Diff-on-Diff implementation in R

Introduction
This document walks through the Difference-in-Difference analysis of thr miniimum legal drinking age example
as it is presented in Angrist and Pischke’s Mastering Metrics book. The code here is R code which is based
on the excellent work by Jeffrey Arnold.
This document assumes that you have access to Chapter 5 of Angist and Pischke’s book. In here we will
replicate most of the results presented in that chapter relating to the investigation of whether changes in the
minimum legal driving age (MLDA) impact numbers of traffic fatalities in young people.
Look at this wikipedia page to get an overview of the (useful) patchwork of MLDA legislation in the US.

MLDA Difference-in-Difference
Load necessary libraries.
library(tidyverse)
library(stargazer)
library(readxl)
library(ggplot2)
library(clubSandwich)

Setup your working directory and load the data which are saved in deaths.Rdata.
setwd("YOUR_WORKING_DIRECTORY")

load("deaths.Rdata")

The data contain the number of deaths due to various reasons, reported by US State and Year.
deaths %>% filter(state == "5", year == "1980", agegr == "18-20 yrs") %>%
select(state,year,agegr, count,dtype, pop) %>% print()

## # A tibble: 6 x 6
## state year agegr count dtype pop
## <fct> <dbl> <fct> <dbl> <fct> <dbl>
## 1 5 1980 18-20 yrs 173 all 125804
## 2 5 1980 18-20 yrs 73 MVA 125804
## 3 5 1980 18-20 yrs 16 suicide 125804
## 4 5 1980 18-20 yrs 17 homicide 125804
## 5 5 1980 18-20 yrs 40 other external 125804
## 6 5 1980 18-20 yrs 27 internal 125804
So, for instance, in State 5 (and we don’t know which state this is) we can see that in 1980 altogether 173
18-20 year olds passed away, 73 of which due to a Motor Vehicle accident (MVA).
A few important variables are
• count, this is the count of Deaths for the indicated year, state and reason for death

1
 • mrate, this is the mortatlity rate (per 100,000) for the indicated year, state and reason for death
• year, gives the year for which the data are
• state, a numeric variable which indicates which US state the data are from
• dtype, stands for death type
• agegr, this is the age group
• pop, this is the population, in the respective age group.
• legal, described in detail below

Identifying States

In the dataset states are represented by numbers. This is a little annoying. So let’s add the state names. But
which number belongs to which state. It is very typical that countries and states have particular numbers
associated to them to assist in identifying states. For instance Utah should be represented by the number 49.
But how do you know that it is these number are called FIPS (Federal Information Processing Standard)
codes?
If you look at a [https://www.census.gov/geographies/reference-files/2017/demo/popest/2017-fips.html]
listing of these codes (and how the match to state names) you will find a couple of peclarities. In general the
numbers go from 1 (Alabama) to 56 (Wyoming), but a few numbers are missing, in particular 3, 7, 12, 43 and
52 are not allocated to states. So if we were to look at our state numbers and these numbers are also missing
then we can be pretty certain that we are looking at FIPS codes and then we can use the information in the
above link to match our numbers to state names.
unique(deaths$state)

## [1] 1 2 4 5 6 8 9 10 11 12 13 15 16 17 18 19 20 21 22 23 24 25 26 27 28
## [26] 29 30 31 32 33 34 35 36 37 38 39 40 41 42 44 45 46 47 48 49 50 51 53 54 55
## [51] 56
## 51 Levels: 1 2 4 5 6 8 9 10 11 12 13 15 16 17 18 19 20 21 22 23 24 25 26 ... 56
Success!!! We have numbers from 1 to 56 and none of the above listed numbers appears. We have FIPS codes.
In some sense this is not 100% proof and you could check whether population numbers which we have match
to reported population numbers in 1980, but we can trust that the authors have done the sensible thing and
used FIPS codes.
Now we need to import the spreadsheet which lists these codes and the associated state abbreviations (2
letter codes) and state names. As it is a little tedious to find an easily accessible file which includes all these
information I have prepared states.xlsx for your convenience.
state<-read_excel("states.xlsx")
state$FIPS <- as.factor(state$FIPS)
names(state)

## [1] "Cstate" "NAME" "FIPS"

This leaves us with a Table which contains state number (FIPS), state name (NAME) and a two letter code
(Cstate).
Now we use state in deaths and FIPS in state to merge the State appreviation (Cstate) and their names
(NAME) into our dataset.
deaths <- merge(deaths,state,by.x = "state", by.y = "FIPS",x.all = TRUE)

This has added the variables Cstate and NAME to our dataframe.

