0:00:00.400,0:00:03.050
从原因到结果的路径

0:00:03.050,0:00:05.539
既黑暗且危险的

0:00:05.539,0:00:08.900
但是计量经济学的武器非常强大

0:00:09.211,0:00:13.700
当目睹平行趋势时[br]我们掌握了双重差分法

0:00:14.450,0:00:16.850
♪ [] ♪

0:00:20.100,0:00:21.423
计量经济学大师在寻找

0:00:21.423,0:00:24.800
令人信服的[br]「其他条件不变的比较」

0:00:25.100,0:00:29.419
理想的对比是

0:00:29.419,0:00:30.600
看起来相似的处理组和对照组[br]形成对照

0:00:30.600,0:00:34.630
但有时这种可比性是难以捉摸的

0:00:34.630,0:00:36.805
在没有处理的情况下

0:00:36.805,0:00:40.100
当处理组及对照组类似地演变时

0:00:40.100,0:00:42.307
即使起点不同

0:00:42.307,0:00:44.900
也有望进行因果推断

0:00:45.600,0:00:48.400
针对平行演化的武器

0:00:48.664,0:00:50.886
大师说的「平行趋势」

0:00:50.886,0:00:53.233
叫做「双重差分法」…

0:00:53.233,0:00:54.420
- 双重差分法...

