To Whom It May Concern,
I'm having a problem performing the
causal inference portion of CEM with my data. Right now I have data with
no missing values that categorizes citizen responses to political
questions on voting behavior (which party bloc voted for, measured on an
ordinal discrete scale), and political attitudes (two barometer
questions, also measured on an ordinal discrete scale), as well as a
number of background covariates (age, education, political interest,
knowledge, religion, gender, etc.). My treatment variable is a 1 or 0
depending on whether a respondent comes from an ethnically diverse
district or not .
In running my CEM model, everything compiled just fine, however when trying to estimate the SATT, I get the following error:
Error in att(mat1, as.factor(ONEPAR) ~ ELF + MUSLIM + CHRISTIAN + AGE + :
please, only use `lm' or `glm'
I
suppose I'm getting this because I chose a forest model for my
estimation, but I don't see why this should be a problem. Just for
reference here is my matching code and my att code.
AGECUT <- c(0,29.5, 39.5, 49.5, 59.5, 69.5, 79.5)
mat1
<- cem(treatment="TREAT1", data=data2, drop=c("RESPNO", "BACKCHK",
"URBRUR", "PROVINCE", "DISTRICT", "ELF", "DIVISION", "LOCATION",
"DISCUSS", "AUTH", "PATR", "ONEPAR","MULTPAR", "FREEVOTE", "FREEEXP",
"SINGLEAD", "ELECT", "REPRESENT", "SALIENCE", "TRUST1", "TRUST2",
"TRUST3", "REFORM1", "REFORM2", "Q88C_KEN", "VOTE", "RATIONALE",
"PATRON", "FRVOTE", "FREEXP", "ELEC", "MULTIPAR", "REFORM", "OTHERREL",
"TREAT", "TREATMED2"), cutpoints=list(AGE=AGECUT))
mat1
est1 <- att(mat1, ONEPAR~ELF+MUSLIM+CHRISTIAN+AGE+KNOW+INTEREST+EDUC+ETHMAJ+MALE, data=data2, model="forest")
est1
If
you have any idea of what I'm doing incorrectly I'd appreciate the
help. My model runs fine using automated CEM via MatchIt, so I'm not
sure what the problem is here. Finally, I'm getting some weird results
with MatchIt -- namely, almost all the covariates on which I matched are
turning up as significant. Any ideas as to what could be motivating
this troubling anomaly? Note it is my understanding (possibly incorrect)
that if treatment and control units are matched exactly, there should
be no difference between treatment and control on matched covariates,
and thus these measures should not significantly affect the outcome in
the analysis.
Sincerely,
Ashley Anderson
--
Ph.D. Candidate
Harvard University Government Dept
--
Ph.D. Candidate
Harvard University Government Dept