Popular in Economics

Central Monetary Authorities

Macroeconomic modelling of CBDC: a critical review

European Central Bank Ulrich Bindseil

Central Monetary Authorities

Financing Private Credit

Federal Reserve Bank of New York Nina Boyarchenko

Economics

Failing Banks

Massachusetts Institute of Technology Emil Verner

Economic History

Golden Fetters

UC Berkeley Barry Eichengreen

Economic Inequalities

Visions of Inequality

City University of New York Branko Milanovic

Economics

School Finance Equalization Increases Intergenerational Mobility

Yale University Barbara Biasi

Economics

Olivier Blanchard on U.S. Pandemic-Era Inflation

Peterson Institute for International Economics Olivier Blanchard

xtevent: Estimation and Visualization in the Linear Panel Event-Study Design

View more insights in Social Sciences - Economics - Central Monetary Authorities - Econometrics

View more figures from Federal Reserve Bank of Philadelphia

View the page of Simon Freyaldenhoven

Read the Study

Next in Central Monetary Authorities

Measurement and Theory of Core Inflation

Argia Sbordone | Federal Reserve Bank of New York

Time-varying Persistence of House Price Growth

C.Y. Choi | Federal Reserve Bank of Dallas

How can the Stata package xtevent be used to create event-study plots and estimate policy effects?

Image courtesy of interviewee. December 2, 2024

Log-in or Sign-up to Faculti
Currently viewing this subject insight as a guest. You have insight(s) remaining for this month. Login to view 8000+ figures on the platform.

Generating Workspace...

Export as PDF

This video discusses the XD events data package and its contributions to estimating linear panel models and visualizing results in event studies. The package allows researchers to estimate the dynamic effects of policies on outcomes using panel datasets, exemplified by examining the effect of minimum wage changes on employment. It employs various estimators, including two-way fixed effects and other advanced methods, catering to different research contexts.

Additionally, the package enhances visualization through improved event study plots, incorporating normalization techniques for better interpretation, and confidence bands for hypothesis testing over the coefficient path. It also includes diagnostics to assess pre-trend assumptions and estimation window adequacy. The overall aim is to provide a versatile and informative tool for researchers in econometrics, facilitating both estimation and visualization of dynamic treatment effects in policy analyses.

Language:

Export as PDF

Export as PDF
Here are five key points from the video regarding the XD events data package:

1. **Purpose of the Package**: The XD events data package facilitates the estimation of linear panel models across a variety of settings while providing enhanced visualization of results through event study plots. This aims to help researchers analyze the dynamic effects of policy changes on outcomes, such as the impact of minimum wage changes on employment.

2. **Panel Data Application**: The package is designed to work with panel datasets, allowing researchers to observe how outcomes, like employment levels, change in response to policy events over time. It emphasizes the importance of comparing trends in treated states (e.g., New Jersey after a minimum wage increase) with control states to isolate the causal impact of policies.

3. **Flexible Estimation Options**: The package supports a range of estimators that address various research convideos, including controlling for unit-specific trends, addressing pre-trends, and implementing instrumental variable (IV) estimators. This flexibility allows researchers to choose appropriate methods based on their specific identifying assumptions and research questions.

4. **Visualization Enhancements**: In addition to estimation, the package improves visualization techniques for event study plots. Key enhancements include proper normalization of coefficients, inclusion of confidence bands for hypothesis testing of the entire treatment path, and the addition of smooth paths to assess underlying trends in data related to policy effects.

5. **Broad Applicability**: The XD events package is versatile and applicable to various types of policies beyond just binary or staggered adoptions. It is designed to accommodate heterogeneous treatment effects and improve the robustness of results, thus addressing common limitations in existing literature on event study methodology.

Export as PDF
Here are some key quotes from the video, highlighting the main points:

1. **Purpose of the XD Events Data Package**:
- "First, it enables the estimation of linear panel models in a fairly broad class of settings. And second, it visualizes the corresponding results in the form of an event study plot, and includes several enhancements to these plots compared to what is currently standard practice."

2. **Convideo of the Research**:
- "What got us started is... Estimation in the linear panel event study design... ideas and concepts we put forth in that previous paper are implemented in this package."

3. **Dynamic Effects of Policy**:
- "Interested in learning the dynamic effect of a policy on some outcome... for example, the effect of the minimum wage on employment."

4. **Estimating Employment Changes**:
- "What we do is we're gonna track how employment changes around the change in the minimum wage."

5. **Causal Relationships**:
- "Economists have devised many different estimators to address this... together, there's more causal relationship that really thinks about the impact of the minimum wage on employment."

6. **Workhorse Estimator**:
- "The workhorse estimator in these models... compares how the employment changes in New Jersey around the change of policy and compare that to how it changes in other states."

7. **Control for Differences**:
- "It allows for something that's essentially an extrapolation of pret trends... estimators that try to account for that and adjust for these pret trends."

8. **Broad Application of the Package**:
- "The package was fairly broad in terms of what policies it covers... it compliments packages that are more specific to certain settings."

9. **Estimating Cohort-Specific Effects**:
- "Under staggered adoption, XD event allows for estimation of these sort of cohort specific effects."

10. **Visualization Contribution**:
- "What the package will do is create a visualization of your results... in the form of an event study plot... several enhancements to these plots compared to what is currently standard practice."

11. **Normalization in Plots**:
- "The normalization we suggest is a good default... we’re gonna plot everything relative to the period before the policy is passed."

