You may want to read this article first: What is a Sufficient Statistic?

## What is the Rao-Blackwell Theorem?

The **Rao-Blackwell theorem **(sometimes called the *Rao–Blackwell–Kolmogorov theorem* or *Rao-Blackwellization*) is a way to improve the efficiency of initial estimators. Estimators are observable random variables used to estimate quantities. For example, the (observable) sample mean is an estimator for the (unknown) population mean.

More formally, the theorem is defined as follows (from Sahai & Ojeda):

Suppose g(X) is an unbiased estimator of a scalar parametric function h(θ) and T

_{i}(X), i = 1,2,…,p, are jointly sufficient for (θ); then there exists an estimator u(T), depending on the data only through the sufficient statistics, such that E[u(T)] = h(θ) and Var[u(T)] ≤ Var[g(X)].

For a proof of the theorem, read Prof. Charles Elkan’s Intuition, Lemmas and Start of Proof.

Note: Although Rao-Blackwellization can improve an estimator, it doesn’t always produce a uniformly minimum variance unbiased estimator.

## Sufficiency

Improvement in efficiency is obtained by taking the statistic’s conditional expectation with respect to a sufficient statistic (assuming one exists). A sufficient statistic is a statistic that summarizes all of the information in a sample about a chosen parameter. In other words, a statistic is “sufficient” if it retains all of the information about the population that was contained in the original data points. According to statistician Ronald Fisher, “…no other statistic that can be calculated from the same sample provides any additional information as to the value of the parameter.”

Rice (2006) states the basic idea is that **“…if an estimator is not a function of a sufficient statistic, it can be improved.”** The improved statistic is optimal by any measure, including the mean squared error criterion or any other standard.

## Applications

The Rao-Blackwell theory has many applications, including estimation of prediction error and producing estimates from sample survey data. For example, observations in adaptive sampling are found in sequence; each new observation depends on one or more characteristics from prior observations. Improved estimators can also be found by taking an average of estimators over every possible order .

**References**

Elkan, C. Rao-Blackwell Theorem: Intuition, Lemmas and Start of Proof. Retrieved July 5, 2017 from: http://cseweb.ucsd.edu/~elkan/291winter2005/lect04.pdf

Fisher, R.A. (1922). “On the mathematical foundations of theoretical statistics”. Philosophical Transactions of the Royal Society A 222: 309–368.

John Rice (2006). Mathematical Statistics and Data Analysis, Page 3

Hardeo Sahai, Mario M. Ojeda. Analysis of Variance for Random Models, Volume 2: Unbalanced Data.

If you prefer an online interactive environment to learn R and statistics, this *free R Tutorial by Datacamp* is a great way to get started. If you're are somewhat comfortable with R and are interested in going deeper into Statistics, try *this Statistics with R track*.

*Facebook page*and I'll do my best to help!