Increased efficiency of analyses: cumulative logistic regression vs ordinary logistic regression

Area of Science:

• Biostatistics
• Epidemiology
• Dental Research

Background:

• Collapsing continuous or ordinal variables into dichotomous categories leads to information loss.
• This information loss can reduce statistical power and increase the risk of Type II errors.
• The study addresses the impact of variable specification on detecting effects in statistical inference.

Purpose of the Study:

• To illustrate the power increase from using ordinal versus dichotomous specifications for continuous responses.
• To test the hypothesis that poor diabetes control increases alveolar bone loss risk over 2 years.

Main Methods:

• Employed ordinary logistic regression (OLR) and cumulative logistic regression (CLR) modeling.
• Analyzed data from 359 subjects, including 21 with non-insulin-dependent diabetes mellitus (NIDDM).
• Compared dichotomous and 4-category ordinal specifications for radiographic bone loss.

Main Results:

• OLR with dichotomous outcome showed non-significant effects for diabetes control.
• CLR with ordinal outcome revealed a significant effect for poorer diabetes control on bone loss.
• CLR provided a significantly better model fit (P < 0.05) and converged for interaction analysis, unlike OLR.

Conclusions:

• Ordinal logistic regression offers increased power, more precise estimates, and better model fit than dichotomous specifications.
• Poorer controlled NIDDM significantly increases the risk of severe alveolar bone loss progression.
• Using ordinal specifications is crucial for accurate risk estimation in studies with continuous or ordinal responses.