Greg Faletto
LightNews LightNews
gregoryfaletto.com
Greg Faletto
@gregoryfaletto.com
410 followers 740 following 700 posts
Statistician, Data Scientist.
Posts Replies Media Videos
gregoryfaletto.com
whoa major heel turn
A screenshot of a social-media post showing the display name “SE Gyges,” the handle “@segyges.bsky.s…,” and a circular profile image with a gray geometric pattern. Below the name and handle, the post text reads: “i think thielism is good.”
December 7, 2025 at 7:56 PM
gregoryfaletto.com
It's crazy how many people think Biden failed to cancel any student loan debt. $190 billion! www.nasfaa.org/news-item/35...
**Alternative text:** A screenshot of a news article titled “Biden Administration Announces ‘Final’ Student Loan Debt Relief Approvals.” The byline reads: “By Maria Carrasco, NASFAA Staff Reporter.” The article explains that shortly before the Trump administration takes office, the Biden administration announced its final approvals of student loan forgiveness through the Income-Based Repayment (IBR) plan and borrower-defense claims. It states that the Department of Education approved $600 million in forgiveness for 4,550 borrowers under IBR and additional forgiveness for 4,100 former DeVry University students, with no total amount specified for that portion. The article concludes by noting that the administration has approved $188.8 billion in forgiveness for 5.3 million borrowers since taking office.
December 6, 2025 at 5:52 PM
gregoryfaletto.com
Total inference cost is [cost per token] times [tokens used]. AI companies want the first number to go down (ideally a lot) and the second number to go up (ideally a lot). So total inference cost isn't a great metric.

First number is in fact going down dramatically (the y-axis is on a log scale).
Source: https://hai.stanford.edu/ai-index/2025-ai-index-report/research-and-development A line chart shows the steep decline in inference prices for various large language models from 2022 to 2025. The vertical axis uses a log scale for price per million tokens, and the horizontal axis shows publication dates. Multiple colored lines represent benchmarks such as GPT-3.5-level MMLU, GPT-4-level HumanEval, PhD-level GPQA, and LMSYS Chatbot Arena Elo. All lines slope sharply downward over time. Models labeled GPT-3.5, GPT-4-0314, GPT-4o-2024-05, Claude-3.5-Sonnet-2024-06, Llama-3.1-Instruct-8B, DeepSeek-V3, Phi 4, and Gemini-1.5-Flash-8B appear along the trend. The chart illustrates dramatic cost reductions, from double-digit dollars per million tokens in 2022 to well under one dollar by late 2024.
November 30, 2025 at 7:54 PM
gregoryfaletto.com
Most people make at least some conscious choices at points in their life in the hopes of eventually getting a more desirable job, like education choices.

"It's hard to do that," "some people are born with advantages," and "there are wide disparities in wages" are all things I would have agreed with
A subway ad for Apex Technical School shows two panels. On the left, large text reads “Choose your trade at Apex,” with photos of two people working: one using a power tool and another leaning over an engine. On the right, a colorful panel displays the Apex Technical School logo, icons representing different trades, a list of benefits such as hands-on training and day/evening classes, the school’s address and phone number, and a “Career training” badge.
November 24, 2025 at 9:22 PM
gregoryfaletto.com
Andy Rooney
November 21, 2025 at 4:03 PM
gregoryfaletto.com
doesn't really matter if you're dead.

Uber reports 153 people died in Ubers in 2021 - 2022 www.uber.com/us/en/about/...
A data table titled “2017–2022 motor vehicle fatality data, Uber-related and United States” compares Uber-related fatality metrics with national metrics across three two-year periods: 2017–2018, 2019–2020, and 2021–2022. Columns list: Year; number of Uber-related fatalities; frequency of fatalities by number of trips; percentage of trips involving fatalities; Uber’s fatality rate per 100 million vehicle miles traveled; the national fatality rate per 100 million miles; and the rate change from the prior period for both Uber and the national figures. Values show Uber fatalities of 107, 101, and 153 in the three periods, with fatality frequencies ranging from roughly 1 in 22 million to 1 in 12 million trips. Uber’s fatality rate increases from 0.58 to 0.62 to 0.87 across periods, while national rates rise from 1.15 to 1.22 to 1.35. Rate-change columns show increases for later periods (e.g., +7% and +40% for Uber; +6% and +11% nationally), with dashes for the baseline period.