2
Data Summaries

Let’s see what the data look like

str(deaths)

## 'data.frame': 24786 obs. of 17 variables:

## $ state : Factor w/ 51 levels "1","2","4","5",..: 1 1 1 1 1 1 1 1 1 1 ...
## $ year : num 1970 1971 1972 1973 1974 ...
## $ legal1820 : num 0 0 0 0 0 ...
## $ dtype : Factor w/ 6 levels "all","MVA","suicide",..: 1 1 1 1 1 1 1 1 1 1 ...
## $ agegr : Factor w/ 3 levels "15-17 yrs","18-20 yrs",..: 1 1 1 1 1 1 1 1 1 1 ...
## $ count : num 224 241 270 258 224 207 231 219 234 176 ...
## $ pop : num 213574 220026 224877 227256 229025 ...
## $ age : num 16 16 16 16 16 ...
## $ legal : num 0 0 0 0 0 0 0 0 0 0 ...
## $ beertaxa : num 1.37 1.32 1.28 1.2 1.08 ...
## $ beerpercap : num 0.6 0.66 0.74 0.79 0.83 ...
## $ winepercap : num 0.09 0.09 0.09 0.1 0.16 ...
## $ spiritpercap: num 0.7 0.76 0.78 0.79 0.81 ...
## $ totpercap : num 1.38 1.52 1.61 1.69 1.8 ...
## $ mrate : num 104.9 109.5 120.1 113.5 97.8 ...
## $ Cstate : chr "AL" "AL" "AL" "AL" ...
## $ NAME : chr "Alabama" "Alabama" "Alabama" "Alabama" ...
summary(deaths)

## state year legal1820 dtype

## 1 : 486 Min. :1970 Min. :0.0000 all :4131
## 2 : 486 1st Qu.:1976 1st Qu.:0.0000 MVA :4131
## 4 : 486 Median :1983 Median :0.0000 suicide :4131
## 5 : 486 Mean :1983 Mean :0.3399 homicide :4131
## 6 : 486 3rd Qu.:1990 3rd Qu.:0.7001 other external:4131
## 8 : 486 Max. :1996 Max. :1.0000 internal :4131
## (Other):21870
## agegr count pop age
## 15-17 yrs:8262 Min. : 0.00 Min. : 15023 Min. :15.95
## 18-20 yrs:8262 1st Qu.: 14.00 1st Qu.: 65106 1st Qu.:16.02
## 21-24 yrs:8262 Median : 38.00 Median : 174899 Median :19.01
## Mean : 91.42 Mean : 255117 Mean :19.17
## 3rd Qu.: 96.00 3rd Qu.: 309315 3rd Qu.:22.47
## Max. :2639.00 Max. :2023616 Max. :22.79
##
## legal beertaxa beerpercap winepercap
## Min. :0.0000 Min. :0.01275 Min. :0.600 Min. :0.0600
## 1st Qu.:0.0000 1st Qu.:0.09102 1st Qu.:1.150 1st Qu.:0.1700
## Median :0.0000 Median :0.15524 Median :1.310 Median :0.2700
## Mean :0.4466 Mean :0.23138 Mean :1.319 Mean :0.3046
## 3rd Qu.:1.0000 3rd Qu.:0.28169 3rd Qu.:1.440 3rd Qu.:0.4000
## Max. :1.0000 Max. :1.98454 Max. :2.280 Max. :1.1100
## NA's :288
## spiritpercap totpercap mrate Cstate
## Min. :0.3700 Min. :1.21 Min. : 0.00 Length:24786
## 1st Qu.:0.7000 1st Qu.:2.12 1st Qu.: 12.07 Class :character
## Median :0.8700 Median :2.49 Median : 21.35 Mode :character

