Frank, I've looked at how IVEware imputes within a specified boundary and it's actually the opposite - the values tend to cluster in the middle approx 50% of the specified range, with rare imputations out to either extreme of the range. In fact, compared to known values, the program tends to overestimate low values and underestimate high values, so that the drift was towards the middle. Even with this issue, there appeared to be little influence in results of final analyses. Overall, I've found the program quite useful. Regarding the imputation of dates, I agree with Rod. Duration is much more likely to be imputed correctly, and should likely be normalized before entered into the MI model (assuming it is skewed). I have had some issues with collinearity of terms if 2 very closely related terms are included in the MI process (noted in your 2 values for start dates and 2 values for stop dates).
Craig >>> "Frank E Harrell Jr" <[email protected]> 08/31/06 9:28 AM >>> Paul Miller wrote: > Hi Rod, > > Thanks for your input. Since I wrote initially, I've had a couple of > ideas. The first one is similar to the duration variable approach you > suggested. The idea would be to introduce a duration variable into the > imputation dataset that would be calculated using cases with complete > data for start and stop. Then the stop date could be constrained to > equal start + duration. Or possibly the stop could just be directly > calculated as start + duration. > > The second idea involves creating a set of 4 variables: MIN_START, > MAX_START, MIN_STOP, and MAX_STOP. These can generally be created using > the limited date information that I have available. If, for example, I > know that a person started taking a drug in 2004 but nothing else, I can > calculate the minimum start as 01/01/04 and the maximum start as > 12/31/04. Then I can tell IVEware to constrain the imputed value to be > between these two dates. I've been playing with this approach a little > earlier today, and, so for, it seems to be working quite well. So now > I'm just hoping that the duration approach can also be successfully > implemented. > > Thanks, > > Paul I'm curious how IVEware does such constraints. Does it result in boundary problems, i.e., a mass of imputed values at the boundary of the constraint? Frank Harrell > > Paul J. Miller, Ph.D. > Research Scientist and Statistician > Ontario HIV Treatment Network > 1300 Yonge St., Suite 308 > Toronto, Ontario M4T 1X3 > Phone: (416) 642-6486 ext 232 > Fax: (416) 640-4245 > > -----Original Message----- > From: Roderick A. Rose [mailto:[email protected]] > Sent: Thursday, August 31, 2006 11:48 AM > To: Paul Miller; [email protected] > Subject: Re: [Impute] Imputing "Plausible" Start and Stop Dates for HIV > Antiretroviral Drugs > > Paul, > > My recommended solution is made under the (perhaps incorrect) assumption > > that what you are mainly interested in is the interval between the start > > and stop dates and not the actual stop and start dates themselves. Let > the > start date equal zero in every case (so it doesn't have to be imputed) > and > the interval is a count of days (or another unit) between zero and the > stop > date. You impute this interval. I've not used IVEware, so I'm not sure > this > will completely eliminate the problem (e.g., you might end up with > negative > intervals if the bounds statement really doesn't work well). > > Regarding the second issue of plausibility, I am curious if it is > necessary > to have precision in days; if you know it happened in May 1998, you can > err > on the side of the least undesirable bias (by making it either May 31 or > > May 1). This is an alternative to ignoring the known value and letting > it > impute a completely new and possibly unrelated value. (Or do both and > see > what happens, as many of us probably do). > > Best, > > Rod -- Frank E Harrell Jr Professor and Chair School of Medicine Department of Biostatistics Vanderbilt University _______________________________________________ Impute mailing list [email protected] http://lists.utsouthwestern.edu/mailman/listinfo/impute -------------- next part -------------- An HTML attachment was scrubbed... URL: http://lists.utsouthwestern.edu/pipermail/impute/attachments/20060831/4606f3f0/attachment.htm
