Return to Video

工具变量(IV)的简介

  • 0:00 - 0:05
    因果推断之路径既黑暗又危险
  • 0:05 - 0:08
    但是计量经济学是很厉害的武器
  • 0:08 - 0:12
    当自然界给你带来偶然的随机分配时
  • 0:12 - 0:16
    使用气势汹汹与灵活多變的
    工具变量进行攻击
  • 0:19 - 0:21
    []
  • 0:24 - 0:26
    随机试验是完成
    “其他条件不变”的比较
  • 0:26 - 0:29
    的最可靠途径
  • 0:29 - 0:33
    但我们经常无法使用
    这个功能强大的工具
  • 0:33 - 0:37
    但是有时候,随机是偶然发生的
  • 0:37 - 0:41
    这时候我们转向工具变量
  • 0:41 - 0:42
    —简称IV
  • 0:42 - 0:45
    工具变量
  • 0:45 - 0:48
    今天的课堂是IV两节课的第一节
  • 0:49 - 0:53
    我们的第一节IV课
    从学校的故事开始
  • 0:53 - 0:54
    []
  • 0:54 - 0:56
    特许学校是一些公立学校
  • 0:56 - 1:00
    不受日常学区监督
    与教师工会合同约束
  • 1:01 - 1:04
    特许学校能否提高成绩
  • 1:04 - 1:05
    是美国教育改革史上
  • 1:05 - 1:08
    最重要的问题之一
  • 1:08 - 1:13
    最受欢迎的特许学校的申请人数
    远多于学位
  • 1:13 - 1:16
    因此抽奖运决定了
    谁家孩子可获录取
  • 1:17 - 1:21
    在学生争夺机会时需要面对很多风险
  • 1:21 - 1:25
    正如获奖纪录片“等待超人”中
  • 1:25 - 1:28
    所描述的那样
  • 1:28 - 1:30
    等待结果时会产生很多种情绪
  • 1:30 - 1:33
    别哭,你会让妈妈哭的
    好吗?
  • 1:37 - 1:41
    特许学校真的能提供更好的教育吗?
  • 1:41 - 1:43
    评论家肯定会说"不是的"
  • 1:43 - 1:47
    他们会争辩说特许学校
    能夠招募更好
  • 1:47 - 1:50
    更聪明或更主动的学生
    因此以后结果的差异
  • 1:50 - 1:52
    反映了选择性偏差
  • 1:53 - 1:55
    等一下,这个似乎很容易
  • 1:55 - 1:58
    在抽奖活动中
    我们会随机选择优胜者
  • 1:58 - 2:00
    因此只比较赢家和输家
    很明显的
  • 2:00 - 2:02
    在正确的轨道上,卡马尔
  • 2:02 - 2:04
    但是特许学校的抽签安排
  • 2:04 - 2:08
    不会强迫孩子们进入
    或离开特定的学校
  • 2:08 - 2:11
    他们随机分配了特许学校的学位
  • 2:12 - 2:13
    有些孩子很幸运
  • 2:13 - 2:15
    有些孩子不是
  • 2:15 - 2:17
    如果我们只是想知道特许学校
  • 2:17 - 2:19
    所带来的影响
  • 2:19 - 2:22
    我们可以将其视为随机试验
  • 2:23 - 2:25
    但是,我们只对特许学校
    就学的影响
  • 2:25 - 2:27
    感兴趣
  • 2:27 - 2:28
    而对录取不感兴趣
  • 2:29 - 2:32
    并非所有获录取的学生
    都会接受学位
  • 2:32 - 2:37
    IV将被录取为特许学校学生的影响
  • 2:37 - 2:40
    转变为实际就读特许学校的影响
  • 2:40 - 2:42
    - 太酷了
    - 哦,太好了
  • 2:46 - 2:49
    让我们看一个例子
  • 2:49 - 2:52
    这是一所执行知识就是力量专案
    的特许学校,或简称为KIPP
  • 2:53 - 2:55
    这所KIPP特许学校位于林恩
  • 2:55 - 2:59
    一座位于麻省海边的
    褪色工业城镇
  • 2:59 - 3:02
    这所学校的申请者多于学位
  • 3:02 - 3:06
    因此他们要抽签来挑选学生
  • 3:06 - 3:12
    从2005年到2008年
    共有371名四年级以及五年级生
  • 3:12 - 3:15
    参加了KIPP林恩的抽签
  • 3:15 - 3:19
    当中253名学生KIPP获录取
  • 3:19 - 3:22
    118名学生没有录取
  • 3:22 - 3:26
    一年后,获录取者的数学分数
  • 3:26 - 3:28
    比未获录取者更高
  • 3:28 - 3:30
    我们并不是试图弄清楚
  • 3:30 - 3:34
    获录取后是否会提高
    你的数学水平
  • 3:34 - 3:38
    我们想知道参加KIPP
    是否会使你的数学成绩改进
  • 3:39 - 3:46
    在253位获录取者中
    实际上只有199位到KIPP上学
  • 3:46 - 3:49
    其他学生选择了传统的公立学校
  • 3:50 - 3:56
    同样,在118名未被录取的学生中
    事实上有一些最终参加了KIPP
  • 3:56 - 3:57
    他们后来也获录取
  • 3:57 - 4:00
    那么,实际上参加KIPP
  • 4:00 - 4:02
    对考试成绩有何影响呢?