3
## Mean :0.9964 Mean :2.62 Mean : 37.33
## 3rd Qu.:1.1200 3rd Qu.:2.89 3rd Qu.: 43.73
## Max. :4.4500 Max. :6.92 Max. :623.02
##
## NAME
## Length:24786
## Class :character
## Mode :character
##
##
##
##
Let’s look at the data for the state of Minnesota (MN) and Arkansas (AR) and compare the time-series of
death rates from motor vehicle accidents (dtype == "MVA") for the three age groups.
deaths_sel <- deaths %>% filter(Cstate %in% c("MN","AR"), dtype == "MVA")

g1 <- ggplot(deaths_sel, aes(y=mrate, x=year, color = agegr)) +

geom_line() +
ggtitle("Death rates (per 100,000) for MVA") +
facet_grid(.~Cstate)
g1

Death rates (per 100,000) for MVA

AR MN

75

agegr
mrate

15−17 yrs
18−20 yrs
50
21−24 yrs

25

1970 1980 1990 1970 1980 1990

year

Further look at the death rates for the 18-20 year group for different causes of death (excluding “all”).
deaths_sel <- deaths %>%
filter(Cstate %in% c("AR","MN"), agegr == "18-20 yrs") %>%

4
 filter(!(dtype == "all"))

g2 <- ggplot(deaths_sel, aes(y=mrate, x=year, color = dtype)) +

geom_line() +
ggtitle("Death rates (per 100,000), different reasons, 18-20 yrs") +
facet_grid(.~Cstate)
g2

Death rates (per 100,000), different reasons, 18−20 yrs

AR MN

75

dtype
MVA
mrate

50 suicide
homicide
other external
internal
25

0
1970 1980 1990 1970 1980 1990
year

The same graphs for Alabama (AL) and Michigan (MI) are here:
deaths_sel <- deaths %>% filter(Cstate %in% c("AL","MI"), dtype == "MVA")

g1 <- ggplot(deaths_sel, aes(y=mrate, x=year, color = agegr)) +

geom_line() +
ggtitle("Death rates (per 100,000) for MVA") +
facet_grid(.~Cstate)
g1

5
 Death rates (per 100,000) for MVA
AL MI

60

agegr
mrate

15−17 yrs
18−20 yrs

40 21−24 yrs

20

1970 1980 1990 1970 1980 1990

year

Further look at the death rates for the 18-20 year group for different causes of death (excluding “all”).
deaths_sel <- deaths %>%
filter(Cstate %in% c("AL","MI"), agegr == "18-20 yrs") %>%
filter(!(dtype == "all"))

g2 <- ggplot(deaths_sel, aes(y=mrate, x=year, color = dtype)) +

geom_line() +
ggtitle("Death rates (per 100,000), different reasons, 18-20 yrs") +
facet_grid(.~Cstate)
g2

6
 Death rates (per 100,000), different reasons, 18−20 yrs
AL MI

60

dtype
MVA
mrate

40 suicide
homicide
other external
internal
20

0
1970 1980 1990 1970 1980 1990
year

Diff-in-Diff, Two States

Let us look at two states only for now. Alabama (AL) and Arkansas (AR). Throughout the sample period
Arkansas had a minimum legel drinking age (MLDA) of 21. Alabama, however, reduced the MLDA, in 1975,
to 19 (from 21) and only in 1985 increased it back to 21.
Hence we could look at the impact of the reduction of the MLDA in Alabama, using Arkansas as a control
state. Let us select the data for these two states, up to 1983 and for the motor vehicle accidents (dtype =
"MVA").
AP_data_1 <- deaths %>% filter(Cstate %in% c("AL","AR")) %>%
filter(year <= 1983, agegr == "18-20 yrs", dtype == "MVA") %>%
arrange(Cstate,year)

This leaves us with 14 observation for each of the two states (1970 to 1983). Let’s illustrate thes data
graphically:
gdd <- ggplot(AP_data_1, aes(y=mrate, x=year, color = Cstate)) +
geom_line() +
ggtitle("Death rates (per 100,000) for MVA")
gdd

7
 Death rates (per 100,000) for MVA

80

Cstate
mrate

AL
AR
60

40

1972 1976 1980

year

Looking at this picture it is not so obvious whether there has been a (causal) effect of the reduction of the
MLDA in Alabama on death rates through motor vehicle accidents for 18-20 year olds. But just graphical
analysis does not always unveil the effects, which is why we will often look at the problem through a regression
lense.
The way in which we would use this dataset to estimate a simple Diff-in-Diff is as follows.
The general specification is:
Yst = α + βT REATs + γP OSTt + δ(T REATs × P OSTt ) + est
we need to specify the variables
• Yst ,
• T REATS , and
• P OSTt
Yst is the mrate in state s at time t. T REATs is a dummy variable which takes a value 1 (or TRUE) if we
have an observation from Alabama, the state in which the policy change (lowering of MLDA) happened, and
0 (or FALSE) otherwise. P OSTt is another dummy variable which is 1 (or TRUE) for periods affected by the
policy change, i.e. 1975 and onwards, and 0 (or FALSE) otherwise.
AP_data_1 <- AP_data_1 %>%
mutate(treat = (Cstate == "AL"),
post = (year >= 1975),
treatpost = treat*post)

