Week_07_1_Difference_in_Difference.Rmd

---
title: "Difference in Differences"
subtitle: "The Most Commonly Applied Research Design in Modern Econometrics"
date: "Updated `r Sys.Date()`"
output:
  xaringan::moon_reader:
    self_contained: TRUE
    css: [default, metropolis, metropolis-fonts]
    lib_dir: libs
    # Run xaringan::summon_remark() for this
    #chakra: libs/remark-latest.min.js
    nature:
      highlightStyle: github
      highlightLines: true
      countIncrementalSlides: false
---


```{r setup, include=FALSE}
options(htmltools.dir.version = FALSE) 
knitr::opts_chunk$set(echo = FALSE, warning = FALSE, message = FALSE, fig.width = 8, fig.height = 6)
library(tidyverse)
library(gganimate)
library(estimatr)
library(magick)
library(dagitty)
library(ggthemes)
library(directlabels)
library(ggdag)
library(fixest)
library(jtools)
library(scales)
library(Cairo)
theme_metro <- function(x) {
  theme_classic() + 
  theme(panel.background = element_rect(color = '#FAFAFA',fill='#FAFAFA'),
        plot.background = element_rect(color = '#FAFAFA',fill='#FAFAFA'),
        text = element_text(size = 16),
        axis.title.x = element_text(hjust = 1),
        axis.title.y = element_text(hjust = 1, angle = 0))
}
theme_void_metro <- function(x) {
  theme_void() + 
  theme(panel.background = element_rect(color = '#FAFAFA',fill='#FAFAFA'),
        plot.background = element_rect(color = '#FAFAFA',fill='#FAFAFA'),
        text = element_text(size = 16))
}
theme_metro_regtitle <- function(x) {
  theme_classic() + 
  theme(panel.background = element_rect(color = '#FAFAFA',fill='#FAFAFA'),
        plot.background = element_rect(color = '#FAFAFA',fill='#FAFAFA'),
        text = element_text(size = 16))
}
```

# Check-in

- We're thinking through ways that we can identify the effect of interest without having to control for everything
- One way is by focusing on *within variation* - if all the endogeneity can be controlled for or only varies between-individuals, we can just focus on within variation to identify it
- Pro: control for a bunch of stuff
- Con: washes out a lot of variation! Result can be noisier if there's not much within-variation to work with
- Also, this requires no endogenous variation over time
- That might be a tricky assumption! Often there are plenty of back doors that shift over time

---

# Difference-in-Differences

- Today we will talk about difference-in-differences (DID), which is a way of using within variation in a more deliberate way in order to identify the effect we want
- All we need is a treatment that *goes into effect* at a particular time, and we need a group that is *treated* and a group that is *not*
- Then, we compare the within-variation for the treated group vs. the within-variation for the untreated group 
- Voila, we have an effect!

---

# Difference-in-Differences

- Because the requirements to use it are so low, DID is used a *lot*
- Any time a policy is enacted but isn't enacted everywhere at once? DID!
- Plus, the logic is pretty straightforward
- Notice it doesn't even get a full chapter in the textbook? That's not because it's unimportant - it's because it's easy!
- Here we'll cover the concept, next time some example studies

---

# Difference-in-Differences

- The question DID tries to answer is "what was the effect of (some policy) on the people who were affected by it?"
- We have some data on the people who were affected both before the policy went into effect and after
- However, we can't just compare before and after, because we have a back door path - Time! Things change over Time for reasons unrelated to Treatment

```{r, dev = 'CairoPNG', fig.height=4}
dag <- dagify(Treatment ~ Time,
              Outcome ~ Treatment + Time,
              coords=list(
                x=c(Treatment = 1, Time = 2, Outcome = 3),
                y=c(Treatment = 1, Time = 2, Outcome = 1)
              )) %>% tidy_dagitty()
ggdag_classic(dag,node_size=10) + 
  theme_void_metro() + 
  expand_limits(x=c(.5,3.5))

```

---

# Difference-in-Differences

- Why not just control for Time? We can certainly measure that and we've controlled for it before!
- But we can't! It would wash out all the variation!
- You're either Before and Untreated, or After and Treated. So if you control for Time, you're comparing people with the same values of Time - who must also have the same values of Treatment! So you can't compare Treated and Untreated to get the effect

