Over the last decade increasing use has been made of individual household data to analyse the gains and losses from tax reform. Much attention has been paid to the econometric estimation of models of household responses to taxes. But these models yield valid estimates of the welfare consequences of tax changes only when the implied preference orderings are well behaved. This paper discusses the nature of such conditions in detail. Where there are rionlinearities in the budget constraint then two sets of "primal" and "dual" conditions must be satisfied. The analysis of these conditions yields suggestions for the specification of behavioural models and the use of individual-specific information in the observed data.