A rather peculiar and challenging question on X validated, concerning the absolute impossibility of a conditional expectation, given a non-sufficient statistic, being still a statistic (i.e., being independent on the parameter θ). Inspired from the following except from Hogg and Craig. Namely, could there exist a specific function φ(·) such that E[φ(Y¹)|Y³] does not depend on the parameter θ? I could not find a satisfactory explanation right away (and the question remains unanswered!)
After posting this entry, I thought anew that cases when the unbiased estimator φ(Y¹) is not a bijective transform of Y¹ would work as a counter-example, since Y³=φ(Y¹) is not sufficient but E[φ(Y¹) |φ(Y¹) ]=φ(Y¹) is not involving θ… And this case does not exhibit a paradox in that the variance does not decrease any further.
Xi'an's Og 