Anonymous posted on Thursday, July 21, 2005 - 9:45 am
I'm setting up a DTSA in mplus. A few people in my dataset have an intermittent missing value. For example: 0 0 . 0 0 1. Is it necessary to censor the person at the second assessment and discard the later three observed waves? Or is it appropriate to allow full information maximum likelihood to account for the missing assessment.
bmuthen posted on Thursday, July 21, 2005 - 1:36 pm
Here is the opinion of Katherine Masyn:
Although there may be information lost, the easiest thing is to censor those with intermittent missing at the last observation before the first missing assessment. If there are only a "few" people in the data set with this kind of missingness, that shouldn't be a problem.
It's really not appropriate to put in all the observed data and then use FIML b/c if the missing value is "1" then that person should not be contributing to the risk set of any subsequent times periods.
I'm looking into multiple imputation for a data set that has many missing assessments like this. However, this will still require the appropriate reformatting of each imputed data set before fitting the survival models.
Anonymous posted on Monday, July 25, 2005 - 8:40 am
Thanks for the response. I have one follow-up question. Since type=missing is always used to specify a DTSA, how does mplus know to treat the periods of time after an event or after censoring as "no longer at risk" rather than using FIML to estimate those values?
FIML doesn't estimate the missing values. I think it is easier to think of FIML as disregarding the missing values - not including them in the likelihood computations. FIML under MAR is just a principle for using all non-missing values.
Dear Dr. Muthens: I am new to Mplus. I have a question about how to get the HRs and their confidence intervals from the DTSA model. 1. Is there any existing command that directly procoduces HRs? 2. If not, can I use the constraint command to generate another parameter indicating the exponential of the coefficients? (I tried the constraint command, but found that the CIs for the new parameters were not the same as the CIs of the HRs calculated by hand based on the coefficients) 3. Any other convenient ways?
1. No 2. Yes - that's the way to do it. The confidence limits are not the same because Mplus does a delta method to estimate Exp(beta) asymptotic distribution and based on that it produces CI. Your method and the Mplus method are asymptotically the same and it is not known in general which method is better for finite sample size. One could conduct simulations to figure out which method is better. 3. No