James Steele

@jamessteeleii.bsky.social

792 Followers 535 Following

I'm just a guy who's an academic for fun

A wrist with a tattoo depicting the implicative syllogism of modus tollens as: T→P ¬P ∴¬T
A person relaxes on a wooden garden bench with a tabby cat curled on their lap while eating a red ice lolly. The person gazes ahead as the cat looks directly toward the camera. The bench sits beside a garden bed of leafy plants, with a lawn in the foreground and a wooden fence and shed lit by warm evening sunlight behind.
Screenshot of a page from https://meehl.umn.edu/sites/meehl.umn.edu/files/files/179faustmeehl2002.pdf discussing Einhorn's (1972) study of radiologists' judgments of biopsy slides. The highlighted passage states that although the pathologists generated useful data, they did not make proper use of it themselves. Statistical analysis identified a subset of rated variables that predicted survival time with modest accuracy, whereas the radiologists' overall severity ratings did not predict outcomes. The passage argues that the experts struggled to identify which observations were predictive and how to combine them effectively, illustrating the value of statistical models over unaided expert judgment.
Cover of The Cat magazine, Summer 2026 edition from Cats Protection. A white-and-tabby kitten stands on a log against a blurred green background. Cover lines include “Protect every cat: Why neutering is so important,” “Behind the scenes at a cat centre,” and “In conversation with Bob Mortimer.”
This study is, to our knowledge, one of the largest to compare the effects of low and high volume RT interventions upon hypertrophy in previously trained participants. We found statistical equivalence between both conditions and both main effects of time, and any interaction effects for condition by time, are likely small. More broadly, this study further corroborates the linear-log theory of adaptation, that the effects of RT in trained persons should be expected to be small, and that current studies in the field of RT are woefully underpowered to be able to detect their effects, let alone test between intervention comparisons.
A presenter stands at a lectern during the Perspectives on Scientific Error conference, pointing to a projected slide titled “Precise Deductions” that shows hypothesis ranges and function-form diagrams.
Apple health screenshot showing 44772 steps and 23.8 miles walked
Screenshot of Apple health showing 41496 steps and 18.9 miles walked.
Forest plot showing study- or group-level mean resting energy expenditure (REE) in kcal/day with corresponding standard errors. Each horizontal line represents the SE around a mean estimate (shown as a point), with estimates ordered from lowest to highest REE along the vertical axis. Mean REE values range approximately from 900 to 2,600 kcal/day. The distribution forms an S-shaped pattern, with most estimates clustered between about 1,300 and 1,800 kcal/day and greater variability at higher REE values.
A screenshot of Apple health showing 43027 steps and 23.3 miles walked
A screenshot of Apple health showing 35473 steps and 19.8 miles walked.
A man sits at a small folding table outdoors on a university campus, holding a mug and smiling, in the style of the “Change My Mind” meme. A sign taped to the front of the table reads “CHANGE MY MIND.” Another sign on the tabletop reads: “Most quant vs qual methods dichotomies as presented are fallacious and more to do with differences in ontological and epistemological differences of the researchers themselves.” Trees, benches, and campus buildings are visible in the background.
The image shows the abstract for my talk "The value of strong theory in intervention research: an example from the field of exercise science" at the upcoming 8th Perspectives on Scientific Error Workshop - you can find it in the program here: https://docs.google.com/document/d/1rt9ToVs1EkEuTbWod4st2b6StUQr5IhyTP1jhGskr4E/edit?usp=sharing
Multi-panel horizontal bar chart showing self-admitted rates (percent on x-axis) of engaging in various questionable research practices (QRPs) at least once. Each panel represents a distinct QRP, including excluding data based on results, failing to report all conditions, falsifying data, HARKing, hiding demographic moderators, opportunistic p-value rounding, optional stopping, selective outcome reporting, selectively reporting performed analyses, selectively reporting what “worked,” selectively including covariates, and selectively switching analyses. Colored horizontal bars indicate estimates from individual studies. In each panel, a vertical red line marks the meta-analytic estimate across studies, and a surrounding red shaded band represents the 95% confidence interval for that estimate. Prevalence varies widely across practices and studies, with higher rates for selective reporting behaviors and HARKing, and consistently low rates for falsifying data. Alt text generated by chatgpt.
Multi-panel bar chart showing self-admitted rates (percentages on the x-axis) of engaging in various questionable research practices (QRPs) at least once, based on multiple published studies (color-coded by study). Each panel corresponds to a specific QRP, including: excluding data based on impact on results, failing to report all conditions, falsifying data, HARKing, hiding demographic moderators, opportunistically rounding p-values, optional stopping, selectively reporting outcomes, selectively reporting performed analyses, selectively reporting what “worked,” selectively including covariates, and switching analyses selectively. Within each panel, horizontal bars represent individual study estimates of the percentage of respondents admitting to that practice. A vertical red line in each panel indicates the meta-analytic estimate across studies for that QRP. Overall, reported prevalence varies widely by practice and study, with some QRPs (e.g., selective reporting and HARKing) showing higher admission rates than others (e.g., falsifying data).
