The Role of the Propensity Score in Fixed Effect Models

Arkhangelsky, Dmitry; Imbens, Guido W.

doi:10.3386/w24814

Cited by 24 publications

(19 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…1 The success of experiments appears to endanger the role of observational studies, whether qualitative or quantitative, as these studies can never meet the stringent criteria imposed by their randomized kin (Beck 2006;Gerber, Green, and Kaplan 2014;Gerstein, McMurray, and Holman 2019). As a result, previously popular methods like large-N time-series cross section models have come under criticism for failing to either estimate average treatment effects (ATEs) (Samii 2016;Arkhangelsky and Imbens 2018;Gibbons, Serrato, and Urbancic 2017), the causal criterion of the potential outcomes framework, or to account for missing variables and over-time dynamics (Plümper and Troeger 2019).…”

Section: The New Intellectual Battlefield: Causalitymentioning

confidence: 99%

“…On one side, Deaton and Cartwright (2018) argue that the emphasis on RCTs as a cure-all for causal inference is overblown because researchers often ignore the known limitations of their samples by reference to randomization. While some support Deaton and Cartwright, including Gelman (2018) and Sampson (2018), others argue that recent research on understanding treatment heterogeneity and the application of experimental results to novel problems mitigate Deaton and Cartwright's concerns (Arkhangelsky and Imbens 2018;Ioannidis 2018). This brewing dispute has all the hallmarks of a noteworthy battle of the minds, although it could create yet another methodological minefield that many researchers fear to tread upon.…”

Section: The New Intellectual Battlefield: Causalitymentioning

confidence: 99%

See 1 more Smart Citation

Holistic Causal Learning with Causal Graphs: A Credible Method for Study Design and Preregistration in the Social Sciences

Kubinec¹

2020

Preprint

View full text Add to dashboard Cite

I argue that social scientific practice would be improved if we abandoned the notion that a gold standard for causal inference exists. Instead, we should consider the three main inference modes--experimental approaches and the counterfactual theory, large-N observational studies, and qualitative process-tracing--as observable indicators of an unobserved latent construct, causality. Under ideal conditions, any of these three approaches provide strong, though not perfect, evidence of an important part of what we mean by causality. I also present the use of statistical entropy, or information theory, as a possible yardstick for evaluating research designs across the silver standards. Rather than dichotomize studies as either causal or descriptive, the concept of relative entropy instead emphasizes the relative causal knowledge gained from a given research finding.

show abstract

Section: The New Intellectual Battlefield: Causalitymentioning

confidence: 99%

Section: The New Intellectual Battlefield: Causalitymentioning

confidence: 99%

Holistic Causal Learning with Causal Graphs: A Credible Method for Study Design and Preregistration in the Social Sciences

Kubinec¹

2020

Preprint

View full text Add to dashboard Cite

show abstract

“…Recent works by Yang (2018), He (2018), and Lee, Nguyen, and Stuart (2019) use methods based on propensity scores, but they require parametric modeling assumptions. Arkhangelsky and Imbens (2019) also use a propensity-score based approach and assume that the conditional distributions of the covariates and the treatment are in the exponential family. Finally, Zetterqvist, Vansteelandt, Pawitan, and Sjölander (2016) propose a doubly robust estimator that is consistent for the treatment effect so long as either the outcome model or the propensity score is correctly specified, but not necessarily both.…”

Section: Prior Work and Our Contributionmentioning

confidence: 99%

“…Specifically, we use predictions from three types of generalized linear models and a neural network. The three types of generalized linear models include generalized linear regression with cluster dummies, generalized linear regression with cluster means, and generalized linear regression with cluster-level random effects, each of which was frequently used in relevant works (Arpino & Mealli, 2011;Li et al, 2013;Arkhangelsky & Imbens, 2019;Li et al, 2020). A neural network uses a rectifier activation function, commonly known as ReLU, with 2 hidden layers and each layer has J +p x neurons; see the CURobustML package for more details.…”