Have a look at the variables to confirm that treatpost only takes the value 1 for AL from 1975 onwards.
It is the estimator to δ which is indicative of any impact the policy has on teh outcome variable (assuming

8
the parallel trend assumption can be justified).
dd_2state <- lm(mrate~treat+post+treatpost, data = AP_data_1)
stargazer(dd_2state, type = "text")

##
## ===============================================
## Dependent variable:
## ---------------------------
## mrate
## -----------------------------------------------
## treat -4.546
## (6.540)
##
## post -14.978**
## (5.767)
##
## treatpost 4.500
## (8.156)
##
## Constant 68.932***
## (4.624)
##
## -----------------------------------------------
## Observations 28
## R2 0.299
## Adjusted R2 0.211
## Residual Std. Error 10.340 (df = 24)
## F Statistic 3.408** (df = 3; 24)
## ===============================================
## Note: *p<0.1; **p<0.05; ***p<0.01
While δ̂ is positive, it is not statiystically significant and, on the basis of these data only, we would not
attribute any adverse consequences to the lowering of the MLDA.
However, we only looked at two states, and disregarding information coming from other states is unwise.

Diff-in-Diff, Many States, variable timing

The question now arises how we would cater for multiple states when implementing a Diff-in-Diff approach
using a regression.
Two complications will need to be taken into account.
1) Different states may implement policies at different times. For instance, Wyoming lowered the MLDA
to 19 as did Alabama, but did so two years earlier, in 1973. This raises the question how the P OSTt
variable shoudl be defined.
2) Some states may have implemented the policy in a slightly different way. As an example, Michigan
lowered the MLDA in 1972 and it lowered the MLDA to 18 (rather than 19), therefore potentially
affecting a larger proportion of 18-20 Year olds. This raises a question about how the T REATs variable
should be defined.
The solution here is to define a new variable, here LEGALst which takes the role of (T REATs × P OSTt ).
This variable takes the value 0 if 0 % of a specific age group (we are interested in the 18-20 year old group)
in state s is allowed to drink legally in a particular year t, say when the MLDA is 21. It takes the value 1 if
100% of the respective age group is allowed to legally drink alcohol, say when the MLDA is 18. It will take

9
values between 0 and 1 to reflect that only a portion of the 18-20 year olds may be legally allowed to drink
alcohol (say, if the MLDA is 19) and/or if they were only allowed to drink alcohol legally for a part of the
respective year, say if the law changed half way through the year.
This variable will be at the core and the coefficient estimate relating to this variable will deliver an estimate
of the causal effect (assuming the parallel trends assumption can be justified).
What happens to the T REATs and P OSTt variables? They will typically be replaced by state- and time
fixed-effects respectively. These two variables controled for differences between the states unrelated to the
policy implementation (T REATs ) and the differences across time which applied universally to all states
(P OSTt ). The fixed effects will do the same thing, merely allowing for more than two states and more than
two time periods.
In our example we have data for 51 states (K = 51) and for 14 years (T = 14). We would include 50
state dummy variables (treating one state as the base case, here Alaska, AK) and 13 year dummy variables
(typically treating the first year as the base case, here 1970). In order to do this we require the state variable
(say the two letter appreviation variable Cstate) and a year variable as factor variables as it is very easy to
add such dummy variables for variables of the factor type.
deaths <- deaths %>% mutate(year_fct = factor(year),
Cstate = factor(Cstate))

We now subset the data as in Angrist and Pischke. Use all 51 states, only look at the “18-20 yrs” age group,
for years up to and including 1983 and for starters look at (dtype == "all").
AP_data <- deaths %>% filter(year <= 1983, agegr == "18-20 yrs", dtype == "all") %>%
arrange(Cstate,year)

The model is now represented as:

Y = α + ΣW Y k 1983 τ
k=AL βk ST AT E + Στ =1971 γτ Y EAR + δLEGAL + e

The index for the ST AT E K variable runs alphabetically from “AL” (Alabama) to “WY” (Wyoming). The
state first in the alphabet , “AK” (Alaska), is the base case Observations are state-years normally indexed by
st but to avoid a too laden notation they are not showing. To illustrate what the dummy variables do let’s
look at the following table which shows a few observations and their respective values for selected variables.

