A Brief Assessment of Neighbourhood Effect on Neonatal Mortality: Translation of Area Level Variance in the Odds Ratio Scale in Multilevel Logistic Regression

Tuesday, 19 August 2014
Exhibit hall (Dena'ina Center)
Bhaskar Thakur, PhD , All India Institute of Medical Sciences, New Delhi, India
Vishnubhatla Sreenivas, PhD , All India Institute of Medical Science, New Delhi, India
Sadanand Dwivedi, PhD , All India Institute of Medical Science, New Delhi, India
Arvind Pandey, PhD , INDIAN COUNCIL OF MEDICAL RESEARCH, NEW DELHI, India
INTRODUCTION:

 Logistic regression is frequently used in epidemiological and public health research to measure the binary outcome.  Unlike linear regression analysis, logistic regression does not agree the interpretational attribute of the normal model.  It is always contend the dissatisfactory of existing measures while quantifying results from multilevel logistic regression model. The variability at different levels is not directly comparable. Quantifying area-level variance in a meaningful way is a challenge in multilevel logistic regression

METHODS:  

We obtained individual and district level information on the binary outcome neo-natal mortality from District Level Household Survey-3. The exploration of data structure confesses the consideration of only two-level structure in analysis, conceptualized as children nested within districts. Estimations of Variance Component Model (empty model) and Random Intercept Model in multilevel logistic analysis were carried out. The latent variable method converts the individual level variance from the probability scale to the logistic scale to compute the intra-class correlation. The median odds ratio translates the area level variance on the odds ratio scale. 

RESULTS:  

The median odds ratio was equal to 1.60, in the empty model which shows  if a person moves from one district to another district with a higher probability of  neonatal mortality, their risk of mortality (in median) will increase by 1.6 times, when randomly picking out two persons in different districts.  After adjusting the individual effect in the random intercept model, this ratio reduced to 1.54.  Area level variance and Intra-class correlation were 0.246 and 0.067 in the empty model as well as 0.210 and 0.059 in the subsequent model respectively.

CONCLUSIONS:  

The usual odds ratio are not proper interpretable for district-level covariates because it is impossible to make comparison within district. As MOR quantifies cluster variance in terms of odds ratios, it is comparable to the fixed effects odds ratio and can be useful in epidemiological studies.