WEBVTT

00:00:00.000 --> 00:00:02.550
♪ [音乐] ♪

00:00:03.800 --> 00:00:05.800
- [旁白] 欢迎收看“诺奖得主畅谈系列”

00:00:07.040 --> 00:00:08.100
在这一集中

00:00:08.100 --> 00:00:11.570
Josh Angrist 和 Guido Imbens
与 Isaiah Andrews

00:00:11.570 --> 00:00:14.600
将讨论计量经济学领域是如何发展的

00:00:16.100 --> 00:00:18.750
- [Isaiah] 那么，Guido 和 Josh

00:00:18.750 --> 00:00:21.500
你们都是开发经济学实证研究工具的先驱

00:00:21.500 --> 00:00:23.174
所以我想了解你对这个领域的发展方向的想法

00:00:23.174 --> 00:00:25.300


00:00:25.709 --> 00:00:28.079
经济学、计量经济学，整个领域

00:00:28.510 --> 00:00:31.302
首先，我很想听听你是否觉得

00:00:32.171 --> 00:00:35.200
局部平均处理效应框架的方式

00:00:35.200 --> 00:00:38.510


00:00:38.800 --> 00:00:42.187
对经济学中的新经验法如何发展和传播

00:00:42.187 --> 00:00:44.300
或它们应如何发展和传播有任何榜样

00:00:44.560 --> 00:00:45.960
- [Josh] 这是个好问题

00:00:46.610 --> 00:00:47.790
你先说吧

00:00:47.790 --> 00:00:49.240
[笑声]

00:00:49.700 --> 00:00:52.940
- [Guido] 是的，所以我认为重要的是

00:00:52.940 --> 00:00:58.550
要提出令人信服的案例

00:00:58.550 --> 00:01:02.207
其中问题清晰

00:01:02.400 --> 00:01:05.720
且方法普遍适用

00:01:06.253 --> 00:01:07.560
有一件事我...

00:01:08.192 --> 00:01:12.000
当我回顾较新的文献时

00:01:12.200 --> 00:01:16.700
我真的很喜欢回归不连续性文献

00:01:16.700 --> 00:01:19.670
其中显然有一堆非常有说服力的例子

00:01:19.670 --> 00:01:23.378
这让人们可以更清晰地思考

00:01:23.378 --> 00:01:27.200
更仔细地研究方法问题

00:01:27.400 --> 00:01:28.800
有着一些清晰的应用方法

00:01:28.800 --> 00:01:30.600
然后允许你思考：

00:01:30.600 --> 00:01:33.600
“哇，这些类型的假设在这里看起来合理吗？

00:01:33.600 --> 00:01:38.000
我们不喜欢早期论文中的哪些方面？

00:01:38.500 --> 00:01:39.802
我们如何改进这些方面？”

