"用随机化在发展经济学研究中的应用：诺贝尔奖得主著作"

需积分: 10 81 浏览量更新于2024-01-18 收藏 1.04MB PDF 举报

"Ester_Duflo.pdf"是诺贝尔经济学奖得主的著作，适合学经济学的同学阅读。本文以"Using Randomization in Development Economics Research: A Toolkit"为题，发表于NBER Technical Working Paper Series的第333篇技术工作论文中，作者是Esther Duflo, Rachel Glennerster和Michael Kremer。该论文于2006年12月发表在NBER官网上。本论文致谢T.Paul Schultz编辑以及Abhijit Banerjee、Guido Imbens和Jeffrey Kling等人的广泛讨论，也致谢David Clingingsmith、Greg Fischer、Trang Nguyen和Heidi Williams等人关于杰出研究的贡献。论文主要介绍了在发展经济学研究中使用随机化的方法。随机化是一种实验设计技术，通过在实验中随机分配参与者或实验条件，以消除因果关系中的干扰因素。随机化试验在医学研究中得到了广泛应用，然而在经济学领域并不常见。本文旨在向经济学研究人员介绍随机化的概念和运用，并展示了在发展经济学研究中随机化的实际应用。论文首先详细介绍了随机化试验的基本原理和设计。作者强调了随机分配的重要性，它能够有效地帮助建立因果关系，并避免结果的偏差。随后，作者通过实际案例展示了随机化试验在发展经济学研究中的应用。论文包括了多个案例研究，涵盖了教育、卫生、农业等领域。在每个案例中，作者详细解释了研究问题、实验设计、数据收集和分析方法。通过这些案例，作者展示了随机化试验在发展经济学研究中的价值和效果。值得一提的是，本文还介绍了一些常见的随机化试验设计和方法，如随机分群试验、随机限制试验、合成控制法等。作者强调了这些方法的灵活性和适用性，可以根据研究问题和资源的情况来选择合适的设计和方法。总的来说，本文是一篇介绍随机化在发展经济学研究中应用的重要著作。它提供了一个完整的工具包，帮助经济学研究人员理解和运用随机化试验的原理和方法。通过案例研究，读者可以更好地了解随机化试验在解决实际问题中的优势和局限性。对于对经济学研究方法感兴趣的同学和学者来说，这本书是一本非常有价值的参考资料。

Fixed

eﬀects generalizes diﬀerence-in-diﬀerences estimates when there is more than one time

period or more than one treatment group. The ﬁxed eﬀects estimates are obtained by regress-

ing the outcome on the control variable, after controlling for year and group dummies. Both

diﬀerence-in-diﬀerences and ﬁxed eﬀect estimates are very common in applied work. Whether

or not they are convincing depends on whether the assumption of parallel evolution of the out-

comes in the absence of the treatment is convincing. Note in particular that if the two groups

have very diﬀerent outcomes before the treatment, the functional form chosen for how outcomes

evolved over time will have an important inﬂuence on the results.

2.4 Comparing Experimental and Non-Experimental Estimates

A growing literature is taking advantage of randomized evaluation to estimate a program’s

impact using both experimental and non-experimental methods and then test whether the non-

experimental estimates are biased in this particular case. LaLonde’s seminal study found that

many of the econometric procedures and comparison groups used in program evaluations did not

yield accurate or precise estimates and that such econometric estimates often diﬀer signiﬁcantly

from experimental results (Lalonde 1986). A number of subsequent studies have conducted such

analysis focusing on the performance of propensity score matching (Heckman, Ichimura, and

Todd 1997, Heckman, Ichimura, Smith, and Todd 1998, Heckman, Ichimura, and Todd 1998,

Dehejia and Wahba 1999, Smith and Todd 2005). Results are mixed, with some studies ﬁnding

that non-experimental methods can replicate experimental results quite well and others being

more negative. A more comprehensive review by Glazerman, Levy, and Myers (2003) compared

experimental non-experimental methods in studies of welfare, job training, and employment

service programs in the United States. Synthesizing the results of twelve design replication

studies, they found that retrospective estimators often produce results dramatically diﬀerent

from randomized evaluations and that the bias is often large. They were unable to identify any

strategy that could consistently remove bias and still answer a well-deﬁned question.