12. **Uniform Confidence Bands**:
- "We also allow plots of uniform confidence bands... help you get, depict basically what kind of treatment paths are plausible."

13. **Pre-Trends Test**:
- "There's a test for pret trends... suggestive of some specification and maybe you ought to rethink your research design or modeling choice."

14. **Conclusions on XD Event Package**:
- "XD event is a versatile package that includes many competing estimators in this linear panel events design... cover many situations."

These quotes encapsulate the main functionalities, goals, and methodologies of the XD Events Data Package and its significance in the convideo of econometric research on policy impacts.

X-TIMESTAMP-MAP=LOCAL:00:00:00.000,MPEGTS:90000

1
00:00:00.025 --> 00:00:02.005
The XD events data package,

2
00:00:02.025 --> 00:00:04.165
and it's accompanying note published in the data

3
00:00:04.165 --> 00:00:05.765
journal really do two things.

4
00:00:06.495 --> 00:00:08.485
First, it enables the estimation

5
00:00:08.485 --> 00:00:11.965
of linear panel models in a fairly broad class of settings.

6
00:00:12.785 --> 00:00:15.805
And second, it visualize the, the corresponding results,

7
00:00:15.805 --> 00:00:17.525
in the form of a event study plot,

8
00:00:18.025 --> 00:00:21.645
and includes several enhancements to these plots compared to

9
00:00:21.645 --> 00:00:23.285
what is currently a standard practice.

10
00:00:24.025 --> 00:00:26.205
And so what what got us started with is

11
00:00:26.205 --> 00:00:29.085
what give you some background for the last World Congress,

12
00:00:29.305 --> 00:00:30.605
of the Econometric Society.

13
00:00:30.785 --> 00:00:33.525
We wrote an article called Visualization Identification

14
00:00:33.525 --> 00:00:36.485
and Estimation in the linear panel event study design.

15
00:00:37.425 --> 00:00:39.285
And that was published in the corresponding issue

16
00:00:39.285 --> 00:00:40.485
of advanced economic,

17
00:00:40.485 --> 00:00:42.285
advances in economics and econometrics.

18
00:00:42.665 --> 00:00:45.805
And a lot of the ideas and conte concepts we put forth in

19
00:00:45.805 --> 00:00:48.125
that previous paper are implemented in this package.

20
00:00:48.785 --> 00:00:50.925
And so in the companion article to this, to this package,

21
00:00:51.025 --> 00:00:52.565
we describe that implementation in more

22
00:00:52.565 --> 00:00:54.245
detail compared to our previous work.

23
00:00:55.385 --> 00:00:58.165
Um, so when we first start with the first part of the paper,

24
00:00:58.165 --> 00:01:00.405
which is about estimation of, of these models.

25
00:01:02.025 --> 00:01:05.085
And so the setting that we're interested in here is we're

26
00:01:05.085 --> 00:01:06.725
interested in learning the dynamic effect

27
00:01:06.725 --> 00:01:08.285
of a policy on some outcome.

28
00:01:08.905 --> 00:01:11.365
So what we have is at our disposal is some panel dataset.

29
00:01:11.625 --> 00:01:13.885
So to give you a concrete example,

30
00:01:14.345 --> 00:01:15.805
we might be interested in the effect

31
00:01:15.865 --> 00:01:17.565
of the minimum wage on employment.

32
00:01:18.025 --> 00:01:19.925
So we'd like to learn what happens,

33
00:01:19.925 --> 00:01:22.045
to employment when we change the level of the minimum wage,

34
00:01:22.045 --> 00:01:23.085
when we change the minimum wage.

35
00:01:24.065 --> 00:01:25.445
And suppose to answer that question,

36
00:01:25.445 --> 00:01:27.045
we have data on both minimum wage levels

37
00:01:27.145 --> 00:01:28.885
and employment numbers across the United States.

38
00:01:28.985 --> 00:01:32.325
So sort of at the state level say, to, to,

39
00:01:32.325 --> 00:01:34.085
to estimate that that model, to learn

40
00:01:34.085 --> 00:01:35.725
that in fact we might write down a,

41
00:01:35.805 --> 00:01:37.605
a linear panel, model.

42
00:01:38.505 --> 00:01:39.885
And many researchers have done this.

43
00:01:39.885 --> 00:01:41.045
That's sort of the, the workforce

44
00:01:41.105 --> 00:01:42.205
way to estimate these models.

45
00:01:42.585 --> 00:01:44.765
And it's akin to an event study where roughly speaking,

46
00:01:44.765 --> 00:01:46.765
what we do is we're gonna track how employment changes

47
00:01:47.145 --> 00:01:48.565
around the change in the minimum wage.

48
00:01:48.585 --> 00:01:51.925
So to make it even more concrete, April 1st, 1992,

49
00:01:52.145 --> 00:01:53.645
New Jersey increased their minimum wage.

50
00:01:54.385 --> 00:01:56.925
And so what we would do is we're gonna see like what happens

51
00:01:56.925 --> 00:01:58.805
to employment in the following months, what happens in,

52
00:01:58.985 --> 00:02:02.975
you know, may, June, July, say, that's sort

53
00:02:02.975 --> 00:02:05.935
of the general idea that you might say, you know, in 1992,

54
00:02:05.995 --> 00:02:08.375
the entire US economy is coming out of a recession,

55
00:02:08.675 --> 00:02:09.975
so maybe they'll bias the results.

