Jonathan Roth教授分享计量经济学研究心得 Jonathan Roth on Doing Econometric Research
本文最初于 2025 年 1 月 1 日 发布于微信公众号 Impactful Research；2026 年 4 月 28 日 同步至本网站。
Originally published on the WeChat official account Impactful Research on 2025-01-01; mirrored to this website on 2026-04-28.
来源：Google图文
这个公众号的第二十三篇文章，我们很荣幸邀请到布朗大学的Jonathan Roth教授分享他对计量经济学研究的心得和建议。
以下是Jonathan Roth教授分享对计量经济学研究的心得和建议。
本文正文内容约三千字，全文阅读需约6分钟
#本期访谈主要问题
1. 您是如何发展您对DID计量经济学的研究兴趣的？
2. 在写作和修改这些文章中，您遇到最大的挑战是什么？
3. 在您看来，让这些论文有影响力的主要原因是什么？
4. 对于对理论计量感兴趣的学生，您有什么建议吗？
Q1：您是如何发展您对DID计量经济学的研究兴趣的？
Q1:How did you develop your research interests and agenda in the econometrics of DID?
在研究生的前几年，我曾以为自己想成为一名劳动经济学家。我跑了很多双重差分（DiD）回归，当我每次在回归分析中按下“回车”键时，我总是担心事前趋势是否看起来很好。当然，对于某些设定，事前趋势看起来不错，而对于另一些设定则不行。人们往往倾向于关注那些平行趋势看起来成立的情况。但选择那些看起来不错的图表，忽略其他不好的情况，让我觉得有点不对劲，于是我开始思考“如果只选择那些预期趋势好的结果，会发生什么？” 这就是我在AER: Insights上发表论文的主题（Roth, 2022）[1]。我很快发现，相比劳动经济学，我更喜欢处理应用计量经济学问题（而且我做得也更好），于是在博士的第四或第五年，我就转向了计量经济学的研究。
In my first couple years of grad school, I thought I wanted to be a labor economist. I found myself running a bunch of DiDs, and every time I clicked ‘enter’ on a regression, there was always this suspense of whether the pre-trends would come out looking good or not. And, of course, for some specifications they looked pretty good and for some they didn’t, and there’s a tendency to focus on the ones where it looks like parallel trends holds. But choosing the plots that looked good and ignoring the other ones didn’t quite sit right with me, so that’s what got me thinking about “what happens if you select on having good pre-trends”, which is the topic of my AER:I paper (Roth, 2022) [1] . I pretty quickly found that I enjoyed working on applied metrics questions a lot more than doing labor economics (and I was much better at it), so I switched to doing econometrics in about my fourth or fifth year of grad school.
Q2：在写作和修改这些文章中，您遇到最大的挑战是什么？
Q2:What was the greatest challenge during the writing and revision of these papers?
写应用计量经济学论文时有一些很大的挑战。一个常见的挑战是，你有两个不同的读者群体——计量经济学和应用经济学研究者，他们的需求不同。你需要写得足够有技术深度，以说服计量经济学家（他们很可能是你的审稿人）相信你做了有技术价值的工作，同时又要简单到应用研究者可以理解并使用的程度。 通常，提供一个“简单的例子”能帮助每个人理解，然后再给出一个更一般的结果，这样既能吸引计量经济学家，又能让应用经济学研究者感到有用。
There are a few big challenges in writing applied econometrics papers.A common one is that you have two audiences, econometricians and applied researchers, and they want different things. You have to write with enough technical sophistication to convince the econometricians (who are likely to be your referees) that you did something technically interesting while also making it simple enough that applied people can understand and use it. It often helps to have a “simple example” that everyone can follow and then a more general result that appeals to econometricians.
