studies.