00:01:39.802 --> 00:01:44.210
因此，我认为有着明确的应用方法

00:01:44.210 --> 00:01:46.400
来激发这些文献是非常有帮助的

00:01:46.800 --> 00:01:48.050
- Guido，我很高兴
你提到了回归不连续性

00:01:48.050 --> 00:01:49.382


00:01:49.382 --> 00:01:53.300
我认为 IV 和 RD、

00:01:54.700 --> 00:01:57.060
工具变量和回归不连续性之间有很多互补性

00:02:00.506 --> 00:02:03.260
回归不连续性的许多计量经济学应用
曾经被称为“模糊” RD

00:02:03.260 --> 00:02:04.520


00:02:04.520 --> 00:02:07.230


00:02:07.230 --> 00:02:11.620
它在截止时不是离散的或确定的

00:02:11.620 --> 00:02:14.900
而只在速率或强度上有着变化

00:02:14.900 --> 00:02:17.737
LATE 框架帮助我们理解这些应用方法

00:02:17.737 --> 00:02:18.740


00:02:18.740 --> 00:02:21.140
并为我们提供了一个清晰的解释

00:02:21.140 --> 00:02:25.000
比如在我与 Victor Lavy 的论文中

00:02:25.000 --> 00:02:28.100
我们使用了 Maimonides 规则
班级规模截断

00:02:28.430 --> 00:02:30.030
那么你这里了解到了什么？

00:02:30.290 --> 00:02:31.820
当然，你可以用线性常数效应模型
来回答这个问题

00:02:31.820 --> 00:02:33.900


00:02:34.200 --> 00:02:36.310
但事实证明我们并不局限于此

00:02:36.310 --> 00:02:39.889
RD 仍然非常强大和有启发性

00:02:40.630 --> 00:02:43.092
即使在这种情况和类型规模下

00:02:43.092 --> 00:02:45.866
截止和感兴趣的变量之间的相关性是局部的

00:02:45.866 --> 00:02:49.133


00:02:49.133 --> 00:02:51.000
甚至可能不是那么强

00:02:52.000 --> 00:02:54.999
所以肯定有着平行发展

00:02:54.999 --> 00:02:56.400
这也很有趣...

00:02:57.253 --> 00:02:59.780
当我们在读研究生时，没有人谈论回归不连续性设计

00:02:59.780 --> 00:03:01.220


00:03:01.220 --> 00:03:02.843
这是其他社会科学家感兴趣的东西

00:03:02.843 --> 00:03:05.300


00:03:05.800 --> 00:03:09.507
它与 LATE 框架一起成长

00:03:09.507 --> 00:03:11.927
我们都曾基于这两种应用方式和方法进行工作

00:03:11.927 --> 00:03:14.565


00:03:14.565 --> 00:03:18.377
而且能看到它的发展并变得如此重要

00:03:18.377 --> 00:03:19.800
我感到非常兴奋

00:03:20.000 --> 00:03:21.767
我认为，这是朝着可靠的识别策略

00:03:21.767 --> 00:03:26.086
因果效应的普遍演变的一部分

00:03:26.086 --> 00:03:27.441


00:03:29.393 --> 00:03:30.642
使计量经济学更多地关注
因果问题而不是模型

00:03:30.642 --> 00:03:33.300


00:03:33.640 --> 00:03:34.650
就未来而言

00:03:34.650 --> 00:03:37.660
我认为 LATE 帮助促进的一件事是

00:03:37.660 --> 00:03:42.008
朝着更具创造性的随机试验迈进

00:03:42.008 --> 00:03:44.400
其中有着一些有趣的东西

00:03:45.500 --> 00:03:48.460
它不可能或直接能被简单地关闭或开启

00:03:48.460 --> 00:03:50.700


00:03:51.000 --> 00:03:54.584
但你可以鼓励或阻止它

00:03:54.584 --> 00:03:58.200
因此，例如，你能通过经济援助
来补贴学校教育

00:03:59.000 --> 00:04:02.080
所以现在我们有了
一个完整的框架来解释这一点

00:04:03.600 --> 00:04:07.113
它打开了对以前似乎不可能的事情
进行随机试验的大门

00:04:07.113 --> 00:04:09.265


00:04:10.300 --> 00:04:12.471


00:04:14.500 --> 00:04:17.864
我们在麻省理工学院
Blueprint实验室中

00:04:17.864 --> 00:04:21.160
大量使用了这一点

00:04:22.360 --> 00:04:26.600
我认为，我们正在以
非常有创意的方式利用随机分配

00:04:28.100 --> 00:04:31.395
- [Isaiah] 与此相关，你是否看到

00:04:31.395 --> 00:04:34.445
有助于计量经济学研究的特定因素？

00:04:34.445 --> 00:04:38.443
你已经提到它与
实际出现的问题有明确的联系

00:04:38.443 --> 00:04:40.300


00:04:40.300 --> 00:04:42.862
并且经验实践通常是一个好主意

00:04:43.290 --> 00:04:45.000
- 这不是一个好主意吗？

00:04:45.700 --> 00:04:50.112
我经常发现自己坐在
一个计量经济学理论研讨会上

00:04:50.700 --> 00:04:52.500
比如哈佛麻省理工学院的研讨会

00:04:53.400 --> 00:04:56.350
我会想，“这个人在解决什么问题？

00:04:56.350 --> 00:04:57.960
谁面对这个问题？”