---

# Difference-in-Differences

- The solution is to bring in a *control group* who is Untreated in both Before and After periods
- Now the diagram looks like this
- There's more to control for (we have to control for Group differences too - isolate within variation), but this ALLOWS us to control for Time

```{r, dev = 'CairoPNG'}
dag <- dagify(Treatment ~ Time + Group,
              Outcome ~ Treatment + Time + Group,
              coords=list(
                x=c(Treatment = 1, Time = 3, Outcome = 3, Group = 1),
                y=c(Treatment = 2, Time = 2, Outcome = 1, Group = 1)
              )) %>% tidy_dagitty()
ggdag_classic(dag,node_size=10) + 
  theme_void_metro() + 
  expand_limits(x=c(.5,3.5))

```


---

# Difference-in-Difference

- We control for Group by isolating within variation (comparing After to Before)
- Then we control for Time by comparing those two sources of Within variation
- We ask *how much more increase* was there in Treatment than control?
- The Control change is probably the increase you could have expected regardless of Treatment
- So anything on *top of that* is the effect of Treatment
- Now we've controlled for Group and Time, identifying the effect!

---

# Example

- Let's say we have a pill that's supposed to make you taller
- Give it to a kid Adelaide who is 48 inches tall
- Next year they're 54 inches tall - a six inch increase! But they probably would have grown some anyway without the pill. Surely the pill doesn't make you six inches taller.
- SO we compare them to their twin Bella, who started at 47 inches but we DON'T give a pill to
- Next year that twin is 51 inches tall - a four inch increase. So Adelaide probably would have grown about 4 inches without the pill. So the pill boosted her by $(54 - 48) - (51 - 47) = 6 - 4 = 2$ additional inches
- "Adelaide, who was Treated, grew by two *more inches* than Bella, who was Untreated, did over the same period of time, so the pill made Adelaide grow by two inches"
- That's DID!

---

# Example

```{r}
df <- data.frame(Person = c('Adelaide','Adelaide','Bella','Bella'),
                 Time = factor(c('Before','After','Before','After'), levels = c('Before','After')),
                 Height = c(48, 54, 47, 51))
ggplot(df, aes(x = Time, y = Height, group = Person, label = Person)) +
  geom_line(color = 'black', size = 1) + 
  geom_point(color = 'red', size = 3) + 
  geom_dl(method = 'last.bumpup', position = position_nudge(x = .05)) + 
  theme_metro() + 
  annotate(geom = 'text', label = 'Before Difference', x = .65, y = 47.5, color = 'red') +
  annotate(geom = 'text', label = 'After Difference', x = 2.35, y = 52.5, color = 'blue') + 
  geom_segment(aes(x = 1, y = 47, xend = 1, yend = 48), color = 'red', size = 1.5) + 
  geom_segment(aes(x = 2, y = 51, xend = 2, yend = 54), color = 'blue', size = 1.5)

```

---

# Example

```{r}
df <- data.frame(Person = c('Adelaide','Adelaide','Bella','Bella'),
                 Time = factor(c('Before','After','Before','After'), levels = c('Before','After')),
                 Height = c(48, 54, 47, 51))
ggplot(df, aes(x = Time, y = Height, group = Person, label = Person)) +
  geom_line(color = 'black', size = 1, alpha = .6) + 
  geom_point(color = 'red', size = 3) + 
  geom_dl(method = 'last.bumpup', position = position_nudge(x = .05)) + 
  theme_metro() + 
  annotate(geom = 'text', label = 'Before Difference', x = .65, y = 47.5, color = 'red') +
  annotate(geom = 'text', label = 'After Difference', x = 2.35, y = 52.5, color = 'blue') + 
  annotate(geom = 'text', label = 'Before Difference', x = 1.65, y = 51.5, color = 'red') +
  annotate(geom = 'text', label = 'Difference-in-Difference', x = 1.6, y = 53, color = 'purple') + 
  geom_segment(aes(x = 1, y = 47, xend = 1, yend = 48), color = 'red', size = 1.5) + 
  geom_segment(aes(x = 2, y = 51, xend = 2, yend = 54), color = 'blue', size = 1.5) + 
  geom_segment(aes(x = 1.9, y = 51, xend = 1.9, yend = 52), color = 'red', size = 1.5) + 
  geom_segment(aes(x = 1.9, y = 52, xend = 1.9, yend = 54), color = 'purple', size = 1.5)  

```

---

# Difference-in-Differences

What *changes* are included in each value?