November 20, 2025 at 2:41 AM
gregoryfaletto.com
Maybe, but this table wouldn't tell you either way
Meme of rapper Drake expressing preferences. In the top panel, Drake turns away with a disgusted expression, rejecting the label “DIFFERENCE-IN-DIFFERENCES.” In the bottom panel, he smiles and points approvingly at the label “DIFFERENCES.” The meme humorously suggests a preference for simple differences over the statistical method “difference-in-differences.” (it's a sarcastic meme, this table is just showing differences from a handful of cities. At the very least you'd want to look at difference-in-differences over a larger set of cities, and even that would be considered pretty crude by contemporary research standards.)
November 12, 2025 at 8:44 PM
gregoryfaletto.com
the world if non-VC-and-C-suite technology brothers, sisters, and non-binary siblings ran everything
The image depicts a futuristic cityscape with sleek, modern architecture and advanced technology. Tall glass skyscrapers with curved designs rise in the background alongside domed and geometric structures made of reflective materials. Numerous flying vehicles move through the sky, while streamlined trains glide on elevated tracks. In the foreground, a person walks along a landscaped green area near a reflective pool, accompanied by a small robotic dog. The scene is bright and sunny, with a clean, high-tech atmosphere suggesting an advanced, utopian future.
October 29, 2025 at 5:40 AM
gregoryfaletto.com
Or (*cough*) maybe you believe *some* TEs are not unique (e.g. 2 cohorts starting treatment consecutively have the same initial TE) & want to use this to reduce variance, but rather than picking the structure yourself & risking bias you want to use ML to learn it from data arxiv.org/abs/2312.05985
Figure 1 in Faletto (2025) (https://arxiv.org/abs/2312.05985). The image shows a triangular grid plotted on a coordinate system, where the horizontal axis (labeled *t*) represents time and the vertical axis (labeled *r*) represents cohorts. The grid displays points connected by lines, forming a right triangle aligned along the diagonal from the lower left to the upper right. Each point is labeled with a symbol of the form **τ̂<sub>rt</sub>**, representing estimated marginal average treatment effect terms. The points appear as follows: * Along *r = 2*: τ̂₂₂, τ̂₂₃, τ̂₂₄, τ̂₂₅, τ̂₂₆ * Along *r = 3*: τ̂₃₃, τ̂₃₄, τ̂₃₅, τ̂₃₆ * Along *r = 4*: τ̂₄₄, τ̂₄₅, τ̂₄₆ * Along *r = 5*: τ̂₅₅, τ̂₅₆ Each horizontal line connects treatment effects for a given cohort across time, while the diagonal connects τ̂₂₂, τ̂₃₃, τ̂₄₄, and τ̂₅₅. The figure caption explains that this visualization illustrates which estimated treatment effect terms τ̂<sub>rt</sub> are penalized toward each other in the FETWFE model, where *T = 6* and *R = {2,…,5}*. The method assumes that linked treatment effects are close together or equal.
October 26, 2025 at 4:51 PM
gregoryfaletto.com
The short explanation is: DID assumes "parallel trends": a time-invariant selection bias in outcomes between eventually-treated and never-treated units. So you can use pre-treatment observations to estimate the selection bias. (Here β1 is selection bias, β2 is the "seasonality effect", β3 is TE.)
Difference-in-differences diagram. Pre-intervention means: treatment 50 (β₀+β₁), comparison 35 (β₀). Post-intervention: comparison 55 (β₀+β₂); treatment observed 85 (β₀+β₁+β₂+β₃). A dashed green line shows the treatment counterfactual at 70 (β₀+β₁+β₂). Gray lines connect pre to post for observed trends; x-axis labeled “Pre-intervention” and “Post-intervention.”
