Alexander
LightNews LightNews
banner
etotheipie.bsky.social
Alexander
@etotheipie.bsky.social
440 followers 1.2K following 2K posts
Texas A&M Statistics PhD Student. Sex Critical—Trans Rights. Climate Physics. Chess. He/Him/They/Them.
Posts Replies Media Videos
etotheipie.bsky.social
I have another thing to say about this, specifically relating to the graphic. This is NOT the accommodations effect in raw GPA scale. This is the rotated GPA effect, removing the decreasing trend prior to accommodation. The attached graph shows non-rotated effects on raw GPA scale.
Figure 6(a), "Event Study of Accommodations Approval [on Semester GPA]", and 6(b), "Event Student of Accommodations Approval (At Least 3 Pre/Post Periods)". The caption reads: "Notes: This figure shows event study plots of the effect of accommodations approval on current semester GPA for students who were approved for accommodations in the first semester they applied. Controls include student and college entry cohort x semester of application FE. The gray line is the pre-treatment period trend. The sample includes students from Fall 2021-Winter 2025." Notably, the pre-accommodation trend is down, starting at +0.14 GPA for semesters ≤4 before approval semester, and reducing to +0 at semester prior to the one accommodation is approved. The "time zero" point is then the last semester where a student did not have an approved accommodation. Unlike the later Figure 7, which shows the effects after removing the pre-accommodation trend, the GPA effect in 6(a) is ≤0.1 in all post-approval semesters, slightly dampening out past 3 semesters; and in 6(b) it decreases to negative "effect" at 4+ semesters post-approval.
November 15, 2025 at 7:40 PM
etotheipie.bsky.social
Food fighters unite
“Soup for my family” guy, Daunte Wright. He’s wearing a black balaclava and dark, puffy winter coat, holding a can of soup and winking directly at the camera.
November 6, 2025 at 8:16 PM
etotheipie.bsky.social
Death playing the Claw Machine game, standing with his grinning skull face at his selection: “Dick Cheney? Eh not bad”
November 4, 2025 at 12:04 PM
etotheipie.bsky.social
Kay being like “I had no idea this was dog shampoo, it was not clearly labeled” will forever be the funniest damn thing to me.
A picture of the yellow and white colored dog shampoo bottle Jonathan Kay used for several months. Specifically, the picture is a zoom-in on the Labrador Retriever dog on the front of the bottle, with pale golden fur and a happy panting face. The shampoo has coconut scents; if I recall correctly, it is an Arm and Hammer product.
November 3, 2025 at 6:39 PM
etotheipie.bsky.social
The Hex codes look all right. Here are the RGB codes, from top-left going across.
R code showing color palette using listed hex codes, along with printed "col2rgb" results. The hex codes are respectively: #EBA843 #2F3C5F #DA4B34 #3B7EA4 #E16C54 #5C4C9A #E5EEEE #010101 #606161 and the RGB codes are respectively: #EBA843 235 168 67 #2F3C5F 47 60 95 #DA4B34 218 75 52 #3B7EA4 59 126 164 #E16C54 225 108 84 #5C4C9A 92 76 154 #E5EEEE 229 238 238 #010101 1 1 1 #606161 96 97 97
September 29, 2025 at 2:59 PM
etotheipie.bsky.social
Anakin and Padme 4-panel meme: 1. [Anakin] Expert bodies in the executive branch cannot be given deference on interpreting and implementing laws 2. [Padme, elated] Oh so that's Congress' power and it can create bodies like that, right? 3. [Anakin stares] 4. [Padme, concerned] That's Congress' power and it can create bodies like that, right?
September 22, 2025 at 9:28 PM
etotheipie.bsky.social
Does anyone remember that Trump sued the WSJ for reporting on the Epstein birthday letter, and JD Vance complained that nobody would ever prove their claims? Now it's been released, aaaaand?

I feel scandals like this—the lying bluster—sometimes get overlooked in the chaos that is the Trump admin.
Tweet from JD Vance on July 24, responding to Chris Hayes who asked "The contention is that the book is entirely fake? Or Trump being a contributor is fake, or all that is real but someone created a totally fake Trump letter and stuck it in the real book?" Vance: "The contention--which is pretty clear if you just read what I said--is that it's bogus for the WSJ to publish a hit piece without showing us the letter. I have no idea if the book exists--WSJ won't show it to us. I have no idea if the letter exists--WSJ won't show it to us. What I find absurd is the idea that Donald Trump was writing poems to Epstein, and I find it equally absurd that a major American paper would attack the President of the United States without revealing the basis for the attack. We all know what's going to happen. They're going to dribble little details out for days or weeks in an effort to assassinate the president's character. They won't show us this book or allow us to refute it until they've wrung every bit of fake news out of the story. And everyone will just move on from the fact that the WSJ is acting like a Democrat SuperPAC. It's disgraceful, and it's why the president sued."