Slide reading with Steele Research Ltd logo in bottom right corner. In the centre it reads: End of year reflections, or… I just finished a cut, here’s how it went, and here’s my progress* over time All data and code for this post can be found in the accompanying GitHub repository https://github.com/jamessteeleii/blackpill_progress.git * Terms and conditions apply ;-) ------ Note, this is the first slide from a carousel of slides made for Instagram. If you have visual impairment and would like to have a copy of the slides which contain mostly text, and a description of any images or plots, then please drop me a DM. Thanks and Happy New Year!
“Kiki, a brown tabby cat with white paws, sits on a narrow wooden wall shelf, looking to the left. Behind her is a framed, hand-painted portrait showing Kiki alongside her two previous cat companions, Cosmic and Bigsey. On either side are framed William Morris cat-themed prints, creating a cozy, art-filled home setting.” Alttext produced by chatgpt
Screenshot from article reading: We implemented a robust Bayesian model-averaged meta-analysis (RoBMA), which effectively adjusts for publication bias(Bartoš et al. 2022). This approach combines selection modelsand PET-PEESE via Bayesian model averaging, yielding bias-adjusted, model-averaged effect sizes and Bayes-factor evidence.By integrating different presumptions about effect size, hetero-geneity, and publication bias from prior distributions, we wereable to reliably determine the effect size (Berkhout et al. 2023;Gronau et al. 2021). In essence, the RoBMA approach does notassume a single fixed level of publication bias; instead, it aver-ages over multiple bias scenarios with associated prior proba-bilities (with a higher prior weight on the possibility of no bias).The Appendix describes the model settings of RoBMA, whichinclude a discrete mixture of publication bias models rangingfrom no bias to high bias.We also applied frequentist bias correction techniques (Egger'sregression test and PET-PEESE) independently of the RoBMAto cross-verify the evidence of publication bias (Duval andTweedie 2000; Egger et al. 1997; Stanley and Doucouliagos 2014).The PET-PEESE was done as a separate analysis on the aggre-gated data, not within the Bayesian model. Small-study effectsbias was considered present when the regression asymmetrytest produced a p-value of 0.05 or less.
A gorgeous tabby cat (Kiki) laid on a tartan blanket atop a leather armchair. A blue cuddly toy rabbit sits behind her. She is looking at the camera.
A multi-panel figure summarizing mean resting energy expenditure (REE) estimates in people with and without polycystic ovary syndrome (PCOS). Left panel: A forest plot shows study-specific conditional estimates for REE. Each study is listed on the y-axis. For each study, blue points and intervals represent the control group and gold points and intervals represent the PCOS group. Most studies show slightly higher REE in PCOS, though with overlapping uncertainty intervals. Top-right panel: Two density plots show global grand mean REE estimates by condition. The control distribution (blue) peaks around 1442 kcal (95% quantile interval ~1334–1553), and the PCOS distribution (gold) peaks around 1472 kcal (95% quantile interval ~1359–1587). Bottom-right panel: A grey density plot shows the overall contrast between conditions (PCOS minus control). The distribution centers around a positive difference of about 30 kcal, with a 95% quantile interval spanning approximately –47 to 113 kcal, indicating compatibility with both small positive and small negative differences. A dashed vertical line marks zero difference. A legend indicates colors for Control (blue) and PCOS (gold). Text notes that point estimates and 95% quantile intervals are reported.
Abstract text which reads: Context: Polycystic ovary syndrome (PCOS) is common in reproductive-age women, who often have higher BMI classification. This is assumed to stem from lower resting energy expenditure (REE), influencing lifestyle intervention guidelines. However, evidence for reduced REE in women with PCOS compared with those without is inconsistent. Objective: To systematically search and meta-analyse the existing literature to estimate and describe the difference in REE between women with and without PCOS. Data Sources: A systematic search was conducted using PubMed, Medline and Web of Science databases of published research from January 1990 to January 2025. Study Selection: Studies that measured REE in women living with PCOS, both with and without control arms of women without PCOS, were included. Data Extraction: Bibliometric, demographic, and REE data was extracted by one investigator and checked in triplicate. Data Synthesis: Thirteen studies were included in a Bayesian arm-based multiple condition comparison (i.e., network) type meta-analysis model with informative priors to compare both mean REE, and between person variation in REE, between women with and without PCOS. Mean REE differed between groups by 30 kcal/day [95% quantile interval: -47 to 113 kcal/day] and the contrast ratio for between person standard deviations was 0.98 [95% quantile interval: 0.71 to 1.33]. Conclusions: These findings indicate that REE does not meaningfully differ between women with and without PCOS. Group-level differences in resting energy expenditure are small, insignificant, or not physiologically relevant.