An empiric study of ecological inference

Aggregate-level (macro) data are sometimes used when examining healthy care issues. Although they may be more accessible than individual (micro) observations, their interpretation is subject to ecological bias which in most cases is not measurable. This paper examines the implications of using aggregate-level data by conducting two separate analyses (micro and macro). Using as a database hospital episodes of care for the North Carolina Medicare aged population, regression models are developed from an examination of geographic grouping effects to explore the impact of extended care services, skilled nursing facility, and home health agency care on acute care hospital days. Specific problems encountered are: variable definition, collinearity, variance reduction, dilution of effect, spurious correlation, and observation influence. Stronger collinear (correlation among independent variables) relations occur at the macro-level than at the micro-level and spurious macro-correlations result from model specification and definition of interaction effects.