For example if you want to observe and measure some attribute of classes at public schools in your county, it might be infeasible to send data collectors to all of 15 schools, but the marginal cost of measuring additional classrooms at the same school once you're there is minimal. So, how many schools should you visit and how many class rooms per school given a budget and assumptions on inter and intra school variation?

We had had a group assignment to estimate the average circumference of trees on campus. Our initial plan something like 1) get a map of campus and split it into zones 2) sample zones randomly 3) everyone goes to a few (small) zones and tries to roughly map out the trees there 4) sample again from those trees and physically measure them. This would mean running around campus for at least a few days if we wanted to an honest job. And it was a rainy spring in North Carolina.

However, one of my group mates had a stroke of brilliance and decided to email the grounds department. To our surprise they were able to provide us with a full list of every known tree on campus as well as GIS data with locations. So we were able to do a legitimate simple random sample which was optimally efficient from in terms of both variance and time-in-rain.

In conclusion I'm pro list-of-individual-trees.

The 16e is the smallest phone in the current lineup, which is bigger than both.

I really like the 13 mini, I had been on Android since since maybe iphone 4. Not sure what I'll do when my current phone dies, I don't like having a big phone in my pocket.

[0] https://spark.apache.org/docs/latest/sql-programming-guide.h...

Somewhat analogous to how the pandas api can be used in pyspark via pyspark pandas, but from the api -> implementation direction rather than (pandas) implementation -> api -> implementation maybe?

https://arstechnica.com/cars/2024/04/tesla-stops-cybertruck-...

This is a huge problem and in my opinion is mostly due to bad incentive structures and bad statistical/methodological education. I'm sure there are plenty of cases where there is intentional or at least known malpractice, but I would argue that most bad research is done in good faith.

When I was working on a PhD in biostatistics with a focus on causal inference among other things, I frequently helped out friends in other departments with data analysis. More often than not, people were working with sample sizes that are too small to provide enough power to answer their questions, or questions that simply could not be answered by their study design. (e.g. answering causal questions from observational data*).

In once instance, a friend in an environmental science program had data from an experiment she conducted where she failed to find evidence to support her primary hypothesis. It's nearly impossible to publish null results, and she didn't have funding to collect more data and had to get a paper out of it.

She wound up doing textbook p-hacking; testing a ton of post-hoc hypotheses on subsets of data. I tried to reel things back but I couldn't convince her to not continue because "that's how they do things" in her field. In reality she didn't really have a choice if she wanted to make progress towards her degree. She was a very smart person, and p-hacking is conceptually not hard to understand, but she was incentivized to not understand it or to not look at her research in that way.

* Research in causal inference is mostly about rigorously defining the (untestable) causal assumptions you must make and developing methods to answer causal questions from observational data. Even if an argument can be made that you can make those assumptions in a particular case, there is another layer of modeling assumptions you'll end up making depending on the method you're using. In my experience it's pretty rare that you can really have much confidence that your conclusions about a causal question if you can't run a real experiment.

(Bandits are often presented as a loop where you control the policy collecting data and update it frequently, but that does not have to be the case.)

Recommender systems and search/counterfactual learning-to-rank can be thought of as an extension to counterfactual bandits as well.

Makes me wonder if I'm missing the point of the fairly simple post?

I agree, changing the model changes the estimates, because the parameters you are estimating change.

However, given one misspecified model, the parameters of that model are still well defined, though they may not have the interpretation they would if the model was correctly specified. As OP called it, this is the "best fit line", and is a projection of the truth onto your model. E.g. for a simple linear regression of Y on X, where the true conditional mean of Y given X is not linear, there is still some "true" best line. This line depends also on the distribution of X, though it would not if the model was correct. Estimates from linear regression will converge to the parameters of this line, though using the usual standard errors will be wrong.

There's a very general theorem or corollary that covers this in Asymptotic Statistics by van der Vaart. I think in the chapter about M estimators, right around where MLEs are covered, but I don't have it in front of me.