I know that a complete sufficient statistic $T$ is such that 1) $T$ is sufficient for $\theta$, unknown parameter and 2) $T$ is complete.

So, is it always the case? If the answer is not, what could be an example of a complete but not sufficient statistic? Instead is it true in exponential family distributions?


For completeness you need to consider only the distribution of the statistic $T$ (more correctly, the family of distributions indexed by $\theta$), whereas for sufficiency you need to consider the complete likelihood for $\theta$ as a function of the sample $X$; so it's trivial to come up with examples where $T$ isn't sufficient but is still complete:

  • The constant statistic $T(X)=7$.

  • Throw some of your data away & calculate your (previously complete & sufficient) test statistic on what's left. Suppose you've four observations $X_1,\ldots,X_4$ from a Gaussian with mean $\theta$ (& known standard deviation): $T(X)=\frac{X_1+X_2}{2}$ isn't sufficient for $\theta$ but is complete.

  • Twice the sample mean from a random variable uniformly distributed between $0$ & $\theta$, $T(X)=\frac{2\sum_{i=1}^n X_i}{n}$ (the method-of-moments estimator), is complete, but we know the sufficient statistic for $\theta$ is the sample maximum $X_{(n)}$.

See Complete sufficient statistic for a more interesting example—of a sufficient statistic that isn't complete.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.