October 26, 2025 at 6:52 AM
gregoryfaletto.com
This later paper by a similar group was an easier entry point for me and also lays out this specific topic more clearly than the 2019 one IMO
The image shows the title and citation information from the top of an academic journal article. Text reads: **JOURNAL OF COMPUTATIONAL AND GRAPHICAL STATISTICS** **2021, VOL. 30, NO. 2, 503–517** *[https://doi.org/10.1080/10618600.2020.1831930](https://doi.org/10.1080/10618600.2020.1831930)* Below that, the article title appears in large bold blue text: **Local Linear Forests** The author line lists: Rina Friedberg¹, Julie Tibshirani², Susan Athey³, and Stefan Wager³ — with superscript letters **a**, **b**, and **c** in blue following the names. The image shows a section of an academic paper titled **“2. Local Linear Forests”** in bold blue text. The passage explains that local linear forests use random forests to generate weights, which can act as kernels for local linear regression. It defines training data ((X_1, Y_1), \ldots, (X_n, Y_n)) with (Y_i = \mu(X_i) + \epsilon_i), and discusses using a random forest to estimate the conditional mean function (\mu(x_0) = \mathbb{E}[Y \mid X = x_0]) at a test point (x_0). The text describes random forests as ensemble methods that average tree predictions. For each tree (T_b) in a forest of (B) trees, the leaf (L_b(x_0)) has a predicted response (\hat{\mu}*b(x_0)), defined as the average response of all training data points assigned to that leaf. The overall prediction is given by [ \hat{\mu}(x_0) = (1/B) \sum*{b=1}^B \hat{\mu}_b(x_0). ] The image shows a section of an academic paper containing text and mathematical equations about random forests viewed as adaptive weight generators. The paragraph cites *Hothorn et al. (2004)*, *Meinshausen (2006)*, and *Athey, Tibshirani, and Wager (2019)* — all references appear in blue. The equations express how the random forest prediction (\hat{\mu}(x_0)) can be written as weighted averages: [ \hat{\mu}(x_0) = \frac{1}{B} \sum_{b=1}^B \sum_{i=1}^n Y_i \frac{1{X_i \in L_b(x_0)}}{|L_b(x_0)|} = \sum_{i=1}^n Y_i \frac{1}{B} \sum_{b=1}^B \frac{1{X_i \in L_b(x_0)}}{|L_b(x_0)|} = \sum_{i=1}^n \alpha_i(x_0) Y_i, ] where the forest weight (\alpha_i(x_0)) is defined as [ \alpha_i(x_0) = \frac{1}{B} \sum_{b=1}^B \frac{1{X_i \in L_b(x_0)}}{|L_b(x_0)|}. ] Equation (4) appears on the right side of the last expression. The text notes that for each (i), (0 \le \alpha_i(x_0) \le 1), and the weights sum to 1 if at least one tree includes a nonempty cell containing (x_0).
October 18, 2025 at 4:10 PM
gregoryfaletto.com
Wait, what exactly is this paper adding that isn't in Athey et al (2019)? (They cite it but don't really spell out the novelty compared to that paper explicitly)
“To avoid this issue, we cast forests as a type of adaptive locally weighted estimators that first use a forest to calculate a weighted set of neighbors for each test point 𝑥 x, and then solve a plug-in version of the estimating equation (1) using these neighbors. Section 2.1 gives a detailed treatment of this perspective.” The image shows a section of a research paper discussing forest-based algorithms for learning adaptive, problem-specific weights. The text cites *Hothorn et al. (2004)* and *Meinshausen (2006)*, both in blue, and presents a mathematical definition for weights (\alpha_i(x)). The equations shown are: [ \alpha_{bi}(x) = \frac{1({X_i \in L_b(x)})}{|L_b(x)|}, \quad \alpha_i(x) = \frac{1}{B} \sum_{b=1}^B \alpha_{bi}(x). ] The text explains that a set of (B) trees is grown, each with leaves (L_b(x)) representing training examples that share a leaf with (x). These weights sum to 1 and define the forest-based adaptive neighborhood of (x). The image shows a line of text from an academic paper. It reads: “Finally, for the special case of regression trees, our weighting-based definition of a random forest is equivalent to the standard ‘average of trees’ perspective taken in Breiman (2001): If we estimate the conditional mean function.” The image shows a mathematical passage from an academic paper defining the conditional mean function for regression trees. It reads: [ \mu(x) = \mathbb{E}[Y_i \mid X_i = x], ] as identified in (1) using [ \psi_{\mu(x)}(Y_i) = Y_i - \mu(x), ] then we see that [ \sum_{i=1}^n \frac{1}{B} \sum_{b=1}^B \alpha_{bi}(x)(Y_i - \hat{\mu}(x)) = 0 ] if and only if [ \hat{\mu}(x) = \frac{1}{B} \sum_{b=1}^B \hat{\mu}_b(x), ] where [ \hat{\mu}*b(x) = \sum*{{i : X_i \in L_b(x)}} Y_i / |L_b(x)| ] is the prediction made by a single CART regression tree.