0:00:54.420,0:00:56.900
- ...或简称为DD[br]- 好的

0:00:56.900,0:00:59.987
- 现在让我们看看 DD

0:00:59.987,0:01:02.888
如何帮助我们了解美国历史上

0:01:02.888,0:01:04.370
最重要的经济事件之一

0:01:05.300,0:01:08.300
- 现在我们一起回顾大萧条的情況—

0:01:08.800,0:01:12.200
我国有史以来最严重的经济灾难

0:01:13.100,0:01:16.200
在 1933 年失业率达到 25%—

0:01:16.600,0:01:19.114
这是之前或之后从未见过的水平

0:01:19.473,0:01:22.100
数百万国民失去了家园或土地

0:01:22.600,0:01:24.737
自杀率飙升

0:01:24.737,0:01:26.766
贫困的家庭依靠施食处和面包生产线

0:01:26.766,0:01:28.155
来避免挨饿

0:01:29.400,0:01:31.589
- 经济学家们

0:01:31.589,0:01:34.000
就大萧条的原因展开了激烈的争论

0:01:34.000,0:01:36.983
然而，大多数经济学家都同意[br]这个难题的关键部分

0:01:36.983,0:01:39.458
是银行大规模倒闭

0:01:39.800,0:01:41.900
这是施行存款保险制度之前的年代

0:01:42.100,0:01:46.513
因此，如果银行破产[br]你的储蓄也会化为乌有

0:01:46.513,0:01:47.672
- 取消你的帐户？

0:01:47.672,0:01:48.892
- 对啊，我想取消我的帐户

0:01:48.892,0:01:50.616
我不会在这家银行留下一分钱

0:01:52.600,0:01:56.100
- 面对银行业危机，央行有一项选择

0:01:56.400,0:01:58.524
随意地放贷给陷入困境的银行

0:01:58.524,0:02:01.100
或者袖手旁观拒绝贷款

0:02:01.500,0:02:05.440
借钱给有困难的银行叫「易钱」

0:02:05.440,0:02:08.100
拒贷叫「紧钱」

0:02:10.200,0:02:12.872
- 货币学派的代表人物[br]米尔顿弗里德曼和安娜施瓦茨

0:02:12.872,0:02:14.882
把大萧条称为

0:02:14.882,0:02:16.350
「大收缩」

0:02:16.800,0:02:18.262
指责美联储

0:02:18.262,0:02:21.200
就国家摇摇欲坠的金融机构[br]实施紧缩政策

0:02:21.200,0:02:24.000
是一项错误的政策

0:02:24.360,0:02:25.743
他们争辩说

0:02:25.743,0:02:27.895
易钱可让许多银行继续营业

0:02:27.895,0:02:29.700
从而缩短大萧条的时期

0:02:30.400,0:02:32.110
但其他人不同意

0:02:32.110,0:02:33.769
如果银行因为其不明智的贷款决定

0:02:33.769,0:02:35.954
而资不抵债

0:02:35.954,0:02:38.900
那么救助只会鼓励更多的愚蠢行为

0:02:39.600,0:02:42.965
经济学家把这个问题称为「道德风险」

0:02:42.965,0:02:46.100
今天人们仍就救助和道德风险[br]继续进行辩论

0:02:46.500,0:02:48.599
如果金融巨头雷曼兄弟

0:02:48.599,0:02:51.500
被允许在大衰退前夕倒闭

0:02:52.000,0:02:54.703
在理想的世界里[br]我们将会通过对随机选择的地区

0:02:54.703,0:02:58.400
应用不同的美联储政策[br]来回答这个问题

0:02:59.000,0:03:00.250
但是通过使用双重差分法

0:03:00.250,0:03:02.119
来比较不同货币政策的跨领域趋势

0:03:02.119,0:03:06.300
我们仍然可以学到很多事情

0:03:10.810,0:03:12.522
- 这怎么可能呢？

0:03:12.522,0:03:15.623
所有美国银行不是实施[br]相同的美联储政策吗？

0:03:15.623,0:03:17.400
- 对啊[br]- 好问题

0:03:17.700,0:03:21.484
联邦储备系统分为 12 个区

0:03:21.484,0:03:23.860
由12家地区性的联邦储备银行组成

0:03:24.301,0:03:27.467
今天，美联储政策是在国家层面制定的

0:03:27.467,0:03:31.973
但在 1930 年代，地区性的[br]联邦储备银行几乎可以随心所欲

0:03:31.973,0:03:33.276
- 啊，真有趣

0:03:33.276,0:03:35.500
- 这就是最棒的地方

0:03:35.500,0:03:39.302
在1930 年代，管理第六区的[br]亚特兰大联邦储备银行

0:03:39.302,0:03:41.473
遵循「易钱」政策

0:03:41.473,0:03:45.400
用手推车运送现金[br]去拯救破产的金融机构

0:03:45.900,0:03:48.816
而管理第八区的[br]圣路易斯联邦储备银行

0:03:48.816,0:03:50.668
则采取了紧缩的资金政策

0:03:51.100,0:03:53.900
「让愚蠢的人倒下吧！」[br]他们在圣路易斯说

0:03:54.300,0:03:58.225
因此，货币政策的自然实验诞生了

0:03:58.701,0:04:02.000
更好的是，这是州内的实验

0:04:02.000,0:04:04.347
第 6 区和第 8 区之间的边界

0:04:04.347,0:04:06.762
穿过密西西比州中部

0:04:07.300,0:04:09.321
密西西比北部实施紧钱政策

0:04:09.321,0:04:11.870
而密西西比南部则实施易钱政策

0:04:11.870,0:04:15.200
但是两个地区施行相同的[br]州法律和银行法规

0:04:15.520,0:04:16.853
- 密西西比州第 6 区

0:04:16.853,0:04:19.985
是处理组

0:04:19.985,0:04:23.100
在金融危机期间可以获得「易钱」

0:04:23.800,0:04:25.091
密西西比州第 8 区

0:04:25.091,0:04:27.800
是对照组

0:04:27.800,0:04:30.225
在金融危机期间实施「紧钱」政策

0:04:31.300,0:04:34.040
1930 年是

0:04:34.040,0:04:35.400
自然实验的关键年份

0:04:35.800,0:04:37.439
考德威尔公司

0:04:37.439,0:04:40.377
是一个位于南方的庞大金融帝国

0:04:40.377,0:04:41.987
垮台了

0:04:42.700,0:04:46.000
银行业建立在信心和信任的基础上

0:04:46.500,0:04:49.151
