聚束效应Bunching分析方法, 政策评估中的新宠
凡是搞计量经济的，都关注这个号了
邮箱：econometrics666@sina.cn
所有计量经济圈方法论丛的程序文件, 微观数据库和各种软件都放在社群里.欢迎到计量经济圈社群交流访问.
由@因果推断研究小组 撰写
聚束效应（Bunching）分析方法，聚束分析方法起初是用于税收相关研究，但是后面已经成为像DID、RDD、SCM一样的经典政策评估方法了。下面这个是Annual review of Economics上关于bunching的文献综述，作者是Henrik Jacobsen Kleven，对此政策评估感兴趣的可以好好阅读一下。当然，也可以到计量社群交流探讨这个方法。
Recent years have seen the development of a new empirical approach in economics: the bunching approach. This approach uses bunching around points that feature discontinuities in incentives to elicit behavioral responses and estimate structural parameters. The approach was initially developed to estimate behavioral responses to taxes and transfers, but is now finding applications in other areas and settings. This review provides a guide to bunching estimation, discusses its strengths and weaknesses, draws links to other literatures, and ponders directions for future research.
The literature distinguishes between two conceptually different bunching designs. One type of design is based on kink points—discrete changes in the slope of choice sets—and was developed by Saez (2010) and Chetty et al. (2011). The other type of design is based on notch points—discrete changes in the level of choice sets—and was developed by Kleven &Waseem (2013). In the context of taxes and transfers, the distinction corresponds to whether the discontinuity occurs in the marginal tax rate or in the average tax rate. Kinks and notches offer different empirical advantages and challenges, as discussed below, and they tend to feature in different types of settings. Although kinks are commonly observed in income redistribution policies (such as graduated income tax systems), notches are ubiquitous across a wide range of other tax and nontax settings.
The emergence of the bunching approach is closely linked to another recent development in applied research: the increased use of administrative data. Because of the local nature of bunching responses—moving to specific points from nearby regions—estimating bunching precisely requires large data sets with very little measurement error. We rarely see any bunching in survey data due to small sample sizes and measurement error. With access to big administrative data sets, conversely, simply plotting the raw data can often reveal bunching and provides prima facie evidence of a causal effect of the incentive in question. A key question, however, is what we can learn from such responses in terms of structural and more externally valid parameters.
I argue that two broad lessons have emerged from the bunching literature to date. First, although bunching provides compelling nonparametric evidence of a behavioral response, moving from observed bunching to a structural parameter that can be used to predict the effects of policy changes is difficult. This is particularly true in the context of labor supply—the context for which the bunching approach was initially developed—due to a range of optimization frictions that attenuate bunching and are difficult to observe and model.1 These frictions include aspects such as hours constraints, search costs, inattention, and uncertainty. Such frictions imply that any evidence of sharp bunching in earnings likely results from tax evasion or tax avoidance rather than real labor supply responses. Indeed, several applications of the bunching approach explicitly consider evasion and avoidance as their main objects of interest. Second, these difficulties of estimating structural elasticities do not invalidate the bunching approach, but they imply that the approach may be better used in different ways than initially intended. This includes studying different outcomes than labor supply (some that are less subject to optimization friction), and it includes using bunching for other purposes than to obtain price elasticities for policy prediction. I provide many examples of such alternative uses of the bunching approach below.
The bunching literature is tied to an earlier literature estimating labor supply in the presence of kinked budget sets, namely the nonlinear budget set approach pioneered by Burtless & Hausman (1978) and Hausman (1981). This literature estimated labor supply using models that predict bunching at kink points even though no bunching was found in the survey data they used, an issue that was debated by Heckman (1983) and Hausman (1983). The way that theory and data were reconciled in those studies was by allowing for measurement error in the data and optimization error by households through the modeling of the error term. Although access to administrative data largely resolves the problem of measurement error, it does not reduce the scope for optimization error in shaping bunching, as highlighted above. The study of optimization frictions in the recent bunching literature is closely related to the debate about how to model the error term in the nonlinear budget set approach. Bunching designs are related to two other research designs often used in empirical work: the regression discontinuity (RD) and the regression kink (RK) designs as laid out, for example, by Imbens & Lemieux (2008) and Card et al. (2015). RD and RK designs essentially exploit notched and kinked incentives, respectively, but in situations in which the assignment variable—the variable that determines whether the agent is above or below the relevant threshold—is not subject to choice or manipulation.
Bunching designs consider the opposite case, in which the assignment variable is a direct choice. In this sense, whenever we observe discrete jumps in incentives at specific thresholds, it is potentially possible to use either RD/RK designs or bunching designs, depending on the manipulability of the assignment variable. A complication in practice is that the manipulability of the assignment variable may not always be clearly determined, especially in situations with optimization frictions.
The article proceeds as follows. Section 2 describes the relationship between the bunching literature and the traditional nonlinear budget set approach, Section 3 lays out the theory underlying bunching estimation, Section 4 describes the empirical implementation and challenges of bunching approaches, Section 5 discusses applications across a wide range of topics, and Section 6 concludes.
下面这个是由@慵懒不是懒 群友推荐的发表在Annual review of Economics的完整文献综述，作者是Henrik Jacobsen Kleven，对此政策评估感兴趣的可以下载参考阅读。
群友长按以上二维码即可下载这个PDF版本文献综述。
2年，计量经济圈公众号近1000篇文章，
Econometrics Circle
干货系列：能源环境 | 效率研究 | 空间计量 | 国际经贸 | 计量软件 | 商科研究 | 机器学习 | SSCI | CSSCI
计量经济圈组织了一个计量社群，有如下特征：热情互助最多、前沿趋势最多、社科资料最多、社科数据最多、科研牛人最多、海外名校最多。因此，建议积极进取和有强烈研习激情的中青年学者到社群交流探讨，始终坚信优秀是通过感染优秀而互相成就彼此的。