Document Type

Working Paper

Date of Original Version



Voter preferences and valuation of public goods are often estimated using aggregated votes matched with Census data at the same spatial scale. However, this method may yield biased estimates for two reasons we examine in this paper: using Census data ignores the selection process of who votes, and relying on comparisons between aggregated units makes models susceptible to omitted variable bias. To assess bias, we use both Monte Carlo simulation and a case study regarding a statewide environmental bond referendum for which we have collected aggregate data and individual exit poll data. Our results confirm the two sources of bias and show that aggregate model regression coefficients can be incorrect in magnitude and even sign. We conclude that using aggregate data will likely lead to incorrect assessment of valuation and distributional impacts of public good provision.