There is wide agreement that one of the major problems in the estimation of wage equations, and in testing theories of wages, is the quality of the data available. In particular, the data used in almost all studies involve indirect or unsatisfactory measures of a variety of relevant factors, including working conditions. In this paper we exploit an unusual data set that includes measures of characteristics of jobs generated by observation (rather than self report, or through inference from the Dictionary of Occupational Titles). With these data we estimate a wage equation in which compensating differentials play a significant role in wage determination and, once working conditions and other characteristics are effectively controlled, all but one measure of previous labour market experience had no effect on the wage rate. Copyright Fondazione Giacomo Brodolini and Blackwell Publishers Ltd 1997.