Closed ijyliu closed 3 years ago
I've done some interpolation, which helps a little. (I only want to do internal interpolation/between known observations, because other kinds are sketchier...)
Another thing we can do: figure out, mathematically, the maximum rectangular panel we can build (ie in terms of area). We could also kind of get at this using a graph.
Basically, sort this picture so that the longest black bars are on the left. This will create a curve/triangle. Then, try to draw the rectangle that maximizes the area within this triangle.
If we were to set missing values at the mean, could that prevent effects in the regression
Grouped lasso?
This is also a problem for the measurement error idea- I have to throw out a ton of countries because they are missing even one indicator!
Just a heads up, but we are going to have a real missing data problem with these panel datasets because you can't do PCA with NaN values.
http://pbil.univ-lyon1.fr/members/dray/files/articles/dray2015a.pdf
We can try to impute as much as we can to save observations.