Transformed Variables in Survey Sampling

It can happen, especially in economic surveys, that we are interested in estimating the population mean or total of a variable Y, based on a sample, when a linear model seems appropriate, not for Y itself, but for a strictly monotone transformation of Y. In the present paper, we mainly focus on the important case where the transformation is logarithmic, but some new ideas introduced are not limited to that case. Currently available methods, based on the lognormal distribution, are reviewed, and two new methods introduced, one based on the idea of “smearing” (Duan, 1983), which do not require the lognormal assumption. Theoretical biases and variances are given, with suggestions for sample design and variance estimation, and a practical measure for reducing sensitivity to deviant points is suggested. We evaluate and compare the different estimators we describe in an extensive empirical study based on four economic populations taken from the UK Monthly Wages and Salaries Survey.