October 18, 2025 at 7:24 AM
gregoryfaletto.com
Yeah it really sucked when Biden didn't get his way on the child tax credit. en.wikipedia.org/wiki/Build_B...
Screenshot of a Wikipedia article section discussing Senator Joe Manchin’s opposition to extending the child tax credit. The text explains that Senate Democrats supported the enhanced credit, citing reductions in child poverty and food insecurity. Manchin demanded stricter work requirements and criticized reliance on temporary funding. He rejected claims that he wanted to eliminate the credit, calling them false, but expressed concerns about misuse of funds. The Census Bureau is cited as reporting that most families used the payments for essentials like food, clothing, utilities, and education. Numerous reference citations in blue are interspersed throughout the text.
October 1, 2025 at 4:18 PM
gregoryfaletto.com
same flavor as
Headline styled as a debate introduction. It reads: "Point/Counterpoint: This War Will Destabilize The Entire Mideast Region And Set Off A Global Shockwave Of Anti-Americanism vs. No It Won’t." It's from this 2003 Onion article about the Iraq war https://theonion.com/this-war-will-destabilize-the-entire-mideast-region-and-1819594296/
September 3, 2025 at 5:19 PM
gregoryfaletto.com
It's that "age as a predictor" plot from like a month ago! Looks great!
A man sitting in a chair with his legs stretched out, holding a drink can and a cigarette in one hand while pointing with the other. He appears to be reacting strongly to something off-screen.
August 25, 2025 at 3:08 PM
gregoryfaletto.com
This isn't thought through IMO. Author's argument is *in favor of* exposing oneself to those kinds of posts.

It only makes sense if being influenced by writing is something that only happens to other people--not the reader (who she's trying to influence with her writing), and certainly not *her*.
A screenshot of a Twitter post embedded in an article. The article text above the post discusses persuasion, noting that persuadable people are often bystanders who rarely participate in online debates but may be influenced by what they see. It asks whether readers will leave such bystanders exposed only to certain arguments. The Twitter post is from an account called *The Transformed Wife (@godlywomanhood)* and reads: “Men’s role is to provide. Women’s role is to nurture. Our society has lost SO much (especially the children) because most women no longer nurture. They would rather provide. 😟” The post shows engagement metrics: 63 comments, 67 retweets, 210 likes, 6 bookmarks, and a share icon.
August 24, 2025 at 6:48 PM
gregoryfaletto.com
No biggie, but feel free to use it, and if you like it leave a review!
Screenshot of the ChatGPT interface showing the "Alt Text Writer 5" GPT menu open. The dropdown menu lists options including Model, New chat, About, Edit GPT, Hide from sidebar, Copy link, Review GPT, and Report GPT. The "Review GPT" selection is highlighted. On the right, the Alt Text Writer page displays its name, creator "Gregory M Faletto," and description explaining it generates W3C-compliant alt text when an image is uploaded.
August 18, 2025 at 5:06 AM
gregoryfaletto.com
Um I didn't check all the details and some of the photos are obviously off, but I think GPT 5 did pretty well.

