I am very new to statistics and longitudinal analysis, so this question may sound very basic.
Consider a dataset where outcome(y) is a binary variable [wheezing , yes/no]. Each child is exposed to some intervention (A) which is a continuous variable. This intervention is very irregular, this is an unbalanced dataset. Some kids are exposed to intervention once, some kids twice and so on.
I have 1235 kids this dataset, only 20 have outcome measured more than once. The outcome for remaining 1235-20=1215 kids have outcome measured only once. I have a lot of observations for the intervention, almost 15,000 for these 1235 kids. My question is, should I include the repeated outcome measured for these 20 kids ? Although only 20 kids have binary outcome(y) measured more than once ? What are the implications of including repeated measure outcome when we have repeated measure for only small subset of subjects ? What are the consequences of dropping these repeated y from these 20 kids ?
Best Answer
With only about 0.02 of the kids having > 1 measurement it is not worth the effort to model this longitudinally. But to have any confidence in the result it is important to (1) know the reasons the intervention was given more than once in a child, (2) what determined the timing of the first intervention, and (3) what is the distribution in days of the time gap between enrollment (and collection of baseline variables) and intervention. If this distribution has any width it is advisable to include days until intervention/wheezing assessment as a covariate if the intervention is more or less pre-planned and is not a time-dependent covariate itself.
Were you to ever model data like these longitudinally I'd suggest a first-order Markov process using logistic regression as discussed here. There is nothing about the problem that suggests a mixed effects model with its oversimplified within-patient correlation assumption. I would remove
mixed-modelandlme4-nlmeas tags.