Multilevel Modeling Trouble

[Ari Fodeman]

Hey y'all, I'm trying to predict student-level outcomes by school-level policies, but run into a snag on the appropriate way to structure the model, b/c two of the school-level policies vary by grade. I'm using wideformat, where each row is a student observation by their school's observation. I'm wondering:

1) Should I keep each school-level policy as individual variables (e.g., Q25_4th, Q25_5th... Q25_12th), or composite them into a single variable with the relevant response for that student's school's grade? (e.g., Q25)

2) Should I use Grade as a control variable, predicting these grade-varying variables, and even the student-level outcome directly? Should I use Grade as a cross-classified cluster?* Should I use Grade as a grouping variable and run the model as multigroup?

We have ~15,187 students across ~199 schools and 9 grade levels (4th - 12th)

For this analysis, we have:

-One student-level outcome variable, physical activity

-Several school-level variables about various phys ed school policies

^Two of these school-level variables also vary by grade level, e.g.:

"How many weeks were students required to attend phys ed in the school year?" for Elementary (4th + 5th), 6^th - 12^th grade

*I tried a cross-classified model, but got back an error message (this was in MPlus btw) that my two troublesome predictors varied by both School and Grade, which I thought was the point of cross-classifying them, but perhaps its only the DVs that are allowed to be in both classes' models, not the IVs? Trouble is These two school policies vary not only by Grade, but of course also by School.