除此之外，我还做了很多关于敏感性分析和边界方法的工作，尤其是针对违反平行趋势假设的情况。从理论角度来看，尽可能少的假设总是很有吸引力，但这通常意味着结果的边界会很宽。因此，需要你在假设的严格性和结果的信息量之间进行权衡，找到合适的平衡点是一个挑战。 例如，在我和 Ashesh Rambachan 的合作研究中，研究者需要选择一个敏感性分析参数 M，这个参数决定了相对于事前趋势，违反平行趋势的严重程度(Rambachan和Roth, 2023) [2]。这需要对以下问题进行一些反思：“可能的混杂因素是什么？它们可能是什么样子的？”我认为人们往往不习惯思考这些问题。比起单纯地跑一个统计检验，让人们思考这些困难的问题总是很有挑战性（尤其是那些可能会让他们结果有偏的东西！），但是我认为使用任何一种有原则的敏感性分析方法，我们都必然需要考虑可能出现的问题。
Beyond that, I’ve worked a lot on sensitivity analysis and bounding approaches for violations of parallel trends. And from a theoretical perspective it’s always appealing to assume very little. But that often means that the resulting bounds will be wide.So there’s a tradeoff between what you’re willing to impose and how informative your results are, and it’s a challenge to come up with the right balance. In my work with Ashesh Rambachan, for example, the researcher has to choose a sensitivity analysis parameter M that determines how bad the violations of parallel trends can be relative to the pre-trends (Rambachan and Roth, 2023) [2]. This requires some introspection on “what are the possible confounding factors? What might they look like?”, which I think people often aren’t used to thinking about. It’s always challenging to get people to have to think hard about something (especially something that might bias their results!), instead of just running a statistical test, even though I think any principled approach to sensitivity analysis will necessarily require thinking about what could have gone wrong rather than just running a statistical test.
Q3：在您看来，让这些论文有影响力的主要原因是什么？
Q3:From your perspective, what are the main reasons that make these papers impactful?
DID方法非常流行，大约四分之一 的NBER 工作论文中都使用了这种方法。所以，即使是对实践的微小改进，也能对许多论文产生影响。我认为我最有影响力的论文是那些解决了在实证研究中常见问题的论文。人们通常对这些问题有一些直觉，但如果你能通过统计学方法把它们形式化，或者提供一些理论上的见解，帮助他们更好地理解这些问题，那这个研究就可能产生很大的影响。 一个例子就是我之前提到的敏感性分析；人们经常担心违反平行趋势的问题，因此我们需要有一种规范的敏感性分析方法来说明违反的程度有多严重才能改变结论。另一个例子是我和 Kevin Chen 合作的论文 “Logs with Zeros”(Chen和Roth, 2024) [3]。实证研究中经常出现的问题是，有人想对结果变量取对数，但变量中包含零值。我想人们大致有一种直觉，认为在取对数之前加 1 的常见做法有些问题，但这篇论文正式地论述了为什么这种做法存在问题，并提出了一些实际的替代方案，我认为这对大家是有帮助的。
Well, DID is extremely popular; it’s used in something like a quarter of NBER working papers. So even small improvements to practice can impact a lot of papers.I think my most impactful papers have been the ones that address a problem that comes up in empirical work all the time. People often have some intuition about the problem, but if you can formalize it in a statistical procedure or shed some theoretical light that helps them understand it better, that can have a lot of impact. One example is the sensitivity analysis I mentioned earlier; people are often worried about violations of parallel trends, so having a formal way of doing sensitivity analysis to say how bad the violations would have to be to change the conclusions is useful. Another example is my “Logs with Zeros” paper with Kevin Chen (Chen and Roth, 2024) [3]. An issue that comes up all the time in empirical work is someone wants to take the log of the outcome but it has zeros. I think people kind of had the intuition that the common practice of adding 1 to the outcome before taking the log was a bit sketchy, but that paper formalized why it was problematic and shared some practical alternatives, which I think was helpful.
Q4：有些学生也许认为计量经济学理论是一个非常有趣但是门槛很高（例如数学和统计学）的领域，对于对理论计量感兴趣的学生，您有什么建议吗？
Q4:Some students might think that the econometrics theory is an area which is extremely interesting but with very high entry barrier (e.g., math and statistics), do you have any advice for students who are interested in theoretical econometrics?
