How wrong is wrong? Monte-Carlo simulation experiments on measuring the neighborhood at the wrong scale

Open Access
Nau, Claudia L
Graduate Program:
Doctor of Philosophy
Document Type:
Date of Defense:
June 18, 2012
Committee Members:
  • Stephen Matthews, Dissertation Advisor
  • Stephen Matthews, Committee Chair
  • Glenn A Firebaugh, Committee Member
  • Barrett Alan Lee, Committee Member
  • Aleksandra B Slavkovic, Special Member
  • Murali Haran, Special Member
  • neighborhood
  • measurement
  • scale
  • census
  • tract
  • Monte-Carlo
  • simulation
  • BMI
  • health
“Neighborhood” is a fluid concept and its meaning can change depending on the outcome, exposure and social group under consideration. Administrative units such as census tracts, block-groups and blocks can therefore hardly fit the demands of the concept. Nevertheless, census units, in particular census tracts, have remained the single most common neighborhood proxy in neighborhood effects research, mostly because of their availability and practicality. While new sophisticated approaches to operationalizing the neighborhood are being developed, “how much” and under what circumstances the use of census units affects the results of neighborhood effects research remains to be assessed. Here, I present a Monte Carlo simulation experiment based on census data from Los Angeles County and parameters derived from models estimated on data from the Los Angeles Family and Neighborhood survey (LA.FANS), to assess the sensitivity of neighborhood effects to measuring the neighborhood at the “wrong” scale of the census geography. Through strategic manipulation of the simulation scenarios, I find that scale effects can introduce substantial downward bias into neighborhood effects estimates. The magnitude of the bias depends, however strongly on how much the distribution of neighborhood characteristics changes when being aggregated at different scales. Census tracts constitute the most robust measurement in all scenarios. Block-groups perform similarly to tracts and can therefore not serve as “middle ground” between tracts and blocks. Because block-groups are in average about one third of the size of tracts, boundary issues are unlikely to affect the prediction of neighborhood effects. The neighborhood level variance and thus, the commonly used intra-class correlation coefficient are highly sensitive to mis-measurement, with bias possible in both directions. This dissertation develops recommendations for neighborhood effects researchers on how to minimize the risk of bias if the scale of the neighborhood effect is not known or cannot be modeled due to data limitations.