56
00:02:09.995 --> 00:02:11.255
In other words, you'll see sort

57
00:02:11.255 --> 00:02:13.055
of employment going up throughout that spring.

58
00:02:13.275 --> 00:02:15.695
But maybe that's not because of the change in minimum wage.

59
00:02:15.695 --> 00:02:17.895
That's really just because the entire US was in a good

60
00:02:17.895 --> 00:02:20.615
trajectory, so to speak, in terms of employment numbers.

61
00:02:21.235 --> 00:02:23.735
And so economists have devised many different estimators

62
00:02:23.875 --> 00:02:26.095
to address this and similar concerns together,

63
00:02:26.095 --> 00:02:27.295
there's more causal relationship

64
00:02:27.295 --> 00:02:28.375
that really thinks about the impact

65
00:02:28.375 --> 00:02:30.655
of the minimum wage, on employment.

66
00:02:31.635 --> 00:02:33.295
The workhorse estimator in these models,

67
00:02:33.455 --> 00:02:35.415
I think is generally speaking two way fixed effects.

68
00:02:35.875 --> 00:02:38.255
Uh, and what that does is, is essentially compares

69
00:02:38.315 --> 00:02:41.055
how the employment changes, in New Jersey

70
00:02:41.235 --> 00:02:43.135
around the change of policy

71
00:02:43.195 --> 00:02:44.615
or around the, the policy implementation

72
00:02:44.955 --> 00:02:47.375
and compare that to how it changes in other states.

73
00:02:47.955 --> 00:02:49.655
So for example, you might say, okay,

74
00:02:49.655 --> 00:02:51.295
it was trending in one direction, New Jersey,

75
00:02:51.315 --> 00:02:53.375
but at the same time, how was it trending in Pennsylvania?

76
00:02:53.835 --> 00:02:55.375
And the only difference between these two is

77
00:02:55.375 --> 00:02:56.815
that New Jersey was, was subject

78
00:02:56.815 --> 00:02:58.495
to this change in, in the minimum wage.

79
00:02:59.305 --> 00:03:01.605
Uh, and in fact that is, if you're familiar with the card

80
00:03:01.605 --> 00:03:03.245
and Krueger paper, that's a very famous paper

81
00:03:03.245 --> 00:03:05.125
that eventually won David Carter's Nobel Prize.

82
00:03:05.275 --> 00:03:06.765
That is precisely the,

83
00:03:06.765 --> 00:03:08.085
the experiment they run in that paper.

84
00:03:08.785 --> 00:03:11.525
And so the package, enables estimation

85
00:03:11.525 --> 00:03:13.925
of these type models, but also more complicated estimators

86
00:03:13.925 --> 00:03:15.525
that go beyond these simple comparisons.

87
00:03:16.185 --> 00:03:18.485
Um, so these include things like controlling

88
00:03:18.485 --> 00:03:19.765
for unit specific trends.

89
00:03:19.865 --> 00:03:22.045
So if you think New Jersey was in a different trajectory

90
00:03:22.265 --> 00:03:23.565
for some other reason altogether

91
00:03:23.565 --> 00:03:25.325
that spring, you can take that out.

92
00:03:26.065 --> 00:03:27.405
Uh, it also allows for something

93
00:03:27.405 --> 00:03:29.485
that's essentially an extrapolation of pret trends.

94
00:03:29.625 --> 00:03:31.925
So there's a large literature that that sort of thinks about

95
00:03:32.395 --> 00:03:35.085
what happens or what, what could be causing,

96
00:03:36.135 --> 00:03:38.235
drifts in the outcome, in this case employment sort

97
00:03:38.235 --> 00:03:39.515
of prior to the, to the policy.

98
00:03:39.645 --> 00:03:41.315
Maybe there was already something already happening

99
00:03:41.315 --> 00:03:42.435
before the policy was passed.

100
00:03:42.975 --> 00:03:45.075
And so we have, there's estimators that try to account for

101
00:03:45.075 --> 00:03:46.915
that and adjust, adjust for these pret trends,

102
00:03:46.915 --> 00:03:48.635
essentially controlling for event time trends.

103
00:03:49.935 --> 00:03:52.035
The package also includes various IV estimators.

104
00:03:52.035 --> 00:03:53.995
So it's sort of, it's pretty broad,

105
00:03:54.135 --> 00:03:55.995
and includes many alternative estimators

106
00:03:55.995 --> 00:03:57.275
that are available in these settings.

107
00:03:58.165 --> 00:03:59.635
Which estimator is appropriate

108
00:03:59.635 --> 00:04:00.715
will really depend on the context.

109
00:04:00.815 --> 00:04:02.435
So that's a question I think I need to stress

110
00:04:02.455 --> 00:04:03.755
as sort of an economic question.

111
00:04:04.095 --> 00:04:05.755
Uh, different identifying assumptions

112
00:04:05.905 --> 00:04:07.195
will lead to different estimators.

113
00:04:07.295 --> 00:04:09.795
So in previous work, we've written down,

114
00:04:09.885 --> 00:04:11.115
we've written a paper where we sort

115
00:04:11.115 --> 00:04:12.275
of think a little bit about that relationship.

