Question on variance of estimated parameters

Stuart Klugman ( stuart.klugman@drake.edu )
Tue, 27 Oct 1998 12:17:30 -0600

With less than 24 hours to go, I thought it reasonable to cease "lurking" and expand on my role as back up moderator to answer some of the last minute questions.

Any method of moments problem with one parameter is amenable to the type of question asked. For the Pareto, with lambda known (call it t), the moment equation is t/(a-1) = x with a for alpha and x for x-bar. The solution is a-hat = t/x + 1. Note that for any moment estimation, parameter-hat = g(x), some function of x-bar.

The general approximation for the variance of a function of a random variable is [g'(mu)]^2*Var(x) where mu = E(x). For the Pareto problem, g'(x) = -t/x^2. Also, mu = t/(a-1) and so g'(mu) = -(a-1)^2/t. Because a is unknown, it must be estimated, and the formula given in Jennifer Charlonne's question results.

This is slightly different from the formula/process for functions of parameters. The reason is that we assume that the parameter estimate is unbiased and so mu=parameter and we wind up with g'(parameter). In the above example, we are looking at a function of x-bar, which is not the parameter estimate.

------------------------------------------------------------------------------------------------
Stuart Klugman, FSA
Principal Financial Group Professor of Actuarial Science
Drake University
2507 University Avenue
Des Moines, IA 50311 USA
ph: 515-271-4097
e-mail: Stuart.Klugman@drake.edu