# Sufficient statistics and randomized estimator in TPE

I've been reading Lehmann and Casella's Theory of Point Estimation 2nd Edition (TPE). In Chapter 1 Section 6 (pp.32-33), they introduce the idea "randomized estimator". Their explanation is, if $$X \sim \mathcal{P} = \{ P_\theta \mid \theta\in \Theta \}$$ and $$T$$ is sufficient for $$\mathcal{P}$$, since $$T$$ contains all the information on $$\theta$$, one can generate $$X'$$ only from the information $$T=t$$ such that $$X$$ and $$X'$$ have the same unconditional distribution, so do $$\delta(X)$$ and $$\delta(X')$$. They call $$\delta (X')$$ a randomized estimator.

I am struggling to make a sense of this statement. For example, $$X_1,...,X_n \sim N(\mu, 1)$$, iid, then $$\bar{X}$$ is sufficient for $$\mu$$. If we want to estimate $$\mu$$ by $$\delta(X) = \bar{X}$$, then $$\bar{X} \sim N(\mu, n^{-1})$$. Upon looking on the realization of the sufficient statistics, $$\bar{X} = \bar{x}$$, we can generate new sample by $$X_1', \dots, X_n' \sim N(\bar{x}, 1)$$. Then, the unconditional distribution of $$\delta(X') = \bar{X}'$$ is $$N(\mu, (2n)^{-1})$$, so the unconditional distributions do not coincide. What part I am doing wrong?

• Would this answer fit your question? – Xi'an May 9 at 17:05
• Thanks! Your answer and the suggested link helped me understand the concept. – user789100 May 10 at 3:37

Randomised estimators are defined in a more general setting than in this case and in particular need not be connected with sufficiency. In full generality, a randomised decision rule is a decision rule that returns a random decision for a given observation (or dataset). To reproduce the quote from Lehmann and Casella :

The starting point of a statistical analysis, as formulated in the preceding sections, is a random observable $$X$$ taking on values in a sample space $$X$$, and a family of possible distributions of $$X$$. It often turns out that some part of the data carries no information about the unknown distribution and that $$A$$ can therefore be replaced by some statistic $$T = T (A)$$ (not necessarily real-valued) without loss of information. A statistic $$T$$ is said to be sufficient for $$A$$, or for the family $$V = \{P_\theta,\ \theta\in\Omega\}$$ of possible distributions of $$A$$, or for $$\theta$$, if the conditional distribution of $$A$$ given $$T = t$$ is independent of $$\theta$$ for all $$t$$.

This definition is not quite precise and we shall return to it later in this section. However, consider first in what sense a sufficient statistic $$T$$ contains all the information about $$\theta$$ contained in $$A$$. For that purpose, suppose that an investigator reports the value of $$T$$, but on being asked for the full data, admits that they have been discarded. In an effort at reconstruction, one can use a random mechanism (such as a pseudo-random number generator) to obtain a random quantity $$X'$$ distributed according to the conditional distribution of $$X$$ given $$t$$. (This would not be possible, of course, if the conditional distribution depended on the unknown $$\theta$$.) Then the unconditional distribution of $$X’$$ is the same as that of $$X$$ , that is, $$P_0 (X' \in A) = P_\theta (X \in A)\quad\text{for all }A,$$ regardless of the value of $$\theta$$. Hence, from a knowledge of $$T$$ alone, it is possible to construct a quantity $$X'$$ which is completely equivalent to the original $$X$$. Since $$X$$ and $$X'$$ have the same distribution for all $$\theta$$ , they provide exactly the same information about $$\theta$$ (for example, the estimators $$\delta(X)$$ and $$\delta(X')$$ have identical distributions for any $$\theta$$).

The estimator $$\delta(X')$$ is possibly random given the observed realisation $$t$$ of $$T(X)$$. Each time $$\delta(X')$$ is considered, unless $$\delta(X')=\delta(X)$$ with probability one, a different realisation occurs. This means that $$\delta(X')$$ is a random variable for the observed realisation of the original data $$X$$, rather than a deterministic value, which explains for the following quote where the notion of randomised estimator is introduced.

The construction of $$X’$$ is, in general, effected with the help of an independent random mechanism. An estimator $$S(X')$$ depends, therefore, not only on $$T$$ but also on this mechanism. It is thus not an estimator as defined in Section 1, but a randomized estimator. Quite generally, if $$X$$ is the basic random observable, a randomized estimator of $$g(\theta)$$ is a rule which assigns to each possible outcome $$x$$ of $$X$$ a random variable $$Y(x)$$ with a known distribution. When $$X = x$$, an observation of $$Y(x)$$ will be taken and will constitute the estimate of $$g(\theta)$$. The risk, defined by (1.10), of the resulting estimator is then $$\int_{\mathcal X}\int_\mathcal{Y} L(\theta , y)dP_Y(y|X=x) dP_X(x;\theta,$$ where the probability measure in the inside integral does not depend on $$\theta$$.

In the special case of the Normal distribution proposed in the question,

1. the new sample $$(X_1^\prime,\ldots,X_n^\prime)$$ is generated conditional on $$T(X^\prime)=\bar X$$ and therefore $$\bar{X^\prime}=\bar X$$. In particular, if $$\delta(X)=\bar X$$, then $$\delta(X^\prime)=\bar X$$
2. the proposed estimator $$\delta(X^\prime)$$, namely the sample mean, is then non-randomised, which definitely cancels the appeal of the example! If instead, the median of the sample was considered as the estimator of the mean $$\mu$$, the estimator $$\delta(X^\prime)$$ would then be truly randomised, since the new normal sample $$X^\prime$$ would differ from $$X$$ and the median would remain random conditional on $$\bar X$$. (With expectation $$\bar X$$, though.)