- Untreated Before: Untreated Group Mean
- Untreated After: Untreated Group Mean + Time Effect
- Treated Before: Treated Group Mean
- Treated After: Treated Group Mean + Time Effect + Treatment Effect
- Untreated After - Before = Time Effect
- Treated After - Before = Time Effect + Treatment Effect
- DID = (Treated After - Before) - (Untreated After - Before) = Treatment EFfect

---

# Concept Checks

- Why do we need a control group? What does this let us do?
- What do we need to assume is true about our control group?
- In 2015, a new, higher minimum wage went into effect in Seattle, but this increase did not occur in some of the areas surrounding Seattle. How might you use DID to estimate the effect of this minimum wage change on employment levels?

---

# Difference-in-Difference

- Of course, this only uses four data points
- Usually these four points would be four means from lots of observations, not just two people in two time periods
- How can we do this and get things like standard errors, and perhaps include controls?
- Why, use OLS of course!

---

# Difference-in-Differences

- We can use what we know about binary variables and interaction terms to get our DID

$$Y_{it} = \beta_0 + \beta_1After_t + \beta_2Treated_i + \beta_3After_tTreated_i + \varepsilon_{it}$$
where $After_t$ is a binary variable for being in the post-treatment period, and $Treated_t$ is a binary variable for being in the treated group

```{r}
data.frame(Person = c('Adelaide','Adelaide','Bella','Bella'),
                 Time = factor(c('Before','After','Before','After'), levels = c('Before','After')),
                 Height = c(48, 54, 47, 51)) %>%
  mutate(After = Time == 'After', 
         Treated = Person == 'Adelaide')
```

---

# Difference-in-Differences

- How can we interpret this using what we know?

$$Y_{it} = \beta_0 + \beta_1After_t + \beta_2Treated_i + \beta_3After_tTreated_i + \varepsilon_{it}$$

- $\beta_0$ is the prediction when $Treated_i = 0$ and $After_t = 0$ $\rightarrow$ the Untreated Before mean!
- $\beta_1$ is the *difference between* Before and After for $Treated_i = 0$ $\rightarrow$ Untreated (After - Before)
- $\beta_2$ is the *difference between* Treated and Untreated for $After_t = 0$ $\rightarrow$ Before (Treated - Untreated)
- $\beta_3$ is *how much bigger the Before-After difference* is for $Treated_i = 1$ than for $Treated_i = 0$ $\rightarrow$ (Treated After - Before) - (Untreated After - Before) = DID!

