Still pondering this one, but I wanted to offer a couple of thoughts on the issue of standard errors and coverage based on some on-the-fly simulations I did in the middle of the night. (1 / 8)

https://twitter.com/RobinGomila/status/1309865417243086848

The issue with standard errors when using OLS to model a binary outcome with categorical predictors (which is a key assumption in this paper) is that the errors are subject to groupwise heteroskedasticity, which can result in either upward or downward bias in the SEs. (2 / 8)

I am looking forward to checking this one out. Giving this a quick look, it seems like the focus is overwhelmingly on the question of bias in the point estimates. (1 / 5)

https://twitter.com/RobinGomila/status/1309865417243086848

The question of bias in the standard errors is addressed using robust standard errors. In other words, it is addressed in the calculation of the SEs, but it is not addressed as part of the simulations. (2 / 5)

My latest post over at @BlogBroadstreet is up! Building on my earlier post on the endogeneity of historical data, I discuss the connection between the creation of historical data and the organization of political space in the American West. (1 / 4)

broadstreet.blog/2020/09/25/the… The main idea is that entities such as states and counties are not neutral units of observation. Political actors not only had a material incentive to continually divide units, they had an incentive to divide units particular ways. (2 / 4)

I'm a bit behind on promoting my @BlogBroadstreet pals, but wanted to highlight this one right away for anyone who has been watching the recent discussion around the appearance of the Turner thesis in economics.

https://twitter.com/jaredcrubin/status/1308036576216145921

It is tempting to want to describe this as history vs. economics, but I don't think that particular framing lends itself to good faith intellectual engagement. I would urge interested parties to do two thing...

I'm behind on repping my new pals over at @BlogBroadstreet, but I thought I would take a minute to spread the word about all the new posts this week!
broadstreet.blog First up, we've got @jaredcrubin with his second post in a series on culture and the economy. In this post, Jared discusses work on trust and kinship in historical political economy, with emphasis on recent research in the field of economics.

broadstreet.blog/2020/09/07/fro…

UVA is going forward with its plans to do face-to-face instruction. From what I can tell, however, most classes are online. I don't see how this is anything other than an exercise in waiting for people to get COVID. From the university:

"We care about the health and well-being not just of our faculty, staff, and students, but of our neighbors in the Charlottesville region. There are no easy answers here, and there are no risk-free paths."

My first post for @BlogBroadstreet is up! I discuss the relationship between quantitative research and historical sociology, before moving on to what I refer to as the endogeneity of historical data.

broadstreet.blog/2020/08/28/the… When I talk about the endogeneity of historical data, what I mean is that the story of how seemingly neutral quantitative data came to be is often bound up with the story that the researcher is trying to tell.

Working my way through @36ftv and just got to the DP4 episode, which is the first point at which the hosts respond to listener comments on prior episodes. I felt so vindicated to learn that I wasn't the only one who took issue with the hot take on Half Step from the DP1 episode. My super-specific beef was not so much that one of the hosts didn't love Half Step, but that they would pan Half Step in the DP1 episode and love Sugaree in the DP3 episode.

I'm super excited to teach this piece by @WesternBruce, Deirdre Bloome, and Christine Percheski: journals.sagepub.com/doi/abs/10.117… There's so many things to love about this paper. I think the thing that I like most is that they show that variation in the residual variance (i.e. heteroskedasticity) is substantively interesting.

Getting ready to teach marginal effects for non-linear probability models. We don't average over heterogenous effects when using interactions, so why do we insist on doing it with a logit, probit, etc.? In the face of observable non-linearity, the AME is less than ideal. @vaiseys If you want marginal effects, the first thing would be to make a histogram of the case-level marginals and see whether the AME is a substantively reasonable summary of what is happening in the data. If it is, then go ahead with the AME.

I have a lot of feelings about this thread, from the wording of the original survey to the subsequent debate about the merits of the LPM.

https://twitter.com/brendan642/status/1201693938706460672

Some people seem to want to interpret the results of the logit version of the survey to mean that the prevalence of odds ratios in the context of logistic regression have destroyed our understanding of what it means for something to be more likely.

If you follow the link in fn.15, you will find that authors appear to have missed the fact that coefficient on the offset term is being constrained to 1.
github.com/lilleymatthew/…
(1/4) As their own footnote suggests, including a control for the log of population is not the only correct way of controlling for population in a negative binomial model. (2/4)

The data for the Trump rallies replication are available now. I took a quick look and found that, yes, the effect of Trump rallies becomes insignificant once you control for the log of the county population. github.com/lilleymatthew/… It is not clear to me that this is how one should correct for population if the goal is simply to correct for size. If you include population as an exposure, the effect of Trump rallies is still significant, though much smaller than it is an a model without an exposure term.

There appears to be A LOT of action around this post, which is being cited as if it were an actual study. I feel like the real study here involves watching political polarization around the dissemination of quasi-facts.
reason.com/2019/09/06/no-… As I made clear on another thread, I have no reason to doubt the reported findings. I went looking for the data because I thought it would be useful for teaching about exposure in a count model.

@treycausey Got it. I don't pay super close attention to this, by the most reasonable methodological critiques are not unique to big data. @treycausey The last one should read "my sense is that the most reasonable..."