causalinf

PEACEKEEPING WORKOUT (based on  King, Gary;Zeng, Langche, 2007, 
                      "Replication data for: When Can History be Our Guide? 
                      The Pitfalls of Counterfactual Inference", 
                      https://hdl.handle.net/1902.1/DXRXCFAWPK, 
                      Harvard Dataverse, V4, 
                      UNF:3:DaYlT6QSX9r0D50ye+tXpA== [fileUNF] )
# CONSIDER USING THE JUPYTER NOTEBOOK WITH R-SERVER KERNEL (NEVER R-SAGE KERNEL)
foo <- read.csv("https://course-resources.minerva.kgi.edu/uploaded_files/mke/00086677-3767/peace.csv")
                
# extract relevant columns
foo <- foo[, c(6:8, 11:16, 99, 50, 114, 49, 63, 136, 109, 126, 48, 160, 142, 10)]

# remove 2 rows with missing data (there are better ways to handle missing data)
foo <- foo[c(-19, -47), ]

# check that all missing data is gone...
which(is.na(foo) == TRUE)

# take a peek at the data set (identify the columns)
head(foo)

### Outcome is "pbs2s3": "democracy" and "peace" within 2 years after the end of the war
### codebook is here: http://www.nyu.edu/gsas/dept/politics/faculty/cohen/codebook.pdf

### Treatment indicator is "untype4": "multidimensional peacekeeping/peacebuilding"

### How many treated units? How many controls? How do you feel about SUTVA?



# STEP 1:
# Take the logistic regression (the ‘original model’) that appears in Gary King's 
# paper with Langche Zheng--it's available here: 

# https://dataverse.harvard.edu/dataset.xhtml?persistentId=hdl:1902.1/DXRXCFAWPK
# go to PAGE 2... download/open "replication.do" ... 

# Find the logistic regression model near the top, on the line starting with the word
# "logistic". Let this be your propensity score model, ignoring references to "cluster"
# (i.e., the model ends with 'decade'.)

# Call your propensity score model "glm1". 
# e.g, glm1 <- glm(untype4 ~ wartype + logcost + ... , data = foo, family = binomial)

# TAKE NOTE OF: 
# (a) the treatment effect (ATT), and the p-value of the treatment effect;
# (b) which conflicts in the treatment group get matched to which 'control' conflicts
#     (using $mdata OR $index.treated and $index.control)
# (c) the lowest p-value obtained from balance tests, after running MatchBalance on: 
#     wartype, logcost, wardur, factnum, factnum2, trnsfcap, untype4, treaty, 
#     develop, exp, decade...
#     SHORTCUT! COPY/PASTE THE BELOW (after matching, filling in the blank):
mb  <- MatchBalance(pbs2s3 ~  wartype + logcost + wardur + factnum + 
                              factnum2 + trnsfcap + untype4 + 
                              treaty + develop + exp + decade, 
                              data=foo, match.out = FILL_IN_BLANK, nboots=500)




# STEP 2:
# Perform genetic matching, with X = to all the variables included in MatchBalance
# pop.size should be at least 200, max.generations should be at least 20,
# and wait generations should be at least 10. Estimand is "ATT".

# Take note of the same 3 results as above. Check the rgenoud output
# to see if the genetic algorithm improves fitness over time.




# STEP 3:
# Augment your STEP 2 analysis by including your propensity score in "X"...
# e.g., X <- cbind(X, glm1$fitted)
# Take note of the same 3 results. See if your balance improves with inclusion
# of the propensity score as a matching variable.  See if your impact
# estimate changes. See if you can improve your results (e.g., balance)
# by changing the matching variables, or the matching parameters (e.g., "M")