正文
Deaton获诺贝尔经济学奖之后，结构模型与随机实验的争论又重现江湖。最近Boston College 的Arthur Lewbel教授在自己的facebook上对一争论发表了自己的看法。
对于墙外的同学，可以直接看他的主页：
https://www.
facebook.com/arthur.lew
bel/posts/10153781244691414
。对于墙内的同学，我在这里将原文复制过来，并做简单翻译。
Deaton argues that the virtues of randomized experiments over detailed survey data have been greatly exaggerated in the modern development literature. I agree. I don't usually discuss my real work on Facebook, however, below are some paragraphs from a survey article I'm writing on identification. I'd be interested in hearing what people of think of this, especially the last paragraph:
Deaton认为在现在的发展经济学文献中，随机实验的好处被大大高估了。我（Lewbel）同意这一点，我在Facebook上不经常讨论我的实际工作，以下是我写的关于识别问题的文章的一些段落，我特别感兴趣大家对这些段落，特别是最后几段有怎样的看法。
The main advantage of structural models over causal models is that, as the name implies, they contain information about underlying structure. Structural models can incorporate, or be used to test, behavioral restrictions implied by economic theory, and can likewise gain identification from incorporating such restrictions, even when external sources of randomization are not present.
正如名字所意味的那样，结构模型最大的好处在于这些模型包含着经济现象背后的结构信息。结构模型可以包含或者用来检验经济学理论所隐含的行为约束，也可以在外部随机性不存在的情况下，通过这些约束来达到识别的目的。
The main disadvantage of structural modeling is that any structure we write down is likely to be misspecified, and one generally does not know the extent to which misspecification corrupts the interpretation and applicability of the structural model estimates. Causal models can of course also be misspecified, e.g., in the LATE model even if an instrument Z is randomly assigned, the population may contain defiers or SUTVA may be violated. The relative plausibility of structural vs causal assumptions depends on context. More subtly, both structural and causal models generally require covariates, which in turn require functional forms that may either be parametrically misspecified (the popularity of linear probability models in reduced form methods is a case in point), or are nonparametrically specified and thereby suffer the curse of dimensionality.
而结构模型的主要缺点在于任何写下来的结构都有可能是错误设定的，而人们很难知道这些错误的设定对于结构模型的估计和应用有多大的影响。当然，因果模型也可能被错误设定，比如在局部平均处理效应（LATE）中，即使工具变量是随机的，总体中也可能会存在defiers或者SUTVA不满足。所以结构模型假设与因果模型假设的合理性依赖于问题的背景。而更微妙的是，结构模型和因果模型一般都需要协变量，而这又要求函数形式的设定，而函数形式有可能是错误设定的，当然也可以非参数设定，而这又会导致维数诅咒。
Still, causal models generally invoke fewer assumptions regarding behavior and outcomes than structural models. But a big issue in causal modeling is external validity. This is the question of relevance of an object like LATE to any situation other than the particular data set at hand. For example, if the set of compliers for a given instrument Z differs greatly from the general population, then the ATE of people we actually treat may differ greatly from the identified LATE. Alternatively, if Z is a policy variable, then compliers may indeed be a subpopulation of interest. But generally, by not assuming any structure (which in the best case would be derived from economic theory), we cannot say much about how any causal effect we identify in one environment would differ from the causal effect in any other, even only slightly different, environment.
当然，因果模型一般需要更少的假设，但是因果模型的一个严重的问题是外部有效性问题。这个问题牵涉到诸如LATE到一般情形推广、从手上的数据到任意情形的推广。比如说，在LATE中，针对一个给定的工具，compliers可能与总体有很大差别，那么实际的平均处理效应和我们估计出来的局部处理效应就会有很大差别。当然，如果工具Z是一个政策变量，那么compliers可能刚好是我们需要关注的群体。但是一般来讲，如果没有假设任何的结构，我们并不知道从一个环境中识别出的因果效应与另外一个环境中的因果效益有多大的不同。
