Introduction to

Difference in Differences
Using Stata

Chuck Huber
StataCorp
[email protected]

Oceania Stata Conference 2023

February 9, 2023
 Download Website
• You can download all of the slides, datasets and
do-files here:

https://tinyurl.com/StataDID
 Outline
• The Question
• An Intuitive Introduction
• Two Period, Two Groups Model
• Repeated Cross-Sectional Panel Data
• Longitudinal Panel Data
• More information
 Outline
• The Question
• An Intuitive Introduction
• Two Period, Two Groups Model
• Repeated Cross-Sectional Panel Data
• Longitudinal Panel Data
• More information
A Conceptual Introduction
A Conceptual Introduction
A Conceptual Introduction
A Conceptual Introduction
A Conceptual Introduction
A Conceptual Introduction
A Conceptual Introduction
A Conceptual Introduction
A Conceptual Introduction
 Outline
• The Question
• An Intuitive Introduction
• Two Period, Two Groups Model
• Repeated Cross-Sectional Panel Data
• Longitudinal Panel Data
• More information
The Two Period, Two Group Model
 The Examples Dataset

. use MLDA21, clear

(Minimum Legal Drinking Age (MLDA) from Angrist and Pischke, 2014 (type 'notes'))

. describe FIPS state year MLDA21_year MLDA21 mrate_agegrp beertax

Variable Storage Display Value

name type format label Variable label

FIPS byte %10.0g State ID (Federal Information Processing Standard)

state str15 %15s * State
year int %9.0g * Year
MLDA21_year int %10.0g Year in which MLDA21 law was enacted
MLDA21 byte %15.0g treat83 Pre- or Post- MLDA21 Law
mrate_agegrp double %9.2f * Mortality rate (deaths/100000) in the 18-20 age group
beertax double %9.2f Beer tax
 The Examples Dataset
. summarize FIPS state year MLDA21_year MLDA21 mrate_agegrp beertax

Variable Obs Mean Std. dev. Min Max

FIPS 1,242 27.13043 14.33362 1 51

state 0
year 1,242 1983 7.792018 1970 1996
MLDA21_year 1,242 1973.239 21.16342 1933 1988
MLDA21 1,242 .5764895 .4943138 0 1

mrate_agegrp 1,242 52.14309 21.93833 8.231131 189.9421

beertax 1,226 .2367918 .236822 .012747 1.984536
. list state MLDA21_year year MLDA21 /// . // List the data for Iowa
> if inlist(FIPS, 17) /// . list state MLDA21_year year MLDA21 ///
> , sepby(state) nolabel noobs abbrev(18) > if inlist(FIPS, 19) ///
> , sepby(state) nolabel noobs abbrev(18)

state MLDA21_year year MLDA21

state MLDA21_year year MLDA21
Illinois 1980 1970 0
Illinois 1980 1971 0 Iowa 1986 1970 0
Illinois 1980 1972 0 Iowa 1986 1971 0
Illinois 1980 1973 0 Iowa 1986 1972 0
Illinois 1980 1974 0 Iowa 1986 1973 0
Illinois 1980 1975 0 Iowa 1986 1974 0
Illinois 1980 1976 0 Iowa 1986 1975 0
Illinois 1980 1977 0 Iowa 1986 1976 0
Iowa 1986 1977 0
Illinois 1980 1978 0
Iowa 1986 1978 0
Illinois 1980 1979 0
Iowa 1986 1979 0
Illinois 1980 1980 1
Iowa 1986 1980 0
Illinois 1980 1981 1 Iowa 1986 1981 0
Illinois 1980 1982 1 Iowa 1986 1982 0
Illinois 1980 1983 1 Iowa 1986 1983 0
Illinois 1980 1984 1 Iowa 1986 1984 0
Illinois 1980 1985 1 Iowa 1986 1985 0
Illinois 1980 1986 1 Iowa 1986 1986 1
Illinois 1980 1987 1 Iowa 1986 1987 1
Illinois 1980 1988 1 Iowa 1986 1988 1
Illinois 1980 1989 1 Iowa 1986 1989 1
Illinois 1980 1990 1 Iowa 1986 1990 1
Illinois 1980 1991 1 Iowa 1986 1991 1
Illinois 1980 1992 1 Iowa 1986 1992 1
Illinois 1980 1993 1 Iowa 1986 1993 1
Illinois 1980 1994 1 Iowa 1986 1994 1
Illinois 1980 1995 1 Iowa 1986 1995 1
Illinois 1980 1996 1 Iowa 1986 1996 1
 Setting Up The Data
. keep if inrange(year, 1979, 1981)
(1,104 observations deleted) Keep the data for 1979, 1980, and 1981