(I'm in the habit of checking these kinds of posts because usually they don't replicate on near-SOTA--seem to typically either be based on free or outdated models)
A grid of official portrait photographs of U.S. presidents from 1929 to the present, each labeled with name and years in office. Top row: Herbert Hoover (1929–1933), Franklin D. Roosevelt (1933–1945), Harry S. Truman (1945–1953), Dwight D. Eisenhower (1953–1961). Second row: John F. Kennedy (1961–1963), Lyndon B. Johnson (1963–1969), Richard Nixon (1969–1974), Gerald R. Ford (1974–1977). Third row: Jimmy Carter (1977–1981), Ronald Reagan (1981–1989), George H. W. Bush (1989–1993), Bill Clinton (1993–2001). Bottom row: George W. Bush (2001–2009), Barack Obama (2009–2017), Donald Trump (2017–2021), Joe Biden (2021–present).
August 9, 2025 at 12:42 AM
gregoryfaletto.com
Darn, hasn't propagated to my Plus account yet on web or mobile 😔
A cartoon scene showing Squidward from behind as he peers through six horizontal window blinds. Outside, SpongeBob SquarePants and Patrick Star stand on a light-colored surface, arms raised and mid-jump, as if dancing or cheering.
August 7, 2025 at 10:00 PM
gregoryfaletto.com
...so when do I get to try GPT-5? 😥
A dark-mode screenshot of the ChatGPT model-selection dropdown. At the top it reads “ChatGPT o3” with a downward arrow. Under a “Models” heading are four options: GPT-4o – Great for most tasks o3 – Uses advanced reasoning (checked) o4-mini – Fastest at advanced reasoning o4-mini-high – Great at coding and visual reasoning Below them is an “Other models >” button, which is expanded to reveal two more entries: GPT-4.1 – Great for quick coding and analysis GPT-4.1-mini – Faster for everyday tasks
August 7, 2025 at 9:54 PM
gregoryfaletto.com
Finally, there's now an S3 class for the output of fetwfe(), which works with functions like print() and summary(). You'll notice the output looks nicer now. See, for example, the getting started vignette cran.r-project.org/web/packages...
Screenshot of R code and console output showing the use of the `fetwfe` package to estimate a Fused Extended Two-Way Fixed Effects model on panel data and print a summary that includes an overall average treatment effect (ATT), cohort-specific treatment effects (CATT) for three cohorts, and model details (number of units, time periods, treated cohorts, covariates, features, selected size, and tuning parameter). --- The image is divided into two main sections: 1. **R code:** * `library(fetwfe)` * A call to `fetwfe()` with arguments: * `pdata = pdata` (panel dataset) * `time_var = "time"` (time variable) * `unit_var = "unit"` (unit identifier) * `treatment = "treated"` (treatment dummy indicator) * `response = "response"` (response variable) * A call to `summary(result)` to display the estimation results. 2. **Console output (Summary of Fused Extended Two-Way Fixed Effects):** * **Overall ATT:** 29.4540 (SE = 5.6503, 95% CI = \[18.3796, 40.5283]) * **CATT (preview) table:** * Cohort 11: Estimated TE = 13.93837, SE = 0.50308, 95% CI = \[12.95236, 14.92438] * Cohort 12: Estimated TE = 19.65173, SE = 0.41869, 95% CI = \[18.83111, 20.47235] * Cohort 7: Estimated TE = 56.13475, SE = 0.37109, 95% CI = \[55.40744, 56.86207] * **Model Details:** * Units (N): 30 * Time periods (T): 20 * Treated cohorts (R): 3 * Covariates (d): 0 * Features (p): 55 * Selected size: 17 * Lambda\*: 0.8412 This image illustrates both the code needed to run the estimator and the key numerical results of the fixed-effects analysis.
July 5, 2025 at 8:39 PM
gregoryfaletto.com
Second, if you want to compare these estimators, there's now a set of functions that make it easy to conduct simulation studies. Check out the new vignette for details.

cran.r-project.org/web/packages...
