Estimating the Union/Non-Union Wage Differential: A Statistical Issue

Abstract

Prior to the availability of micro-data sets for the United States which indicated the union or non-union status of individuals, almost all estimates of the conditional average union/non-union wage differential, given union membership, were made using mean average wages from a sample of industries or occupations. The statistical and economic methodology employed in such studies is basically the same. It involves (a) the construction of a statistical relationship between the average wage and the extent of unionization and (b) the use of a maintained hypothesis about the economic structure of union/non-union wages under which the resulting estimate of the average conditional differential is unbiased or, at least, consistent. We show below that this incomplete data method has produced biased estimates of the union/non-union wage differential when applied to data from the United Kingdom. This is an important matter. In Britain a number of estimates of the differential have been made using this incomplete data method. A consensus has emerged that the differential for manual workers is around 25 per cent.' We show below that this estimate is quite likely to be biased by as much as 50 per cent. It is worthwhile stressing at this point that we are not referring to (potentially) biased estimates of the structural effect of unionism on wages,2 but rather to (observably) biased estimates of the conditional effect of unionism on wages.3 Before US estimates based on individual data became available, there was some consensus that the average differential was between 20 and 30 per cent.4 Micro-data estimates now put the US figure at around 12 per cent.5 There are no comparable micro-data estimates for Britain. Hence, it is important to know if the incomplete data estimates for Britain may be subject to the same biases as apparently arise when the procedure is used in the United States. In the next section we outline the differences between the (conventional) incomplete data method and some straightforward complete data methods. In the last section we apply both methods to industry average data from the New Earnings Survey (1974) and examine the resulting biases.