---

# Graphically

```{r, dev='CairoPNG', echo=FALSE, fig.width=8,fig.height=7}
df <- data.frame(Control = c(rep("Control",150),rep("Treatment",150)),
                 Time=rep(c(rep("Before",75),rep("After",75)),2)) %>%
  mutate(Y = 2+2*(Control=="Treatment")+1*(Time=="After") + 1.5*(Control=="Treatment")*(Time=="After")+rnorm(300),state="1",
         xaxisTime = (Time == "Before") + 2*(Time == "After") + (runif(300)-.5)*.95) %>%
  group_by(Control,Time) %>%
  mutate(mean_Y=mean(Y)) %>%
  ungroup()

df$Time <- factor(df$Time,levels=c("Before","After"))

#Create segments
dfseg <- df %>%
  group_by(Control,Time) %>%
  summarize(mean_Y = mean(mean_Y)) %>%
  ungroup()

diff <- filter(dfseg,Time=='After',Control=='Control')$mean_Y[1] - filter(dfseg,Time=='Before',Control=='Control')$mean_Y[1]

dffull <- rbind(
  #Step 1: Raw data only
  df %>% mutate(state='1. Start with raw data.'),
  #Step 2: Add Y-lines
  df %>% mutate(state='2. Explain Y using Treatment and After.'),
  #Step 3: Collapse to means
  df %>% mutate(Y = mean_Y,state="3. Keep only what's explained by Treatment and After."),
  #Step 4: Display time effect
  df %>% mutate(Y = mean_Y,state="4. See how Control changed Before to After."),
  #Step 5: Shift to remove time effect
  df %>% mutate(Y = mean_Y 
                - (Time=='After')*diff,
                state="5. Remove the Before/After Control difference for both groups."),
  #Step 6: Raw demeaned data only
  df %>% mutate(Y = mean_Y 
                - (Time=='After')*diff,
                state='6. The remaining Before/After Treatment difference is DID.'))



p <- ggplot(dffull,aes(y=Y,x=xaxisTime,color=as.factor(Control)))+geom_point()+
  guides(color=guide_legend(title="Group"))+
  geom_vline(aes(xintercept=1.5),linetype='dashed')+
  scale_color_colorblind()+
  scale_x_continuous(
    breaks = c(1, 2),
    label = c("Before Treatment", "After Treatment")
  )+xlab("Time")+
  #The four lines for the four means
  geom_segment(aes(x=ifelse(state %in% c('2. Explain Y using Treatment and After.',"3. Keep only what's explained by Treatment and After."),
                            .5,NA),
                   xend=1.5,y=filter(dfseg,Time=='Before',Control=='Control')$mean_Y[1],
                   yend=filter(dfseg,Time=='Before',Control=='Control')$mean_Y[1]),size=1,color='black')+
  geom_segment(aes(x=ifelse(state %in% c('2. Explain Y using Treatment and After.',"3. Keep only what's explained by Treatment and After."),
                            .5,NA),
                   xend=1.5,y=filter(dfseg,Time=='Before',Control=='Treatment')$mean_Y[1],
                   yend=filter(dfseg,Time=='Before',Control=='Treatment')$mean_Y[1]),size=1,color="#E69F00")+
  geom_segment(aes(x=ifelse(state %in% c('2. Explain Y using Treatment and After.',"3. Keep only what's explained by Treatment and After."),
                            1.5,NA),
                   xend=2.5,y=filter(dfseg,Time=='After',Control=='Control')$mean_Y[1],
                   yend=filter(dfseg,Time=='After',Control=='Control')$mean_Y[1]),size=1,color='black')+
  geom_segment(aes(x=ifelse(state %in% c('2. Explain Y using Treatment and After.',"3. Keep only what's explained by Treatment and After."),
                            1.5,NA),
                   xend=2.5,y=filter(dfseg,Time=='After',Control=='Treatment')$mean_Y[1],
                   yend=filter(dfseg,Time=='After',Control=='Treatment')$mean_Y[1]),size=1,color="#E69F00")+
  #Line indicating treatment effect
  geom_segment(aes(x=1.5,xend=1.5,
                   y=ifelse(state=='6. The remaining Before/After Treatment difference is DID.',
                            filter(dfseg,Time=='After',Control=='Treatment')$mean_Y[1]-diff,NA),
                   yend=filter(dfseg,Time=='Before',Control=='Treatment')$mean_Y[1]),size=1.5,color='blue')+
  #Line indicating pre/post control difference
  geom_segment(aes(x=1.5,xend=1.5,
                   y=ifelse(state=="4. See how Control changed Before to After.",
                            filter(dfseg,Time=='After',Control=='Control')$mean_Y[1],
                            ifelse(state=="5. Remove the Before/After Control difference for both groups.",
                                   filter(dfseg,Time=='Before',Control=='Control')$mean_Y[1],NA)),
                   yend=filter(dfseg,Time=='Before',Control=='Control')$mean_Y[1]),size=1.5,color='blue')+
  labs(title = 'The Difference-in-Difference Effect of Treatment \n{next_state}')+
  transition_states(state,transition_length=c(6,16,6,16,6,6),state_length=c(50,22,12,22,12,50),wrap=FALSE)+
  ease_aes('sine-in-out')+
  exit_fade()+enter_fade() + 
  theme_metro_regtitle()

animate(p,nframes=150)
```

---

# Design vs. Regression

- Like with fixed effects, there is a distinction between *regression model* and *research design*
- We have a model with an interaction term
- Not all models with interaction terms are DID!
- It's DID because it's an interaction between treated/control and before/after
- If you don't have a before/after, or you don't have a control group, that same setup may tell you something interesting but it won't be DID!

