Sophie E. Hill
LightNews LightNews
banner
sophieehill.bsky.social
Sophie E. Hill
@sophieehill.bsky.social
4.1K followers 2.7K following 880 posts
PoliSci PhD student @ Harvard / 🇬🇧🏳️‍🌈 / Creator of MyLittleCrony.com
Signal: @sehill.11
Posts Replies Media Videos
sophieehill.bsky.social
They can't even spell the name of the subject of the article correctly!

I've read AI slop that's higher quality than this.
"The one thing everyone [in the Trump administration] agrees on is the need to dramatically pressure the UK over the Online Safety Act," the source said. The decision to revoke Mr Ahemd's visa would not be without precedent.
November 10, 2025 at 5:25 PM
sophieehill.bsky.social
Utterly pathetic client journalism in The Telegraph — accepting the Trump administration's claim that CCDH (an anti-disinformation charity) is "complicit in censoring American citizens". No quotes, no qualification.

@telegraphnews.bsky.social

www.telegraph.co.uk/us/news/2025...
Trump to deport boss of Starmer-linked charity White House to take aim at organisation accused of stifling free speech Revoking Imran Ahmed's visa could 'send a message' over online censorship Connor Stringer Washington Correspondent 09 November 2025 10:30am GMT| The White House is poised to revoke the visa of the boss of a campaign group founded by Sir Keir Starmer's chief of staff, Morgan McSweeney. Advertisement More Stories A source with knowledge of the discussions said: "They want to send a message that they're not having it. They're not happy. They are very actively considering visa restrictions on him." The move is the latest example of Mr Trump's willingness to clamp down on British nationals complicit in censoring American citizens.
November 10, 2025 at 5:18 PM
sophieehill.bsky.social
i don't know what's funnier: the error in this paper or my incredible pun

pubpeer.com/publications...
Something fishy? That final paragraph contains a very odd reference to the "World Aquaculture Society". What does this have to do with mental health and well-being during Covid-19 in higher education in G20 countries?
November 10, 2025 at 12:13 PM
sophieehill.bsky.social
waking up and being told that Valerie Vaz was describing My Little Crony to Jacob Rees-Mogg in the House of Commons...

10/10, peak lockdown fever-dream experience
November 8, 2025 at 7:20 PM
sophieehill.bsky.social
What a throwback!

Had to fire up mylittlecrony.com and find this company in the web, for old times' sake!
November 8, 2025 at 7:11 PM
sophieehill.bsky.social
the authors say they fixed all the incorrect confidence intervals and i just have one question:

would you say the number 69 is closer to 93 or to 6? 🧐
consumption of combined LNCSs in the highest tertiles was associated with a faster decline in memory (second tertile: β = −0.069, 95%CI −0.093 to −0.006; third tertile: β = −0.088, 95% CI −0.148 to −0.028) and global cognition (second tertile: β = −0.052, 95% CI −0.096 to −0.008; third tertile: β = −0.063, 95% CI −0.105 to −0.021)
November 7, 2025 at 1:31 AM
sophieehill.bsky.social
This is pure chaos. There's an estimate that looks correct in the text but incorrect in the appendix (upper bound is missing a minus). So what did the authors do? They removed the (correct) value from the text, and left the incorrect value in the table unchanged!
November 7, 2025 at 1:29 AM
sophieehill.bsky.social
Call me old-fashioned, but I think if you have to correct one of the key results cited in your *abstract*, then the correction actually does affect the overall conclusions of your study!
Original version of abstract, with problematic estimate highlighted Corrected version of estimate with problematic estimate omitted In the Research Article “Association Between Consumption of Low- and No-Calorie Artificial Sweeteners and Cognitive Decline: An 8-Year Prospective Study” by Gomes Gonçalves et al.,1 there were errors in the CIs for some estimates presented in the article and the supplementary eTables. The article and its accompanying supplementary material have been replaced by corrected versions. The original version with the changes highlighted is available online alongside the corrected article. The corrections do not affect the overall conclusions of the study. The authors regret the errors.
November 6, 2025 at 11:21 PM
sophieehill.bsky.social
anyway, i figured out the problem. they standardized the betas but not the confidence intervals. they have made this mistake before with log-transformations...

pubpeer.com/publications...
Explaining the estimate/CI discrepancies These discrepancies appear to result from inconsistent standardization: that is, the estimates were standardized but the confidence intervals were not. Let's consider the 3 examples of estimates outside their CIs highlighted in point #1: Suppose that the CI interval represents unstandardized values. Then the midpoint ci_mid is the unstandardized β. To standardize a β, we multiply it by SD(X)/SD(Y) The outcome Y in this case is Delayed Recall Memory. We know from Table 1 that this variable has a standard deviation of 1.9 For the X variables: we can extract SD(Age) from Table 1. For the binary variables (women, education), we can extract the proportions reported in the abstract (prop women = 0.545, prop educated = 0.43) and then use the formula S D p r o p = ( p ∗ ( 1 − p ) ) SD ​prop ​​ =√ ​(p∗(1−p)) ​ ​​ These values are included in the table below as var_sd The final column ci_mid_std shows the standardized βs (i.e. ci_mid * var_sd / 1.9). These line up very closely with the βs reported in the text. var estimate ci_mid var_sd ci_mid_std Female sex 0.243 0.944 0.498 0.247 Age -0.110 -0.044 4.900 -0.113 Education 0.205 0.799 0.495 0.208
November 4, 2025 at 7:24 PM
sophieehill.bsky.social
that's right.

