Skip to main content

Causal Inference Course

Discussion 7: Matching Lab

Download the R Markdown. file for today’s lab. Submit your work on Canvas before you leave discussion! Solutions will be posted here later.

If you finish the exact matching exercise early, work through this R Markdown Notebook with further examples (download the .Rmd file here).

Data and R Libraries

The first exercise and problem set 4 use the lalonde dataset from the following paper:

Dehejia, R. H. and Wahba, S. 1999. Causal Effects in Nonexperimental Studies: Reevaluating the Evaluation of Training Programs. Journal of the American Statistical Association 94(448):1053–1062.

The paper compares methods for observational causal inference to recover an average causal effect that was already known from a randomized experiment. You do not need to read the paper; we will just use the study’s data as an illustration. We’ll load the data into R with the first code block.

knitr::opts_chunk$set(echo = TRUE)
library("tidyverse")
library("MatchIt")
data("lalonde") # this step loads the data

To learn about the data, type ?lalonde in your R console.

2. Example: Exact Matching with low-dimensional confounding

Our goal is to estimate the effect of job training treat on future earnings re78 (real earnings in 1978), among those who received job training (the average treatment effect on the treated, ATT).

2.1. Using matchit() to conduct a matching

For this part, we assume that three variables comprise a sufficient adjustment set: race, married, and nodegree. We use matchit with:

  • a formula treat ~ race + married + nodegree
  • method = "exact" to conduct exact matching, which matches two units only if they are identical along race, married, and nodegree
  • data = lalonde since we are using the lalonde data
  • estimand = "ATT" since we are targeting the average treatment effect on the treated (ATT)

We then use the summary() function to see how many control units and how many treatment units were matched.

exact_low <- matchit(treat ~ race + married + nodegree,
                 data = lalonde,
                 method = "exact",
                 estimand = "ATT")
# Note: There are multiple correct ways to extract the numbers below
summary(exact_low)$nn

Question: How many control units were matched? How many treated units?

Answer:

2.2. Effect estimate

Here, we estimate a linear regression model using the match data from 2.1 using the lm() function with the formula re78 ~ treat + race + married + nodegree. We pass weights that come from the matching. Notice that for this piece, we have passed the matched data match.data(exact_low). The coefficient in front of the variable treat in the linear regression is our estimated effect.

fit <- lm(re78 ~ treat + race + married + nodegree,
          data = match.data(exact_low),
          w = weights)
print(round(coef(fit)["treat"],2))

Question: What is the estimated effect of job training on earnings?

Answer:

2.3. Assessing the Match: Balance of Covariates

In matching, one thing we care about is balance across covariates. In other words, we want to see that the distributions of different covariates are about the same between the treatment and the control groups. We can check how well the balancing has been done with the summary() function.

  • interactions: check interaction terms too? (T or F)
  • un: show statistics for unmatched data as well? (T or F)
summary(exact_low, interactions = F, un = F)$sum.matched

Question: What do you notice about the means of different covariates for the treated versus control groups?

Answer:

In this case, we basically have perfect balance. This doesn’t always happen. Depending on the method and parameters you use, you could have “bad” matches where the covariates are unbalanced. If you conduct a matching and the covariate balance doesn’t look good, try another matching procedure!

3. Try it Yourself: Exact matching with high-dimensional confounding

You will use the results from this section in Problem Set 4.

3.1. Using matchit() to conduct a matching

Now suppose the adjustment set needs to also include 1974 earnings, re74. The adjustment set for this part is race, married, nodegree, and re74. Repeat exact matching as above.

# Your code goes here

Question: How many control units were matched? How many treated units?

Answer:

3.2. Assessing the Match: Examining matched units

Look at the re74 values in the full data and among the matched units.

Here is one way to do this:

  1. Use the select() function to get the re74 column in the full data. Pass this to the summary() function to look at descriptive statistics of the re74 values in the full data.
  2. Use the select() function to get the re74 column in the matched data. Pass this to the summary() function to look at descriptive statistics of the re74 values in the full data. You can get the matched data using the match.data function.
  • Examples of using the summary function are here.
  • Examples of using the select() function are here

Full data:

# your code goes here

Matched data:

# your code goes here

Explain what happened: What do you notice? What is different about the values of re74 in the full data versus the matched data? Explain what happened and why it happened. Briefly interpret the result from 3.2: what is the drawback of using exact matching in this setting?