00:04:57.960 --> 00:04:59.800
而且，你知道…

00:05:01.600 --> 00:05:04.700
如果我问的话
有时会出现令人尴尬的沉默

00:05:04.900 --> 00:05:08.300
或者可能会有一个相当牵强的场景

00:05:08.800 --> 00:05:11.600
我想看看这个工具在哪里有用

00:05:12.500 --> 00:05:14.765
有一些是纯粹的基础工具

00:05:14.765 --> 00:05:16.250
我确实明白这一点

00:05:16.250 --> 00:05:21.735
有些人正在研究的概念基础

00:05:22.600 --> 00:05:25.300
它变得更像数理统计

00:05:25.800 --> 00:05:27.653
我的意思是，我记得一个较早的例子

00:05:27.653 --> 00:05:29.920
那就是我很难理解的随机等连续性的概念

00:05:29.920 --> 00:05:32.500


00:05:32.500 --> 00:05:35.070
我的论文顾问之一，Whitney Newey

00:05:35.070 --> 00:05:36.479
使用它取得了很好的效果

00:05:36.479 --> 00:05:38.821
我当时试图理解这一点

00:05:40.600 --> 00:05:42.034
这真的很基础

00:05:42.034 --> 00:05:45.200
驱动它的不是一个应用方法

00:05:45.890 --> 00:05:47.300
至少不是立即的

00:05:48.600 --> 00:05:53.200
但是大多数事情不是这样的
所以应该有着问题

00:05:53.800 --> 00:05:59.247
我认为这取决于这类事情的卖家

00:06:00.480 --> 00:06:02.250
因为有着机会成本、时间和注意力

00:06:02.250 --> 00:06:05.295
以及理解事情的努力

00:06:05.980 --> 00:06:07.200
卖家要说：“嘿，我正在解决这个问题，这里有一组结果表明它很有用，这是我得到的一些见解。”

00:06:07.400 --> 00:06:08.900


00:06:09.400 --> 00:06:12.900


00:06:12.900 --> 00:06:15.200


00:06:16.097 --> 00:06:18.280
- [以赛亚]正如你所说，乔希，已经朝着更多地考虑经济学中的因果关系和经济学中的实证工作的方向发展。

00:06:18.280 --> 00:06:20.700


00:06:20.700 --> 00:06:22.900


00:06:22.900 --> 00:06:26.570
这种观点的传播有什么让你感到惊讶的后果，或者你认为是经验经济学发展方式的不利因素吗？

00:06:26.570 --> 00:06:28.347


00:06:28.705 --> 00:06:31.500


00:06:31.500 --> 00:06:34.322
- 有时我看到有人做静脉注射，他们得到的结果似乎难以置信地大——这是通常的情况。

00:06:34.322 --> 00:06:38.304


00:06:38.800 --> 00:06:40.200


00:06:42.500 --> 00:06:45.220
因此，这可能是一些相对较小的干预造成的非常大的因果影响，这些干预是随机的，或者您可以证明存在一个好的设计。

00:06:45.220 --> 00:06:48.600


00:06:49.100 --> 00:06:52.260


00:06:52.260 --> 00:06:54.490


00:06:54.900 --> 00:06:57.205
然后当我看到这一点时，我认为我很难相信这种相对较小的干预会产生如此大的影响。

00:06:57.944 --> 00:07:00.101


00:07:00.101 --> 00:07:02.030


00:07:02.030 --> 00:07:03.720


00:07:04.100 --> 00:07:06.277
作者有时会诉诸局部平均治疗效果定理并说，“嗯，这些编译器——他们在某些方面很特别，他们只是从这种干预中受益匪浅。”

00:07:06.277 --> 00:07:08.690


00:07:08.690 --> 00:07:11.066


00:07:11.066 --> 00:07:12.700


00:07:13.300 --> 00:07:15.800


00:07:15.800 --> 00:07:17.600


00:07:18.100 --> 00:07:21.175
我不愿意从表面上看。

00:07:21.175 --> 00:07:23.750
我经常认为，当影响太大时，这是因为排除限制失败了，所以你真的没有正确的内生变量来衡量这个结果。

00:07:24.300 --> 00:07:26.780


00:07:26.780 --> 00:07:29.456


00:07:29.456 --> 00:07:31.380


00:07:32.000 --> 00:07:35.700
因此，我不太高兴看到一个通用的异质性论点被用来为我认为可能是一个更深层次的问题的事情辩解。

00:07:36.937 --> 00:07:39.692


00:07:40.022 --> 00:07:41.760


00:07:41.760 --> 00:07:43.800


00:07:45.190 --> 00:07:47.358
- [Guido] 我认为它在简化形式和结构方法之间的讨论中扮演了一个不幸的角色，我觉得这不太正确。

00:07:47.358 --> 00:07:50.083


00:07:50.083 --> 00:07:51.700


00:07:51.700 --> 00:07:55.510


00:07:56.090 --> 00:07:58.810
工具变量假设是关于行为的核心结构假设——它们来自经济……考虑代理人的经济行为，不知何故它被推向了一个我认为不是很有帮助的方向。

