Missing data and -xtlogit-

[Laura Myles]

Hi Listers,

I have data from 400 participants who were interviewed every 2 months for a whole year (6 'waves'). At each wave, they were asked "since our last call, have you started exercising" (coded as yes/no) and to rate their wellbeing (continuous measure). I am interested in exploring the association between their wellbeing score and whether they have started exercising.

I have some missing data mostly due to participants not attending all 'waves'. I opted for analysing the data using a mixed model approach with -xtlogit-. Based on previous posts, this approach is tolerant of missing data but I am wondering if I am still at risk of biased estimates and how to address this. In particular, I was wondering if inverse probability weighting can be used with -xtlogit- if participants' characteristics were to explain missing data or if alternative approaches may be recommended.

Thanks in advance!

[Carlo Lazzaro]

Laura:
as usual, the first step to take is diagnosing the mechanism(s) underlying the missingness in your data: is missingness informative or not?

[Laura Myles]

Thanks Carlo Lazzaro -I cannot completely exclude that those who missed the waves/had missing data were not exercising' however, looking at my data I found that several measures taken at baseline (e.g. age, education) are associated with the variable coding whether data are missing during the follow-up waves. I take this to indicate that data may be missing at random.

Is IPW a sensible option and how can I do so within -xtlogit-?

[Carlo Lazzaro]

Laura:
if your missing values are MAR, I would go -mi-.

[Laura Myles]

Thanks Carlo - with repeated measures, can I use MI on the wide data and then transform to long for the analysis?

[Carlo Lazzaro]

Laura:
see -mi set- entry in Stata .pdf manual.