CState Year mrate LEGAL ... Y ear1975 Y ear1976 ... ST AT E AL ST AT E AR ...

AL 1974 147.9 0.0 ... 0 0 ... 1 0 ...
AL 1975 147.9 0.294 ... 1 0 ... 1 0 ...
AL 1976 132.6 0.665 ... 0 1 ... 1 0 ...
... ... ... ... ... ... ... ... ... ... ...
AR 1976 137.5 0.0 ... 0 1 ... 0 1 ...
AR 1977 111.5 0.0 ... 0 0 ... 0 1 ...

Fortunately we do not have to create all the year and state dummy variables. R will do this internally.
The crucial variabke which indicates the “strength” of the policy (reduction of minimum legal drinking age)
is the variable LEGAL. It expresses the proportion of 18-20 year olfd which are allowed to legally dring alcohol.
YOu can see in the table that this propotion was 0 in Arkansas in 1976 and 1977 (as the MLDA was 21). In
Alabama the story is different. In 1974 it was also 0 as the MLDA was still 21. In 1976 the proportion was
0.665 or about 67% as the MLDA had been reduced to 19 (2/3s of 18-20 year olds were allowed to legally
drink). The value for 1975 (LEGAL = 0.294) is a combination of the fact that 2/3s of 18-20 year olds were
allowed to legally drink, but that legislation was in effect for only about 5 months (0.294 is approximately
0.665 * (5/12)).
The above model is then implemented in R as follows:

10
mod1<- lm(mrate ~ legal + state + year_fct, data = AP_data)

Yes that is it! state (we also could have used Cstate) and year_fct are factor variables. R therefore
recognises that these are categorical variables and the appropriate way to include these is by internally
creating the ST AT E k and Y EARτ dummy variables.
With this under our belt we can tackle recreating the values in Angrist and Pischke’s tables 5.2 and 5.3.

Table 5.2

This first regression will estimate the above Diff-in-Diff specification for all causes of death in 18-20 year olds.
We use the lm function to estimate the regression.
mod1 <- lm(mrate ~ legal + state + year_fct, data = AP_data)
stargazer(mod1, keep = "legal", type="text") # keep = "legal" only reports legal

##
## ===============================================
## Dependent variable:
## ---------------------------
## mrate
## -----------------------------------------------
## legal 10.804***
## (3.138)
##
## -----------------------------------------------
## Observations 714
## R2 0.821
## Adjusted R2 0.804
## Residual Std. Error 17.339 (df = 649)
## F Statistic 46.561*** (df = 64; 649)
## ===============================================
## Note: *p<0.1; **p<0.05; ***p<0.01
This is the first entry in Table 5.2 (column (1) “All deaths”). By using the keep = "legal" option in the
call to the stargazer function we ensure that only the coefficient we are really interested in is reported.
Let’s implement the additional state time-trends (as for column (2)). We obtain these by adding state:year
as “a” regressor. Here we use the year variable (not the year_fct variable). This is important as, when we
want a time trend, we need R to recognise that year is a numeric variable (as it is). Adding state:year
actually adds 50 (51 minus one base state) time trends.
mod2 <- lm(mrate ~ legal + year_fct + state + state:year, data = AP_data)
stargazer(mod1, mod2, dep.var.caption = "All Deaths", keep = "legal", type="text")

##
## ======================================================================
## All Deaths
## --------------------------------------------------
## mrate
## (1) (2)
## ----------------------------------------------------------------------
## legal 10.804*** 8.467**
## (3.138) (3.724)
##
## ----------------------------------------------------------------------