考德威尔的垮台引发了恐慌

0:04:49.151,0:04:53.049
一下子导致了大规模的银行挤兑

0:04:53.049,0:04:55.114
存款人想拿回他们的钱

0:04:55.114,0:04:58.200
导致银行破产并关门大吉

0:05:01.000,0:05:03.321
我们将会使用双重差分法

0:05:03.321,0:05:06.614
来衡量相反的货币政策的影响

0:05:06.614,0:05:09.164
以应对考德威尔危机

0:05:12.500,0:05:16.279
这幅图按年份绘制了

0:05:16.279,0:05:18.762
密西西比州第8区和第6区的[br]银行数目

0:05:19.273,0:05:21.413
我们从 1929 年开始

0:05:21.413,0:05:24.200
这是考德威尔垮台前一年

0:05:24.200,0:05:27.565
在8区有169家银行开业

0:05:27.565,0:05:31.242
而在6区有141家银行开业

0:05:31.242,0:05:32.574
在接下来的一年里

0:05:32.574,0:05:37.000
我们在两个地区都看到[br]类似的少数银行倒闭

0:05:37.400,0:05:40.135
运营银行数目的变化

0:05:40.135,0:05:42.168
非常相似—

0:05:42.168,0:05:44.600
这就是平行趋势的样子

0:05:45.500,0:05:48.783
于1930 年 11 月，考德威尔倒下

0:05:48.783,0:05:50.400
恐慌开始了

0:05:51.200,0:05:53.917
施行紧钱政策的 8 区银行

0:05:53.917,0:05:55.300
倒闭频发

0:05:55.800,0:05:58.712
但是施行易钱政策的6区银行

0:05:58.712,0:06:00.247
倒闭速度较慢

0:06:00.800,0:06:03.000
这一时期的分化趋势

0:06:03.000,0:06:06.602
可能归因于易钱与紧钱的分别

0:06:06.602,0:06:12.020
在1931年7月，8区放弃紧钱政策

0:06:12.020,0:06:14.420
在两个区都施行易钱政策

0:06:14.700,0:06:16.900
平行趋势得以恢复

0:06:17.300,0:06:19.229
在反事实的世界里

0:06:19.229,0:06:22.420
6 区银行亦施行紧钱政策的话

0:06:22.420,0:06:23.800
那会发生什么事呢？

0:06:24.201,0:06:28.547
如果我们将第 8 区的趋势[br]推断到第 6 区

0:06:28.547,0:06:29.900
看起来就像这样子

0:06:30.290,0:06:33.095
因此有效的「易钱」政策

0:06:33.095,0:06:36.462
使 6 区偏离 8 区趋势

0:06:36.462,0:06:38.900
所默示的路径

0:06:41.244,0:06:44.475
「易钱」政策拯救了多少家银行呢？

0:06:44.475,0:06:49.021
这个表的第一行报告了

0:06:49.021,0:06:50.162
第 6 区处理组的数据

0:06:50.162,0:06:54.237
第二行报告了第 8 区对照组的数据

0:06:54.237,0:06:57.494
第一列显示了 1930 年危机开始之前

0:06:57.494,0:07:00.613
营运银行的数目

0:07:00.613,0:07:03.673
第二列显示 1931 年的数目

0:07:03.903,0:07:05.607
这是每个地区在危机期间

0:07:05.607,0:07:08.663
采取不同货币政策时的

0:07:08.663,0:07:10.237
关键时期

0:07:10.237,0:07:14.347
最右边的一栏报告了区内的变化

0:07:14.347,0:07:20.456
第 6 区有 14 家银行倒下[br]而第 8 区有 33 家银行倒下

0:07:20.456,0:07:23.973
政策效果的数学公式很简单

0:07:24.400,0:07:28.430
我们从第 6 区运营银行的变化中

0:07:28.810,0:07:32.200
减去第 8 区运营银行的变化

0:07:32.800,0:07:35.700
因而得名「双重差分法」

0:07:37.000,0:07:41.736
-14 minus -33 equals 19.\]

0:07:42.500,0:07:46.800
We estimate that 19 banks[br]were saved by easy money.

0:07:47.300,0:07:50.730
In practice, tables and figures[br]like those shown here

0:07:50.730,0:07:52.912
are the beginning[br]rather than the end

0:07:52.912,0:07:54.453
of a DD analysis.

0:07:55.400,0:07:57.100
The problem of how to gauge

0:07:57.100,0:08:00.047
the statistical significance[br]of DD estimates

0:08:00.047,0:08:02.264
turns out to be exceedingly tricky,

0:08:02.264,0:08:05.600
and a regression is typically[br]part of the solution.

0:08:09.200,0:08:12.336
The key assumption[br]behind a valid DD analysis

0:08:12.336,0:08:14.894
is that of parallel trends.

0:08:15.369,0:08:17.842
Recall the principle[br]of ceteris paribus.

0:08:17.842,0:08:20.922
Our ideal comparison[br]would have the two districts

0:08:20.922,0:08:24.023
experience an identical[br]business environment,

0:08:24.023,0:08:25.997
except for one factor:

0:08:25.997,0:08:27.916
easy or tight money.

0:08:29.200,0:08:32.348
Both districts would have[br]identical types of customers

0:08:32.348,0:08:35.300
who would go bankrupt[br]at exactly the same rate.

0:08:35.700,0:08:38.600
The skill of their employees[br]would be equal, and so on.

