17-regression-discontinuity.Rmd

```{r 17_setup, include=FALSE}
knitr::opts_chunk$set(echo=TRUE, eval=FALSE, fig.align="center")
```

# (PART) Quasi-Experimental Designs {-}

# Regression Discontinuity

## Learning Goals {-}

- Implement and interpret the results of a regression discontinuity study

<br>

Slides from today are available [here](https://docs.google.com/presentation/d/1FMxqK7NdqGn8RKGYem-D60FAhzUE5FL3doc8a3qjfzw/edit?usp=sharing).



<br><br><br>



## Exercises {-}

**You can download a template RMarkdown file to start from [here](template_rmds/17-regression-discontinuity.Rmd).**

### Context and setup {-}

We'll look at data from a study by [Carpenter and Dobkin (2009)](http://masteringmetrics.com/wp-content/uploads/2015/01/Carpenter-and-Dobkin-2009.pdf) on the effect of alcohol consumption on mortality rates. 

The variables are as follows:

- `agecell`: Age of individual (the study focuses on adults between the ages of 19-22). This is the assignment (forcing, running) variable.
- `all`: Overall mortality rate
- `alcohol`: Mortality rate for alcohol-related causes
- `homicide`:Mortality rate for homicides
- `suicide`: Mortality rate for suicide
- `mva`: Mortality rate for car accidents
- `drugs`: Mortality rate for drug-related causes (alcohol excluded)
- `externalother`: Mortality rate for other external causes

```{r}
library(tidyverse)

alco <- read_csv("https://raw.githubusercontent.com/DS4PS/pe4ps-textbook/master/data/RegDisc2.csv")

# We create the treatment variable and a centered version of the assignment variable
alco <- alco %>%
    mutate(
        treatment = agecell > 21,
        age_centered = agecell - 21
    )
```

### Main analysis {-}

**Step 1:** Clarifying sharp vs. fuzzy regression discontinuity design

- Is this a sharp or a fuzzy RD design? Explain.





**Step 2:** Checking the assignment variable

Check the distribution of the assignment (forcing, running) variable (`agecell`). Are there any apparent bumps in the distribution? What would bumps in the distribution suggest?

```{r}
# Distribution of the assignment / forcing / running variable (age)
```





**Step 3:** Covariate balance

While we don't have any covariates available in this dataset, what do you think would be important confounders to include?





**Step 4:** Visually check for a treatment effect

Make plots of the assignment variable and the `mva`, `alcohol`, and `drugs` outcomes.

- Either the original or centered version is fine for these plots, but it's probably more useful to look at the original version.
- Include smoothing trend lines.
- Roughly eyeball the treatment effect.

```{r}
# Visually check for a treatment effect

```





**Step 5:** Build models to estimate the treatment effects

Based on your visualizations, determine whether a model with or without interaction between treatment and the assignment variable would be appropriate. Fit these models to estimate the treatment effects, and interpret the appropriate coefficients.

```{r}
# Modeling to estimate treatment effects

```





**Step 6:** Sensitivity analyses

What aspects of regression discontinuity analyses are uncertain (lend themselves to many analysis choices)? What sensitivity analyses might be useful to run?