Comparison of Oridinary and Multilevel Logistic Regression in Determining Goiter Related Factors in Children

The aim of this research was comparison of multilevel and ordinary logistic regression in determining goiter related factors in children, based on hierarchical structure of the data.

Data of 35747 schoolchildren, aged 8-10 years, was collected in 2007 by the Research Institute of Endocrine Sciences, Shahid Beheshti University of Medical Sciences. Stratified and cluster random sampling methods were employed in 30 provinces across the country. Goiter status, sex, area of residence, altitude from sea level and status of Iodized salt storage in households were documented. Multilevel logistic regression model was applied and the results compared to those obtained using ordinary logistic regression.

The study showed that 5.74% of the pupils suffered from goiter. The variance of the 2nd level (districts) was estimated as 0.941 (SE=0.092) in the two-level model (p<0.001), in which no significant relation was seen between goiter and altitude from sea level (OR=1.27 95%CI: 0.78, 1.62). In the three-level model, the variance in the third level (provinces) was 1.031 (SE=0.287) (p<0.001). In the final three-level model, only sex and area of residence were found to be significant.

Results showed that, compared to two and three level logistic regression models, the variances of parameter estimates are underestimated in the ordinary logistic regression.