WEBVTT

00:00:00.400 --> 00:00:03.050
从原因到结果的路径

00:00:03.050 --> 00:00:05.539
既黑暗且危险的

00:00:05.539 --> 00:00:08.900
但是计量经济学的武器非常强大

00:00:09.211 --> 00:00:13.700
当目睹平行趋势时
我们掌握了双重差分法

00:00:14.450 --> 00:00:16.850
♪ [] ♪

00:00:20.100 --> 00:00:21.423
计量经济学大师在寻找

00:00:21.423 --> 00:00:24.800
令人信服的
「其他条件不变的比较」

00:00:25.100 --> 00:00:29.419
理想的对比是

00:00:29.419 --> 00:00:30.600
看起来相似的处理组和对照组
形成对照

00:00:30.600 --> 00:00:34.630
但有时这种可比性是难以捉摸的

00:00:34.630 --> 00:00:36.805
在没有处理的情况下

00:00:36.805 --> 00:00:40.100
当处理组及对照组类似地演变时

00:00:40.100 --> 00:00:42.307
即使起点不同

00:00:42.307 --> 00:00:44.900
也有望进行因果推断

00:00:45.600 --> 00:00:48.400
针对平行演化的武器

00:00:48.664 --> 00:00:50.886
大师说的「平行趋势」

00:00:50.886 --> 00:00:53.233
叫做「双重差分法」…

00:00:53.233 --> 00:00:54.420
- 双重差分法...