Cook, Shadish, and Wong (2006) conducts a comparison of randomized and non-randomized

studies, most of which were implemented in educational settings, and arrives at a more nuanced

conclusion. He ﬁnds that experimental and non-experimental results are similar when the non-

experiment technique is a regression discontinuity or “interrupted time series” design (diﬀerence-

in-diﬀerences with long series of pre-data), but that matching or other ways to control for

observ

ables does not produce similar results. He concludes that well designed quasi-experiments

(regression discontinuity designs in particular) may produce results that are as convincing as

those of a well-implemented randomized evaluation but that “You cannot put right by statistics

what you have done wrong by design.” While Cook’s ﬁndings are extremely interesting, the

level of control achieved by the quasi-experiments he reviews (in terms, for example, of strictly

following threshold rules) is such that for developing countries these designs may actually be

less practical than randomized evaluations.

We are not aware of any systematic review of similar studies in developing countries, but

a number of comparative studies have been conducted. Some suggest omitted variables bias

is a signiﬁcant problem; others ﬁnd that non-experimental estimators may perform well in cer-

tain contexts. Buddlemeyer and Skoﬁas (2003) and Diaz and Handa (2006) both focus on

PROGRESA, a poverty alleviation program implemented in Mexico in the late 1990s with a

randomized design. Buddlemeyer and Skoﬁas (2003) use randomized evaluation results as the

benchmark to examine the performance of regression discontinuity design. They ﬁnd the per-

formance of such a design to be good, suggesting that if policy discontinuities are rigorously

enforced, regression discontinuity design frameworks can be useful. Diaz and Handa (2006)

compare experimental estimates to propensity score matching estimates, again using the PRO-

GRESA data. Their results suggest that propensity score matching does well when a large

number of control variables is available.

In contrast, several studies in Kenya ﬁnd that estimates from prospective randomized eval-

uations can often be quite diﬀerent from those obtained using a retrospective evaluation in

the same sample, suggesting that omitted variable bias is a serious concern. Glewwe, Kremer,

Moulin, and Zitzewitz (2004) study an NGO program that randomly provided educational ﬂip

charts to primary schools in Western Kenya. Their analysis suggests that retrospective es-

timates seriously overestimate the charts’ impact on student test scores. They found that a

diﬀerence-in-diﬀerences approach reduced but did not eliminate this problem.

Miguel and Kremer (2003) and Duﬂo, Kremer, and Robinson (2006) compare experimental

and non-experimental estimate of peer eﬀects in the case of the take up of deworming drug and

fertilizer adoption, respectively. Both studies found that the individual’s decision is correlated

with the decisions of their contacts. However, as Manski (1993) has argued, this could be due to

many factors other than peer eﬀects, in particular, to the fact that these individuals share the

same

environment. In both cases, randomization provided exogenous variation in the chance

that some members of a particular network adopted the innovation (deworming or fertilizer,

respectively). The presence of peer eﬀect can then be tested by comparing whether others in

the network were then more likely to adopt as well (we will come back to this speciﬁc method

of evaluating peer eﬀects in section 6.3). Both studies ﬁnd markedly diﬀerent results from the

non-experimental results: Duﬂo, Kremer, and Robinson (2006) ﬁnd no learning eﬀect, while

Miguel and Kremer (2003) ﬁnd negative peer eﬀects. Furthermore, Miguel and Kremer (2003)

run on the non-experimental data a number of speciﬁcations checks that have been suggested

in the peer eﬀects literature, and all these checks support the conclusion that peer eﬀects are in

fact positive, suggestive that such checks may not be suﬃcient to erase the speciﬁcation bias.