---

# Example

- The Earned Income Tax Credit was increased in 1993. This may increase chances single mothers (treated) return to work, but likely not affect single non-moms (control)
- Does this program help moms get back to work?

```{r, echo = TRUE}
df <- read.csv('http://nickchk.com/eitc.csv') %>%
  mutate(after = year >= 1994,
         treated = children > 0)
df %>% 
  group_by(after, treated) %>%
  summarize(proportion_working = mean(work))
```

---

# Example

- We can do it by just comparing the points, like we did with Adelaide and Bella
- This will give us the DID estimate: The EITC increase increases the probability of working by 4.7 percentage points
- But not standard errors, or the ability to include controls easily

```{r, echo = TRUE}
means <- df %>% 
  group_by(after, treated) %>%
  summarize(proportion_working = mean(work)) %>%
  pull(proportion_working)
(means[4] - means[2]) - (means[3] - means[1])
```

---

# Let's try OLS!
```{r, echo = TRUE}
did_reg <- feols(work ~ after*treated, data = df)
etable(did_reg, digits = 3)
```

---

# Concept Checks

- There are four coefficients on the previous slide. Interpret them carefully in a sentence each.
- Why can the $After \times TreatedGroup$ interaction stand in for $Treated$ ?

---

# More!!

- You can recognize $After_t$ and $Treated_i$ as *fixed effects*
- $Treated_i$ is a fixed effect for group - we only need one coefficient for it since there are only two groups
- And $After_t$ is a fixed effect for *time* - one coefficient for two time periods
- You can have more than one set of fixed effects like this! Our interpretation is now within-group *and* within-time
- (i.e. comparing the within-group variation across groups)

---

# More!!

- We can extend this to having more than two groups, some of which get treated and some of which don't
- And more than two time periods! Multiple before and/or multiple after
- We don't have a full set of interaction terms, we still only need the one, which we can now call $CurrentlyTreated_{it}$
- **If you have more than two groups and/or more than two time periods, then this is what you should be doing**

---

# More!!

- Let's make some quick example data to show this off, with the first treated period being period 7 and the treated groups being 1 and 9, and a true effect of 3

```{r}
set.seed(10600)
```

```{r, echo = TRUE}
did_data <- tibble(group = sort(rep(1:10, 10)),
                   time = rep(1:10, 10)) %>%
  mutate(CurrentlyTreated  = group %in% c(1,9) & time >= 7) %>%
  mutate(Outcome = group + time + 3*CurrentlyTreated + rnorm(100))
did_data
```

---

# More!!

```{r, echo = TRUE}
# Put group first so the clustering is on group
many_periods_did <- feols(Outcome ~ CurrentlyTreated | group + time, data = did_data)
etable(many_periods_did)
```


---

# Downers and Assumptions

- So... does this all work?
- That example got pretty close to the truth of 3 but who knows in other cases!
- What needs to be true for this to work?

---

# Does it Work?

- This gives us a causal effect as long as *the only reason the gap changed* was the treatment
- If Adelaide would have grown six inches *anyway*, then the gap would have grown by 2, pill or not. The pill did nothing in that case! But we don't know that and mistakenly say it helped her grow by 2
- In fixed effects, we need to assume that there's no uncontrolled endogenous variation across time
- In DID, we need to assume that there's no uncontrolled endogenous variation *across this particular before/after time change*
- An easier assumption to justify but still an assumption!