0:08:39.200,0:08:43.498
Perfect ceteris paribus comparisons[br]would allow us to clearly see

0:08:43.498,0:08:46.594
the causal effect[br]of different Fed policies.

0:08:46.594,0:08:49.019
In this case, that's not possible.

0:08:49.019,0:08:53.600
But the idea of parallel trends[br]is based on a similar concept.

0:08:53.600,0:08:57.364
If we see that the two regions[br]experience similar trends

0:08:57.364,0:08:59.654
in the number of banks over time,

0:08:59.654,0:09:01.294
in the absence of treatment,

0:09:01.294,0:09:04.407
we can assume[br]they are good comparisons.

0:09:04.407,0:09:07.434
We see that the two districts[br]move in parallel,

0:09:07.434,0:09:09.800
both before the crisis and after,

0:09:10.100,0:09:12.400
when they have the same Fed policy.

0:09:13.100,0:09:16.028
The only time the districts[br]behave differently

0:09:16.028,0:09:18.300
is when the Fed policy is different.

0:09:19.400,0:09:20.599
In view of this,

0:09:20.599,0:09:24.268
Fed policy is a likely cause[br]of diverging trends

0:09:24.268,0:09:26.700
from 1930 to 1931.

0:09:27.800,0:09:30.106
But we should also check[br]for other changes

0:09:30.106,0:09:31.703
unique to northern Mississippi.

0:09:31.703,0:09:33.200
- [Man] Huh?[br]- What do you mean?

0:09:33.500,0:09:35.396
- [Teacher] Imagine that bad tornadoes

0:09:35.396,0:09:39.100
hit northern but not[br]southern Mississippi in 1930.

0:09:39.600,0:09:41.950
These tornadoes devastate farms,

0:09:41.950,0:09:44.300
causing farmers[br]to default on loans,

0:09:44.551,0:09:46.800
which drives their banks[br]out of business.

0:09:47.400,0:09:49.438
Then the 6th and 8th districts

0:09:49.438,0:09:52.272
would differ in not one[br]but two ways:

0:09:52.700,0:09:55.047
Fed policy and weather.

0:09:55.047,0:09:58.219
And we'd have trouble[br]identifying Fed policy

0:09:58.219,0:10:01.590
as the causal factor[br]behind increased bank failures

0:10:01.590,0:10:02.600
in the 8th.

0:10:02.600,0:10:04.248
- [Man] Ceteris is not paribus.

0:10:07.200,0:10:09.014
- DD credibility lives or dies

0:10:09.014,0:10:10.980
with the claim that the only reason

0:10:10.980,0:10:13.794
northern Mississippi[br]was special in 1930

0:10:13.794,0:10:16.161
is differing regional Fed policy.

0:10:16.600,0:10:20.530
We're in DD heaven with strong,[br]visual evidence of parallel trend.

0:10:21.277,0:10:25.549
- In general, the first step[br]in evaluating whether to use DD

0:10:25.549,0:10:30.200
is usually this type of visual[br]confirmation of parallel trends

0:10:30.200,0:10:31.700
outside of the period,

0:10:31.700,0:10:34.784
when we expect to see[br]a treatment effect.

0:10:35.094,0:10:36.985
The treatment in our example

0:10:36.985,0:10:39.835
is easy money[br]in the face of bank failures.

0:10:40.500,0:10:45.000
Metrics masters use DD[br]to explore effects of many policies,

0:10:45.800,0:10:47.900
like the minimum legal drinking age,

0:10:48.500,0:10:52.200
and environmental changes,[br]like access to clean water.

0:10:52.867,0:10:54.200
In our next video,

0:10:54.500,0:10:57.035
we'll see an example[br]of how regression is used

0:10:57.035,0:10:59.200
to implement a DD approach.

0:11:00.800,0:11:02.183
- [Narrator] Are you a teacher?

0:11:02.183,0:11:05.788
Click to explore ways[br]to use these videos in class.

0:11:05.788,0:11:08.847
If you're a learner,[br]make sure this video sticks

0:11:08.847,0:11:11.200
by taking a few quick[br]practice questions.

0:11:11.600,0:11:14.200
Or if you're ready,[br]click for the next video.

0:11:14.600,0:11:17.093
You can also check out[br]MRU's website

0:11:17.093,0:11:20.193
for more courses,[br]teacher resources, and more.

0:11:20.193,0:11:21.692
♪ [music] ♪