Future research along these lines would be valuable, since comparative studies can be used

to assess the size and prevalence of biases in retrospective estimates and provide more guidance

which methodologies are the most robust. However, as discussed below, these types of studies

have to be done with care in order to provide an accurate comparison between diﬀerent meth-

ods. If the retrospective portions of these comparative studies are done with knowledge of the

experimental results, there is a natural tendency to select from plausible comparison groups

and methodologies in order to match experimental estimates. To address these concerns, future

researchers should conduct retrospective evaluations before the results of randomized evalua-

tions are released or conduct blind retrospective evaluations without knowledge of the results of

randomized evaluations or other retrospective studies.

2.5 Publication Bias

2.5.1 Publication bias in non-experimental studies

Uncertainty over bias in the reported results from non-experimental studies is compounded

by publication bias, which occurs when editors, reviewers, researchers, or policymakers have a

preference for results that are statistically signiﬁcant or support a certain view. In many cases,

as we just reviewed, researchers will have many possible choices of how to specify empirical

models, and many of these might still be subject to remaining omitted variable bias.

Consider an example in which the true treatment eﬀect is zero, but each non-experimental

technique yields an estimated treatment eﬀect equal to the true eﬀect plus an omitted variable

bias

term that is itself a normally distributed random variable with mean zero. Unfortunately,

what appears in the published literature may not reﬂect typical results from plausible speciﬁca-

tions, which would be centered on zero, but may instead be systematically biased.

In any study, a number of choices will need to be made about how the analysis would

most appropriately be conducted, which method should be used, which control variable should

be introduced, which instrumental variable to use. There will often be legitimate arguments

for a variety of diﬀerent alternatives. Researchers focus their time and eﬀort on completing

studies that seem to produce statistically signiﬁcant results, those that do not end up in the

“ﬁle drawer.”

Some researchers may inappropriately mine various regression speciﬁcations for one that pro-

duces statistically signiﬁcant results. Even researchers who do not deliberately search among

speciﬁcations for those that yield signiﬁcant results may do so inadvertently. Consider a re-

searcher undertaking a retrospective study for which there are several potentially appropriate

speciﬁcations, not all of which are known to the researcher before beginning the analysis. The

researcher thinks of one speciﬁcation and runs a series of regressions. If the results are statisti-

cally signiﬁcant and conﬁrm what the researcher expected to ﬁnd, it is perhaps likely that he or

she will assume the speciﬁcation is appropriate and not spend much time considering possible

alternatives. However, if the regression results are not statistically signiﬁcant or go against what

the researcher expected to ﬁnd, he or she is perhaps more likely to sp end substantial amounts

of time considering other possible speciﬁcations. In addition to generating too many false pos-

itives in published papers, this type of speciﬁcation searching probably leads to under-rejection

of commonly held views.

Even if a researcher introduces no bias of this kind, the selection by journals of papers with

signiﬁcant results introduces another level of publication bias. Moreover, citation of papers

with extreme results by advocates on one side or another of policy debates is likely to compound

publication bias with citation bias. The cumulative result of this process is that even in cases

when a program has no eﬀect, strongly positive and/or strongly negative estimates are likely to

be published and widely cited.

A growing bo dy of available evidence suggests publication bias is a serious problem within

the economics literature. DeLong and Lang (1992) devise a test to determine the fraction of

un-rejected null hypotheses that are false. They note that under the null, the distribution of test

剩余88页未读，继续阅读

caesor9806

粉丝: 0
资源: 1

"用随机化在发展经济学研究中的应用：诺贝尔奖得主著作"

Methane-diphosphonic Acid Tetraethyl Ester.pdf

led-tcp-mastithub

机器人移动天轨step_三维3D设计图纸.zip

轿厢式电梯sw12可编辑_三维3D设计图纸.zip

车牌识别-基于Pytorch实现的MTCNN+LPRNet两阶段轻量级高性能车牌识别算法-附项目源码-优质项目实战.zip

SSM+JSP爱旅行平台答辩PPT.ppt

基于微信小程序的考研助手小程序答辩PPT.pptx

基于Java语言的新闻资讯项目设计源码

草莓采摘机器人sw22可编辑+仿真视频_三维3D设计图纸.zip

基于Java+Springboot+Vue+Mysql实现的音乐推荐管理系统【毕业设计/课程设计】

最新资源