My multilevel model’s 2 levels consist of land surface images and meteorological readings from different places. The readings are level 2, with 7 variables and 1 observation per group. The images are level 1, with 5 variables (different images) where each pixel is an observation. The images are of size 1000 x 1000, which means that per group, there are 1 million observations.
The predictand (left) is taken by a thermal camera, while the predictor images are all taken by a more accurate normal camera, and some of them show these features (like trees) vividly.
The results of the model (on the right) after cross validation are wildly inaccurate numerically, but perfectly capture every spatial detail in the image, clearly having a massive bias toward the first level.
My question is 2-fold:
- How do I show the model’s bias towards one level?
- How many observations is, in fact, too many? Should I cut the images down in size?