Thread by @ProfPaulPoast

Good question!

Lots of ways to measure "attention": cites, media outreach, speaking engagements, placed on syllabus, or policy engagement

Sometimes these go together, but not always/often: paper could generate buzz, but not be cited; paper could be cited a lot because its wrong

https://twitter.com/cullenhendrix/status/1195019503916048384

So which paper of mine do I feel "under performed" or, stated differently, "flew under the radar"?

One of my methods pieces, namely my 2013 @polanalysis piece w/ Walter Mebane

cambridge.org/core/journals/…

The title probably didn't help: I mean it's just loaded with words & phrases that make you go 🙄😴

"Causal Inference"

"Ignorability"

"Identification"

"nonrandom assignment"

...aaaaannnnddd I'm asleep.

But the point of the paper was a useful one and influenced some of the analysis in my 2012 @IntOrgJournal piece on alliances, trade, and issue linkage

cambridge.org/core/journals/…

A key goal of the @polanalysis piece was to generate greater awareness in (and possibly use of) of "Manski Bounds", developed by the economist Charles Manski, in political science

The idea is this: without any modeling (e.g. linear regression) can you figure out a range of values capturing the relationship between variable X and variable Y? 🤔

The range of relationship values could be quite wide and, without assumptions, nearly always include 0.

Let's be honest, social scientists aren't too keen of "null results" (hello, publication bias).

nature.com/articles/s4156…

Manski's overall goal, in his words, was to encourage social scientists to "embrace uncertainty"

In other words, he wants social scientists to be cool with null results. Again, that's a tough case to make.

The first step is to derive the bounds...

...and then progressively adding assumptions. These assumptions can include:

1) monotonicity (the effect is never negative; the effect is never positive)

2) no missing data

3) so long as we include relevant covariates, X is "as if" randomly assigned

The last two assumptions are what we typically mean by "ignorability assumptions": we can IGNORE missing data, selection, or the fact that the data were NOT generated by an experiment.

The more assumptions you add, the more precisely you can IDENTIFY the effect (technically, the Average Effect) of X on Y.

Of course, these are strong assumptions.

So our paper provides a way, building on the work of Francesca Molinari, to compute Manski bounds in the presence of missing data

amstat.tandfonline.com/doi/abs/10.119…

...and we show how to compute sensitivity analysis on those bounds, meaning what happens if you nudge the data by assuming more or less "missingness" in the data

We then discuss the application of these bounds to civil war onset due to repression...

...by drawing on this working paper by Sambanis and Zinne

citation.allacademic.com/meta/p_mla_apa…

I also wrote up simple R script & Stata Do files that compute the bounds for your data.

paulpoast.com/missing-treatm…

In the end, it hasn't really taken off.To be honest, one barely sees applications of Manski bounds in Social Science research, even though he's projected to win a @NobelPrize in economics in the near future.

So I guess this paper "underperforming" isn't all that bad 😉

[END]