ANOTHER paper published by authors associated with the ELSA-Brasil team with estimates outside their confidence intervals.

no-one noticed this??
chart showing how far these reported estimates are outside their CIs
November 4, 2025 at 7:24 PM
sophieehill.bsky.social
my spidey sense is tingling...
In model 1 of multiple linear regression, adiponectin levels showed to be associated with delayed recall memory (Table 3), while prediabetes with learning memory (β −0.080; p=0.015) and delayed recall memory (β −0.083; p=0.012). After adjustments (model 2), adiponectin but not prediabetes maintained independently associated with delayed recall memory. In the same model, significant associations of delayed recall memory with the female sex [β 0.243 (95% CI 0.675; 1.212)], age [β −0.110 (95% CI −0.071; −0.018)], and education level [β 0.205 (95% CI 0.535; 1.063)] were also detected. The inclusion of the interaction variable adiponectin*HOMA-IR did not change the association between adiponectin levels with improved delayed recall memory.
November 4, 2025 at 7:24 PM
sophieehill.bsky.social
ahhh beautiful! an interlocking cat bridge.

yes, he has a very fluffy sister. (she is much cleverer than him and prefers hunting squirrels to munching cardboard.)
two cats cuddling
November 4, 2025 at 2:36 AM
sophieehill.bsky.social
no i'm afraid it's a genuinely unhinged personal project.

(my friend's baby had his gallbladder removed so ... naturally ... i thought he might like an Emotional Support Gallbladder Rattle)

that was an old picture. this is how it turned out:
November 4, 2025 at 2:21 AM
sophieehill.bsky.social
he made a little headrest right over the picture of the Felix cat on the box 🥹
Barney resting his head on his newly carved headrest
November 4, 2025 at 2:11 AM
sophieehill.bsky.social
bonding with my cat by engaging in our shared hobby (whittling)

he uses teeth, i use knife

together we create debris and vanquish stress
Barney the cat looking up from chomping on a cardboard cat food box. there are little scraps of chewed cardboard on the floor. latest whittling project. pear-shaped wooden object (you'll honestly never guess), next to whittling knife and wood shavings.
November 4, 2025 at 2:07 AM
sophieehill.bsky.social
don't interrupt me, i'm doing important work!!*

*reverse engineering how some researchers used the wrong formula AND the wrong constants to calculate eGFR in a clinical trial

pubpeer.com/publications...
different eGFR formulas scatterplots showing correlations between the author's eGFR and the calculated eGFR under different formulas
November 3, 2025 at 4:53 PM
sophieehill.bsky.social
tag yourself, i'm-

Fickle, Feeble, For Pity's Sake

Flimsy, Fishy, For Sale

Favourless, Flavourless, For Your Information

Freaky, Flaky, Four Score And Seven Years Ago
photo of Mike Tapp with union jack background
November 3, 2025 at 9:15 AM
sophieehill.bsky.social
I regret to inform you it's now Week 2 of Reform-curious Labour MP Mike Tapp trying to soft launch his new slogan ("Firm, Fair, For Britain")
screenshot of Mike Tapp instagram video with slogan in caption. Oct 26 Tweet by Mike Tapp including the slogan
November 3, 2025 at 9:15 AM
sophieehill.bsky.social
I fear that interacting with this account may be, erm, grasping at straws?
November 2, 2025 at 6:37 PM
sophieehill.bsky.social
Even in the era of zero-cost LLMs, some scientists are still crafting data for their clinical trials by hand... and getting published in the @bmj.com!

Inspiring.

pubpeer.com/publications...
screenshot of Pubpeer comment. Out of the 396 rows in the dataset, 131 rows have non-missing values for cholesterol. All of these 131 rows are "clones", meaning that they have an identical match in the data across all of these lab panel variables: WBC, Hb, Plt, BUN, Cr, Na, BS, TOTALCHO, LDL, HDL, TG. For example, one clone group looks like this: <table> Looking at the observation ids for each clone group reveals a clear pattern: they are in groups of 4 (except for one), and follow a clear pattern (n, n+101, n+202, n+303). <table>
November 2, 2025 at 5:58 PM
sophieehill.bsky.social
A couple of days ago Neurology accepted this letter to the editor which offers a good critique and underscores this point:

"Emphasis should remain on reducing intake of conventional sugar, given its well-established risks for brain and metabolic health."

