Dealing with earnings bracket responses in household surveys - How sharp are midpoint imputations?