Data set up Longitudinal Multiple Membership Model

[GerineLodder]

I have a dataset with the levels Time, Child, Classroom and School. My question is about how to set up the data for analyses in MLWIN for longitudinal multiple membership models.

Not all children are in my data at all timepoints (varying between all to one, in all possible combinations, such as: in the classroom at T1 and T3 but not at T2). In addition, children can change classrooms from time to time, thus, they mix (i.e., composition changes and mixes, for instance some children may be in the same classroom at T1 and T3, but not at T2).

In following the manual, I figured out that I can basically create a column "classroomT1", and put the classroom name of the classroom the child was first present there. For instance, if a child is absent at T1, in class 2001 at T2 and class 3004 at T3, entry would be:

classT1.......ClassT2........ClassT3
2001..........3004...........0

For this child, weights would be (.5/.5/0).

I now am confused about how I should enter data. I.e., I also have a column time. Suppose the child scores 5 at T2 and 7 at T3, what should my data look like?

Should it be:

ID......Classt1.....CLasst2......classT3.......Time......Score
1.......2001........3004.........0.............1.........5
1.......2001........3004.........0.............2.........7
1.......2001........3004.........0.............3.........MISSING

Or:

ID......Classt1.....CLasst2......classT3.......Time......Score
1.......2001........3004.........0.............1.........MISSING
1.......2001........3004.........0.............2.........5
1.......2001........3004.........0.............3.........7


Or something else entirely?

I would be very happy with your response!

[GerineLodder]

Just incase, a bit more information on the data:

Data were collected accross multiple waves, of which I will use 3 (named t1, t2, t3 here), which are all 1 year apart.
The data were collected in dutch primary schools. In the Dutch system, different grades are usually mereged into one classroom with one teacher.
For instance, oftentimes children in grade 3&4 are together in 1 classroom, and children in group 5&6 are together in another classroom.
If next year, a child moves from grade 4 to grade 5, about half of the children in their classrom change (i.e., the other grade 4 children remain, but whereas the classroom was shared with grade 3 children last year, now they are grouped together with the former grade 5 children (who are now in grade 6).

This is one complexity to the data.

Another complexity is that not every classroom is included in every wave.
That is, we collected data only in grades 3-6. This means that each year, new grade 3 children enter, whereas children who were in grade 6 at T1 are only included at T1.

A minor complexity is that some children skip a grade and some children have to redo a grade.

Finally, we have children who are simply absent during the time of data collection at each time point.

Because of these complications, I believe I need multiple membership models rather than cross-clasified models.

My question regarding setup of the data is described in the post above.

[GeorgeLeckie]

Hi Gerine,

There is a lot of information here.

However, as far as I can see your scores are the level-1 units in cross-classified data structure rather than a multiple membership structure. You have a separate score for each child in each wave. So each score is associated with one child, one class and one school. Put differently children nearly always change classes and schools between years, not within years.

Best wishes

George

[GerineLodder]

Hi George,

Thank you for your response.
You are right, every observation belongs to 1 child, and for that observation, the child belongs to 1 classroom and 1 school.
For the next observation (i.e., 1 year later), the observation belongs to the same child, but that child may then be in a classroom that is different from the classroom at T1 (i.e., different members).
They are still in the same school though. So school is constant for each child between timepoints, but classroom is not.

Does that make a difference?

[GeorgeLeckie]

Hi Gerine,

Its all about the level-1 observations. Each observation belongs to one child, one class and one school. The data are therefore cross-classified rather than multiple membership. So what you want to do is fit cross-classified models rather than multiple membership models.

I hope that helps

George

[GerineLodder]

Thank you for your clear answer

I have another question if I may:
In my dataset, we always measure grades 5-8 for certain schools, irrespective of what children are in those grades. The reason for that is that the data are collected as part of a randomized control trial for effectivity of an intervention program that is administered in those grades. In the dutch system, children leave school after grade 8.

As we follow these classrooms over multiple years, this means that there are many children joining or leaving.
For instance, in year 2, all children who were in grade 8 are no longer in the dataset, whereas all children who used to be in grade 4 now join.
Thus, for some children we may have data at T1 only, for some at T1, T2, for some at T2, T3, for some only T3 etc.
The missings are because the children could not have filled out the data even if they wanted to.

What would you advice in this situation? Can MLwin analyze all data, or only complete data?

[GeorgeLeckie]

Hi Gerine,

MlwiN like all standard software will listwise delete the data (drop rows where there are one or more missing values in the response or covariates) prior to fitting any specified model. So it depends on what you intend to do with your data. If you want to regress grade 8 score on grade 5, 6 and 7 scores then you can only do this for the subset of students who are observed in all four grades. If you want to regress scores on a time trend (i.e., growth curve analysis) you would reshape the data to long form and this way you would include all students even if they were only observed in one grade.

Best wishes

George

[GerineLodder]

Hi George,

I think this is why we initially chose for a multilevel model with multiple membership, because it seemed like we could include all data that way (with weighing based on how many timepoint children were present). i might be totally off though...

[GeorgeLeckie]

Hi Gerine,

If you haven't already done so, I recommend that you read the Multiple Membership module of our LEMMA online course for a good overview of what these models can and can't do.

Hope that helps

George

[GerineLodder]

Thanks, I have read those, but I find it hard to apply this knowledge to my model to be honest

best

Gerine