An image showing a screenshot of R code in a monospaced font. The code reads: ``` coefs <- genCoefs(R = 3, T = 4, d = 2, density = 0.1, eff_size = 2, seed = 2025) result_piped <- coefs |> simulateData(N = 60, sig_eps_sq = 1, sig_eps_c_sq = 1) |> fetwfeWithSimulatedData() cat("Estimated Overall ATT from piped workflow:", result_piped$att_hat, "\n") #> Estimated Overall ATT from piped workflow: -0.7045089 true_tes_piped <- coefs |> getTes() # Print the true overall treatment effect cat("True Overall ATT:", true_tes_piped$att_true, "\n") #> True Overall ATT: -0.6666667 ``` The code first generates coefficient values with `genCoefs`, then pipes them through `simulateData` and `fetwfeWithSimulatedData`, printing an estimated average treatment‐on‐the‐treated (ATT) of approximately –0.7045. It then uses `getTes()` on the same coefficients to retrieve and print the true ATT of –0.6667.
July 5, 2025 at 8:39 PM
gregoryfaletto.com
First of all, I added an implementation of extended two-way fixed effects (@jmwooldridge.bsky.social 2021), etwfe(), with inputs and outputs aligned with fetwfe().

There's also now betwfe(), which implements a bridge-penalized (includes lasso and ridge regression) version of etwfe().
A screenshot of an R console showing code and output for an extended two-way fixed effects analysis. The code loads the `fetwfe` and `did` packages, loads the `mpdta` dataset, and transforms it into `pdata` with `attgtToFetwfeDf`, specifying outcome `lemp`, time variable `year`, unit identifier `countyreal`, treatment onset `first.treat`, and covariate `lpop`. Then it runs `etwfe(pdata, time_var="time_var", unit_var="unit_var", treatment="treatment", response="response", covs="lpop")`. Below is the “Extended Two-Way Fixed Effects Results”: * **Overall Average Treatment Effect (ATT):** – Estimate: –0.0452 – Std. Error: 0.0145 – 95 % CI: \[–0.0736, –0.0167] * **Cohort Average Treatment Effects (CATT):** | Cohort | Estimated TE | SE | 95 % CI low | 95 % CI high | | ------ | ------------ | ---------- | ----------- | ------------ | | 2004 | –0.08762696 | 0.03555885 | –0.15732102 | –0.01793290 | | 2006 | –0.02127833 | 0.02128938 | –0.06300475 | 0.02044809 | | 2007 | –0.04595453 | 0.01633327 | –0.07796715 | –0.01394190 | * **Model Details:** – Units (N): 500 – Time periods (T): 5 – Treated cohorts (R): 3 – Covariates (d): 1 – Features (p): 29
July 5, 2025 at 8:39 PM
gregoryfaletto.com
July 5, 2025 at 5:17 AM
gregoryfaletto.com
this rules. Haven't taken the plunge to pay for Gemini access, been making the most of my free daily 2.5 Pro prompts for math. The CLI preview access is a super helpful workaround. I just put my prompt in a text file along with the pdf I had uploaded in a question I had asked in the web interface.
Screenshot of a dark-themed macOS Terminal window running “Gemini.” Across the top, large ASCII-art letters spell “GEMINI” in a pastel gradient from blue (left) through purple to pink (right). Beneath, white text on black lists four “Tips for getting started,” such as asking questions, being specific, creating a GEMINI.md file, and using /help. The prompt then shows a user command instructing Gemini to load a PDF on Oracle efficient variable selection and a question.txt file for feedback. A bordered status box follows, with a green checkmark and the message “ReadManyFiles Will attempt to read and concatenate files using patterns…,” indicating the files are being processed from the user’s Documents directory. A terminal window displaying feedback on a mathematical proof. The text comments on the generalization of Theorem 1 from Kock (2013), noting that the new assumptions are reasonable within cited literature. It includes a detailed breakdown under the heading "Proof Structure and Logic." Step 1 discusses bounding an error term and using assumptions to eliminate a cross-term, and confirms the validity of the argument. Step 2 begins with a comment on a modification of Kock's Lemma 1. The content includes inline LaTeX code for mathematical notation.
June 28, 2025 at 7:03 AM