.
. list state MLDA21_year year MLDA21 ///
> if year!=1980 & inlist(FIPS, 17,18,19) ///
> , sepby(state) nolabel noobs abbrev(18)

state MLDA21_year year MLDA21

Illinois 1980 1979 0 The law went into effect

Illinois 1980 1981 1 during the time interval
Indiana 1934 1979 1 The law was in effect
Indiana 1934 1981 1 during the entire interval
Iowa 1986 1979 0 The law was never in effect
Iowa 1986 1981 0 during the entire interval
 Setting Up The Data
. // Create a variable for time
. generate time = (year>1980)

.
. list state MLDA21_year year MLDA21 time ///
> if year!=1980 & inlist(FIPS, 17,18,19) ///
> , sepby(state) nolabel noobs abbrev(18)

state MLDA21_year year MLDA21 time

Illinois 1980 1979 0 0

Illinois 1980 1981 1 1 Time equals 0 prior to
the intervention in 1980
Indiana 1934 1979 1 0 and time equals 1 after
Indiana 1934 1981 1 1
the intervention in 1980
Iowa 1986 1979 0 0
Iowa 1986 1981 0 1
 Setting Up The Data
// Create a variable for treatment status
gen treated = .
replace treated = 0 if MLDA21_year >1980
replace treated = 0 if MLDA21_year <=1980 & year<1980
replace treated = 1 if MLDA21_year <=1980

. list state MLDA21_year year MLDA21 time treated ///

> if year!=1980 & inlist(FIPS, 17,18,19) ///
> , sepby(state) nolabel noobs abbrev(18)

state MLDA21_year year MLDA21 time treated

Illinois 1980 1979 0 0 1 Treated equals 1 if the

Illinois 1980 1981 1 1 1 MLDA21 law was in effect
Indiana 1934 1979 1 0 1 in that state during 1980 or
Indiana 1934 1981 1 1 1 sooner. Treated equals 0
Iowa 1986 1979 0 0 0 if the law went into effect
Iowa 1986 1981 0 1 0
after 1980.
 Setting Up The Data
. // Create a variable for the interaction of time and treatment
. gen TimeTreat = time*treated

.
. list state MLDA21_year year MLDA21 time treated TimeTreat ///
> if year!=1980 & inlist(FIPS, 17,18,19) ///
> , sepby(state) nolabel noobs abbrev(18)

state MLDA21_year year MLDA21 time treated TimeTreat

Illinois 1980 1979 0 0 1 0

Illinois 1980 1981 1 1 1 1

Indiana 1934 1979 1 0 1 0

Indiana 1934 1981 1 1 1 1

Iowa 1986 1979 0 0 0 0

Iowa 1986 1981 0 1 0 0

TimeTreat equals time x treated.

It will be the interaction term when we use regress.
It will be the treatment variable when we use didregress!!!!
 DiD Using regress
. regress mrate_agegrp time treated TimeTreat, vce(cluster FIPS)

Linear regression Number of obs = 138

F(3, 45) = 5.76
Prob > F = 0.0020
R-squared = 0.0233
Root MSE = 22.879

(Std. err. adjusted for 46 clusters in FIPS)

Robust
mrate_agegrp Coefficient std. err. t P>|t| [95% conf. interval]

time -3.92868 2.045822 -1.92 0.061 -8.049176 .1918168

treated 6.549406 6.480692 1.01 0.318 -6.503378 19.60219
TimeTreat -4.611695 3.32785 -1.39 0.173 -11.31433 2.090939
_cons 59.82451 3.878595 15.42 0.000 52.01262 67.6364

The Average Treatment Effect Among the Treated (ATET) is the coefficient
for TimeTreat.