11
## Observations 714 714
## R2 0.821 0.848
## Adjusted R2 0.804 0.820
## Residual Std. Error 17.339 (df = 649) 16.616 (df = 599)
## F Statistic 46.561*** (df = 64; 649) 29.412*** (df = 114; 599)
## ======================================================================
## Note: *p<0.1; **p<0.05; ***p<0.01
The standard errors which are calculated in this routine are the usual standard errors which are not corrected
for heteroskedaticity of any kind. And in fact, when you compare these to those in Angrist and Pischke,
you can see that they are notably smaller. In situations like this it is commonly appropriate to calculate
cluster-robust standard errors, i.e. standard errors which recognise that there may be clusters of data which
exhibit different error variances. The natural cluster variable here is the state variable. In other words we
want to allow for error variances which differ across states. To achieve this we use the vcovCL function from
the sandwich package. This will be fed into the function coeftest from the lmtest package.
library(sandwich)
library(lmtest)
mod1_cr <- coeftest(mod1,vcovCL(mod1, cluster = ~ state))

# Create vectors with cluster robust standard errors

# for use in stargazer below
mod1_cr_se <- sqrt(diag(vcovCL(mod1, cluster = ~ state)))
mod2_cr_se <- sqrt(diag(vcovCL(mod2, cluster = ~ state)))

In mod1_ct you now find all the (unchanged) coefficient estimates, the new standard errors and as a result
new t-statistics and p-values for the null hypothesis that the respective population coefficients are 0, for mod1.
Let’s just see the first few rows which also contain the legal variable.
mod1_cr[1:4,]

## Estimate Std. Error t value Pr(>|t|)

## (Intercept) 153.66915 1.9813614 77.557356 0.000000e+00
## legal 10.80414 4.5922045 2.352713 1.893458e-02
## state2 107.88660 1.1828382 91.209937 0.000000e+00
## state4 37.80343 0.6717837 56.273219 7.753281e-252
You can see that the new standard error to legal is 4.59 as in Table 5.2 in Angrist and Pischke. And in fact
you can work this new standard error into the ‘stargazer display
# keep = "legal" only reports legal
stargazer(mod1, keep = "legal", type="text", se=list(mod1_cr_se))

##
## ===============================================
## Dependent variable:
## ---------------------------
## mrate
## -----------------------------------------------
## legal 10.804**
## (4.592)
##
## -----------------------------------------------
## Observations 714
## R2 0.821
## Adjusted R2 0.804
## Residual Std. Error 17.339 (df = 649)

12
## F Statistic 46.561*** (df = 64; 649)
## ===============================================
## Note: *p<0.1; **p<0.05; ***p<0.01
or for models mod1 and mod2
stargazer(mod1,mod2, keep = "legal", type="text",
se=list(mod1_cr_se,mod2_cr_se)) # keep = "legal" only reports legal

##
## ======================================================================
## Dependent variable:
## --------------------------------------------------
## mrate
## (1) (2)
## ----------------------------------------------------------------------
## legal 10.804** 8.467*
## (4.592) (5.098)
##
## ----------------------------------------------------------------------
## Observations 714 714
## R2 0.821 0.848
## Adjusted R2 0.804 0.820
## Residual Std. Error 17.339 (df = 649) 16.616 (df = 599)
## F Statistic 46.561*** (df = 64; 649) 29.412*** (df = 114; 599)
## ======================================================================
## Note: *p<0.1; **p<0.05; ***p<0.01
Now we repeat this analysis for the different reasons of Death, always reporting the cluster robust standard
errors:
# MVAs
AP_data <- filter(deaths, year <= 1983, agegr == "18-20 yrs", dtype == "MVA")
mod1 <- lm(mrate ~ legal + state + year_fct, data = AP_data)
mod1_cr_se <- sqrt(diag(vcovCL(mod1, cluster = ~ state)))

mod2 <- lm(mrate ~ legal + year_fct + state + state:year, data = AP_data)

mod2_cr_se <- sqrt(diag(vcovCL(mod2, cluster = ~ state)))

stargazer(mod1, mod2, dep.var.caption = "MVA", keep = "legal",

type="text",se=list(mod1_cr_se,mod2_cr_se))

##
## ======================================================================
## MVA
## --------------------------------------------------
## mrate
## (1) (2)
## ----------------------------------------------------------------------
## legal 7.592*** 6.644**
## (2.496) (2.656)
##
## ----------------------------------------------------------------------
## Observations 714 714
## R2 0.798 0.836
## Adjusted R2 0.778 0.805