00:00:54.420 --> 00:00:56.900
- ...或简称为DD
- 好的

00:00:56.900 --> 00:00:59.987
- 现在让我们看看 DD

00:00:59.987 --> 00:01:02.888
如何帮助我们了解美国历史上

00:01:02.888 --> 00:01:04.370
最重要的经济事件之一

00:01:05.300 --> 00:01:08.300
- 现在我们一起回顾大萧条的情況—

00:01:08.800 --> 00:01:12.200
我国有史以来最严重的经济灾难

00:01:13.100 --> 00:01:16.200
在 1933 年失业率达到 25%—

00:01:16.600 --> 00:01:19.114
这是之前或之后从未见过的水平

00:01:19.473 --> 00:01:22.100
数百万国民失去了家园或土地

00:01:22.600 --> 00:01:24.737
自杀率飙升

00:01:24.737 --> 00:01:26.766
贫困的家庭依靠施食处和面包生产线

00:01:26.766 --> 00:01:28.155
来避免挨饿

00:01:29.400 --> 00:01:31.589
- 经济学家们

00:01:31.589 --> 00:01:34.000
就大萧条的原因展开了激烈的争论

00:01:34.000 --> 00:01:36.983
然而，大多数经济学家都同意
这个难题的关键部分

00:01:36.983 --> 00:01:39.458
是银行大规模倒闭

00:01:39.800 --> 00:01:41.900
这是施行存款保险制度之前的年代

00:01:42.100 --> 00:01:46.513
因此，如果银行破产
你的储蓄也会化为乌有

00:01:46.513 --> 00:01:47.672
- 取消你的帐户？

00:01:47.672 --> 00:01:48.892
- 对啊，我想取消我的帐户

00:01:48.892 --> 00:01:50.616
我不会在这家银行留下一分钱

00:01:52.600 --> 00:01:56.100
- 面对银行业危机，央行有一项选择

00:01:56.400 --> 00:01:58.524
随意地放贷给陷入困境的银行

00:01:58.524 --> 00:02:01.100
或者袖手旁观拒绝贷款

00:02:01.500 --> 00:02:05.440
借钱给有困难的银行叫「易钱」

00:02:05.440 --> 00:02:08.100
拒贷叫「紧钱」

00:02:10.200 --> 00:02:12.872
- 货币学派的代表人物
米尔顿弗里德曼和安娜施瓦茨

00:02:12.872 --> 00:02:14.882
把大萧条称为

00:02:14.882 --> 00:02:16.350
「大收缩」

00:02:16.800 --> 00:02:18.262
指责美联储

00:02:18.262 --> 00:02:21.200
就国家摇摇欲坠的金融机构
实施紧缩政策

00:02:21.200 --> 00:02:24.000
是一项错误的政策

00:02:24.360 --> 00:02:25.743
他们争辩说

00:02:25.743 --> 00:02:27.895
易钱可让许多银行继续营业

00:02:27.895 --> 00:02:29.700
从而缩短大萧条的时期

00:02:30.400 --> 00:02:32.110
但其他人不同意

00:02:32.110 --> 00:02:33.769
如果银行因为其不明智的贷款决定

00:02:33.769 --> 00:02:35.954
而资不抵债

00:02:35.954 --> 00:02:38.900
那么救助只会鼓励更多的愚蠢行为

00:02:39.600 --> 00:02:42.965
经济学家把这个问题称为「道德风险」

00:02:42.965 --> 00:02:46.100
今天人们仍就救助和道德风险
继续进行辩论

00:02:46.500 --> 00:02:48.599
如果金融巨头雷曼兄弟

00:02:48.599 --> 00:02:51.500
被允许在大衰退前夕倒闭

00:02:52.000 --> 00:02:54.703
在理想的世界里
我们将会通过对随机选择的地区

00:02:54.703 --> 00:02:58.400
应用不同的美联储政策
来回答这个问题

00:02:59.000 --> 00:03:00.250
但是通过使用双重差分法

00:03:00.250 --> 00:03:02.119
来比较不同货币政策的跨领域趋势

00:03:02.119 --> 00:03:06.300
我们仍然可以学到很多事情

00:03:10.810 --> 00:03:12.522
- 这怎么可能呢？

00:03:12.522 --> 00:03:15.623
所有美国银行不是实施
相同的美联储政策吗？

00:03:15.623 --> 00:03:17.400
- 对啊
- 好问题

00:03:17.700 --> 00:03:21.484
联邦储备系统分为 12 个区

00:03:21.484 --> 00:03:23.860
由12家地区性的联邦储备银行组成

00:03:24.301 --> 00:03:27.467
今天，美联储政策是在国家层面制定的

00:03:27.467 --> 00:03:31.973
但在 1930 年代，地区性的
联邦储备银行几乎可以随心所欲

00:03:31.973 --> 00:03:33.276
- 啊，真有趣

00:03:33.276 --> 00:03:35.500
- 这就是最棒的地方

00:03:35.500 --> 00:03:39.302
在1930 年代，管理第六区的
亚特兰大联邦储备银行

00:03:39.302 --> 00:03:41.473
遵循「易钱」政策

00:03:41.473 --> 00:03:45.400
用手推车运送现金
去拯救破产的金融机构

00:03:45.900 --> 00:03:48.816
而管理第八区的
圣路易斯联邦储备银行

00:03:48.816 --> 00:03:50.668
则采取了紧缩的资金政策

00:03:51.100 --> 00:03:53.900
「让愚蠢的人倒下吧！」
他们在圣路易斯说

00:03:54.300 --> 00:03:58.225
因此，货币政策的自然实验诞生了

00:03:58.701 --> 00:04:02.000
更好的是，这是州内的实验

00:04:02.000 --> 00:04:04.347
第 6 区和第 8 区之间的边界

00:04:04.347 --> 00:04:06.762
穿过密西西比州中部

00:04:07.300 --> 00:04:09.321
密西西比北部实施紧钱政策

00:04:09.321 --> 00:04:11.870
而密西西比南部则实施易钱政策

00:04:11.870 --> 00:04:15.200
但是两个地区施行相同的
州法律和银行法规

00:04:15.520 --> 00:04:16.853
- 密西西比州第 6 区

00:04:16.853 --> 00:04:19.985
是处理组

00:04:19.985 --> 00:04:23.100
在金融危机期间可以获得「易钱」

00:04:23.800 --> 00:04:25.091
密西西比州第 8 区

00:04:25.091 --> 00:04:27.800
是对照组

00:04:27.800 --> 00:04:30.225
在金融危机期间实施「紧钱」政策

00:04:31.300 --> 00:04:34.040
1930 年是

00:04:34.040 --> 00:04:35.400
自然实验的关键年份

00:04:35.800 --> 00:04:37.439
考德威尔公司

00:04:37.439 --> 00:04:40.377
是一个位于南方的庞大金融帝国

00:04:40.377 --> 00:04:41.987
垮台了