The MLDA21 law caused a decrease of 4.6 people per 100,000 population
in the 18-20 year old age group between 1979 and 1981.
 DiD Using regress
. regress mrate_agegrp i.time##i.treated, vce(cluster FIPS)

Linear regression Number of obs = 138

F(3, 45) = 5.76
Prob > F = 0.0020
R-squared = 0.0233
Root MSE = 22.879

(Std. err. adjusted for 46 clusters in FIPS)

Robust
mrate_agegrp Coefficient std. err. t P>|t| [95% conf. interval]

1.time -3.92868 2.045822 -1.92 0.061 -8.049176 .1918168

1.treated 6.549406 6.480692 1.01 0.318 -6.503378 19.60219

time#treated
1 1 -4.611695 3.32785 -1.39 0.173 -11.31433 2.090939

_cons 59.82451 3.878595 15.42 0.000 52.01262 67.6364

 DiD Using regress
. regress mrate_agegrp i.year TimeTreat i.FIPS, vce(cluster FIPS)

Linear regression Number of obs = 138

F(2, 45) = .
Prob > F = .
R-squared = 0.8714
Root MSE = 10.187

(Std. err. adjusted for 46 clusters in FIPS)

Robust
mrate_agegrp Coefficient std. err. t P>|t| [95% conf. interval]

year
1980 -.5155614 2.882514 -0.18 0.859 -6.321242 5.29012
1981 -4.18646 3.313667 -1.26 0.213 -10.86053 2.487608

TimeTreat -4.611695 4.083392 -1.13 0.265 -12.83607 3.612679

FIPS
2 10.37779 2.88e-13 3.6e+13 0.000 10.37779 10.37779
4 3.268883 2.88e-13 1.1e+13 0.000 3.268883 3.268883
5 5.474611 1.361131 4.02 0.000 2.733153 8.216069

(FIPS output truncated)

 DiD Using regress
. regress mrate_agegrp i.year TimeTreat i.FIPS, vce(cluster FIPS)

Linear regression Number of obs = 138

F(2, 45) = .
Prob > F = .
R-squared = 0.8714
Root MSE = 10.187

(Std. err. adjusted for 46 clusters in FIPS)

Robust
mrate_agegrp Coefficient std. err. t P>|t| [95% conf. interval]

year
1980 -.5155614 2.882514 -0.18 0.859 -6.321242 5.29012
1981 -4.18646 3.313667 -1.26 0.213 -10.86053 2.487608

TimeTreat -4.611695 4.083392 -1.13 0.265 -12.83607 3.612679

FIPS
2 10.37779 2.88e-13 3.6e+13 0.000 10.37779 10.37779
4 3.268883 2.88e-13 1.1e+13 0.000 3.268883 3.268883
5 5.474611 1.361131 4.02 0.000 2.733153 8.216069
DiD Using regress
 DiD Using didregress

didregress (ovar omvarlist) ///

(tvar[, continuous]), ///
group(groupvars) ///
[time(timevar) options]

didregress (mrate_agegrp) ///

(TimeTreat), ///
group(FIPS) ///
time(year)

TimeTreat = Time x Treat !!!!

 DiD Using didregress
. didregress (mrate_agegrp)(TimeTreat), group(FIPS) time(year)

Number of groups and treatment time

Time variable: year

Control: TimeTreat = 0
Treatment: TimeTreat = 1

Control Treatment

Group
FIPS 33 13

Time
Minimum 1979 1981
Maximum 1979 1981

Difference-in-differences regression Number of obs = 138

Data type: Repeated cross-sectional

(Std. err. adjusted for 46 clusters in FIPS)

Robust
mrate_agegrp Coefficient std. err. t P>|t| [95% conf. interval]

ATET
TimeTreat
(1 vs 0) -4.611695 4.083392 -1.13 0.265 -12.83607 3.612679

Note: ATET estimate adjusted for group effects and time effects.
Checking the Parallel Trends Assumption

estat trendplots
Checking the Parallel Trends Assumption

. estat granger

Granger causality test

H0: No effect in anticipation of treatment

F(1, 45) = 0.02

Prob > F = 0.8871
 didregress With Covariates
. didregress (mrate_agegrp beertax)(TimeTreat), group(FIPS) time(year)