September 12, 2025 at 2:09 AM
etotheipie.bsky.social
This is my letter which I will deliver to Texas A&M President Welsh calling on him to resign from his position, following his politically pressured removal of the Dean of the College of Arts and Sciences and the Head of the English Department, and his unbelievable pretextual reason for doing so.
Dear President Welsh, Your predecessor President Banks resigned in disgrace following revelations that she was directed by the Board of Regents to interfere with the hiring decisions of faculty and department leadership, for political reasons; that she acquiesced to this pressure; and that she subsequently lied to the Faculty Senate about her involvement. The State of Texas' recent dissolution of the Faculty Senate has emboldened the Board of Regents, disgruntled students, political operatives, and now apparently University administration to further interfere with academic staffing decisions and leadership in order to target undesirable faculty members. You now have, it seems, been directed by the Board of Regents, or other political actors, to interfere in this manner; you have acquiesced to this pressure, removing the Dean of my College and the Head of a sister Department, to target an individual faculty member operating professionally and well within her academic purview; and you have given an unbelievable pretext for your decisions, which no member of the the Texas A&M academic community lacks the decency to accept. If you cannot thwart attempts to be made a tool of undoing, you should similarly resign your position. Sincerely, Alexander Coulter (Department of Statistics College of Arts and Sciences Texas A&M University)
September 9, 2025 at 1:59 PM
etotheipie.bsky.social
It is a multiple choice question survey, with this being one of the questions. The premise is the software system would take about 2.7-5.4 billion USD to set up. Seems like a small price to pay for free filing in perpetuity! Intuit Consumer Group (includes TurboTax) revenue in 2023 was $4.9 billion.
Image of question from IRS survey on Direct File system: “While a program operated by the IRS is free to you, setting up and running the program is expected to have an initial cost to the federal government of at least $10-$20 per return processed. Knowing this information, which of the following statements applies to you?” [option 1] I am more likely to choose this IRS free online tax preparation program [option 2] I am neither more or less likely to choose this IRS free online tax preparation program [option 3] I am less likely to use this IRS free online tax preparation program” I have highlighted option 1.
August 30, 2025 at 5:16 PM
etotheipie.bsky.social
Shining The Symbol Onto The Clouds
Anti-Realist Man symbol: a “blackboard R” symbol for the set of real numbers, with a red circle around and strike through it.
August 22, 2025 at 2:29 PM
etotheipie.bsky.social
Comparison of old (left) and new (right) Cracker Barrel logos. The old logo has an older man sitting on a chair, leaning against a large barrel set next to him. The logo contains a bean-shaped outline of stylized “Cracker Barrel” name, including the subtitle “OLD COUNTRY STORE”. The new logo is just “Cracker Barrel” in a similar font styling, in a plainer yellow hexagonal background; the man and barrel and subtitle are gone.
August 22, 2025 at 12:56 AM
etotheipie.bsky.social
“You may not disobey us even if you think our reasoning was bad!”

the reasoning (two pages!):
The two pages (that’s all!) of “reasoning” given by the majority in California v. Dept. Ed, 2025. Link here: https://www.supremecourt.gov/opinions/24pdf/24a910_f2bh.pdf
August 21, 2025 at 9:52 PM
etotheipie.bsky.social
I got an Obamacare Check for $74.45. Thanks Obama!
A letter detailing why BCBSTX is rebating me part of my premiums, because it didn’t spend at least 80% of the premiums it collected in the last several years. The rule is from the Affordable Care Act, “Obamacare”. A picture of the check amount, of $74.45.
August 18, 2025 at 11:38 PM
etotheipie.bsky.social
This premise is basically false. Support for marriage equality is a very recent phenomenon and support percentages are only about 5% off support percentages for abortion. “Much wider”? No.
Polling showing marriage equality was at 26% (ish) among adults in 1997 and has increased to 68% in 2025, dropping in recent years. Polling showing support for abortion rights among US adults has been more steady, at 60% in 1995 to 63% now.