00:04:42.700 --> 00:04:46.000
银行业建立在信心和信任的基础上

00:04:46.500 --> 00:04:49.151
考德威尔的垮台引发了恐慌

00:04:49.151 --> 00:04:53.049
一下子导致了大规模的银行挤兑

00:04:53.049 --> 00:04:55.114
存款人想拿回他们的钱

00:04:55.114 --> 00:04:58.200
导致银行破产并关门大吉

00:05:01.000 --> 00:05:03.321
我们将会使用双重差分法

00:05:03.321 --> 00:05:06.614
来衡量相反的货币政策的影响

00:05:06.614 --> 00:05:09.164
以应对考德威尔危机

00:05:12.500 --> 00:05:16.279
这幅图按年份绘制了

00:05:16.279 --> 00:05:18.762
密西西比州第8区和第6区的
银行数目

00:05:19.273 --> 00:05:21.413
我们从 1929 年开始

00:05:21.413 --> 00:05:24.200
这是考德威尔垮台前一年

00:05:24.200 --> 00:05:27.565
在8区有169家银行开业

00:05:27.565 --> 00:05:31.242
而在6区有141家银行开业

00:05:31.242 --> 00:05:32.574
在接下来的一年里

00:05:32.574 --> 00:05:37.000
我们在两个地区都看到
类似的少数银行倒闭

00:05:37.400 --> 00:05:40.135
运营银行数目的变化

00:05:40.135 --> 00:05:42.168
非常相似—

00:05:42.168 --> 00:05:44.600
这就是平行趋势的样子

00:05:45.500 --> 00:05:48.783
于1930 年 11 月，考德威尔倒下

00:05:48.783 --> 00:05:50.400
恐慌开始了

00:05:51.200 --> 00:05:53.917
施行紧钱政策的 8 区银行

00:05:53.917 --> 00:05:55.300
倒闭频发

00:05:55.800 --> 00:05:58.712
但是施行易钱政策的6区银行

00:05:58.712 --> 00:06:00.247
倒闭速度较慢

00:06:00.800 --> 00:06:03.000
这一时期的分化趋势

00:06:03.000 --> 00:06:06.602
可能归因于易钱与紧钱的分别

00:06:06.602 --> 00:06:12.020
在1931年7月，8区放弃紧钱政策

00:06:12.020 --> 00:06:14.420
在两个区都施行易钱政策

00:06:14.700 --> 00:06:16.900
平行趋势得以恢复

00:06:17.300 --> 00:06:19.229
在反事实的世界里

00:06:19.229 --> 00:06:22.420
6 区银行亦施行紧钱政策的话

00:06:22.420 --> 00:06:23.800
那会发生什么事呢？

00:06:24.201 --> 00:06:28.547
如果我们将第 8 区的趋势
推断到第 6 区

00:06:28.547 --> 00:06:29.900
看起来就像这样子

00:06:30.290 --> 00:06:33.095
因此有效的「易钱」政策

00:06:33.095 --> 00:06:36.462
使 6 区偏离 8 区趋势

00:06:36.462 --> 00:06:38.900
所默示的路径

00:06:41.244 --> 00:06:44.475
「易钱」政策拯救了多少家银行呢？

00:06:44.475 --> 00:06:49.021
这个表的第一行报告了

00:06:49.021 --> 00:06:50.162
第 6 区处理组的数据

00:06:50.162 --> 00:06:54.237
第二行报告了第 8 区对照组的数据

00:06:54.237 --> 00:06:57.494
第一列显示了 1930 年危机开始之前

00:06:57.494 --> 00:07:00.613
营运银行的数目

00:07:00.613 --> 00:07:03.673
第二列显示 1931 年的数目

00:07:03.903 --> 00:07:05.607
这是每个地区在危机期间

00:07:05.607 --> 00:07:08.663
采取不同货币政策时的

00:07:08.663 --> 00:07:10.237
关键时期

00:07:10.237 --> 00:07:14.347
最右边的一栏报告了区内的变化

00:07:14.347 --> 00:07:20.456
第 6 区有 14 家银行倒下
而第 8 区有 33 家银行倒下

00:07:20.456 --> 00:07:23.973
政策效果的数学公式很简单

00:07:24.400 --> 00:07:28.430
我们从第 6 区运营银行的变化中

00:07:28.810 --> 00:07:32.200
减去第 8 区运营银行的变化

00:07:32.800 --> 00:07:35.700
因而得名「双重差分法」

00:07:37.000 --> 00:07:41.736
-14 minus -33 equals 19.\]

00:07:42.500 --> 00:07:46.800
We estimate that 19 banks
were saved by easy money.

00:07:47.300 --> 00:07:50.730
In practice, tables and figures
like those shown here

00:07:50.730 --> 00:07:52.912
are the beginning
rather than the end

00:07:52.912 --> 00:07:54.453
of a DD analysis.

00:07:55.400 --> 00:07:57.100
The problem of how to gauge

00:07:57.100 --> 00:08:00.047
the statistical significance
of DD estimates

00:08:00.047 --> 00:08:02.264
turns out to be exceedingly tricky,

00:08:02.264 --> 00:08:05.600
and a regression is typically
part of the solution.

00:08:09.200 --> 00:08:12.336
The key assumption
behind a valid DD analysis

00:08:12.336 --> 00:08:14.894
is that of parallel trends.

00:08:15.369 --> 00:08:17.842
Recall the principle
of ceteris paribus.

00:08:17.842 --> 00:08:20.922
Our ideal comparison
would have the two districts

00:08:20.922 --> 00:08:24.023
experience an identical
business environment,

00:08:24.023 --> 00:08:25.997
except for one factor:

00:08:25.997 --> 00:08:27.916
easy or tight money.

00:08:29.200 --> 00:08:32.348
Both districts would have
identical types of customers

00:08:32.348 --> 00:08:35.300
who would go bankrupt
at exactly the same rate.

