Elea McDonnell Feit @eleafeit.bsky.social · 16/04/2025Inspired by @emilyriederer.bsky.social, I made some pictures for my seminar on causal inference. They are not labeled. Let's see if you can name that research design. 271
Elea McDonnell Feit @eleafeit.bsky.social · 04/04/2025This paper taught me that the competition between advertisers for the most desirable ad slots is really important. 030
Elea McDonnell Feit @eleafeit.bsky.social · 04/04/2025Chat GPT just gave me the best complement on my writing I've ever gotten. 030
Elea McDonnell Feit @eleafeit.bsky.social · 28/03/2025How frequently do you meet with your students? Once a week just doesn't feel like enough when you are in the "thick" of a project. 010
Elea McDonnell Feit @eleafeit.bsky.social · 25/03/2025I need to answer that email about the revision! I'm on-board, but the email got buried in my inbox. 110
Elea McDonnell Feit @eleafeit.bsky.social · 20/03/2025Or maybe youse guys data is much cleaner than mine. 000
Elea McDonnell Feit @eleafeit.bsky.social · 20/03/2025Maybe my style of interactive debugging is archaic. 100
Elea McDonnell Feit @eleafeit.bsky.social · 20/03/2025Sometimes I just want to know how many zeros and NAs there are in a field and I don't want to write a long dplyr chain to figure that out. 100
Elea McDonnell Feit @eleafeit.bsky.social · 20/03/2025In reviewing dplyr code written by a PhD student, I'm reminded of how useful base R functions like summary() and str() are for inspecting data and detecting errors. 240
Elea McDonnell Feit @eleafeit.bsky.social · 15/03/2025Happy to report that this paper has been accepted at Management Science! (At least that was reported in the ISMS Newsletter, but I can't find the paper on the Management Science website. 020
Reposted by Elea McDonnell FeitElea McDonnell Feit @eleafeit.bsky.social · 02/01/2025Joo and Chiong’s recent working paper provides a Gaussian approximation to the regret function that you can use with ✨any✨ asymptotically normal estimator. This means you can use the minimax-regret criteria with your favorite treatment effect estimators: diff-in-diff, ML estimators. 141
Elea McDonnell Feit @eleafeit.bsky.social · 19/02/2025I know you were responding to @blasimon.bsky.social, but wanted to chime in to say I agree they are very strong parametric assumptions. We tried to spell that out as clearly as we could. 020
Elea McDonnell Feit @eleafeit.bsky.social · 19/02/2025To help marketing reviewers and editors understand the untestable assumptions of causal inference methods, Dominik Papies, Peter Ebbes and I wrote this "menu" as part of our chapter on "Endogeneity and Causal Inference in Marketing". (Preprint: dx.doi.org/10.2139/ssrn...) 140
Elea McDonnell Feit @eleafeit.bsky.social · 19/02/2025An important plea from @lizstuart.bsky.social in today's SCI-OCIS Special Webinar Series: 2102
Elea McDonnell Feit @eleafeit.bsky.social · 12/02/2025#Eagles and #Dragons fly together! 💚🦅🤍 💙🐉💛 Drexel will close on Friday to join the rest of #Philly in celebrating the Champions! 100
Elea McDonnell Feit @eleafeit.bsky.social · 12/02/2025I made it home! Everyone was a little bleary-eyed on Monday. 🦅🦅🦅 000
Reposted by Elea McDonnell FeitAEA Data Editor @aeadata.bsky.social · 03/02/2025We are seeing some impact on reproducibility efforts with the transition to the new US administration. For now, I suggest that you make backups of any data that required you to actively interact with the Federal government, as there may be a delay in response times.... #openscience 29856
Elea McDonnell Feit @eleafeit.bsky.social · 31/01/2025That might do it. I'm finding it really annoying that ChatGPT writes in markdown and it doesn't paste well into Google Docs. 100
Elea McDonnell Feit @eleafeit.bsky.social · 31/01/2025Nice! I’m worried I’m not organized enough for this. Sometimes my content is just a list of links to papers. (This is a seminar for PhD students.) 000
Elea McDonnell Feit @eleafeit.bsky.social · 31/01/2025For my other courses, I create a master Google doc that I share with students which includes links to other course materials. 000
Elea McDonnell Feit @eleafeit.bsky.social · 31/01/2025Toying with the idea of using RStudio + #quarto + github to manage content for my doctoral seminar on causal inference. I'd like to be able to create short "explainer" documents that integrate math + code + hyperlinks. I don't want this to be a major project. Thoughts on design and workflow? 411
Elea McDonnell Feit @eleafeit.bsky.social · 29/01/2025High school civics teaches us that through discussion and compromise we achieve government outcomes that serve as many citizens as possible. 000
Elea McDonnell Feit @eleafeit.bsky.social · 29/01/2025Regardless of where you stand politically, the separation of powers is a fundamental principle of American democracy. The President should pursue policy goals by working with Congress—not through unilateral actions that undermine the constitutional process and sow uncertainty. 130
