If whole exit interviews are missing, it would seem to me that nonresponse weights might be better than imputation. Still, if you wanted to pursue imputation, it might be reasonable to model the sum as an ordinal logit model. Once you have a vector of estimated probabilities for each case, you can then make a random multinomial draw for each case.
If the missingness is in individual items, I would think that you would want to use Gibbs sampling where each item in the scale is assumed to be conditionally Bernoulli distributed given the other items in the scale as well as any available back ground variables. You can then make posterior draws for the missing items and then sum together with your reported items David Judkins Senior Statistician Westat 1650 Research Boulevard Rockville, MD 20850 (301) 315-5970 [email protected] . -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Howells, William Sent: Thursday, September 30, 2004 11:15 AM To: G.K.Balasubramani; [email protected] Subject: RE: [Impute] Modeling and Imputation for MNAR data set I don't know the correct way of modeling as far as imputation model, but I analyze similar data, and one issue that arose was whether to impute the sum of the items or whether to impute the individual items and then sum them up after imputation. We chose the latter. Bill Howells, Wash U, St Louis -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of G.K.Balasubramani Sent: Thursday, September 30, 2004 9:11 AM To: [email protected] Subject: [Impute] Modeling and Imputation for MNAR data set Hi all, I am working on the large data set on Major depression disorder. One of the outcome variable of interest is the Hamilton Depression Rating Scale(its a17 item scale). About 28% of the exit data are missing. I would like to impute the missing data for the outcome varaible. There are several covariates associated with the outcome of the data among which one variable is highly correlated with the outcome variable. What is the correct way of modeling this kind of data and later for imputation?. Thanks in advance for any help and suggestion on this question. Bala University of Pittsburgh -------------- next part -------------- An HTML attachment was scrubbed... URL: http://lists.utsouthwestern.edu/pipermail/impute/attachments/20041001/22775252/attachment.htm From BalaGK <@t> edc.pitt.edu Thu Oct 21 11:05:44 2004 From: BalaGK <@t> edc.pitt.edu (Balasubramani, G.K.) Date: Sun Jun 26 08:25:02 2005 Subject: [Impute] Multiply Imputation - Descriptive Stats Message-ID: <[email protected]> Hello all, This is a basic question in relation to imputation. That is, the imputed data is an outcome variable, which is Hamilton depression rating scale. I am using the threshold to create an indicator of remission or not remission. After I imputed the data (say for 5 times) , how do I show the descriptive statistics? That is, the percentage with remission when data include imputed values. (Ex. Sex with remission , Employment status with remission, etc..). Can I take the mean of the 5 imputed data sets to create the indicator variable for remission? Is there any other way to present the descriptive using the imputed data? Thanks in advance. Bala -------------- next part -------------- An HTML attachment was scrubbed... URL: http://lists.utsouthwestern.edu/pipermail/impute/attachments/20041021/c473ed91/attachment.htm From BalaGK <@t> edc.pitt.edu Thu Oct 21 11:10:38 2004 From: BalaGK <@t> edc.pitt.edu (Balasubramani, G.K.) Date: Sun Jun 26 08:25:02 2005 Subject: [Impute] Imputation for MNAR data set Message-ID: <[email protected]> Hi David, Thanks for your response. The missing value occurs due to dropout not by missing items of the 17 item Hamilton scale. There are only a few subjects whose sum is not taken into account due to one or more items are missing. As you mentioned in the reply, if I model the sum(max 42) of the rating scale value as an ordinal measurement, it may loose the originality of the value of the variable while I moving to the imputation. I have couple of other doubts about your reply, if I want to predict the sum of scores of the missing ness through the available covariates, why would I use the random draw for each case. Also can you please explain me the non-response weights. Thanks Balasubramani,G.K. Epidemiology Data Center University of Pittsburgh Pittsburgh, PA 15261 412-648-2625 Message: 1 Date: Fri, 1 Oct 2004 17:41:51 -0400 From: David Judkins <[email protected]> Subject: RE: [Impute] Modeling and Imputation for MNAR data set To: "'Howells, William'" <[email protected]>, "G.K.Balasubramani" <[email protected]>, [email protected] Message-ID: <[email protected]> Content-Type: text/plain; charset="us-ascii" If whole exit interviews are missing, it would seem to me that nonresponse weights might be better than imputation. Still, if you wanted to pursue imputation, it might be reasonable to model the sum as an ordinal logit model. Once you have a vector of estimated probabilities for each case, you can then make a random multinomial draw for each case. If the missingness is in individual items, I would think that you would want to use Gibbs sampling where each item in the scale is assumed to be conditionally Bernoulli distributed given the other items in the scale as well as any available back ground variables. You can then make posterior draws for the missing items and then sum together with your reported items David Judkins Senior Statistician Westat 1650 Research Boulevard Rockville, MD 20850 (301) 315-5970 [email protected] -------------- next part -------------- An HTML attachment was scrubbed... URL: http://lists.utsouthwestern.edu/pipermail/impute/attachments/20041021/832513ec/attachment.htm From depuy001 <@t> dcri.duke.edu Thu Oct 21 11:19:17 2004 From: depuy001 <@t> dcri.duke.edu (DePuy, Venita) Date: Sun Jun 26 08:25:02 2005 Subject: [Impute] Multiply Imputation - Descriptive Stats Message-ID: <[email protected]> Hi Bala et al - In the varous MI papers we work on in my group, we typically provide baseline descriptive stats for the unimputed group. If that is not an option, consider using either the first imputed sample or the overall imputated values. The overall MI mean for a value is merely the mean of the 5 (or however many) means, one from each dataset. However, you typically want to reporta measure of variance. For the unimputed or 1st imputed sample method, you can just use std dev. For the overall imputed values, you need to use standard errors. Personally, I prefer using unimputed for the baseline descriptives and full imputation values in subsequent analyses . . . but I would say the main deciding factor is the amount of missingness in your data. If it's very large, you will probably want to use imputed values. Hope this helps! Venita -----Original Message----- From: Balasubramani, G.K. To: '[email protected]' Sent: 10/21/2004 12:05 PM Subject: [Impute] Multiply Imputation - Descriptive Stats Hello all, This is a basic question in relation to imputation. That is, the imputed data is an outcome variable, which is Hamilton depression rating scale. I am using the threshold to create an indicator of remission or not remission. After I imputed the data (say for 5 times) , how do I show the descriptive statistics? That is, the percentage with remission when data include imputed values. (Ex. Sex with remission , Employment status with remission, etc..). Can I take the mean of the 5 imputed data sets to create the indicator variable for remission? Is there any other way to present the descriptive using the imputed data? Thanks in advance. Bala <<ATT93287.txt>>