August 12, 2025 at 1:07 AM
etotheipie.bsky.social
A stats post: very strangely, operations with the Matrix package in R are so efficient for large objects that "crossprod"-ing a long dense vector with an entirely dense dgCMatrix ones-vector is faster than base "sum".
R text and output, showing that summing a 1-million long vector is fastest done by crossprod-ing it with a fully dense dgCMatrix vector of 1's, instead of using base "sum" operation. Using "sum" is approximately 1.64 times slower. R code: set.seed(31) library(microbenchmark) library(Matrix) n = 1e6 ones = rep(1, n) onesSparse = sparseMatrix(i = 1:n, j = rep(1, n), x = rep(1, n)) x = rnorm(n) # The following operations are equivalent: sum(x) # -1375.999 crossprod(ones, x) # -1375.999 crossprod(onesSparse, x) # -1375.999 # times microbenchmark("sum" = sum(x), "Xprod" = crossprod(ones, x), "sparse_Xprod" = crossprod(onesSparse, x), times = 10, unit = "rel") # Unit: relative # expr min lq mean median uq max neval # sum 1.652303 1.655201 1.637911 1.644753 1.644046 1.541583 10 # Xprod 1.985430 1.979733 1.958493 1.973892 1.963744 1.826737 10 # sparse_Xprod 1.000000 1.000000 1.000000 1.000000 1.000000 1.000000 10
August 1, 2025 at 5:12 PM
etotheipie.bsky.social
Marge Simpson holding a potato meme, [My para/social online relationships] I just think they’re neat.
July 12, 2025 at 4:24 AM
etotheipie.bsky.social
Here's a graph setting degree = 4. More knots are needed near the split point to accommodate the higher order. Here I've chosen knots at 0.48, 0.5, and 0.52. The choice of knots here is not a science; I wouldn't go as far to say it's an art. It does provide models with nice continuity properties!
Same as plot(s) in previous comment, with general takeaway being the fits are a little more wiggly due to higher order fitting. Curves still diverge starting at first non-boundary knot, 0.48 (bottom right panel).
July 8, 2025 at 2:52 PM
etotheipie.bsky.social
Here's a graph of the splitting idea. The true models are parabolas (or lower order), as well as the fitted models, so there's decent visual overlap. But the same can be done by fitting higher order models, too. What is the easiest way to share this code? I can email it if you want to DM.
4-panel graph illustrating simultaneously fitted diverging spline curves. Data pattern is Gaussian noise around 1 of 2 paths: time [0, 0.5] parabola, time [0.5, 1] either continuation of parabola or flat curve. True curves are co-smooth, so desire fitted curves that are also co-smooth. Top left: data points and true path 1 highlighted in red, showing the continuation in parabola after split at time point 0.5. Top right: data points and true path 2 highlighted in blue, showing the flat path continued from the apex of the parabola after time point 0.5. Bottom left: fitted simultaneous spline curves. Boundary knots are placed at t = 0, 1, and extra knots are placed near the time split point, at 0.48 and 0.52. The two fitted curves - red and blue - represent estimated "lifetime paths" for points that are (or perhaps, would eventually be in time) on path 1 or path 2, respectively. The curves are visually identical in the first half, representing shared path until the break point. The curves are visually separate, respectively a continuation of the original parabola and a flat line. The curves visually smoothly diverge from the identical parent curve. Bottom right: zoomed-in version of graph 3, to show how paths diverge beginning at knot point 0.48. They are identical before that. Curves *must* have this buffer region around 0.5 to smoothly merge/diverge, but this region can be made smaller.
July 8, 2025 at 2:11 PM
etotheipie.bsky.social
From Amstat News—wtf? Sundown towns aren’t some cute folksy thing!
From recent edition of Amstat News, “Top Ten Signs You Are a Texas Statistician: Amstat News continues its entertaining offering by ASA Executive Director Ron Wasserstein, who delivers a special Top 10—one that aired during a recent edition [of…]” A sketch of Texas and a star, with “#01: You don’t just reject the null hypothesis; you tell it to get out of town before sundown.”
June 19, 2025 at 7:12 PM
etotheipie.bsky.social
I don't really understand Table 5 in the Hisle-Gorman et al. 2021 study. It gives two columns of "crude rate of visits per year" before and after initiation of gender-affirming pharmaceutical treatment, but its "All Mental Health Visits" counts don't equal the column sums.