00:07:58.810 --> 00:08:01.510


00:08:01.510 --> 00:08:03.483


00:08:03.483 --> 00:08:05.200


00:08:07.100 --> 00:08:09.900


00:08:10.300 --> 00:08:15.100


00:08:15.100 --> 00:08:17.600


00:08:20.426 --> 00:08:21.663
我认为，最初，我们把事情写下来，描述正在发生的事情。

00:08:22.800 --> 00:08:24.067


00:08:24.067 --> 00:08:26.480


00:08:26.480 --> 00:08:29.783
人们使用了一套方法。

00:08:29.783 --> 00:08:32.111
我们以一种我认为包含相当多洞察力的方式阐明了这些方法的作用。

00:08:32.811 --> 00:08:38.361


00:08:39.100 --> 00:08:42.050
但它不知何故被推到了一个我认为不一定很有帮助的角落。

00:08:42.050 --> 00:08:45.379


00:08:45.379 --> 00:08:48.604
- 在简化形式与结构的语言中，我觉得有点有趣，因为局部平均治疗效果模型，潜在结果模型是非参数结构模型，如果你想像你建议的那样考虑它，Guido .

00:08:48.604 --> 00:08:50.306


00:08:50.306 --> 00:08:52.985


00:08:52.985 --> 00:08:54.154


00:08:54.154 --> 00:08:56.110


00:08:56.110 --> 00:08:58.600


00:08:58.600 --> 00:09:01.129
所以当——

00:09:01.129 --> 00:09:03.505


00:09:03.505 --> 00:09:05.116
- [圭多] 是的。
- [Josh] 当然，这种语言来自我们继承的联立方程框架。

00:09:05.116 --> 00:09:08.371


00:09:08.371 --> 00:09:09.641


00:09:10.400 --> 00:09:11.440
它的优点是人们在你使用它时似乎知道你的意思，但这可能是不同的人听到不同的东西。

00:09:11.440 --> 00:09:14.085


00:09:14.085 --> 00:09:15.164


00:09:16.559 --> 00:09:18.300


00:09:18.300 --> 00:09:20.480
- [圭多] 是的。我认为简化形式已被用于贬义...

00:09:20.480 --> 00:09:22.200


00:09:22.200 --> 00:09:23.540
- [乔希] 有时。 ...这并不是它最初的目的。

00:09:25.104 --> 00:09:28.250


00:09:30.100 --> 00:09:33.090
- [以赛亚] 我想在思考局部平均治疗效果框架的影响时，还有一点让我印象深刻的是，人们通常会在远远超出任何形式的正式结果实际存在的环境中诉诸局部平均治疗效果直觉已确立的。

00:09:33.090 --> 00:09:35.645


00:09:35.645 --> 00:09:37.676


00:09:37.676 --> 00:09:40.000


00:09:40.000 --> 00:09:42.358


00:09:42.358 --> 00:09:44.963


00:09:45.440 --> 00:09:49.180
我很好奇，鉴于你们为在不同环境中建立 LATE 结果所做的所有工作，我很好奇，对此有什么想法吗？

00:09:49.180 --> 00:09:52.390


00:09:52.390 --> 00:09:54.415


00:09:55.360 --> 00:09:57.306
- 我认为在很多情况下直觉确实会让你保持距离，但它会受到一些限制，建立正式的结果可能会有点棘手，然后可能只在特殊情况下才有效，然后你就结束了有很多可能无法完全捕捉直觉的形式。

