Unfork Statistics
LightNews LightNews
banner
btjo.bsky.social
Unfork Statistics
@btjo.bsky.social
400 followers 2.6K following 380 posts
False +ves only occur on true nulls. Srsly.
Stats: https://bsky.app/profile/btjo.bsky.social/post/3kvloovvbcz23
QRPs: https://bsky.app/profile/btjo.bsky.social/post/3kvmompohir24
NNT me: https://bsky.app/profile/btjo.bsky.social/post/3kvkrxghlhv2d
She/they
Posts Replies Media Videos
btjo.bsky.social
Domain knowledge is always critical for the appropriate application of statistical methods. Here, it would have prevented this embarrassing episode occurring at all.
To him, the research seemed off. Elkan used to work as a managing director at Goldman Sachs and as a scientist at Amazon, where he led machine-learning teams. The technology sounded too good to be true to him, given its supposed date of deployment. Also, 3M and Corning were the only U.S. materials companies Elkan knew of that employed a large number of scientists, and he questioned why, if they were inventing so many new materials, they would be keeping it secret.
November 22, 2025 at 2:40 PM
btjo.bsky.social
This is a failure to explain herd immunity properly. Some parents hear only a demand that they vaccinate to protect others. But their (unvaccinated) kids are most likely to be harmed in an outbreak. It's their (unvaccinated) kids who pay the price of herd immunity breaking down. Spell it out.
"The impact of their vaccine choices on others is not necessarily part of their calculation. And why should it be? There are few other arenas in which parents are asked to consider the effects of their choices on other children."
October 1, 2025 at 8:34 AM
btjo.bsky.social
Nice example of the ecological fallacy, with a conclusion so absurd it's hard to understand how it got written down. #stats
A new WalletHub study shows how deeply car loan debt cuts into household budgets across the U.S. The financial site reviewed more than 2,500 cities, measuring median auto-loan balances against median incomes. The results revealed a clear warning: in more than 160 cities, drivers spend half or more of their yearly income on car loans alone. That burden comes as this particular debt keeps climbing nationwide. Some of the hardest-hit cities are in Texas Rio Grande City leads the list, where the median auto-loan balance is nearly equal to median income: $32,173 in debt versus $33,782 in earnings. That’s a 95% debt-to-income ratio. Nearby towns like Donna, Mercedes, San Juan, and Brownsville all land in the danger zone as well. Uvalde, Eagle Pass, and Pharr also show residents committing well over two-thirds of their annual income to car notes.
August 30, 2025 at 11:28 AM
btjo.bsky.social
It really is quite stark, mapped out like this. (Assuming the underlying data is correct, I can't vouch for it but it smells right.)
A useful graphic from NYT charting published claims that there is, or is not, a link between vaccines and autism. 18 of the 27 claiming a link were auhored by David & Mark Geier, with 5 others retracted (including Wakefield's original 1998 paper). None of the 43 showing no link have been retracted, or authored by a Geier. vs 43 claiming
August 20, 2025 at 12:14 PM
btjo.bsky.social
We would never dare shut ours out of the bathroom but he doesn't need a cat-hole, he follows us in. His cat-hole leads from the kitchen to his pantry.
A cute little kitten - white with some black shading - stares up at the camera in front of a cat-sized hole in a door, which leads to his food and litter.
June 28, 2025 at 3:24 PM
btjo.bsky.social
No. In both cases, it depends on whether GPs have the correct gender recorded.
Trans men and non-binary people assigned female at birth Registered with a GP as female If you are aged from 50 years up to your 71st birthday and registered with a GP as female, you will be routinely invited for breast screening. We recommend you consider having breast screening if you have not had chest reconstruction (top surgery) or still have breast tissue. If you are worried about visiting a breast screening unit you can ring to arrange a more suitable appointment. For example, we can arrange for you to have an appointment at the beginning or end of a clinic. Registered with a GP as male If you are registered with a GP as male, you will not be invited for breast screening. If you have not had chest reconstruction (top surgery) and are aged 50 or over we suggest you talk to your GP. They can arrange a referral for you to have mammograms at a hospital near you. You should keep aware of the symptoms of breast cancer, and contact your GP if you notice any unusual changes. If you have had chest reconstruction (top surgery), we advise you have a conversation with your surgeon about the amount of breast tissue you have remaining. If they confirm you still have breast tissue, you can ask your GP to refer you for mammograms at a hospital near you. Trans women and non-binary people assigned male at birth Registered with a GP as a female If you are aged from 50 up to your 71st birthday and registered with a GP as female, you will be routinely invited for screening. Long-term hormone therapy can increase your risk of developing breast cancer so it is important that you consider going for breast screening when you are invited. Registered with a GP as male If you are registered with a GP as male, you will not be invited for breast screening. If you have been on long-term hormone therapy you may be at increased risk of developing breast cancer. Your GP can arrange a referral for you to have mammograms at a hospital near you.
April 10, 2025 at 10:27 AM
btjo.bsky.social
They do invite trans women if the GP has the correct gender recorded.