  • 4:03 - 4:05
    为什么我们不能只衡量
    他们的数学成绩?
  • 4:06 - 4:07
    这是很好的问题
  • 4:07 - 4:09
    你将他们与谁进行比较呢?
  • 4:09 - 4:11
    那些没有参加的学生
  • 4:11 - 4:13
    上学率是随机的吗?
  • 4:14 - 4:16
    - 不是啊
    - 选择性偏差
  • 4:16 - 4:18
    - 对啊
    - 什么?
  • 4:18 - 4:22
    KIPP的录取是随机的,因此我们
    对“其他条件不变”的假设充满信心
  • 4:22 - 4:26
    但上学率不是随机的
  • 4:27 - 4:31
    选择接受录取通知
  • 4:31 - 4:33
    可能是与数学成绩有关的特征
  • 4:33 - 4:36
    例如,有奉献精神的父母
  • 4:36 - 4:39
    更有可能接受录取
  • 4:39 - 4:43
    无论上那间学校
  • 4:43 - 4:44
    他们的孩子的数学成绩
    也有可能更好
  • 4:44 - 4:45
    对啊
  • 4:45 - 4:48
    IV将录取的影响
  • 4:48 - 4:51
    转化为KIPP上学率的影响
  • 4:51 - 4:53
    并就一些获录取者到其他学校上学
  • 4:53 - 4:57
    而一些未被录取者还是设法
    参加了KIPP 而进行调整
  • 4:57 - 5:01
    本质上,IV需要进行不完全的随机化
  • 5:01 - 5:03
    并进行适当的调整
  • 5:04 - 5:07
    怎么样? IV描述了一种连锁反应
  • 5:08 - 5:10
    为什么学校的录取会影响成绩?
  • 5:10 - 5:13
    可能是因为这影响了
    特许学校的上学率
  • 5:13 - 5:17
    而特许学校的上学率
    提高了数学成绩
  • 5:17 - 5:21
    连锁反应的第一个环节
    称之为“第一阶段”
  • 5:21 - 5:24
    是抽签对特许学校上学率的影响
  • 5:24 - 5:28
    第二阶段是在特许学校学
  • 5:28 - 5:30
    以及结果变量之间的关联
  • 5:30 - 5:32
    在这情况下,数学分数
  • 5:33 - 5:36
    工具变量或简称为“工具”
  • 5:36 - 5:40
    是启动链式反应的变量
  • 5:41 - 5:44
    工具变量对结果的影响
  • 5:44 - 5:47
    称为简化式
  • 5:48 - 5:52
    这个链式反应可以用数学表示
  • 5:52 - 5:54
    我们乘以第一阶段
  • 5:54 - 5:56
    即录取者对上学率的影响
  • 5:56 - 5:58
    到第二阶段
  • 5:58 - 6:01
    上学率对分数的影响
  • 6:01 - 6:03
    我们得到简化式
  • 6:03 - 6:06
    获录取对分数的影响
  • 6:07 - 6:12
    简化式和第一阶段是可观察的
    并且易于计算
  • 6:12 - 6:15
    但是,上学率对成绩的影响
  • 6:15 - 6:17
    并未能直接观察到
  • 6:17 - 6:20
    这是我们试图确定的因果关系
  • 6:21 - 6:24
    考虑到我们将在稍后进行讨论的
    一些重要假设
  • 6:24 - 6:26
    我们可以通过将简化式
    除以第一阶段
  • 6:26 - 6:29
    来找出KIPP上学率的影响
  • 6:29 - 6:33
    通过示例,这点将会更加清楚
  • 6:33 - 6:34
    让我们做吧
  • 6:37 - 6:39
    有关衡量的简短笔记
  • 6:39 - 6:42
    我们使用标准差来衡量成就
  • 6:42 - 6:45
    通常用希腊字母sigma (σ) 表示
  • 6:45 - 6:49
    一个σ是从大多数成就分配的
    最低15%
  • 6:49 - 6:52
    到中间位置的巨大变化
  • 6:52 - 6:55
    甚至¼或½ σ 的差异也很大
  • 6:56 - 6:58
    现在我们准备将一些数字
  • 6:58 - 7:02
    插入到前面介绍的方程式中
  • 7:02 - 7:03
    首先,获录取对数学成绩
  • 7:03 - 7:06
    有何影响呢?
