A more general description of missing mechanism is a coarsened mechanism that includes
full data as a special case. So, CAR (Coarsened At Random), CCAR (Coarsened Completely
At Random), etc. are probably better terms...
-----Original Message-----
From: Laaksonen Seppo [mailto:[EMAIL PROTECTED]]
Sent: Thursday, March 29, 2001 12:06 PM
To: Donald B. Rubin; Laaksonen Seppo
Cc: [EMAIL PROTECTED]
Subject: IMPUTE: Re(2): MAR
Thanks Don,
You are right in some sense as that the mathematical meaning is OK, but for
practitioners such as me, Frank E Harrel Jr (he says: "I like the missing at
random conditionally term - MARC") or Werner Wothke ("In my experience, the
MAR-MCAR distinction is confusing at best. For non-statisticians, these two
terms simply
*sound* too similar to make a meaningful distinction. In addition,
people's common-sense understanding of the term "missing at random"
is really MCAR. This is reflected in Seppo's desire to change the
"MAR" terminology to "missing at random conditionally." In other words,
the MAR-MCAR terminology is somewhat of an obstacle to teaching
ML-based missing data methods to non-statisticians.") or many others (I just
was discussing this issue in the EU project meeting on editing and imputation)
this is 'confusing.' Why not clarify these terms so that all understand at
least the principles easily?
Seppo
>Dear Laaksonen,
>
>The term MAR was carefully selected more than a quarter century ago so as
>not to conflict with imprecise existing uses of the term in the
>statistical literature dealing with missing data in experimental design,
>MLE, etc. going back nearly half a century earlier. It may not be perfect
>but it has a well defined mathematical meaning, which was accepted by the
>editorial board of Biometrika of that time. "Unbiased" can also be viewed
>as confusing, and we could try to redefine it to be "expectation unbiased"
>versus "median unbiased" or the even more appealing "posterior unbiased"
>etc., but that effort doesn't seem that fruitful. Or even closer, the
>term "randomized" experiment could be viewed as confusing because it may
>suggest to some no blocking, etc. -- i.e., a completely randomized
>experiment. So maybe we should introduce the concept of the
>"conditionally" randomized experiment?? I think it is best to accept
>accepted definitions. Agree?
>
>Best wishes, Don
>
>
>
>On 29 Mar 2001, Laaksonen Seppo wrote:
>
>> I do not like about the term MAR, missing at random. Of course, when
>> it has been defined, there are no problems. But the direct
>> interpretation of that term is confusing, since missingness is not
>> random in this case but conditionally in some sense. The term should
>> be something like missing at random conditionally (MARC) or MAR
>> according to covariates. I am not fully satisfied to those terms. What
>> do you prefer?
>>
>> Best regards
>> Seppo Laaksonen
>>
>>
>>
>
