How accurate are surveyed preferences for public policies? Evidence from a unique institutional setup