00:08:35.700 --> 00:08:38.600
The skill of their employees
would be equal, and so on.

00:08:39.200 --> 00:08:43.498
Perfect ceteris paribus comparisons
would allow us to clearly see

00:08:43.498 --> 00:08:46.594
the causal effect
of different Fed policies.

00:08:46.594 --> 00:08:49.019
In this case, that's not possible.

00:08:49.019 --> 00:08:53.600
But the idea of parallel trends
is based on a similar concept.

00:08:53.600 --> 00:08:57.364
If we see that the two regions
experience similar trends

00:08:57.364 --> 00:08:59.654
in the number of banks over time,

00:08:59.654 --> 00:09:01.294
in the absence of treatment,

00:09:01.294 --> 00:09:04.407
we can assume
they are good comparisons.

00:09:04.407 --> 00:09:07.434
We see that the two districts
move in parallel,

00:09:07.434 --> 00:09:09.800
both before the crisis and after,

00:09:10.100 --> 00:09:12.400
when they have the same Fed policy.

00:09:13.100 --> 00:09:16.028
The only time the districts
behave differently

00:09:16.028 --> 00:09:18.300
is when the Fed policy is different.

00:09:19.400 --> 00:09:20.599
In view of this,

00:09:20.599 --> 00:09:24.268
Fed policy is a likely cause
of diverging trends

00:09:24.268 --> 00:09:26.700
from 1930 to 1931.

00:09:27.800 --> 00:09:30.106
But we should also check
for other changes

00:09:30.106 --> 00:09:31.703
unique to northern Mississippi.

00:09:31.703 --> 00:09:33.200
- [Man] Huh?
- What do you mean?

00:09:33.500 --> 00:09:35.396
- [Teacher] Imagine that bad tornadoes

00:09:35.396 --> 00:09:39.100
hit northern but not
southern Mississippi in 1930.

00:09:39.600 --> 00:09:41.950
These tornadoes devastate farms,

00:09:41.950 --> 00:09:44.300
causing farmers
to default on loans,

00:09:44.551 --> 00:09:46.800
which drives their banks
out of business.

00:09:47.400 --> 00:09:49.438
Then the 6th and 8th districts

00:09:49.438 --> 00:09:52.272
would differ in not one
but two ways:

00:09:52.700 --> 00:09:55.047
Fed policy and weather.

00:09:55.047 --> 00:09:58.219
And we'd have trouble
identifying Fed policy

00:09:58.219 --> 00:10:01.590
as the causal factor
behind increased bank failures

00:10:01.590 --> 00:10:02.600
in the 8th.

00:10:02.600 --> 00:10:04.248
- [Man] Ceteris is not paribus.

00:10:07.200 --> 00:10:09.014
- DD credibility lives or dies

00:10:09.014 --> 00:10:10.980
with the claim that the only reason

00:10:10.980 --> 00:10:13.794
northern Mississippi
was special in 1930

00:10:13.794 --> 00:10:16.161
is differing regional Fed policy.

00:10:16.600 --> 00:10:20.530
We're in DD heaven with strong,
visual evidence of parallel trend.

00:10:21.277 --> 00:10:25.549
- In general, the first step
in evaluating whether to use DD

00:10:25.549 --> 00:10:30.200
is usually this type of visual
confirmation of parallel trends

00:10:30.200 --> 00:10:31.700
outside of the period,

00:10:31.700 --> 00:10:34.784
when we expect to see
a treatment effect.

00:10:35.094 --> 00:10:36.985
The treatment in our example

00:10:36.985 --> 00:10:39.835
is easy money
in the face of bank failures.

00:10:40.500 --> 00:10:45.000
Metrics masters use DD
to explore effects of many policies,

00:10:45.800 --> 00:10:47.900
like the minimum legal drinking age,

00:10:48.500 --> 00:10:52.200
and environmental changes,
like access to clean water.

00:10:52.867 --> 00:10:54.200
In our next video,

00:10:54.500 --> 00:10:57.035
we'll see an example
of how regression is used

00:10:57.035 --> 00:10:59.200
to implement a DD approach.

00:11:00.800 --> 00:11:02.183
- [Narrator] Are you a teacher?

00:11:02.183 --> 00:11:05.788
Click to explore ways
to use these videos in class.

00:11:05.788 --> 00:11:08.847
If you're a learner,
make sure this video sticks

00:11:08.847 --> 00:11:11.200
by taking a few quick
practice questions.

00:11:11.600 --> 00:11:14.200
Or if you're ready,
click for the next video.

00:11:14.600 --> 00:11:17.093
You can also check out
MRU's website

00:11:17.093 --> 00:11:20.193
for more courses,
teacher resources, and more.

00:11:20.193 --> 00:11:21.692
♪ [music] ♪