Spatial Linear Regression from Census Microdata: Combining Microdata and Small Area Data

Census microdata have become an extremely valuable source of information in social sciences research. These data, however, must have very coarse geographic resolution in order to protect respondent anonymity. Thus the geographic scale of these microdata sources is drastically different from the scale of many spatial processes—particularly neighborhood-scale processes. It is suggested that this difference in geographic scales creates a problem of conclusion validity for regression models which use anonymized microdata: measures of statistical significance are biased in these models. A correction to this problem in which small area data and population-density maps are used to estimate the effects of spatial dependence is presented. Monte Carlo evidence is presented which demonstrates that the conclusion-validity problem may be severe in practice. Further, this evidence shows that the suggested correction with small area data restores conclusion validity to statistical tests.