116
00:04:12.295 --> 00:04:14.515
And so, so say, you know, depending on

117
00:04:15.065 --> 00:04:17.275
what identifying assumption you think is plausible in your

118
00:04:17.275 --> 00:04:18.915
given context, what estimator would

119
00:04:18.915 --> 00:04:20.635
that apply, to be reasonable?

120
00:04:21.935 --> 00:04:23.955
So by, in this paper here, by including

121
00:04:23.955 --> 00:04:26.915
or in this, in this, in this software package,

122
00:04:27.215 --> 00:04:29.395
by including many of these widely used estimators,

123
00:04:29.395 --> 00:04:30.875
we're hoping to cover many situations

124
00:04:31.515 --> 00:04:32.835
a researcher may encounter.

125
00:04:33.055 --> 00:04:35.395
And that sorts of many of the estimators a researcher may

126
00:04:35.395 --> 00:04:36.395
want to use,

127
00:04:36.705 --> 00:04:38.475
when considering to run these events studies.

128
00:04:39.255 --> 00:04:41.515
Uh, so instead of always using say, two-way fixed effects,

129
00:04:41.515 --> 00:04:44.515
which is a standard, you specify what estimator you want

130
00:04:44.515 --> 00:04:45.595
to use, as one of the

131
00:04:45.595 --> 00:04:46.755
options that you put into this package.

132
00:04:48.225 --> 00:04:50.555
Similarly, the package was fairly broad in terms of

133
00:04:50.555 --> 00:04:51.635
what policies it covers.

134
00:04:51.695 --> 00:04:53.475
So it's not the case that it only applies

135
00:04:53.495 --> 00:04:54.875
to what's called a staggered adoption.

136
00:04:54.875 --> 00:04:57.035
So in other words, we don't need the policy to be binary,

137
00:04:57.385 --> 00:04:59.635
they can be fairly generic policies.

138
00:05:00.175 --> 00:05:02.715
Um, and so in that's the thing, it compliments packages

139
00:05:02.715 --> 00:05:05.275
that are more specific to very, to certain settings,

140
00:05:05.495 --> 00:05:07.635
but being sort of fairly broad in that dimension as well.

141
00:05:10.115 --> 00:05:14.015
One final thing I wanna say on estimation is that

142
00:05:14.885 --> 00:05:17.135
there's sort of a large recent literature that thinks about

143
00:05:17.135 --> 00:05:19.335
what happens when treatment effects are heterogeneous

144
00:05:19.675 --> 00:05:21.735
by treatment time part in particular, what I mean by

145
00:05:21.735 --> 00:05:24.695
that is, folks in the literature use the word cohorts

146
00:05:24.935 --> 00:05:26.135
frequently where we say there's some states

147
00:05:26.135 --> 00:05:28.615
that are in early cohort, so they're treated early,

148
00:05:28.615 --> 00:05:30.655
they have sort of minimum wage changes early on in the

149
00:05:30.655 --> 00:05:32.135
sample, and they have some

150
00:05:32.325 --> 00:05:33.895
treatment effect associated with them.

151
00:05:33.915 --> 00:05:36.535
So the effect of the minimum wage is say, large.

152
00:05:36.795 --> 00:05:39.415
And then there's other states that pass the policy later on

153
00:05:39.415 --> 00:05:40.935
the sample where the effect may be small.

154
00:05:40.935 --> 00:05:42.895
So they sort of differ along that dimension.

155
00:05:44.145 --> 00:05:46.975
Under staggered adoption XT event allows for estimation

156
00:05:47.355 --> 00:05:49.055
of these sort of core specific effects.

157
00:05:49.075 --> 00:05:50.655
So you get a different treatment effect for the early

158
00:05:50.655 --> 00:05:51.895
and for the, for the late cohort.

159
00:05:52.515 --> 00:05:54.255
So it allows for heterogeneous treatment effects,

160
00:05:54.915 --> 00:05:56.895
and that's something that's not limited to

161
00:05:57.065 --> 00:05:58.485
Two FX effects estimation,

162
00:05:58.535 --> 00:06:00.125
which is the norm in the literature,

163
00:06:00.345 --> 00:06:03.525
but rather, you know, it provides a way to, create

164
00:06:03.545 --> 00:06:06.885
or to estimate to do heterogenei heterogeneity robust

165
00:06:06.885 --> 00:06:09.525
estimation for other estimators, such as some

166
00:06:09.525 --> 00:06:10.925
of the stuff I mentioned earlier, for example,

167
00:06:11.025 --> 00:06:14.125
as an estimator that extrapolates an existing pret trends.

168
00:06:14.505 --> 00:06:16.645
So that brings us to, visualization,

169
00:06:16.645 --> 00:06:18.525
which is the second contribution of this package.

170
00:06:18.825 --> 00:06:21.405
Um, once you've chose chosen your estimator,

171
00:06:21.795 --> 00:06:23.965
what the package will do is we'll create a visualization

172
00:06:23.965 --> 00:06:26.005
of your results, of your estimates,

173
00:06:26.005 --> 00:06:28.125
and it'll do so in the form of an event study plot.

174
00:06:28.985 --> 00:06:30.965
And for those plots, what the package does,

175
00:06:30.985 --> 00:06:34.525
it includes several enhancements to these plots compared to

176
00:06:34.525 --> 00:06:35.725
what is currently standard practice.

177
00:06:36.305 --> 00:06:37.685
Um, and really the idea is

178
00:06:37.685 --> 00:06:39.925
that these suggestions aim at improving the informativeness

179
00:06:39.925 --> 00:06:42.125
of these plots so they can be a little bit more,

180
00:06:42.625 --> 00:06:43.685
out, out of these plots.

