Re: st: query on right-censored predictors/covariates

[Richard Goldstein <richgold@ix.netcom.com>]

Phil,

sorry for not mentioning that I have uncensored observations as well as
censored observations

thank you for the info about the -ice- options -- the variables are
*not* interval censored -- only the top is censored (e.g., I have the
actual counts if they are less than 10, but have a value of 10 for any
observation that has a count of at least 10; another example is that I
have actual age if less than 90, but anyone who is at least 90 years of
age has the value 90; the count variable is definitely not normally
distributed

Rich

On 6/24/10 12:46 PM, Phil Schumm wrote:
> On Jun 24, 2010, at 10:30 AM, Richard Goldstein wrote:
>> I have a data set in which the outcome is *not* censored but at least
>> 2 of the predictors (e.g., age which is "topcoded") are right censored
>> and am looking for any literature dealing with this; right now I am
>> leaning toward treating this as a missing data problem but I am
>> certainly open to other solutions also
>>
>> re: use of multiple imputation, I would certainly need to constrain
>> the imputed values to be at, or larger than, the current top value; I
>> see no way to do this in either -mi- or -ice-: have I missed something?
> 
> WRT imputation: If you have no uncensored observations in the dataset,
> then this is difficult to handle, since you have to make an assumption
> about what the form of the distribution is beyond the cutoff.  If you're
> willing to make a distributional assumption, then you can handle this
> with -ice-, using the -interval()- option.  More specifically, -ice-
> will generate imputed values for an interval censored variable by
> assuming that the underlying distribution is Normal; this means that you
> can also use this for any variable that can be transformed to
> approximately Normal (i.e., by generating the imputations on the
> transformed scale, and then transforming back to the original).  Note
> that -ice- will even let you impose an upper limit on the imputed values
> (i.e., so that you don't generate any predicted ages that are
> unreasonably high).
> 
> Perhaps if you had another dataset with uncensored versions of the
> variables you are interested in imputing and substantial overlap with
> the other variables in your model, you might be able to do something
> else...
> 
> 
> -- Phil
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/