Y: mixture of normals with: \( f(y) = p\phi(0, c1) + (1-p)\phi(0, c2) \)
\( logit(Pr(D = 1 | T,Y)) = \beta_0 + \beta_1T +\psi_1Y + \psi_2YT \)
k must be large (20~30) in order for variation in p to influence:
This may be why the coverage with lowest performance did not have a strong association with a specific value of p. I wonder if minimizing over p for poor coverage is biasing our estimation of coverage downward. action: see if ci coverage is associated with p when k is high
Risk curves are fun to look at, but it is hard to draw any conclusions from them. Points 1 and 2 are more important than 3 and 4.
Risk by F(y) given trt by q0 and q1. True value of Theta in grey.
Distribution of \( \Delta(y) = Pr(D=1|T=0, Y) - Pr(D-1|T=1, Y) \) by \( F(y) \).
pdf of y colored by p. The vertical black line indicates where \( \Delta(y)=0 \).
This is a plot of p (x-axis) vs \( F(y_{\Delta=0} + 0.005) - F(y_{\Delta=0} - 0.005) \) where \( \Delta(y_{\Delta=0}) = 0 \). This is a very arbitrary measure of proportion of observations near the boundary of interest.