Elea McDonnell Feit @eleafeit.bsky.social · 23/01/2025Hey Science Queen @ekoenig8.bsky.social: I've noticed the snow melting this week, even though it has been well below freezing. How does this happen? Is the snow sublimating? Or does it melt than freeze? TIA! 100
Elea McDonnell Feit @eleafeit.bsky.social · 06/01/2025I'm teaching "Calling BS: The Art of Skepticism in a Data-Driven World" as a honors seminar at Drexel. 010
Elea McDonnell Feit @eleafeit.bsky.social · 04/01/2025Use code to express things you might otherwise express with math. 010
Elea McDonnell Feit @eleafeit.bsky.social · 03/01/2025Where did you get this chart? The highest county-level proportions reported on the Urban Institute dashboard are around 30%. apps.urban.org/features/deb... areapps.urban.orgDebt in America: An Interactive MapThis map shows the geography of debt in America at the national, state, and county levels. How does your community compare? 010
Elea McDonnell Feit @eleafeit.bsky.social · 02/01/2025This is a big contrast to null hypothesis testing, which weights all type-I errors equally and totally ignores type-II errors. All of the above papers show that return-aware strategies use *much* smaller sample sizes than the hypothesis testing you learned in college. 030
Elea McDonnell Feit @eleafeit.bsky.social · 02/01/2025All these decision theoretic approaches to A/B testing are “return-aware” (as Sudijono et al. put it). Statistical errors are weighted according to their effect on profits. 130
Elea McDonnell Feit @eleafeit.bsky.social · 02/01/2025Finally, if you really, really want to keep using p-values, this working paper by Sudijono, Ejdemyr, Lal and Tingley shows how to solve Azevedo et al.’s “A/B Testing Problem” b tuning the hypothesis testing confidence level based on the prior to replicate the Bayes optimal solution. 130
Elea McDonnell Feit @eleafeit.bsky.social · 02/01/2025Azevedo et al. 2020 demonstrate that the prior can affect Bayes optimal strategies. So, I say, “if you’ve got priors, use them.” And if you’ve got data on past A/B tests, you use it to form priors. 120
Elea McDonnell Feit @eleafeit.bsky.social · 02/01/2025Which brings me to Azevedo et al. (2020), who ask: If you have many ideas to test, should you run many small tests or fewer large ones? Solving for Bayes regret, they show it depends on whether your prior has fat tails. If you expect many high-performing ideas, then you should run smaller tests. 120
Elea McDonnell Feit @eleafeit.bsky.social · 02/01/2025Bayes regret or minimax regret, which should you use? Minimax regret has the convenient "feature" that it does not require a prior on the treatment effect. But that could be considered a "bug" for companies that have a large number of past A/B test results and can use them to set priors. 110
Elea McDonnell Feit @eleafeit.bsky.social · 02/01/2025Joo and Chiong’s recent working paper provides a Gaussian approximation to the regret function that you can use with ✨any✨ asymptotically normal estimator. This means you can use the minimax-regret criteria with your favorite treatment effect estimators: diff-in-diff, ML estimators. 141
Elea McDonnell Feit @eleafeit.bsky.social · 02/01/2025I recently came across Manski 2019 that provides an accessible introduction to the minimax regret approach developed in a series of papers by Manski and Tetenov. They focus on applications to clinical trials, but A/B tests are analogous. (Love _The American Statistican_ BTW.) 120
Elea McDonnell Feit @eleafeit.bsky.social · 02/01/2025Some of my followers will know Feit and Berman 2019. (How has it been five years?). Ron and I describe A/B tests as a decision problem where the goal is to maximize profit (minimize Bayes regret) and derive the optimal sample size over priors which we estimate from collections of past A/B tests. 130
Elea McDonnell Feit @eleafeit.bsky.social · 02/01/2025Lots of folks are discovering the relevance of decision theory to the practical problem of analyzing an A/B. A quick🧵 1326
Elea McDonnell Feit @eleafeit.bsky.social · 02/01/2025These figures describing the plasmode data generating process are nice, too. [from bmcmedresmethodol.biomedcentral.com/articles/10.... ] 010
Elea McDonnell Feit @eleafeit.bsky.social · 02/01/2025Methods papers in Marketing Science and Management Science do simulations inspired by real data, but I've never seen it called plasmode simulation. 100
Elea McDonnell Feit @eleafeit.bsky.social · 29/12/2024Description: This course teaches critical thinking skills to combat misinformation. You will learn to identify common forms of misinformation, evaluate evidence and arguments, and critically evaluate claims you encounter in your personal and professional lives. 010