Section: Ensemble Machine Learningmentioning

confidence: 99%

Robust Machine Learning for Treatment Effects in Multilevel Observational Studies Under Cluster-level Unmeasured Confounding

Suk¹,

Kang²

2020

Preprint

View full text Add to dashboard Cite

Recently, machine learning (ML) methods have been used in causal inference to estimate treatment effects in order to reduce concerns for model mis-specification. However, many, if not all, ML methods require that all confounders are measured to consistently estimate treatment effects. In this paper, we propose a family of ML methods that estimate treatment effects in the presence of cluster-level unmeasured confounders, a type of unmeasured confounders that are shared within each cluster and are common in multilevel observational studies. We show through simulation studies that our proposed methods are consistent and doubly robust when unmeasured cluster-level confounders are present. We also examine the effect of taking an algebra course on math achievement scores from the Early Childhood Longitudinal Study, a multilevel observational educational study, using our methods. The proposed methods are available in the CURobustML R package.

show abstract

“…Direct cross-sectional estimation with this sample may suffer from biases due to individual heterogeneity. However, recent econometrics literature suggests that PSM can deal with individual heterogeneity effectively (Arkhangelsky and Imbens, 2018), and thus we combined cross-sectional analyses with PSM and report the results in table C1 (those without using PSM are available upon request). The results still support our conclusions, although support for hypothesis 4b is weaker.…”

Section: Appendix C: Clarifying Data and Analysesmentioning

confidence: 99%

Corrigendum: Marquis, Christopher, and Kunyuan Qiao

2020

Administrative Science Quarterly

View full text Add to dashboard Cite

This article has been corrected since it was initially published OnlineFirst to provide a more clear and accurate description of the sample and to increase the transparency of our analytical approach. Corrections focused on our sample are detailed in the Corrigendum available at https://journals.sagepub.com/doi/ full/10.1177/0001839219846350. The following changes reflect our additional efforts to increase transparency about our data and analyses. All changes indicated in both corrigenda were made before the article was included in the September 2020 print issue of the Administrative Science Quarterly. On page 808, we have added the following text to the paragraph describing our independent variable, communist ideological imprint: ''We note that although in theory variables reflecting entrepreneurs' life history are timeinvariant, errors can lead to observations exhibiting variation over time, as is common in longitudinal survey data. To be in line with theory, we use timeinvariant to describe these variables and random effects estimation for the analyses. We followed the literature on longitudinal survey data and used the conservative approach of including the raw values for our analyses and conducted sensitivity analyses under different potential error scenarios. Thirty percent of the enterprises in our data set never experienced change in this variable through the survey years. Online Appendix C details potential sources of error in our data set, our data-cleaning efforts, and our sensitivity analyses. Our robustness checks show that these data errors do not affect our results.'' On page 810, we added a footnote that reads, ''The combination of PSM and panel data is relatively new and has been implemented in a number of different ways. For our presented analyses, we used a probit model to generate weights, took the average of weights for observations of the same firm, and then ran panel data regressions with firm-year as unit of analysis. In Online Appendix C, we explain some alternative approaches and show that our results are consistent regardless of the approach we use. For example, we dropped observations with missing weight first, assigned the mean value of weights to observations with non-missing weight for each firm, and then used panel data methods. The results of this analysis appear in Online Appendix C, table C2. In addition, we used weights from the first step of PSM directly and then crosssectional analyses that ignore the longitudinal nature of the data (results available upon request).''

show abstract

The Role of the Propensity Score in Fixed Effect Models

Cited by 24 publications

References 15 publications

Holistic Causal Learning with Causal Graphs: A Credible Method for Study Design and Preregistration in the Social Sciences

Holistic Causal Learning with Causal Graphs: A Credible Method for Study Design and Preregistration in the Social Sciences

Robust Machine Learning for Treatment Effects in Multilevel Observational Studies Under Cluster-level Unmeasured Confounding

Corrigendum: Marquis, Christopher, and Kunyuan Qiao

Contact Info

Product

Resources

About