13
## Residual Std. Error 10.960 (df = 649) 10.269 (df = 599)
## F Statistic 39.976*** (df = 64; 649) 26.797*** (df = 114; 599)
## ======================================================================
## Note: *p<0.1; **p<0.05; ***p<0.01
# Suicides
AP_data <- filter(deaths, year <= 1983, agegr == "18-20 yrs", dtype == "suicide")
mod1 <- lm(mrate ~ legal + state + year_fct, data = AP_data)
mod1_cr_se <- sqrt(diag(vcovCL(mod1, cluster = ~ state)))

mod2 <- lm(mrate ~ legal + year_fct + state + state:year, data = AP_data)

mod2_cr_se <- sqrt(diag(vcovCL(mod2, cluster = ~ state)))

stargazer(mod1, mod2, dep.var.caption = "Suicide", keep = "legal",

type="text",se=list(mod1_cr_se,mod2_cr_se))

##
## =====================================================================
## Suicide
## -------------------------------------------------
## mrate
## (1) (2)
## ---------------------------------------------------------------------
## legal 0.591 0.474
## (0.590) (0.795)
##
## ---------------------------------------------------------------------
## Observations 714 714
## R2 0.608 0.637
## Adjusted R2 0.570 0.568
## Residual Std. Error 4.372 (df = 649) 4.381 (df = 599)
## F Statistic 15.747*** (df = 64; 649) 9.218*** (df = 114; 599)
## =====================================================================
## Note: *p<0.1; **p<0.05; ***p<0.01
# Internal
AP_data <- filter(deaths, year <= 1983, agegr == "18-20 yrs", dtype == "internal")
mod1 <- lm(mrate ~ legal + state + year_fct, data = AP_data)
mod1_cr_se <- sqrt(diag(vcovCL(mod1, cluster = ~ state)))

mod2 <- lm(mrate ~ legal + year_fct + state + state:year, data = AP_data)

mod2_cr_se <- sqrt(diag(vcovCL(mod2, cluster = ~ state)))

stargazer(mod1, mod2, dep.var.caption = "Internal", keep = "legal",

type="text",se=list(mod1_cr_se,mod2_cr_se))

##
## ======================================================================
## Internal
## --------------------------------------------------
## mrate
## (1) (2)
## ----------------------------------------------------------------------
## legal 1.333 0.079
## (1.586) (1.933)
##

14
## ----------------------------------------------------------------------
## Observations 714 714
## R2 0.622 0.681
## Adjusted R2 0.585 0.620
## Residual Std. Error 6.365 (df = 649) 6.088 (df = 599)
## F Statistic 16.680*** (df = 64; 649) 11.206*** (df = 114; 599)
## ======================================================================
## Note: *p<0.1; **p<0.05; ***p<0.01

Table 5.3

For the results in this table we add an additional control variable, namely the amount of beer taxes.
# All deaths
AP_data <- filter(deaths, year <= 1983, agegr == "18-20 yrs", dtype == "all")
mod1 <- lm(mrate ~ legal + beertaxa + state + year_fct, data = AP_data)
mod1_cr_se <- sqrt(diag(vcovCL(mod1, cluster = ~ state)))

mod2 <- lm(mrate ~ legal + beertaxa + year_fct + state + state:year, data = AP_data)
mod2_cr_se <- sqrt(diag(vcovCL(mod2, cluster = ~ state)))

stargazer(mod1, mod2, dep.var.caption = "MVA - incl. beertax",

keep = c("legal","beertaxa"), type="text",
se=list(mod1_cr_se,mod2_cr_se))

##
## ======================================================================
## MVA - incl. beertax
## --------------------------------------------------
## mrate
## (1) (2)
## ----------------------------------------------------------------------
## legal 10.983** 10.029**
## (4.691) (4.915)
##
## beertaxa 1.505 -5.525
## (9.071) (32.238)
##
## ----------------------------------------------------------------------
## Observations 700 700
## R2 0.825 0.850
## Adjusted R2 0.807 0.821
## Residual Std. Error 17.230 (df = 635) 16.570 (df = 586)
## F Statistic 46.645*** (df = 64; 635) 29.452*** (df = 113; 586)
## ======================================================================
## Note: *p<0.1; **p<0.05; ***p<0.01
# MVA
AP_data <- filter(deaths, year <= 1983, agegr == "18-20 yrs", dtype == "MVA")
mod1 <- lm(mrate ~ legal + beertaxa + state + year_fct, data = AP_data)
mod1_cr_se <- sqrt(diag(vcovCL(mod1, cluster = ~ state)))

mod2 <- lm(mrate ~ legal + beertaxa + year_fct + state + state:year, data = AP_data)
mod2_cr_se <- sqrt(diag(vcovCL(mod2, cluster = ~ state)))

