Discussion of a Statistical Model of Earnings Estimation

Abstract

The participants in this conference may generally be classified as either professional or academic accountants. As an econometrician, I obviously do not qualify for either of those catagories; rather I suspect that the adjective amateur might be most appropriate in my case. Nevertheless, I hope to make some points which are relevant to the theory and methodology of Greenball's paper on earnings estimation. Noting that diversity in accounting practice hampers cross-firm comparison of accounting earnings, Greenball proposes a model of true earnings which involves three parameters and requires only data on sales, noninterest-bearing liabilities and net transfers to equity-holders, all of which are relatively unambiguous with regard to their measurement. Once the parameters of the model are estimated, the series of implied true earnings estimates is readily computed. Assuming normality for the disturbances in the model, maximum likelihood estimates of parameters are provided by nonlinear least squares by a search over the parameter space. It would have been interesting, simply as a computational matter, to see other procedures for nonlinear estimation attempted, for example, the iterative Gauss-Newton method which is based on the linear approximation to the model given by the Taylor series expansion around parameter values from the previous iteration. Regardless of the computational approach taken, it is easy to compute standard errors for parameter estimates and it would have been useful to have standard errors included in the paper. Under general conditions, maximum likelihood estimates are consistent with asymptotic variance-covariance matrix given by