Spurious interaction as a result of categorization

Area of Science:

• Epidemiology
• Biostatistics
• Clinical Research

Background:

• Continuous variables are frequently converted into categorical variables in applied research.
• These categorized variables are then utilized as exposure variables in regression models.
• This common practice faces statistical objections, with this paper presenting an additional argument against it.

Purpose of the Study:

• To demonstrate that categorizing continuous variables can induce spurious interactions in regression models.
• To provide analytical expressions for spurious interaction occurrence in linear models with normally distributed variables.
• To interpret these findings through the lens of measurement error.

Main Methods:

• Analytical derivations for linear regression models with normally distributed exposure variables.
• Simulation studies to validate analytical results across different variable distributions.
• Examination of interaction effects when two variables are categorized at the same cut point.

Main Results:

• Categorization can lead to spurious interactions in multiple regression models.
• In linear models with two normally distributed exposure variables, spurious interaction arises unless categorized at the median or variables are uncorrelated.
• Simulation results confirm the general effect of categorization, with cut point choice impacting results across distributions.

Conclusions:

• Categorizing continuous exposure variables introduces significant problems, including spurious interaction effects.
• This practice should be discontinued in research.
• Alternative statistical methods should be explored for analyzing continuous exposure variables.