15
stargazer(mod1, mod2, dep.var.caption = "MVA - incl. beertax",
keep = c("legal","beertaxa"), type="text",
se=list(mod1_cr_se,mod2_cr_se))

##
## ======================================================================
## MVA - incl. beertax
## --------------------------------------------------
## mrate
## (1) (2)
## ----------------------------------------------------------------------
## legal 7.588*** 6.888***
## (2.561) (2.664)
##
## beertaxa 3.819 26.882
## (5.394) (20.114)
##
## ----------------------------------------------------------------------
## Observations 700 700
## R2 0.797 0.836
## Adjusted R2 0.777 0.804
## Residual Std. Error 10.953 (df = 635) 10.254 (df = 586)
## F Statistic 38.994*** (df = 64; 635) 26.425*** (df = 113; 586)
## ======================================================================
## Note: *p<0.1; **p<0.05; ***p<0.01
# Suicide
AP_data <- filter(deaths, year <= 1983, agegr == "18-20 yrs", dtype == "suicide")
mod1 <- lm(mrate ~ legal + beertaxa + state + year_fct, data = AP_data)
mod1_cr_se <- sqrt(diag(vcovCL(mod1, cluster = ~ state)))

mod2 <- lm(mrate ~ legal + beertaxa + year_fct + state + state:year, data = AP_data)
mod2_cr_se <- sqrt(diag(vcovCL(mod2, cluster = ~ state)))

stargazer(mod1, mod2, dep.var.caption = "Suicide",

keep = c("legal","beertaxa"), type="text",
se=list(mod1_cr_se,mod2_cr_se))

##
## =====================================================================
## Suicide
## -------------------------------------------------
## mrate
## (1) (2)
## ---------------------------------------------------------------------
## legal 0.448 0.378
## (0.595) (0.768)
##
## beertaxa -3.052* -12.127
## (1.633) (8.818)
##
## ---------------------------------------------------------------------
## Observations 700 700
## R2 0.619 0.648
## Adjusted R2 0.581 0.580

16
## Residual Std. Error 4.314 (df = 635) 4.317 (df = 586)
## F Statistic 16.146*** (df = 64; 635) 9.559*** (df = 113; 586)
## =====================================================================
## Note: *p<0.1; **p<0.05; ***p<0.01
# Internal
AP_data <- filter(deaths, year <= 1983, agegr == "18-20 yrs", dtype == "internal")
mod1 <- lm(mrate ~ legal + beertaxa + state + year_fct, data = AP_data)
mod1_cr_se <- sqrt(diag(vcovCL(mod1, cluster = ~ state)))

mod2 <- lm(mrate ~ legal + beertaxa + year_fct + state + state:year, data = AP_data)
mod2_cr_se <- sqrt(diag(vcovCL(mod2, cluster = ~ state)))

stargazer(mod1, mod2, dep.var.caption = "Internal",

keep = c("legal","beertaxa"), type="text",
se=list(mod1_cr_se,mod2_cr_se))

##
## ======================================================================
## Internal
## --------------------------------------------------
## mrate
## (1) (2)
## ----------------------------------------------------------------------
## legal 1.465 0.877
## (1.609) (1.808)
##
## beertaxa -1.357 -10.309
## (3.069) (11.636)
##
## ----------------------------------------------------------------------
## Observations 700 700
## R2 0.637 0.692
## Adjusted R2 0.600 0.632
## Residual Std. Error 6.251 (df = 635) 5.995 (df = 586)
## F Statistic 17.382*** (df = 64; 635) 11.628*** (df = 113; 586)
## ======================================================================
## Note: *p<0.1; **p<0.05; ***p<0.01

Summary

This document illustrated how to implement a Diff-in-Diff estimation when you are dealing with multiple
groups which implement policies at different times. In this particular case we also allowed for implementations
which differ, not only in timing, but also in terms of intensity.
As usual, the actual estimation was rather straightforward. The real work lies in the collection and handling
of the data as well as in figuring out how to apply the straightword techniques.

17