Table 5. Transgender and gender-diverse youth mental healthcare and psychotropic medication use following initiation of gender affirm- ing pharmaceutical treatment as compared to before initiation Mental healthcare visits (N = 963) Crude rate of visits per year Before After Adjusted* IRR [95% CI] All Mental Health Visits 5.50 6.10 1.04 [0.90-−1.20] Adjustment 0.94 0.83 0.89 [0.67−-1.18] Anxiety 0.98 1.04 1.07 [0.84−1.35] ADHD 0.50 0.20 0.40 [0.27−-0.58] Cognitive 0.02 0.02 0.83 [0.40−1.75] Developmental 0.03 0.01 0.35 [0.16−0.78] Infancy 0.37 0.53 1.02 [0.41−2.54] Impulse 0.001 0.01 0.10 [0.02−0.53] Mood 2.90 2.33 1.12 [0.94−1.35] Personality 0.02 0.03 1.40 [0.44−-4.39] Psychotic 0.13 0.16 0.99 [0.48−2.06] Alcohol Abuse 0.13 0.06 0.66 [0.15−2.87] Substance Abuse 0.05 0.12 1.39 [0.68−2.85] Suicide 0.07 0.12 1.74 [1.18−2.56] Miscellaneous 0.09 0.19 1.45 [0.56−3.60]
May 1, 2025 at 8:32 PM
etotheipie.bsky.social
These couple sentences get to the heart of the matter. The nervous system (and hence the brain) takes in a ton of information, but it seems the "sifting" the brain does is not considered "behavioral throughput". (It's also not clear whether recall is supposed to be sensory or not.)
Excerpt from paper: "For example, the photoreceptor cells in the human eye convert dynamic light input to a continuously varying membrane voltage. Using the approach described in Appendix B, one finds that one human cone photoreceptor can transmit information at ≈ 270 bits/s. The population of 6 million cones in one eye has a capacity of about 1.6 gigabits/s. From that enormous bitstream, the brain of the typist sifts out just the 10 bits/s needed to perform the behavioral task correctly. To represent this degree of filtering, we define the dimension-less “sifting number”: [...]"
December 26, 2024 at 2:39 PM
etotheipie.bsky.social
The article above (written by Janice Turner, with a second op-ed about annoying seals) with all words blocked out except “Janice Turner a Flipping Nuissance”
December 19, 2024 at 8:39 PM
etotheipie.bsky.social
In fact, just checking now, ChatGPT says several false things here. These are canned responses based on propagated falsehoods in introductory texts. Sample size n ≥ 30 is not the minimal sample size, this is simply not true! The rate of CLT convergence depends on the underlying distribution.
ChatGPT: "The minimal sample size required to use a Z-test depends on whether the population standard deviation ( σ σ) is known and if the sample size ensures the sample mean follows a normal distribution. General Guidelines: Known Population Standard Deviation ( σ σ): The Z-test assumes the sampling distribution of the sample mean is approximately normal. For large enough sample sizes ( n ≥ 30 n≥30), the Central Limit Theorem ensures normality regardless of the population distribution. If the population is already normally distributed, smaller sample sizes can be acceptable, even n < 30 n<30. Unknown Population Standard Deviation ( σ σ): If σ σ is unknown, a t-test is more appropriate unless n n is very large ( n ≥ 30 n≥30), in which case the t-distribution approaches the Z-distribution. Practical Considerations: For a Z-test to be valid: The sample size should be n ≥ 30 n≥30 unless the population is normal. If the sample size is smaller ( n < 30 n<30) and the population distribution is not normal, the Z-test is not reliable."
December 15, 2024 at 6:48 PM
etotheipie.bsky.social
calculate sample standard deviations with/without Bessel's correction, erroring out instead of getting the wrong answer. Now, I think that is a much better outcome than simply stating the wrong answer, to be sure. But this is in principal an easy calculation, and it's not clear why it stopped here.
ChatGPT request to find the sd of 20 numbers, separately with and without applying Bessel's correction, which divides by (n-1) instead of (n). ChatGPT says "It seems like I can’t do more advanced data analysis right now. Please try again later. If you'd like, I can guide you on how to calculate the standard deviations manually or with a specific tool like R or Python. Let me know how you'd like to proceed!"
December 15, 2024 at 6:44 PM
etotheipie.bsky.social
It’s actually part of your profile button banner thing! So I think we need a collection of people first.
Screenshot of my profile page, showing what it looks like to select “Starter Packs” on your button banner next to Media/Likes/Feeds. You can Create your own.
November 11, 2024 at 3:51 PM