Given the challenges that you note, I think you may be best off doing some exploration by
hand of the potential comparison schools with the variables that you mentioned to see
whether any candidate matches seem remotely plausible or at least close enough. If the 3
schools are the most extreme on those variables, especially if they are outliers on those
variables, perhaps there are no good matches in the same state.
Perhaps there are states with policies that make them appropriate to use outlier schools
from those states as controls: e.g., synthetic controls for these 3 schools as
composition of several schools in other states.
Janet Rosenbaum, Ph.D.
Assistant Professor of Epidemiology
School of Public Health, SUNY Downstate Medical Center, Brooklyn, NY
janet.rosenbaum(a)downstate.edu
On Apr 16, 2018, at 3:39 PM,
matchit-request(a)lists.gking.harvard.edu wrote:
Matching experts,
I am trying to match a small treatment group of schools (n = 3) with
controls using a ratio of 1:4. I am using Mahal distance.
The co variates are at the school level, but the analyses are planned at
the student level.
The match faces huge challenges to start with.
The three treatment school are among the most extreme in the state with
regard to the covariates:
the highest dropout highest minority, highest poverty and lowest academic
performance.
I am getting warnings as follows:
Warning messages:
1: glm.fit: algorithm did not converge
2: glm.fit: fitted probabilities numerically 0 or 1 occurred .
I wonder if there are ways to work around the lack of convergence and
fitted probabilities at the extreme.
In previous work with this group I tried propensity scoring but that was
equally problematic in term sof std differences.
Thanks for any help.
Bill