You are here

Missing data in longitudinal study

4 posts / 0 new
Last post
newbie's picture
Joined: 09/07/2015 - 07:38
Missing data in longitudinal study

Hello all,
I am pretty new to SEM and I am currently using AMOS to analyse longitudinal data. I have data from the same participants and the same measurement instrument (EPDS for assessing postnatal depression) for 4 measurement occasions. I do have quite a lot of missing data (60% on one occasion as an axample) Some of the participants have not completed the EPDS on any of the 4 measurement occasions, and some have only completed it only 1 out of 4 times. I have spoken to two persons, and they have given me different advise. The first person who is familiar with SEM said to include all participants who have completed the EPDS, even if it was only on one occasion, and let AMOS handle the missing data with ML. The other person (a statistician) said that I should at least exclude the participants who had missing data on three out of four occasions, even if I am using ML estimation.
I am a little confused on what advice to follow, even if I delete participants, I still have more than enough, and it makes sense to exclude participants who only have answered on one occasion as that is not enough data to see how depressive symptoms change over time for that participant, but I have not seen that it is common to use listwise deletion by removing some participants, and then using ML for the rest of the missing data. I do not wish to create a bias by deleting certain participants.
What is your view on this? Would it be very wrong to delete the participants with the most missing data or is it the best solution?

I will be grateful for any answer.

Thank you..

mhunter's picture
Joined: 07/31/2009 - 15:26
Both ways

The people with only one measurement occasion aren't really helping your longitudinal model.

If it were me, I'd run it both ways. The results should be about the same. If they're not, let us know!

Ryne's picture
Joined: 07/31/2009 - 15:12
I disagree with Mike. People

I disagree with Mike. People with one measurement occasion are helping your longitudinal model. Presuming that scores at any timepoint predict dropout, they will help you yield unbiased measures of change. For example, it's common in cognitive studies for people with low cognitive abilities to drop out. If you restrict your model to only those with two or more measurements, your sample will be biased towards high performing individuals. This is especially an issue if dropout could be an important outcome (e.g., your measure predicts morbidity or mortality).

I also agree with Mike: run it both ways. If they disagree, the one with more data, however sparse, is usually correct.

AdminNeale's picture
Joined: 03/01/2013 - 14:09
What Ryne said

I agree with Ryne, single occasion data can reduce sampling bias. This is especially true in a two occasion study, where there is attrition after the first time point. To the extent that attrition at time 2 is predicted by data at time 1, quite different estimates could emerge due to FIML yielding asymptotically unbiased estimates when missingness is of the predicted-by-time-1 variety.

Personally I wouldn't bother running it both ways -- use all the data you have whenever possible.