In jjchern/lalonde: The Lalonde Datasets

knitr::opts_chunk$set(
  warning = FALSE,
  message = FALSE,
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-",
  out.width = "100%"
)

About lalonde

The Lalonde datasets are widely used in the causal inference literature. The current package makes loading such datasets in R easier. I found myself calling the following command

haven::read_dta("http://www.nber.org/~rdehejia/data/nsw_dw.dta")

in several R projects. It might be easier to just type lalonde::nsw_dw.

Data and Source

• NSW Data Files (Lalonde Sample)

  • lalonde::nsw
  • These files contain the treated and control units from the male sub-sample from the National Supported Work Demonstration as used by Lalonde in his paper.

• NSW Data Files (Dehejia-Wahha Sample)

  • lalonde::nsw_dw
  • Based on pre-intervention variables, Dehejia-Wahha extract a further subset of Lalonde's NSW experimental data, a subset containing information on RE74 (earnings in 1974).

• Non-experimental Comparison Data Files:

  • lalonde::psid_controls
  • lalonde::psid_controls2
  • lalonde::psid_controls3
  • lalonde::cps_controls
  • lalonde::cps_controls2
  • lalonde::cps_controls3

• All the datasets are available in txt and dta format from Dehejia's website

Installation

# install.packages("devtools")
devtools::install_github("jjchern/lalonde")

Usage

The datasets print nicely in the tidyverse:

library(tidyverse)

lalonde::nsw

lalonde::nsw_dw

Combine the treatment group from lalonde::nsw_dw with a non-experimental comparison group from the Panel Study of Income Dynamics (PSID):

lalonde::nsw_dw %>% 
    filter(treat == 1) %>% 
    bind_rows(lalonde::psid_controls) %>% 
    select(-data_id) %>% 
    print() -> df

# install.packages("skimr")
skimr::skim(df)

The unadjusted difference in means is -$15,205:

df %>% 
    group_by(treat) %>% 
    summarise(mean_re78 = mean(re78)) %>% 
    print() %>% 
    spread(treat, mean_re78, sep = "_") %>% 
    mutate(diff = treat_1 - treat_0)

The naive estimate is certainly biased, because the treated group looks very different from the control group:

# install.packages("cem")
cem::imbalance(group = df$treat, 
               data  = as.data.frame(df),
               drop  = c("treat", "re78"))

The multivariate imbalanced meaure is close to 1, suggesting an almost complete separation between the treated and control group. The differences in the empirical quantiles of the two distributions also indicate a large amount of imbalance for many covariates. For example, the treated group tends to be younger, has fewer years of education, are less likely to be married, and earns a lot less in 1974 and 1975.

Matching on the covariates can help to create a matching sample in which the matched control group is more comparable to the treated group. Below we call the cem() function to implement an automatic coarsened exact matching (CEM):

cem::cem(treatment = "treat",
         data = as.data.frame(df), 
         verbose = TRUE, 
         keep.all = TRUE, 
         drop = "re78") -> cem
cem

The cem() function includes the automatic cut points:

cem$breaks

Alternatively, we can supply some infomation to aid the CEM process. For example, we can choose to discretize the variable age, educ, re74, re75 in the following way:

cut_age = seq(min(df$age), max(df$age), by = 15)
cut_educ = c(0, 6.5, 8.5, 12.5, 17)
cut_re74 = seq(0, max(df$re74), by = 5000)
cut_re75 = seq(0, max(df$re75), by = 5000)
cem::cem(treatment = "treat",
         data = as.data.frame(df), 
         verbose = TRUE, 
         keep.all = TRUE, 
         drop = "re78",
         cutpoints = list(age = cut_age,
                          educcation = cut_educ,
                          re74 = cut_re74,
                          re75 = cut_re75)) -> mat2
mat2

Is there a way to improve the number of subjects who can be matched?

cem::relax.cem(obj = mat2, data = as.data.frame(df), verbose = FALSE)

jjchern/lalonde documentation built on June 1, 2019, 3:56 a.m.