---

# Parallel Trends

- This assumption - that nothing else changes at the same time, is the poorly-named "parallel trends"
- Again, this assumes that, *if the Treatment hadn't happened to anyone*, the gap between the two would have stayed the same
- Sometimes people check whether this assumption is plausible by seeing if *prior trends* are the same for Treated and Untreated - if we have multiple pre-treatment periods, was the gap changing a lot during that period?
- Sometimes people also "adjust for prior trends" to fix parallel trends violations, or use related methods like Synthetic Control (these are outside the scope of this class)

---

# Prior Trends

- Let's see how that EITC example looks in the leadup to 1994
- They look like the gap between them is pretty constant before 1994! They move up and down but the *gap* stays the same. That's good.

```{r, echo = FALSE, dev = 'CairoPNG'}
df %>%
  mutate(tlabel = factor(treated, labels = c('Untreated','Treated'))) %>%
  group_by(year, treated, tlabel) %>%
  summarize(proportion_working = mean(work)) %>%
  ggplot(aes(x = year, y = proportion_working, color = treated)) + 
  geom_line(size = 1.5) + 
  geom_vline(aes(xintercept = 1993.5), linetype = 'dashed') +
  geom_dl(aes(label = tlabel), method = list('last.bumpup', cex = 1.2)) +
  theme_metro_regtitle() + 
  scale_x_continuous(limits = c(1991,1997)) + 
  scale_y_continuous(labels = function(x) scales::percent(x, accuracy = 1)) +
  guides(color = FALSE) + 
  labs(x = "Year", y = 'Proportion Working')
```


---

# Prior Trends

- Formally, prior trends being the same tells us nothing about parallel trends
- But it can be suggestive. Going back to the height pill example, what if instead of comparing Adelaide and Bella, child twins, we compared Adelaide to *me*? 
- Seeing the gap closing *anyway* in previous years would be a pretty good clue that it's not just the pill

```{r}
df <- data.frame(Person = c('Adelaide','Adelaide','Adelaide','Me','Me','Me'),
                 Time = factor(c('Before 1','Before 2','After','Before 1','Before 2','After'), levels = c('Before 1','Before 2','After')),
                 Height = c(43, 48, 54, 67, 67, 67))
ggplot(df, aes(x = Time, y = Height, group = Person, label = Person)) +
  geom_line(color = 'black', size = 1) + 
  geom_point(color = 'red', size = 3) + 
  geom_dl(method = 'last.bumpup', position = position_nudge(x = .05)) + 
  theme_metro() +
  annotate(geom = 'text', x = 1.5, y = 60, hjust = .5, label = 'The gap was already\nclosing before\nany Treatment...') +
  annotate(geom = 'text', x = 2.5, y = 60, hjust = .5, label = 'telling us that\nnot all of this change\n is just Treatment')

```

---

# Parallel Trends

- Just because *prior trends* are equal doesn't mean that *parallel trends* holds. 
- *Parallel trends* is about what the before-after change *would have been* - we can't see that!
- For example, let's say we want to see the effect of online teaching on student test scores, using COVID school shutdowns to get a Before/After
- As of March/April 2020, some schools had gone online (Treated) and others hadn't (Untreated)
- Test score trends were probably pretty similar in the Before periods (Jan/Feb 2020), so prior trends are likely the same
- But LOTS of stuff changed between Jan/Feb and Mar/Apr, like, uh, Coronavirus, lockdowns, etc. not just online teaching! SO *parallel trends* likely wouldn't hold

---

# Concept Checks

- Go back to the Seattle minimum wage effect example from the first Concept Check slide. Clearly state what the parallel trends assumption means in this context.
- It's possible (although perhaps unlikely) that parallel trends can hold even if it looks like the treatment and control groups were trending apart before treatment went into effect. How is this possible?

---

# And a Warning!

- DID is so nice and simple that it feels like you can get real flexible with it
- But the stuff we're covering in this class - up to TWFE, *relies very strongly* on the assumptions we made. If you break them, the *research design* may hold up, *but the estimator really doesn't* and you need a different estimator

---

# And a Warning! 

- TWFE *does not work if different groups get treated at different times* (try: Callaway & Sant'anna (R **did**) or Wooldridge (R **etwfe**))
- If you need to control for stuff to support parallel trends, *just tossing controls in a TWFE doesn't work* (try **did** again, or Sant'anna and Zhao **DRDID**)
- If the outcome is binary, you can't just run TWFE but with logit/probit (just do LPM)
- If the effect gets stronger or weaker over time, the TWFE effect will be weird (do "dynamic DID", see the textbook)
- For now, stick to DID designs where all the Treated groups are treated in the same period

---

# Swirl

- Enough downers! Let's swirl
- Do the Difference-in-differences swirl