  • 7:06 - 7:10
    KIPP申请人的数学成绩是
  • 7:10 - 7:12
    申请KIPP之前一年中
  • 7:12 - 7:14
    低于州平均值的标准差的三分之一
  • 7:14 - 7:18
    但是一年后,获录取者得分
    达到了州平均水平
  • 7:18 - 7:21
    而未被录取者
  • 7:21 - 7:25
    仍然落后于平均分数-0.36σ
  • 7:26 - 7:30
    获录取者对分数的影响
    是获录取者的分数
  • 7:30 - 7:33
    与未被录取者的分数之间的差异
  • 7:33 - 7:36
    获录取者的平均数学成绩
  • 7:36 - 7:38
    减去未被录取者的平均数学成绩
  • 7:38 - 7:42
    你的答案是0.36σ
  • 7:42 - 7:47
    Next up: what's the effect
    of winning the lottery on attendance?
  • 7:47 - 7:49
    In other words,
    if you win the lottery,
  • 7:49 - 7:52
    how much more likely
    are you to attend KIPP
  • 7:52 - 7:53
    than if you lose?
  • 7:54 - 7:58
    First, what percentage
    of lottery winners attend KIPP?
  • 7:58 - 8:01
    Divide the number of winners
    who attended KIPP
  • 8:01 - 8:05
    by the total number
    of lottery winners -- that's 78%.
  • 8:06 - 8:09
    To find the percentage
    of lottery losers who attended KIPP,
  • 8:09 - 8:12
    we divide the number of losers
    who attended KIPP
  • 8:12 - 8:17
    by the total number
    of lottery losers -- that's 4%.
  • 8:17 - 8:22
    Subtract 4 from 78, and we find
    that winning the lottery
  • 8:22 - 8:26
    makes you 74%
    more likely to attend KIPP.
  • 8:26 - 8:29
    Now we can find
    what we're really after --
  • 8:29 - 8:35
    the effect of attendance on scores,
    by dividing 0.36 by 0.74.
  • 8:35 - 8:38
    Attending KIPP raises math scores
  • 8:38 - 8:42
    by 0.48 standard deviations
    on average.
  • 8:42 - 8:45
    That's an awesome achievement gain,
  • 8:45 - 8:47
    equal to moving
    from about the bottom third
  • 8:47 - 8:50
    to the middle
    of the achievement distribution.
  • 8:50 - 8:51
    - [Student] Whoa, half a sig.
  • 8:51 - 8:54
    - [Instructor] These estimates
    are for kids opting in
  • 8:54 - 8:55
    to the KIPP lottery,
  • 8:55 - 8:58
    whose enrollment status
    is changed by winning.
  • 8:58 - 9:01
    That's not necessarily
    a random sample
  • 9:01 - 9:02
    of all children in Lynn.
  • 9:03 - 9:05
    So we can't assume
    we'd see the same effect
  • 9:05 - 9:07
    for other types of students.
    - [Student] Huh.
  • 9:07 - 9:10
    - But this effect
    on keen for KIPP kids
  • 9:10 - 9:13
    is likely to be a good indicator
    of the consequences
  • 9:13 - 9:16
    of adding additional charter seats.
  • 9:16 - 9:17
    - [Student] Cool.
    - [Student] Got it.
  • 9:20 - 9:23
    - IV eliminates selection bias,
    but like all of our tools,
  • 9:23 - 9:26
    the solution builds on a set
    of assumptions
  • 9:26 - 9:28
    not to be taken for granted.
  • 9:28 - 9:31
    First, there must be
    a substantial first stage --
  • 9:31 - 9:36
    that is the instrumental variable,
    winning or losing the lottery,
  • 9:36 - 9:39
    must really change the variable
    whose effect we're interested in --
  • 9:39 - 9:41
    here, KIPP attendance.
  • 9:41 - 9:45
    In this case, the first stage
    is not really in doubt.
  • 9:45 - 9:48
    Winning the lottery makes
    KIPP attendance much more likely.
  • 9:48 - 9:51
    Not all IV stories are like that.
  • 9:51 - 9:54
    Second, the instrument
    must be as good
  • 9:54 - 9:55
    as randomly assigned,
  • 9:55 - 9:59
    meaning lottery winners and losers
    have similar characteristics.