181
00:06:44.595 --> 00:06:47.165
They generally involve a mix of sort of co codifying

182
00:06:47.165 --> 00:06:49.125
what we consider best practices

183
00:06:49.125 --> 00:06:50.845
and thinking about what are good defaults for some

184
00:06:50.845 --> 00:06:53.125
of these choices you have to make as well

185
00:06:53.125 --> 00:06:55.165
as suggesting some new things that are not currently,

186
00:06:55.225 --> 00:06:56.245
used in the literature.

187
00:06:57.885 --> 00:07:00.025
So the first thing I wanna talk about is, is normalization.

188
00:07:00.025 --> 00:07:01.145
Is the normalization here.

189
00:07:01.205 --> 00:07:03.665
So if you've ever worked with these models, you realize

190
00:07:03.665 --> 00:07:05.665
that you need to do some normalization.

191
00:07:05.725 --> 00:07:07.185
In other words, we have

192
00:07:07.185 --> 00:07:08.705
to choose one coefficient on our plot,

193
00:07:08.705 --> 00:07:09.865
and we have to set that equal to zero.

194
00:07:10.965 --> 00:07:13.705
And so the normalization we suggest is a good default here

195
00:07:13.705 --> 00:07:15.945
is to normalize the coefficient on the first

196
00:07:15.975 --> 00:07:17.145
lead of the policy variable.

197
00:07:17.285 --> 00:07:20.065
Now to put that into sort of more easy to understand terms,

198
00:07:20.325 --> 00:07:22.465
in the case of staggered adoption, what that means is

199
00:07:22.465 --> 00:07:26.625
that we're gonna, plot everything relative to the period

200
00:07:26.625 --> 00:07:27.945
before the policy is passed.

201
00:07:29.445 --> 00:07:31.825
Um, so I think we generally think that's a,

202
00:07:31.825 --> 00:07:33.425
that's a pretty good default in those settings.

203
00:07:33.855 --> 00:07:35.705
Further, to help with the interpretation

204
00:07:35.705 --> 00:07:39.225
of the estimated effects will also include in your figure a

205
00:07:39.225 --> 00:07:40.985
label for the normalized coefficient

206
00:07:40.985 --> 00:07:42.705
that reflects the mean of the dependent variable.

207
00:07:42.885 --> 00:07:45.145
So what that means is that, go back to an earlier example,

208
00:07:45.675 --> 00:07:47.925
there's gonna be sort of an, a number indicating

209
00:07:48.115 --> 00:07:49.765
what was the employment level right

210
00:07:49.765 --> 00:07:51.645
before the minimum wage change happened.

211
00:07:52.505 --> 00:07:54.845
And that's helpful because it helps you assess the magnitude

212
00:07:55.065 --> 00:07:56.565
of the effect into economic terms.

213
00:07:56.705 --> 00:07:59.205
So for example, if you estimate a thousand job losses,

214
00:07:59.955 --> 00:08:02.165
that number is economically very different if you are

215
00:08:02.165 --> 00:08:03.885
working with a sample that includes all of New Jersey.

216
00:08:03.905 --> 00:08:05.605
So you started with several millions of jobs

217
00:08:06.345 --> 00:08:07.565
rather than having a sample

218
00:08:07.565 --> 00:08:08.965
that just specific to one county.

219
00:08:09.145 --> 00:08:10.085
So you started with a

220
00:08:10.085 --> 00:08:11.325
few thousand jobs to start with, right?

221
00:08:11.605 --> 00:08:12.965
A thousand job losses relative

222
00:08:12.965 --> 00:08:15.605
to a few thousand initial jobs is very different in terms

223
00:08:15.885 --> 00:08:17.125
economic magnitude,

224
00:08:17.125 --> 00:08:18.045
than a thousand job losses

225
00:08:18.045 --> 00:08:19.205
when you start with a million jobs.

226
00:08:21.125 --> 00:08:24.105
The next thing we add to the, the plots is

227
00:08:25.805 --> 00:08:28.385
in addition to the standard point-wise confidence intervals,

228
00:08:28.385 --> 00:08:30.025
you might see for all coefficients.

229
00:08:30.285 --> 00:08:33.225
Uh, on the event part, we also allow plots

230
00:08:33.225 --> 00:08:34.545
of uniform confidence bands.

231
00:08:34.545 --> 00:08:36.545
So in particular we use soup t confidence bands.

232
00:08:37.365 --> 00:08:39.225
And what these allow you to do is they allow

233
00:08:39.225 --> 00:08:41.865
for visual tests of hypotheses about the entire coefficient

234
00:08:41.865 --> 00:08:43.825
path rather than a single coefficient.

235
00:08:43.825 --> 00:08:46.025
So the idea here is that, you know,

236
00:08:46.035 --> 00:08:47.745
since you estimated this dynamic model

237
00:08:47.745 --> 00:08:49.705
with all the estimated the entire path,

238
00:08:49.715 --> 00:08:52.145
presumably the entire dynamic treatment path is the object

239
00:08:52.145 --> 00:08:54.905
of interest, not the effect at any given,

240
00:08:54.965 --> 00:08:57.165
Single horizon and the sort

241
00:08:57.165 --> 00:08:59.885
of traditional point wise competence bands are not

242
00:08:59.885 --> 00:09:01.005
valid for that type of interest.