www.neurology.org/doi/10.1212/...
We read with great interest the recent article by Gonçalves et al.1 This cohort analysis is purely observational and cannot establish causality. Reliance on self-reported FFQs introduces substantial exposure misclassification, particularly with sweetener blends. The 8–9-year follow-up seems brief to assess neurodegenerative risk adequately. Reverse causality is plausible; individuals with diabetes or metabolic concerns preferentially consume low and no-calorie sweeteners (LNCS), confounding any direct cognitive associations. Notably, a “low-sweetness” comparator group—excluding both sugar and LNCS—is absent. These limitations—many acknowledged by the authors—suggest the small associations may simply reflect residual confounding. Furthermore, the analysis pools chemically and physiologically different compounds: artificial high-intensity sweeteners (aspartame, saccharin, acesulfame-K) and natural bulk polyols (erythritol, xylitol, sorbitol). “Sweet taste” is the only commonality; interpreting pooled associations is problematic. Where sweeteners are used, compound-specific effects and doses matter and treating all LNCS as a homogeneous class should be avoided. The reported polyol intakes (e.g. xylitol 3 mg and erythritol ≈0.1 mg/day) are far below real-world exposures and unlikely to have any biological effect. By contrast, the classic Turku xylitol studies used several–tens of grams/day without evident cognitive harm.2 Emphasis should remain on reducing intake of conventional sugar, given its well-established risks for brain and metabolic health.
November 2, 2025 at 4:06 PM
sophieehill.bsky.social
The paper has an obvious numerical error IN THE ABSTRACT. And it only goes downhill from there.

This should not be a difficult editorial decision.

github.com/sophieehill/...
Example For example, consider this estimate, reported in the abstract on page 1: “Among participants aged younger than 60 years, consumption of combined LNCSs in the highest tertiles was associated with a faster decline in verbal fluency (second tertile: β = −0.016, 95% CI −0.040 to −0.008 …” It is immediately obvious that the confidence interval ( − 0.040 , − 0.008 ) is not centred around the point estimate ( − 0.016 ) . (-0.040 + -0.008)/2 ## [1] -0.024 Summary Unfortunately, this example is not an isolated issue. Across 9 appendix tables with 256 regression coefficients, I find: 5 cases of estimates outside their reported confidence interval 46 duplicates (i.e. estimate/CI combinations with an identical match in the same table) At least 20 cases of asymmetric confidence intervals, which cannot be explained as rounding errors 17 values rounded to 4 decimal places (all others to 3 decimal places), indicating manual editing of tables Over 50% of all estimates and confidence intervals are multiples of 0.008, indicating possible post-hoc rounding and rescaling not mentioned in the text It is difficult to diagnose exactly what is going on without access to the underlying data and code. However, two broad conclusions can be drawn: There are many errors and inconsistencies in the results which appear to be the result of manual editing. There is not enough information in the paper and appendix about how the results were generated.
November 2, 2025 at 4:06 PM
sophieehill.bsky.social
In the meantime, this study has been covered in the media in over fifty articles, across a dozen countries, with thousands of engagements on social media!
New York Post headline A daily diet soda can make the brain age more quickly, study finds Mail Online Millions at risk of rapid brain aging as six common artificial sweeteners are linked to decline BILD Light and Zero can damage the brain TODAY These 6 Artificial Sweeteners May Be Linked to Cognitive Decline, New Study Finds
November 2, 2025 at 4:06 PM
sophieehill.bsky.social
Error #1

Asymmetric confidence intervals

Error #2

Duplicated p-value (extremely unlikely by chance)
annotated table 1. Asymmetric CIs (highlighted in red) This table reports coefficients from a linear regression with 95% CIs. For standard 95% confidence intervals, the point estimate should equal the midpoint of the CI (i.e., β = (lower + upper)/2). Even allowing for rounding errors, these CIs and estimates are not consistent: Model 1, moderate: 15.57 (10.33 to 30.81) (10.33 + 30.81)/2 = 20.57 Model 3, moderate: 0.04 (-1.19 to 0.78) (-1.19+0.78)/2 = -0.205 Model 2, heavy: -6.09 (-11.42 to 0.48) (-11.42+0.48)/2 = -5.47 Model 2, former heavy: -4.68 (-9.36 to -1.18) (-9.36 + -1.18)/2 = -5.27 2. Duplicate p-values (highlighted in green) The table reports a p-value of 0.089 for two separate estimates (Model 2 and Model 3 for heavy drinkers). This is highly unlikely to occur by chance. Based on the reported CI, the p-value for Model 3 should be closer to 0.15. A p-value of 0.089 looks more reasonable for Model 2, though it is hard to say exactly since the estimate/CI reported here is inconsistent (see above).
November 2, 2025 at 3:05 PM
sophieehill.bsky.social
Round 3

Same paper as round 2. Note these are beta coefficients from a linear regression.
another table
November 2, 2025 at 3:05 PM