Measuring neighbourhood effects non-experimentally : how much do alternative methods matter?

European research attempting to quantify neighbourhood effects has relied almost exclusively on analyses of observational data. No consensus has emerged, perhaps because a variety of statistical procedures have been employed. We investigate this by exploring the degree to which alternative, non-experimental statistical methods yield different estimates of the relationship between neighbourhood income mix and individual work income when applied to the same longitudinal database. We find that results are highly sensitive to the statistical approach employed. Methods controlling for geographic selection bias generally reduce the negative association between low-income neighbours and individual earnings, but substantial differences across models remain. Controlling for both selection and endogeneity produces larger associations and evidence of non-linearity, something that is hidden in models only controlling for selection. All methods suffer shortcomings, so we argue for multi-method investigations to identify robust findings, with instrumental variables and fixed effects on non-mover samples being preferred. In our case, we find a substantial neighbourhood effect, regardless of the method employed.