  • 9:59 - 10:02
    This is the independence assumption.
  • 10:02 - 10:06
    Of course, KIPP lottery wins
    really are randomly assigned.
  • 10:06 - 10:10
    Still, we should check for balance
    and confirm that winners and losers
  • 10:10 - 10:11
    have similar family backgrounds,
  • 10:11 - 10:14
    similar aptitudes and so on.
  • 10:14 - 10:17
    In essence, we're checking
    to ensure KIPP lotteries are fair
  • 10:17 - 10:20
    with no group of applicants
    suspiciously likely to win.
  • 10:21 - 10:24
    Finally, we require
    the instrument change outcomes
  • 10:24 - 10:26
    solely through
    the variable of interest --
  • 10:26 - 10:28
    in this case, attending KIPP.
  • 10:28 - 10:31
    This assumption is called
    the exclusion restriction.
  • 10:33 - 10:38
    - IV only works if you can satisfy
    these three assumptions.
  • 10:38 - 10:40
    - I don't understand
    the exclusion restriction.
  • 10:41 - 10:44
    How could winning the lottery
    affect math scores
  • 10:44 - 10:45
    other than by attending KIPP?
  • 10:45 - 10:47
    - [Student] Yeah.
    - [Instructor] Great question.
  • 10:47 - 10:51
    Suppose lottery winners
    are just thrilled to win,
  • 10:51 - 10:55
    and this happiness motivates them
    to study more and learn more math,
  • 10:55 - 10:57
    regardless of where
    they go to school.
  • 10:57 - 11:00
    This would violate
    the exclusion restriction
  • 11:00 - 11:04
    because the motivational effect
    of winning is a second channel
  • 11:04 - 11:07
    whereby lotteries
    might affect test scores.
  • 11:07 - 11:10
    While it's hard
    to rule this out entirely,
  • 11:10 - 11:13
    there's no evidence
    of any alternative channels
  • 11:13 - 11:14
    in the KIPP study.
  • 11:18 - 11:21
    - IV solves the problem
    of selection bias
  • 11:21 - 11:25
    in scenarios like the KIPP lottery
    where treatment offers are random
  • 11:25 - 11:27
    but some of those offered opt out.
  • 11:28 - 11:32
    This sort of intentional
    yet incomplete random assignment
  • 11:32 - 11:33
    is surprisingly common.
  • 11:33 - 11:36
    Even randomized clinical trials
    have this feature.
  • 11:37 - 11:40
    IV solves the problem
    of non-random take-up
  • 11:40 - 11:43
    in lotteries or clinical research.
  • 11:43 - 11:47
    But lotteries are not the only source
    of compelling instruments.
  • 11:47 - 11:49
    Many causal questions
    can be addressed
  • 11:49 - 11:51
    by naturally occurring
  • 11:51 - 11:54
    as good as randomly
    assigned variation.
  • 11:55 - 11:57
    Here's a causal question for you:
  • 11:57 - 11:59
    Do women who have children
    early in their careers
  • 11:59 - 12:02
    suffer a substantial earnings penalty
  • 12:02 - 12:03
    as a result?
  • 12:03 - 12:05
    After all, women earn less than men.
  • 12:06 - 12:09
    We could, of course, simply compare
    the earnings of women
  • 12:09 - 12:11
    with more and fewer children.
  • 12:11 - 12:14
    But such comparisons are fraught
    with selection bias.
  • 12:15 - 12:17
    If only we could
    randomly assign babies
  • 12:17 - 12:19
    to different households.
  • 12:19 - 12:22
    Yeah, right,
    sounds pretty fanciful.
  • 12:22 - 12:27
    Our next IV story -- fantastic
    and not fanciful --
  • 12:27 - 12:30
    illustrates an amazing,
    naturally occurring instrument
  • 12:30 - 12:32
    for family size.
  • 12:33 - 12:35
    ♪ [] ♪
  • 12:35 - 12:38
    - [Instructor] You're on your way
    to mastering econometrics.
  • 12:38 - 12:40
    Make sure this video sticks
  • 12:40 - 12:43
    by taking a few
    quick practice questions.
  • 12:43 - 12:46
    Or, if you're ready,
    click for the next video.
  • 12:47 - 12:50
    You can also check out
    MRU's website for more courses,
  • 12:50 - 12:52
    teacher resources, and more.
  • 12:52 - 12:54
    ♪ [] ♪
Title:
工具变量(IV)的简介
Description:

more » « less
Video Language:
English
Team:
Marginal Revolution University
Project:
Mastering Econometrics
Duration:
12:57

Chinese, Simplified subtitles

Revisions Compare revisions