243
00:09:01.005 --> 00:09:02.845
They're valid really for these pre-specified

244
00:09:03.145 --> 00:09:04.725
single horizons.

245
00:09:05.105 --> 00:09:07.365
And so these, these uniform bands help you get,

246
00:09:07.865 --> 00:09:10.405
depict basically what kind of treatment paths are plausible

247
00:09:10.405 --> 00:09:11.925
where we talk about the entire object.

248
00:09:14.955 --> 00:09:16.575
Two other things we, we, we put out

249
00:09:16.575 --> 00:09:18.935
with these graphs are a couple of diagnostics.

250
00:09:18.935 --> 00:09:20.295
So there's a test for pret trends.

251
00:09:20.475 --> 00:09:22.775
So it's gonna test whether there's aerodynamics happening

252
00:09:22.775 --> 00:09:24.455
prior to the, prior to the policy

253
00:09:25.155 --> 00:09:27.695
and the test for whether the estimation window is too short

254
00:09:27.995 --> 00:09:30.135
to fully capture the dynamics in the treatment effect.

255
00:09:31.275 --> 00:09:33.735
And the idea is that if you find evidence of either of these

256
00:09:33.795 --> 00:09:35.735
or either sort of things going on before

257
00:09:36.155 --> 00:09:38.655
or things not actually flattening out towards the end

258
00:09:38.655 --> 00:09:40.375
of the graph, so things not becoming flat,

259
00:09:40.605 --> 00:09:42.375
that is suggestive of some specification

260
00:09:42.475 --> 00:09:45.095
and maybe you ought to rethink your, your research design

261
00:09:45.095 --> 00:09:46.175
or, or your modeling choice.

262
00:09:48.565 --> 00:09:50.785
And then one final thing, that we add

263
00:09:51.285 --> 00:09:54.905
to these plots is we're adding one additional path,

264
00:09:55.205 --> 00:09:56.465
to, to event study plot.

265
00:09:57.045 --> 00:09:59.845
And essentially this most smooth

266
00:09:59.905 --> 00:10:04.645
or least wiggly, path that we, that we add here depicts

267
00:10:05.305 --> 00:10:06.965
the, the, the most smooth path

268
00:10:06.965 --> 00:10:09.605
through throughout the entire plot, including the pre-PA

269
00:10:10.035 --> 00:10:11.645
that is not rejected by the data

270
00:10:11.905 --> 00:10:13.445
or not rejected by a joint wall test.

271
00:10:14.225 --> 00:10:16.485
And so the idea is to make, to go back

272
00:10:16.485 --> 00:10:18.565
to my running example, suppose we're looking at the

273
00:10:18.565 --> 00:10:20.005
evolution of employment from January

274
00:10:20.005 --> 00:10:22.685
through July if this minimum wage change happened on April

275
00:10:22.705 --> 00:10:25.325
one, and the idea is that there's a smooth pass,

276
00:10:25.345 --> 00:10:28.565
so things straight line where things start in January sort

277
00:10:28.565 --> 00:10:30.885
of trend upward and they just keep trending throughout the

278
00:10:30.885 --> 00:10:32.925
entire spring, then maybe

279
00:10:32.925 --> 00:10:34.125
something else is going on together.

280
00:10:34.125 --> 00:10:35.805
There are no policy facts, there's no effect

281
00:10:35.805 --> 00:10:36.765
that all the minimum wage policy

282
00:10:36.785 --> 00:10:37.965
or at least we can't rule that out.

283
00:10:37.965 --> 00:10:40.565
And there's just sort of another thing going on around

284
00:10:40.565 --> 00:10:42.405
that time that sort of smoothly evolves over time.

285
00:10:43.825 --> 00:10:46.165
On the other hand, if you try to come up with a sort

286
00:10:46.165 --> 00:10:47.965
of the smoothest path throughout this entire window

287
00:10:47.965 --> 00:10:49.565
and what you come up with is something that's sort

288
00:10:49.565 --> 00:10:52.285
of flat January, February, March,

289
00:10:52.625 --> 00:10:54.445
and then sort of complicated numbers kick in

290
00:10:54.445 --> 00:10:55.605
or jump kicks in even right

291
00:10:55.605 --> 00:10:58.325
after the policy takes effect, then maybe you can think that

292
00:10:58.325 --> 00:11:00.245
of as, as, as some additional evidence

293
00:11:00.435 --> 00:11:02.805
that the policy does indeed have an effect on the outcome in

294
00:11:02.805 --> 00:11:05.285
this case in wage, that it does indeed affect, affect the,

295
00:11:05.585 --> 00:11:07.445
the employment numbers.

296
00:11:09.585 --> 00:11:11.285
And that's pretty much it. So, you know, I think

297
00:11:11.285 --> 00:11:13.125
to conclude, I would say, you know, XD event is a,

298
00:11:13.125 --> 00:11:14.605
is a versatile package that includes

299
00:11:14.605 --> 00:11:15.765
many competing estimators.

300
00:11:15.985 --> 00:11:18.325
Uh, in the, in this linear panel events that I designed,

301
00:11:18.675 --> 00:11:21.245
it's generally enough to hopefully cover many situations.

302
00:11:21.425 --> 00:11:22.805
So it doesn't require second adoption,

303
00:11:22.805 --> 00:11:24.205
it also heterogeneous treatment effects.

