方法对比

	Ecological Fallacy Diagnostics	Multilevel Neighborhood Effects
领域	Social Epidemiology	Social Epidemiology
方法族	Process / pipeline	Regression model
起源年份	1950	2000
提出者	William S. Robinson (ecological correlation); Sander Greenland & Hal Morgenstern (ecological bias theory)	Ana V. Diez Roux; Juan Merlo and colleagues
类型	Diagnostic and design pipeline for detecting and avoiding cross-level inferential bias	Hierarchical regression model for contextual effects on individual health
开创性文献	Robinson, W. S. (1950). Ecological Correlations and the Behavior of Individuals. American Sociological Review, 15(3), 351-357.	Diez Roux, A. V. (2000). Multilevel Analysis in Public Health Research. Annual Review of Public Health, 21, 171-192.
别名	Cross-Level Bias Diagnostics, Ecological Bias Assessment, Aggregation Bias Diagnostics, Ecological Inference Bias Checks	Contextual Effects Models, Hierarchical Neighborhood Health Models, Multilevel Analysis of Place and Health, Variance Partition / MOR Analysis
相关	4	3
摘要	Ecological fallacy diagnostics are the design and analysis tools used to detect, quantify, and avoid the bias that arises when associations measured on groups are mistakenly taken to hold for individuals. The problem was crystallized by W. S. Robinson (1950), who showed that the correlation between, say, immigrant share and illiteracy across U.S. states bore no resemblance to the correlation between being an immigrant and being illiterate among individuals, sometimes even reversing sign. Greenland and Morgenstern (1989) gave the modern account, decomposing ecological bias into within-group confounding, effect modification, and model misspecification, and clarifying that the ecological fallacy is not a single artifact but a family of cross-level biases. As a pipeline, the diagnostics contrast ecological and individual associations, attribute any discrepancy to its sources, model the within-group covariate distribution that aggregate analyses ignore, place bounds on the individual-level quantity, and where possible move to hybrid or multilevel designs that recover individual effects.	Multilevel models of neighborhood effects estimate how the places people live shape their health, over and above who those people are. Individuals are nested within neighborhoods, so their outcomes are not independent: residents of the same area share an environment and tend to be more alike than residents drawn at random. Ana Diez Roux's foundational synthesis showed that ordinary single-level regression ignores this clustering and conflates contextual effects (features of the place) with compositional effects (the mix of people in it), whereas a hierarchical model with neighborhood random effects separates the two. Juan Merlo and colleagues turned the method into an epidemiological toolkit by reframing the random-effect variance as substantively interpretable measures of variation, such as the variance partition coefficient and the median odds ratio, so that a study can report not only whether a neighborhood characteristic matters on average but how much of the health difference between people is attributable to where they live.