Statistical Test Selection for Pediatric Obesity Research with Multiple Diagnostic Categories
Multinomial logistic regression is the best statistical test when analyzing the relationship between pediatric groups and obesity with more than 3 diagnostic criteria. 1
Why Multinomial Logistic Regression is Superior
The outcome variable in pediatric obesity research naturally creates multiple distinct categories (underweight, normal weight, overweight, obesity, severe obesity), requiring a statistical approach that can simultaneously analyze all categories rather than forcing them into binary outcomes. 1
Key Advantages Over Other Tests:
T-test is inappropriate because it only compares means between two groups and cannot handle the polytomous (multiple category) nature of obesity classification 1
Linear regression is unsuitable because BMI categories represent ordinal categorical outcomes, not continuous dependent variables that linear regression requires 1
Binary logistic regression fails because it can only analyze two outcome categories at a time, forcing you to either:
Evidence Supporting Multinomial Approach
Pediatric obesity assessment requires multiple BMI percentile categories including underweight, normal weight, overweight, obesity (BMI ≥95th percentile), and severe obesity (BMI ≥99th percentile or ≥120% of 95th percentile), creating 4-5 distinct diagnostic categories. 2, 1
Different risk factor profiles exist across these multiple weight categories, with severe obesity showing markedly elevated inflammatory markers (C-reactive protein >5 mg/L), oxidized LDL, and clustering of cardiometabolic risk factors compared to lower weight categories. 2
Research demonstrates successful application of multinomial logistic regression to predict weight categories based on BMI percentiles and demographic factors in pediatric populations. 1, 3
Common Pitfalls to Avoid
Never collapse multiple obesity categories into binary outcomes (obese vs. non-obese), as this loses critical information about the gradient of risk across weight categories and reduces statistical power. 1
Avoid running multiple binary logistic regressions for each category comparison, as this approach inflates Type I error rates and doesn't properly account for the relationships between all categories simultaneously. 1
Recognize that prevalence patterns differ significantly across weight categories by demographic factors (race/ethnicity, poverty level, age), making it essential to analyze all categories together rather than separately. 2, 3