304
00:11:24.985 --> 00:11:27.325
Um, and in terms of the output, in terms of the, the,

305
00:11:27.545 --> 00:11:29.445
the figure you, you get back from it,

306
00:11:29.625 --> 00:11:31.445
it includes a slightly richer event study plot

307
00:11:31.445 --> 00:11:32.205
that we think is more

308
00:11:32.205 --> 00:11:33.725
informative compared to what is currently the norm.

309
00:11:33.825 --> 00:11:36.085
So we hope that it provides a useful resource

310
00:11:36.085 --> 00:11:37.325
to make both estimation

311
00:11:37.325 --> 00:11:39.085
and visualization easier in this setting

312
00:11:39.265 --> 00:11:40.805
and perhaps a little bit more informative.

Enhancements in Linear Panel Model Estimation and Visualization

00:00:00 - The XD events data package, and it's accompanying note published in the data journal really do two things. First, it enables the estimation of linear panel models in a fairly broad class of settings. And second, it visualize the, the corresponding results, in the form of a event study plot, and includes several enhancements to these plots compared to what is currently a standard practice. And so what what got us started with is what give you some background for the last World Congress, of the Econometric Society.

00:00:33 - and Estimation in the linear panel event study design. And that was published in the corresponding issue of advanced economic, advances in economics and econometrics. And a lot of the ideas and conte concepts we put forth in that previous paper are implemented in this package. And so in the companion article to this, to this package, we describe that implementation in more detail compared to our previous work. so when we first start with the first part of the paper, which is about estimation of, of these models.

00:01:05 - interested in learning the dynamic effect of a policy on some outcome. So what we have is at our disposal is some panel dataset. So to give you a concrete example, we might be interested in the effect of the minimum wage on employment. So we'd like to learn what happens, to employment when we change the level of the minimum wage, when we change the minimum wage. And suppose to answer that question, we have data on both minimum wage levels and employment numbers across the United States. So sort of at the state level say, to, to,

Analyzing Employment Changes Post-Minimum Wage Increase

00:01:34 - that in fact we might write down a, a linear panel, model. And many researchers have done this. That's sort of the, the workforce way to estimate these models. And it's akin to an event study where roughly speaking, what we do is we're gonna track how employment changes around the change in the minimum wage. So to make it even more concrete, April 1st, 1992, New Jersey increased their minimum wage. And so what we would do is we're gonna see like what happens to employment in the following months, what happens in, you know, may, June, July, say, that's sort

00:02:05 - the entire US economy is coming out of a recession, so maybe they'll bias the results. In other words, you'll see sort of employment going up throughout that spring. But maybe that's not because of the change in minimum wage. That's really just because the entire US was in a good trajectory, so to speak, in terms of employment numbers. And so economists have devised many different estimators to address this and similar concerns together, there's more causal relationship that really thinks about the impact of the minimum wage, on employment. The workhorse estimator in these models,

00:02:35 - and what that does is, is essentially compares how the employment changes, in New Jersey around the change of policy or around the, the policy implementation and compare that to how it changes in other states. So for example, you might say, okay, it was trending in one direction, New Jersey, but at the same time, how was it trending in Pennsylvania? And the only difference between these two is that New Jersey was, was subject to this change in, in the minimum wage. and in fact that is, if you're familiar with the card and Krueger paper, that's a very famous paper that eventually won David Carter's Nobel Prize.

Versatile Estimation Methods for Policy Impact Analysis

00:03:06 - the experiment they run in that paper. And so the package, enables estimation of these type models, but also more complicated estimators that go beyond these simple comparisons. so these include things like controlling for unit specific trends. So if you think New Jersey was in a different trajectory for some other reason altogether that spring, you can take that out. it also allows for something that's essentially an extrapolation of pret trends. So there's a large literature that that sort of thinks about what happens or what, what could be causing,

00:03:38 - of prior to the, to the policy. Maybe there was already something already happening before the policy was passed. And so we have, there's estimators that try to account for that and adjust, adjust for these pret trends, essentially controlling for event time trends. The package also includes various IV estimators. So it's sort of, it's pretty broad, and includes many alternative estimators that are available in these settings. Which estimator is appropriate will really depend on the context. So that's a question I think I need to stress as sort of an economic question. different identifying assumptions will lead to different estimators.

00:04:09 - we've written a paper where we sort of think a little bit about that relationship. And so, so say, you know, depending on what identifying assumption you think is plausible in your given context, what estimator would that apply, to be reasonable? So by, in this paper here, by including or in this, in this, in this software package, by including many of these widely used estimators, we're hoping to cover many situations a researcher may encounter. And that sorts of many of the estimators a researcher may want to use, when considering to run these events studies.

Versatile Estimation with Cohort-Specific Effects

00:04:41 - which is a standard, you specify what estimator you want to use, as one of the options that you put into this package. Similarly, the package was fairly broad in terms of what policies it covers. So it's not the case that it only applies to what's called a staggered adoption. So in other words, we don't need the policy to be binary, they can be fairly generic policies. and so in that's the thing, it compliments packages that are more specific to very, to certain settings, but being sort of fairly broad in that dimension as well.

