textmatching (version 0.1.0)

sim: Simulated Matching Data

Description

A 270 document set along with a prefit topic model that is used to demonstrate the matching functionality.

Format

stm formatted object corresponding to simulated documents

treat

a binary treatment variable

confound

an unknown binary confounding variable

simy

a simulated outcome

Details

This is a set of documents and a prefit topic model used to demonstrate functionality of the package. It is loosely based off of the gender citation example in Roberts, Stewart and Nielsen (2020). The data is simulated such that the true treatment effect is 1 for all units. There is separable 'unobserved' confounding provided by the binary variable confound variables which are themselves based on real data. The outcome simy is purely synthetic.

Note that due to data size limitations on CRAN we only included a subset of the documents and a prefit topic model. Because there are so many fewer documents than were used to fit the original topic model, any model fit with this data would likely look substantially different. The original model was fit on 3201 documents using the treatment as the content covariate with 15 topics. All other settings were at their default.

Because we had to select subsets of the data, we emphasize that the example here isn't reflective of a real problem, its just a way of getting a handle on the objects in the code.

Examples

# NOT RUN {
#We start by assuming that you have run a topic model in stm using
#your treatment variable as a content covariate. This is step 1.

#We have done this already and the following command loads the 
#topic model as well as the documents, vocab and meta data objects.
#See the stm package for more details about these model objects.
data(sim)

#Step 2 is to recalculate the topics as though they were observed
#as treated.  We do this using the refit() function.
refitted <- refit(sim_topics, sim_documents, content_level="1")

#to this we needed to specify the value of the treatment (here "1").  
#If you have forgotten content_levels() will tell you the levels
#for a given stm model. 
content_levels(sim_topics)

#Step 3 is to calculate the projection onto the treatment variable
projection <- project(sim_topics, sim_documents, interactions = FALSE)
#NB: here we have turned off interactions purely for speed during
#CRAN checks.  Consider including them if you believe topic-specific
#word choice is relevant.  See description above.

#Finally Step 4 is to match using CEM or other matching method of your
#choice
matched <- cem_match(refitted,projection=projection, sim_meta$treat,
                     projection_breaks=2)
#note here we use a much weaker match on the projections because the data
#have already been trimmed a lot.

#Now the matched data can be analyzed using standard tools from cem
cem::att(matched, simy ~ treat, data=sim_meta)
#the estimator overestimates a bit but contains the truth in the CI

#We can compare this to the unadjusted difference in means (overestimates)
summary(lm(simy ~ treat, data=sim_meta))
#and the oracle estimator (based on unobserved covariates)
summary(lm(simy ~ treat + confound1 + confound2 + confound3,data=sim_meta))

#Please, be sure to diagnose your matches!!! The key advantage of matching
#is being able to examine matched pairs.  It is always important to read 
#the documents!

# }