的确，计量经济学比经济学的其他领域需要更多的数学和统计学知识。但我认为，门槛并不像你想象的那么高。 当你第一次阅读一篇计量经济学论文时，可能会看到一页又一页的数学公式，心想“我永远做不出这些”。但一旦你对文献有了更多了解，你会发现很多步骤其实是非常标准化的。 里面确实有某些创新可能很难，但这并不是说作者必须从头开始构思一切，他们可以基于已有的文献做某些修改。因此，一旦你对文献稍微熟悉一点，你就会发现自己也能很容易地通过复制其他论文中的标准设定，写出一页又一页的数学公式 :) 对于更偏向应用的计量经济学论文，挑战往往在于提出一个好的问题（或以正确的方式表达问题），而不是证明某个非常难的定理。 所以总的来说，如果你对计量经济学感兴趣，我鼓励你去尝试！它可能没有你想象的那么难，而且也可以非常有趣。
Well, it is true that econometrics requires more math and statistics than some other fields in economics.But I think the barriers are not as high as you might think. When you first pick up an econometrics paper, you might see pages and pages of math and think “I could never produce this myself”. But once you get a bit more familiar with the literature, you’ll see that many of the steps are actually very standard. There’s an innovation somewhere in there that was probably hard, but it’s not like the author had to come up with everything from scratch; they took the existing literature and modified it somewhere.So once you get a little more familiar with the literature, you’ll see that you too can produce pages and pages of math pretty easily by copying the standard set-up from other papers :) For more applied econometrics papers, the challenge is often more in coming up with a good question (or framing the question in the right way), rather than proving something really hard. So the bottom line is if you’re interested in econometrics, I’d encourage you to try it! It might not be as hard as you think. And it can also be really fun.
[1]Roth, Jonathan. “Pretest with caution: Event-study estimates after testing for parallel trends.” American Economic Review: Insights 4, no. 3 (2022): 305-322.
[2]Rambachan, Ashesh, and Jonathan Roth. “A more credible approach to parallel trends.” Review of Economic Studies 90, no. 5 (2023): 2555-2591.
[3]Chen, Jiafeng, and Jonathan Roth. “Logs with zeros? Some problems and solutions.” The Quarterly Journal of Economics 139, no. 2 (2024): 891-936.
学者简介：
Jonathan Roth 是布朗大学经济学系的 Groos Family 助理教授，主要研究领域是计量经济学，尤其是因果推断。他的研究还涉及劳动经济学、机器学习和算法公平等主题。
在加入布朗大学之前，Jonathan曾担任微软首席经济学家办公室的高级研究员。他于 2020 年在哈佛大学获得经济学博士学位，博士论文荣获David A. Wells最佳论文奖。在此之前，他在宾夕法尼亚大学获得数学与经济学的summa cum laude（最高荣誉）学士学位。
Jonathan Roth is the Groos Family Assistant Professor of Economics at Brown University. His primary research interests lie in econometrics, with a focus on causal inference. His work also encompasses topics in labor economics, machine learning, and algorithmic fairness.
Before joining Brown, Jonathan was a senior researcher in the Office of the Chief Economist at Microsoft. He earned his PhD in economics from Harvard University in 2020, receiving the David A. Wells Prize for the best dissertation. Prior to that, he graduated summa cum laude with a BA in mathematics and economics from the University of Pennsylvania.
参考文献：
Roth, Jonathan. “Pretest with caution: Event-study estimates after testing for parallel trends.” American Economic Review: Insights 4, no. 3 (2022): 305-322.
Rambachan, Ashesh, and Jonathan Roth. “A more credible approach to parallel trends.” Review of Economic Studies 90, no. 5 (2023): 2555-2591.
Chen, Jiafeng, and Jonathan Roth. “Logs with zeros? Some problems and solutions.” The Quarterly Journal of Economics 139, no. 2 (2024): 891-936.
| 责任编辑 | 秦雨 张帆 |
| 整理翻译 | 谈明康 张诗怡 |
| 校对 | Jonathan Roth |