5
$\begingroup$

I'm having a hard time trying to relate or understand it in the simplest way (without solving).

"Without solving" in a sense that I don't have to solve for the marginal distribution of T2, if for example there are T1 and T2, and see that it is independent of the parameter. I'm just trying to find a way how to explain it to other people in layman's terms without using any technical statistical terms

$\endgroup$
2
  • 1
    $\begingroup$ Can you clarify what you mean by "(without solving)"? Also, can you include references to where you are trying to learn ancillary statistics from? $\endgroup$ Apr 11 '16 at 0:26
  • $\begingroup$ "Without solving" in a sense that I don't have to solve for the marginal distribution of T2, if for example there are T1 and T2, and see that it is independent of the parameter. So far, I've been using my lecture notes from my class, which only includes the usual definition that can be found online that T=(T1,T2) is sufficient for a parameter, where T1 is conditionally sufficient and T2 is ancillary. I also read Casella's section on Ancillarity. I'm just trying to find a way how to explain it to other people in layman's terms without using any technical statistical terms. $\endgroup$
    – ji-ln
    Apr 11 '16 at 0:32
5
$\begingroup$

The way I think of ancillary statistics is: a data set has numerous sources of randomness. A sufficient statistic carries all the information that I can extract in the data about this randomness. A minimally sufficient statistic is the "smallest" sufficient statistic. Now say that this has the form $S = (T,C)$.

Now I happen to be interested only in some aspects of the randomness ($T$) in the data, and I would rather condition on the uninteresting parts (treat $C$ as actually fixed to $c$). This makes sense to do when $C$ depends only on parameters that I am not interested in, and it is especially useful when $T|C$ has a simple distribution.

For example, I carry out a linear regression. I am interested in how some covariate affects the outcome, and I assume that $Y_i \sim N(\alpha + \beta x_i, \sigma^2)$ for $1 \leq i \leq m$. What I care about is $\alpha, \beta$ and $\sigma^2$. But there are also other sources of randomness: $x$ is probably realization of a random variable $X$ (uninteresting). The sample size is probably random as well (if it is not fixed in advance), and $m$ is a realization of a random variable $M$ (uninteresting). When I carry out my linear regression I condition on these aspects of the data. One reason is that I am not really interested in $X$ and $M$ per se. The second reason is that it would be very complicated to write a likelihood for all sources of randomness $(Y_i, X_i, M)$, $1\leq i\leq M$ at the same time.

If there is a minimal sufficient statistic $S=(T, C)$, and $C$ describes exclusively $X$'s and $M$ then I would rather condition on $C=c$, which is the same as treating $X_i = x_i$ and $M = m$ as fixed.

$\endgroup$
1
  • 1
    $\begingroup$ (+1) Very clear account. Typo in last sentence: "Ife is a minimal". Perhaps consider adding some discussion of the consequences of conditioning on $C$ for the relevance of the hypothetical long-run properties of estimators &c. $\endgroup$ Apr 11 '16 at 10:20

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.