Estimating the pay distribution in the new earnings survey using missing data methods

Abstract

The National Minimum Wage (NMW) was first introduced in Great Britain in April 1999. Measuring the impact of the NMW on levels of pay, and informing policy makers on how to set and change NMW levels require estimation of hourly pay distributions, especially at lower levels of the wage spectrum. Up to now, the hourly pay distribution was estimated using the hourly pay rate, which is derived from total earnings and hours worked. A major problem in producing reliable estimates is the difficulty in obtaining an accurate derived hourly pay, as the component variables are often prone to errors. The manual editing of all records potentially in error will be very costly. In the context of the introduction of the NMW, an alternative measure of hourly pay was introduced into the New Earnings Survey (NES), measuring the hourly pay rate directly. This direct variable appears to give a more accurate measure of hourly pay at the lower end of the distribution than the derived variable. Because the lower end is the main area of interest, we propose to estimate the hourly pay distribution on the basis of the direct hourly pay rate. However, the direct variable is subject to a high amount of missing data since it only applies to employees paid on an hourly basis. In this paper, using data from the 2002 NES, we evaluate alternative methods for imputing the true hourly rate for those employees who do not return a direct hourly pay. It is found that deterministic imputation methods and in particular, predictive regression imputation for hourly rates above £4, show advantages in terms of bias and root mean squared error at the lower end of the distribution.