Number of groups and treatment time

Time variable: year

Control: TimeTreat = 0
Treatment: TimeTreat = 1

Control Treatment

Group
FIPS 32 13

Time
Minimum 1979 1981
Maximum 1979 1981

Difference-in-differences regression Number of obs = 135

Data type: Repeated cross-sectional

(Std. err. adjusted for 45 clusters in FIPS)

Robust
mrate_agegrp Coefficient std. err. t P>|t| [95% conf. interval]

ATET
TimeTreat
(1 vs 0) -5.140093 4.195906 -1.23 0.227 -13.59639 3.316199

Note: ATET estimate adjusted for covariates, group effects, and time effects.
Estimation of Standard Errors
 Estimation of Standard Errors

Must account for the serial correlation of the outcome.

For a large number of groups

By default, didregress uses cluster–robust standard
errors and uses critical values of a t distribution with
G − 1 degrees of freedom, where G is the number
of groups.

For more information: https://www.stata.com/manuals/tedidintro.pdf

 Estimation of Standard Errors
For a small number of groups
wildbootstrap[(wildopts)] computes confidence intervals and p-values with the wild bootstrap. The wild bootstrap is constructed
imposing the null hypothesis that the ATET is 0; that is, it is a restricted wild bootstrap. Confidence intervals are computed
separately from the p-values. The bounds of the confidence interval are computed using a bisection optimization algorithm
described in Methods and formulas in [TE] didregress. wildopts are errorweight(edtype), reps(#), rseed(#), and blocksize(#).

errorweight(edtype) defines the error weight used to draw residuals from the wild bootstrap. edtype is one of rademacher (the
default), mammen, webb, normal, or gamma.

rademacher multiplies the residuals at each bootstrap replication with a randomly generated variable that takes the value
of 1 with probability 0.5 and the value of -1 with probability 0.5. errorweight(rademacher) is the default.

mammen multiplies the residuals at each bootstrap replication with a randomly generated variable that takes the value of 1
- phi with probability phi/sqrt{5} and phi otherwise, where phi = (1+sqrt{5})/2.

webb multiplies the residuals at each bootstrap replication with a randomly generated variable that takes the values
-sqrt{3/2}, -sqrt{2/2}, -sqrt{1/2}, sqrt{1/2}, sqrt{2/2}, and sqrt{3/2}, each with probability 1/6.

normal multiplies the residuals at each bootstrap replication with a randomly generated normal distribution variable with
the first four moments given by 0, 1, 0, and 3.

gamma multiplies the residuals at each bootstrap replication with a randomly generated gamma distribution variable with
shape parameter 4 and scale parameter 1/2.

reps(#) performs # wild bootstrap replications. The default is reps(1000).

rseed(#) sets the random-number seed to #.

blocksize(#) specifies that the wild bootstrap be performed in blocks, with # replications per block. The wild bootstrap
computation requires a matrix with dimensions (# groups) x (# replications). If this is too large, you can reduce the
matrix to (# groups) x (# block size) and loop (# replications)/(# block size) times. When the same random seed is set,
using a different block size does not change the numerical results; it only modifies the computation method. The block
size must be less than or equal to the number of bootstrap replications.
 Outline
• The Question
• An Intuitive Introduction
• Two Period, Two Groups Model
• Repeated Cross-Sectional Panel Data
• Longitudinal Panel Data
• More information
DiD For Repeated Cross-Sectional Data
DiD For Repeated Cross-Sectional Data

use MLDA21, clear

// Create a variable for time

generate time = (year>1980)

// Create a variable for treatment status

gen treated = .
replace treated = 0 if MLDA21_year >1980
replace treated = 0 if MLDA21_year <=1980 & year<1980
replace treated = 1 if MLDA21_year <=1980

