Re: imputation
- From: "Thom" <thom@xxxxxxxxxxxxxxx>
- Date: Wed, 10 Aug 2005 11:41:47 +0100
I'm not an expert on missing values, but I think I'd do it more than one way
and compare the results. If the conclusions don't depend on the choice of
analysis then you have some degree of robustness. If they do - you'll need
more research or help from someone more expert.
Thom
"Duncan Smith" <buzzard@xxxxxxxxxxxxxxxxxxxxx> wrote in message
news:ddb8qb$v14$1@xxxxxxxxxxxxxxxxxxxxxxx
> Thom wrote:
>> In theory imputation should always reduce noise ... I'm not sure throwing
>> away observations would ever help imputation because where imputation
>> works
>> it is using the information about what is missing and how much is
>> missing.
>> Assuming, I think, that the MAR assumption is tenable.
>>
>> Thom
>>
>
> Thom,
> Yes, this is the sort of thing I was wondering about (i.e. whether
> there was some theoretical argument that records should never be
> removed). But I guess it would have to depend on MAR or MCAR
> assumptions. But it's the results of subsequent analyses that really
> concerns me. I normally deal with noisy data where the noise has been
> added deliberately (for disclosure control purposes). It can have
> significant effects on e.g. Type 1 error rates if no adjustment is made
> to the test statistics. Including a record with very few known values
> and many imputed values, just seems to me like adding a little
> information and a little noise; and I don't see that this is
> unquestionably a good idea. The easiest thing for me to do (in terms of
> satisfying referees that I've done something sensible) is probably just
> to impute. But there are apparently 3 types of missing data in the
> dataset I've been supplied with, some can be imputed with certainty
> (done that), many almost certainly take a specific value and the rest
> are (possibly) MAR. The latter two types have been coded identically.
> Lovely. Cheers.
>
> Duncan
.
- References:
- imputation
- From: Duncan Smith
- Re: imputation
- From: Thom
- Re: imputation
- From: Duncan Smith
- imputation
- Prev by Date: Re: Agreement for one subject with gold standard, ordinal data
- Next by Date: A tool for analysis clinical trials
- Previous by thread: Re: imputation
- Next by thread: Modal Divergence
- Index(es):
Relevant Pages
|
