Empirical Exercise 9

In this exercise, we’ll be using data from the paper Islamic Rule and the Empowerment of the Poor and Pious by Dr. Erik Meyersson. The paper uses a regression discontnuity design to estimate the impact of Islamic political parties on women’s equality, comparing municipalities where the Islamic party just won the mayorial election in 1994 to those where the Islamic party just lost. We’ll be replicating Dr. Meyersson’s key results using a replication data set produced by Professor Matias Cattaneo (who uses this paper as an example in one of his books on RD designs).

Getting Started

Before you get started, you need to open Stata and install the commands rd and rdrobust. To find rdcv, type findit rdcv. The pop-up window will only contain one link to a package. Click it. Then click the blue click here to install link. Then, close the window. To find the rd command, type findit Austin Nichols in Stata. When the pup-up window opens, scroll throught the commands (which are listed alphabetically) until you get to the command rd. Click the link to rd from http://fmwww.bc.edu/RePEc/bocode/r, and then click the blue click here to install link. Finally, make sure that you have Dr. Meyersson’s article handy, as you will want to look at the tables and figures as you work through this exercise.

Now create a do file. After the usual preliminaries, include the following code to download the replication data for Dr. Meyersson’s paper:

webuse set https://pjakiela.github.io/ECON379/exercises/E9-RD/
webuse E9-Meyersson-data

The data set contains data on 2,629 Turkish municipalities. Use the summarize and describe commands to familiarize yourself with the data set.

The running variable is margin1994, the Islamic party’s win margin in the 1994 elections (which is negative when the Islamic party didn’t win). What is the mean of this variable? Generate a new variable mayor1994 that is a dummy equal to one in municipalities where the Islamic party won the municipal election in 1994. This is the treatment variable. What is the mean of mayor1994. How many of the 2,629 elections did the Islamic party win?

Empirical Exercise

Question 1

The loop below generates part of Table I:

foreach var of varlist hs_women hs_men mayor1994 {
	    ttest `var', by(mayor)
	}

The foreach var of varlist command tells Stata that you want it to execute the command (or series of commands) in curly braces for each of the listed values of var.

Extend the code so that Stata loops through all of the variables summarized in Table I. What is the t-statistic from the test of the hypothesis that municipalities where the Islamic party won the election in 1994 had the same average household size as municipalities where the Islamic party did not win the 1994 election?

Question 2

Use the histogram command to replicate Figure 2a, the histogram of the Islamic win margin in 1994. You will want to play around with the the options, particularly the bin width, to get as close as possible to what is shown in Figure 2a.

Question 3

What bandwidth does the author use in his RD estimation of the impact of Islamic mayors on women’s educational attainment? (You can find this information in the text of the paper.)

Question 4

The code below replicates the regression results presented in Column 1 of Panel A of Table II (confirm that this is the case):

reg hs_women mayor, cluster(province)

Now add the controls described in the text to get as close as possible to the regression results presented in Column 2 of Panel A of Table II. What is the t-statistic on the mayor1994 variable in your expanded regression?

Question 5

Now implement the RD specification from Column 3 (of Panel A of Table II). Remember that this should include separate linear terms for the running variable above and below the discontinuity. Do you get exactly the same coefficient and standard error as Dr. Meyersson? When you set the bandwidth = 0.24, how many observations are included in the RD estimation?

Question 6

Now implement the RD estimation including controls, as Dr. Meyersson does in Column 4 of Table II. What is the estimated coefficient on the the mayor1994 (ie treatment) variable? It may not be exactly the same as the coefficient reported in the table, but it should be close.

Question 7

In Columns 5 and 6, Dr. Meyersson presents RD estimation using two different alternative bandwidths. Replicate your RD analysis with controls using a very narrow bandwidth: one quarter of the optimal bandwidth used in Columns 3 and 4. Do the results change? What is the estimated impact of having a mayor from the Islamic Party?

Question 8

In Column 7, Dr. Meyersson uses a quadratic function of the running variable in addition to the linear function. In other words, he generates a term that is equal to the Islamic win margin squared for municipalities where the Islamic party won the 1994 election (and equal to zero otherwise), and he generates another term that is equal to the Islamic win margin squared for municipalities where the Islamic party lost the election (and zero otherwise). Generate these terms, and estimate an RD specification that includes them as well as the linear terms, but no controls (so, your specification is different than that reported in Column 7 of Table II). Use the same optimal bandwidth that Dr. Meyersson uses in Columns 3 and 4. What is the estimated impact of having an Islamic party mayor on female educational attainment?

Question 9

To demonstrate the validity of an RD design, we want to show that covariates do not “jump” at the discontinuity. Do this by replicating the simple specification from Question 5, but using the outcome variable voteshare1994 (the Islamic vote share in 1994). If having an Islamic party win the election predicts a covariate, it tells us that there is something wrong with ur RD design. What is the p-value associated with this hypothesis test?

Question 10

Now implement the RD using the rd command, as described in lecture. The outcome variable is (once again) hs_women. Do not include any controls. The RD command reports results for three bandwidths: the Imbens-Kalyanaraman optimal bandwidth (lwald in the Stata output), half of that (lwald50), and two times the optimal bandwidth (lwald200). What is the optimal bandwidth (the first one reported), according to the rd command? How does it compare to the one used by Dr. Meyersson?

Question 11

When you use the rd command, what is the estimated treatment effect of having a mayor from the Islamic party on women’s education (using the optimal bandwidth)?

Question 12

Now use the rdcv command (as described in the lecture notes). What is the rdcv command’s estimate of the optimal bandwidth?