// Create a variable for the interaction of time and treatment

gen TimeTreat = time*treated
. list state MLDA21_year year MLDA21 time treated TimeTreat ///
> if year!=1980 & inlist(FIPS, 17) ///
> , sepby(state) nolabel noobs abbrev(18)

state MLDA21_year year MLDA21 time treated TimeTreat

Illinois 1980 1970 0 0 1 0

Illinois 1980 1971 0 0 1 0
Illinois 1980 1972 0 0 1 0
Illinois 1980 1973 0 0 1 0
Illinois 1980 1974 0 0 1 0
Illinois 1980 1975 0 0 1 0
Illinois 1980 1976 0 0 1 0
Illinois 1980 1977 0 0 1 0
Illinois 1980 1978 0 0 1 0
Illinois 1980 1979 0 0 1 0
Illinois 1980 1981 1 1 1 1
Illinois 1980 1982 1 1 1 1
Illinois 1980 1983 1 1 1 1
Illinois 1980 1984 1 1 1 1
Illinois 1980 1985 1 1 1 1
Illinois 1980 1986 1 1 1 1
Illinois 1980 1987 1 1 1 1
Illinois 1980 1988 1 1 1 1
Illinois 1980 1989 1 1 1 1
Illinois 1980 1990 1 1 1 1
Illinois 1980 1991 1 1 1 1
Illinois 1980 1992 1 1 1 1
Illinois 1980 1993 1 1 1 1
Illinois 1980 1994 1 1 1 1
Illinois 1980 1995 1 1 1 1
Illinois 1980 1996 1 1 1 1
DiD For Repeated Cross-Sectional Data
. didregress (mrate_agegrp)(TimeTreat), group(FIPS) time(year)

Number of groups and treatment time

Time variable: year

Control: TimeTreat = 0
Treatment: TimeTreat = 1

Control Treatment

Group
FIPS 33 13

Time
Minimum 1970 1981
Maximum 1970 1981

Difference-in-differences regression Number of obs = 1,242

Data type: Repeated cross-sectional

(Std. err. adjusted for 46 clusters in FIPS)

Robust
mrate_agegrp Coefficient std. err. t P>|t| [95% conf. interval]

ATET
TimeTreat
(1 vs 0) -3.268054 4.212866 -0.78 0.442 -11.7532 5.217093

Note: ATET estimate adjusted for group effects and time effects.
DiD For Repeated Cross-Sectional Data

estat trendplots
DiD For Repeated Cross-Sectional Data

. estat granger

Granger causality test

H0: No effect in anticipation of treatment

F(10, 45) = 0.95

Prob > F = 0.4952
 Outline
• The Question
• An Intuitive Introduction
• Two Period, Two Groups Model
• Repeated Cross-Sectional Panel Data
• Longitudinal Panel Data
• More information
DiD For Longitudinal Panel Data
 DiD For Longitudinal/Panel Data

. xtset FIPS year, yearly

Panel variable: FIPS (strongly balanced)

Time variable: year, 1970 to 1996
Delta: 1 year
 DiD For Longitudinal/Panel Data
. xtdidregress (mrate_agegrp)(TimeTreat), group(FIPS) time(year)

Number of groups and treatment time

Time variable: year

Control: TimeTreat = 0
Treatment: TimeTreat = 1

Control Treatment

Group
FIPS 33 13

Time
Minimum 1970 1981
Maximum 1970 1981

Difference-in-differences regression Number of obs = 1,242

Data type: Longitudinal

(Std. err. adjusted for 46 clusters in FIPS)

Robust
mrate_agegrp Coefficient std. err. t P>|t| [95% conf. interval]

ATET
TimeTreat
(1 vs 0) -3.268054 4.134048 -0.79 0.433 -11.59445 5.058346

Note: ATET estimate adjusted for panel effects and time effects.
 DiD For Longitudinal/Panel Data

estat trendplots
 DiD For Longitudinal/Panel Data

. estat granger

Granger causality test

H0: No effect in anticipation of treatment

F(10, 45) = 0.99

Prob > F = 0.4650
 Outline
• The Question
• An Intuitive Introduction
• Two Period, Two Groups Model
• Repeated Cross-Sectional Panel Data
• Longitudinal Panel Data
• More information
 Read The Fine Manuals!

didregress https://www.stata.com/manuals/te.pdf
xtdidregress https://www.stata.com/manuals/xt.pdf
More Information
 More Information

Richard Gallenstein’s YouTube video about DiD

https://www.youtube.com/watch?v=rrgxR97QBpc
 More Information
US National Highway Traffic Safety Administration Information

Miron and Tetelbaum, 2009

 Thank you!

Questions?

You can download the slides, datasets, and do-files here:

https://tinyurl.com/StataDID

You can contact me anytime at [email protected]