00:05:14 - there's sort of a large recent literature that thinks about what happens when treatment effects are heterogeneous by treatment time part in particular, what I mean by that is, folks in the literature use the word cohorts frequently where we say there's some states that are in early cohort, so they're treated early, they have sort of minimum wage changes early on in the sample, and they have some treatment effect associated with them. So the effect of the minimum wage is say, large. And then there's other states that pass the policy later on the sample where the effect may be small.

00:05:44 - Under staggered adoption XT event allows for estimation of these sort of core specific effects. So you get a different treatment effect for the early and for the, for the late cohort. So it allows for heterogeneous treatment effects, and that's something that's not limited to Two FX effects estimation, which is the norm in the literature, but rather, you know, it provides a way to, create or to estimate to do heterogenei heterogeneity robust estimation for other estimators, such as some of the stuff I mentioned earlier, for example,

Enhancements in Event Study Plot Visualization

00:06:14 - So that brings us to, visualization, which is the second contribution of this package. once you've chose chosen your estimator, what the package will do is we'll create a visualization of your results, of your estimates, and it'll do so in the form of an event study plot. And for those plots, what the package does, it includes several enhancements to these plots compared to what is currently standard practice. and really the idea is that these suggestions aim at improving the informativeness of these plots so they can be a little bit more,

00:06:44 - They generally involve a mix of sort of co codifying what we consider best practices and thinking about what are good defaults for some of these choices you have to make as well as suggesting some new things that are not currently, used in the literature. So the first thing I wanna talk about is, is normalization. Is the normalization here. So if you've ever worked with these models, you realize that you need to do some normalization. In other words, we have to choose one coefficient on our plot, and we have to set that equal to zero. And so the normalization we suggest is a good default here

00:07:15 - lead of the policy variable. Now to put that into sort of more easy to understand terms, in the case of staggered adoption, what that means is that we're gonna, plot everything relative to the period before the policy is passed. so I think we generally think that's a, that's a pretty good default in those settings. Further, to help with the interpretation of the estimated effects will also include in your figure a label for the normalized coefficient that reflects the mean of the dependent variable. So what that means is that, go back to an earlier example,

Assessing Job Loss Impact and Visualizing Confidence Intervals

00:07:48 - what was the employment level right before the minimum wage change happened. And that's helpful because it helps you assess the magnitude of the effect into economic terms. So for example, if you estimate a thousand job losses, that number is economically very different if you are working with a sample that includes all of New Jersey. So you started with several millions of jobs rather than having a sample that just specific to one county. So you started with a few thousand jobs to start with, right? A thousand job losses relative to a few thousand initial jobs is very different in terms

00:08:17 - than a thousand job losses when you start with a million jobs. The next thing we add to the, the plots is in addition to the standard point-wise confidence intervals, you might see for all coefficients. on the event part, we also allow plots of uniform confidence bands. So in particular we use soup t confidence bands. And what these allow you to do is they allow for visual tests of hypotheses about the entire coefficient path rather than a single coefficient. So the idea here is that, you know,

00:08:47 - with all the estimated the entire path, presumably the entire dynamic treatment path is the object of interest, not the effect at any given, Single horizon and the sort of traditional point wise competence bands are not valid for that type of interest. They're valid really for these pre-specified single horizons. And so these, these uniform bands help you get, depict basically what kind of treatment paths are plausible where we talk about the entire object. Two other things we, we, we put out

Testing Pre-Trends and Estimation Dynamics

00:09:18 - So there's a test for pret trends. So it's gonna test whether there's aerodynamics happening prior to the, prior to the policy and the test for whether the estimation window is too short to fully capture the dynamics in the treatment effect. And the idea is that if you find evidence of either of these or either sort of things going on before or things not actually flattening out towards the end of the graph, so things not becoming flat, that is suggestive of some specification and maybe you ought to rethink your, your research design or, or your modeling choice.

00:09:51 - to these plots is we're adding one additional path, to, to event study plot. And essentially this most smooth or least wiggly, path that we, that we add here depicts the, the, the most smooth path through throughout the entire plot, including the pre-PA that is not rejected by the data or not rejected by a joint wall test. And so the idea is to make, to go back to my running example, suppose we're looking at the

00:10:20 - through July if this minimum wage change happened on April one, and the idea is that there's a smooth pass, so things straight line where things start in January sort of trend upward and they just keep trending throughout the entire spring, then maybe something else is going on together. There are no policy facts, there's no effect that all the minimum wage policy or at least we can't rule that out. And there's just sort of another thing going on around that time that sort of smoothly evolves over time. On the other hand, if you try to come up with a sort of the smoothest path throughout this entire window and what you come up with is something that's sort

Enhancing Policy Impact Analysis with XD Event Package

00:10:52 - and then sort of complicated numbers kick in or jump kicks in even right after the policy takes effect, then maybe you can think that of as, as, as some additional evidence that the policy does indeed have an effect on the outcome in this case in wage, that it does indeed affect, affect the, the employment numbers. And that's pretty much it. So, you know, I think to conclude, I would say, you know, XD event is a, is a versatile package that includes many competing estimators. in the, in this linear panel events that I designed, it's generally enough to hopefully cover many situations.

00:11:39 - it also heterogeneous treatment effects. and in terms of the output, in terms of the, the, the figure you, you get back from it, it includes a slightly richer event study plot that we think is more informative compared to what is currently the norm. So we hope that it provides a useful resource to make both estimation and visualization easier in this setting and perhaps a little bit more informative.

Faculti AI can make mistakes

Research

Policy

Analysis

Data

Originals

Explainers

Podcasts