Multilevel analysis of survey data

Abstract

EditorialUsing the 1930 census, W.S. Robinson 1 showed that although there was a positive correlation between the literacy rate in each of the 48 states in the US and the proportion of the population born outside the US, the correlation was negative when individuals were considered. The reason for this discrepancy was that immigrants tended to settle in states where the native population was more literate. Because of possible ecological fallacy-bias, epidemiologists and public health researcher have felt hesitant about deducing conclusions about individuals using ecological data or data population-level. However, individuals in a population are more than a simple sum of a set of loners. Individuals in a population are living within an ecosystem and there is a complex interaction between the health of the individual and the contextual factors of the environment as well as personal factors.In this issue of International Journal of Public Health, two papers are published, in which the authors combined in their analysis information of individuals with information available on populations. They used multilevel models. These types of models are also known as hierarchical models because of (1) the structure of the data (e. g. students clustered within schools) and (2) the hierarchy within the model with the parameters of the within-population regressions at the bottom, controlled by hyperparameters, the parameters of the upper-level models. In the first paper 2 , an Index of Relative Socio-Economic Disadvantage score (IRDS) was used as an indicator of local area disadvantage. The effect of this contextual measure on health-related quality of life (physical, mental health indicators) and lifestyle factors was estimated after accounting for individual socio-economic factors. The authors found independent associations between local area IRDS and health-related quality of life, obesity and smoking. After adjusting for individual socio-economic factors, no independent association was seen between local area IRDS and a cardiovascular risk index, chronic lung disease symptoms, leisure time physical activity and high risk alcohol use. In the second paper 3 , the potential influence of school environment, defined as the school level opportunity for physical activity, on the physical activity of nine-year old children was studied. It was examined if the association between school opportunity and leisure time physical activity was different by gender and BMI status. The authors found that the association between the physical activity frequency of the children and the schoollevel physical activity opportunity differs by gender and BMI and that overweight boys profit from a school policy and environment favoring physical activity.Survey data are often collected with an inherent multilevel structure (e. g. students within school, patients within hospital or data from cluster sampling, annual series of a survey). Based on such sampling design, a more traditional approach of data analysis in both studies would have considered the varyi...

The Psychological Impact of COVID-19 on Healthcare Providers in Obstetrics: A Cross-Sectional Survey Study

“…Comparisons between the model with fixed coefficients and multilevel models were made using the corrected Akaike information criterion (Tabachnick et al, 2019). This further analysis was performed because of the fact that our survey data had an inherent multilevel structure: HCPs within University hospitals (van Oyen, 2009). Therefore, the defined multivariable logistic regression model with fixed coefficients would not have completely corrected for between-group (University hospitals) differences, potentially relevant due, for example, to the different incidence of COVID-19 cases in the four geographic areas.…”