00:09:57.306 --> 00:10:02.200


00:10:02.800 --> 00:10:04.989


00:10:04.989 --> 00:10:07.600


00:10:08.400 --> 00:10:09.490


00:10:09.490 --> 00:10:12.700


00:10:14.600 --> 00:10:16.540


00:10:16.540 --> 00:10:19.500


00:10:19.900 --> 00:10:21.550
有时我对它们有些不安，它们不一定是我想写的论文，但我确实认为直觉通常确实抓住了问题的一部分。

00:10:21.550 --> 00:10:24.438


00:10:25.148 --> 00:10:27.819


00:10:27.819 --> 00:10:30.912


00:10:33.100 --> 00:10:36.300
我认为，从某种意义上说，我们很幸运，LATE 论文在期刊上得到了处理，所以，实际上，编辑把它缩短了很多，这让我们能够专注于非常清晰、清晰的结果。

00:10:36.900 --> 00:10:39.250


00:10:39.250 --> 00:10:41.766


00:10:42.100 --> 00:10:46.300


00:10:49.924 --> 00:10:51.770
计量经济学文献中有些不幸的趋势是论文越来越长。

00:10:51.770 --> 00:10:53.118


00:10:53.118 --> 00:10:54.870


00:10:54.870 --> 00:10:56.550
- 嗯，你应该能够解决这个问题，伙计。

00:10:56.550 --> 00:10:58.915
- 我正在努力解决这个问题。 [笑声]

00:10:58.915 --> 00:11:01.510
但我认为这是一个例子，很明显，它很短实际上是——

00:11:01.510 --> 00:11:03.097


00:11:03.097 --> 00:11:04.750
- 你应该强制规定没有任何论文可以比 LATE 论文更长。

00:11:04.750 --> 00:11:06.802


00:11:07.269 --> 00:11:09.617
- 那……哇！那可能很棒。

00:11:09.617 --> 00:11:11.685
- 至少没有理论，没有理论论文。

00:11:11.892 --> 00:11:13.339
- 是的，我认为......

00:11:14.500 --> 00:11:16.800
我正在努力让论文更短，我认为今天有很多价值，因为它通常是论文的第二部分实际上并没有让你得到太多进一步理解事物，它确实使事情变得更难阅读。

00:11:16.800 --> 00:11:19.514


00:11:19.514 --> 00:11:21.573


00:11:21.573 --> 00:11:25.049


00:11:25.049 --> 00:11:26.395


00:11:27.000 --> 00:11:29.870


00:11:32.426 --> 00:11:36.179
这可以追溯到我认为应该如何进行计量经济学。

00:11:36.179 --> 00:11:38.070


00:11:38.700 --> 00:11:41.300


00:11:41.500 --> 00:11:43.900


00:11:44.800 --> 00:11:48.900


00:11:48.900 --> 00:11:50.010


00:11:51.100 --> 00:11:54.670


00:11:56.260 --> 00:11:57.750


00:11:57.750 --> 00:12:00.230


00:12:00.230 --> 00:12:03.765


00:12:03.765 --> 00:12:04.994


00:12:04.994 --> 00:12:07.563


00:12:07.563 --> 00:12:09.840


00:12:09.840 --> 00:12:11.600


00:12:11.600 --> 00:12:13.950


00:12:14.600 --> 00:12:16.770


00:12:16.770 --> 00:12:21.000


00:12:21.800 --> 00:12:27.530


00:12:28.000 --> 00:12:29.300


00:12:29.300 --> 00:12:32.030


00:12:32.600 --> 00:12:35.500


00:12:35.500 --> 00:12:37.000


00:12:37.400 --> 00:12:42.362


00:12:43.228 --> 00:12:46.700


00:12:46.700 --> 00:12:49.250


00:12:49.250 --> 00:12:50.800


00:12:52.000 --> 00:12:54.250


00:12:54.250 --> 00:12:56.100


00:12:56.100 --> 00:12:59.736


00:12:59.736 --> 00:13:02.143


00:13:02.143 --> 00:13:04.290


00:13:04.800 --> 00:13:07.000


00:13:07.600 --> 00:13:09.501


00:13:09.920 --> 00:13:12.805


00:13:13.100 --> 00:13:15.400


00:13:15.900 --> 00:13:17.730


00:13:17.730 --> 00:13:19.231


00:13:19.231 --> 00:13:22.616


00:13:22.616 --> 00:13:25.100


00:13:25.230 --> 00:13:27.534


00:13:27.534 --> 00:13:30.900


00:13:31.700 --> 00:13:36.068


00:13:37.900 --> 00:13:41.190


00:13:41.190 --> 00:13:42.900


00:13:43.200 --> 00:13:47.280


00:13:47.280 --> 00:13:51.180


00:13:51.180 --> 00:13:53.272


00:13:54.906 --> 00:13:56.477


00:13:56.477 --> 00:13:59.300


00:13:59.643 --> 00:14:01.992


00:14:01.992 --> 00:14:05.657


00:14:05.657 --> 00:14:07.910


00:14:07.910 --> 00:14:09.557


00:14:10.100 --> 00:14:13.389


00:14:13.389 --> 00:14:14.688


00:14:16.300 --> 00:14:17.414


00:14:19.000 --> 00:14:21.700


00:14:22.120 --> 00:14:23.370


00:14:23.370 --> 00:14:26.780


00:14:27.200 --> 00:14:28.780


00:14:28.780 --> 00:14:32.836


00:14:33.370 --> 00:14:36.325


00:14:36.325 --> 00:14:37.807


00:14:37.807 --> 00:14:39.289


00:14:39.289 --> 00:14:42.100


00:14:44.846 --> 00:14:47.500


00:14:47.500 --> 00:14:49.467


00:14:49.467 --> 00:14:54.212


00:14:54.212 --> 00:14:56.920


00:14:56.920 --> 00:14:58.950


00:14:58.950 --> 00:15:01.140


00:15:01.140 --> 00:15:02.723


00:15:03.900 --> 00:15:07.630


00:15:07.630 --> 00:15:09.430


00:15:09.430 --> 00:15:12.730


00:15:12.730 --> 00:15:15.480


00:15:15.480 --> 00:15:16.800


00:15:17.100 --> 00:15:18.550


00:15:18.550 --> 00:15:20.900


00:15:22.104 --> 00:15:23.732


00:15:23.732 --> 00:15:26.783


00:15:26.783 --> 00:15:32.772


00:15:32.772 --> 00:15:35.595


00:15:35.595 --> 00:15:38.290


00:15:38.290 --> 00:15:40.300


00:15:40.300 --> 00:15:42.300


00:15:42.300 --> 00:15:44.400


00:15:44.400 --> 00:15:45.890


00:15:45.890 --> 00:15:48.200


00:15:49.100 --> 00:15:52.200


00:15:52.200 --> 00:15:53.675


00:15:53.675 --> 00:15:56.390


00:15:56.913 --> 00:15:59.400


00:15:59.980 --> 00:16:02.840


00:16:02.840 --> 00:16:05.112


00:16:05.112 --> 00:16:07.100


00:16:07.380 --> 00:16:08.900


00:16:10.200 --> 00:16:13.929


00:16:13.929 --> 00:16:15.950


00:16:15.950 --> 00:16:17.330


00:16:18.200 --> 00:16:21.241


00:16:21.241 --> 00:16:27.430


00:16:27.430 --> 00:16:31.600


00:16:31.600 --> 00:16:33.274


00:16:35.500 --> 00:16:38.600


00:16:38.600 --> 00:16:42.893


00:16:42.893 --> 00:16:46.330


00:16:46.330 --> 00:16:48.300


00:16:48.600 --> 00:16:51.500


00:16:52.718 --> 00:16:53.868


00:16:53.868 --> 00:16:56.920


00:16:56.920 --> 00:17:00.533


00:17:00.800 --> 00:17:03.240


00:17:03.240 --> 00:17:05.420


00:17:05.420 --> 00:17:07.600


00:17:07.600 --> 00:17:09.700


00:17:09.700 --> 00:17:13.370


00:17:15.110 --> 00:17:18.230


00:17:18.500 --> 00:17:22.900


00:17:23.400 --> 00:17:24.933


00:17:24.933 --> 00:17:27.570


00:17:27.570 --> 00:17:30.300


00:17:31.700 --> 00:17:35.204


00:17:35.204 --> 00:17:38.516


00:17:38.516 --> 00:17:43.400


00:17:44.600 --> 00:17:47.020


00:17:47.020 --> 00:17:48.200


00:17:48.200 --> 00:17:50.500


00:17:50.500 --> 00:17:53.100


00:17:53.700 --> 00:17:56.720


00:17:56.720 --> 00:17:58.300


00:17:59.036 --> 00:18:01.495