Trans women and non-binary people assigned male at birth Registered with a GP as a female If you are aged from 50 up to your 71st birthday and registered with a GP as female, you will be routinely invited for screening. Long-term hormone therapy can increase your risk of developing breast cancer so it is important that you consider going for breast screening when you are invited. Registered with a GP as male If you are registered with a GP as male, you will not be invited for breast screening. If you have been on long-term hormone therapy you may be at increased risk of developing breast cancer. Your GP can arrange a referral for you to have mammograms at a hospital near you.
April 10, 2025 at 10:08 AM
btjo.bsky.social
Trans women are advised to attend breast screening if they are on long-term HRT.

Here is the NHS advice to trans and non-binary people for this and the other screening programmes: www.gov.uk/government/p...
Screening for trans people at a glance Trans women and non-binary people assigned male at birth who are registered with a GP as female: are invited for breast screening are invited for bowel cancer screening do not need cervical screening as they do not have a cervix, but may still get invited are not routinely invited for abdominal aortic aneurysm (AAA) screening but can request screening Trans women and non-binary people assigned male at birth who are registered with a GP as male: are not routinely invited for breast screening but can request screening are invited for bowel cancer screening do not need cervical screening as they do not have a cervix are invited for AAA screening Trans men and non-binary people assigned female at birth who are registered with a GP as female: are invited for breast screening are invited for bowel cancer screening are invited for cervical screening are not invited for AAA screening Trans men and non-binary people assigned female at birth who are registered with a GP as male: are not routinely invited for breast screening but can request screening if they still have breast tissue that needs screening are invited for bowel cancer screening can ‘opt in’ to receive routine cervical screening invitations by contacting their cervical screening provider, for example their GP or practice nurse, sexual health clinic or transgender health clinic are invited for AAA screening but do not have a high risk of AAA Trans men who are pregnant should be offered the same antenatal and newborn screening tests as all other pregnant individuals.
April 10, 2025 at 9:59 AM
btjo.bsky.social
March 30, 2025 at 12:38 PM
btjo.bsky.social
She's learning fast!
March 25, 2025 at 5:05 PM
btjo.bsky.social
First time on a longline today. She can't wait to be off-lead.
March 24, 2025 at 2:14 PM
btjo.bsky.social
Here's Luna not wanting to get up this morning. Thank you so much for getting her safely to us. Six weeks on she is calm, confident, and such a sweet gentle soul. Even the cat is forced to agree. They have zoomies together in the middle of the night. 🥰
A six month old puppy, black with a speckled white belly and socks, sprawling on a duvet, with remote control nearby. She's got her front paws up round her sleepy head.
March 24, 2025 at 11:44 AM
btjo.bsky.social
She's still little! ~5 months old and living on the street until a few weeks ago. Has discovered sofas exist and plans to never leave. Her big brother was also quite little when he came to live with us and is teaching her about belly rubs.

Luna and Nads. 🥰
A pcture of a ~5 month old puppy, black with white chest and speckled forelegs, snuggled into a corner of the sofa with her Blahaj shark plushie. Her ears are perked up but her eyes are saying "please don't make me move". A little white cat staring intently at the camera while he has his belly rubbed.
March 2, 2025 at 12:43 PM
btjo.bsky.social
That's one method of p-hacking! And it's a key part of how Bem managed to 'confirm' his results. He ran a bunch of 'pilot' studies and only continued with the ones that 'worked'. But he rolled the pilot data into the larger studies instead of starting afresh: replicationindex.com/2018/01/05/b...
Figure 1 provides a visual representation of the decline effect in Bem’s studies. The x-axis is the sample size and the y-axis is the cumulative effect size. As sample sizes increase, the cumulative effect size approaches the population effect size. The grey area represents the results of simulation studies with a population effect size of d = .20. As sampling error is random, the grey area is a symmetrical funnel around the population effect size. The blue dotted lines show the cumulative effect sizes for Bem’s studies. The solid blue line shows the average cumulative effect size. The figure shows how the cumulative effect size decreases by more than 50% from the first 5 participants to a sample size of 100 participants. The selection effect is so strong that Bem could have stopped 9 of the 10 studies after collecting a maximum of 15 participants with a significant result. The average sample size for these 9 studies would have been only 7.75 participants. Table 1 shows the one-sided p-values for Bem’s datasets separately for the first 50 participants and for participants 51 to 100. For the first 50 participants, 8 out of 10 tests are statistically significant. For the following 50 participants none of the 10 tests is statistically significant. A meta-analysis across the 10 studies does show a significant effect for participants 51 to 100, but the Test of Insufficient Variance also shows insufficient variance, Var(z) = 0.22, p = .013, suggesting that even these trials are biased by selection for significance (Schimmack, 2015).
December 21, 2024 at 1:34 PM
btjo.bsky.social
With apologies for a serious response to a good joke, particle physics did notice, and then deal with, a replication crisis in the 1970s: www.nature.com/articles/526... #stats
Snippet from the intro to the linked article: Decades ago, physicists including Richard Feynman noticed some- thing worrying. New estimates of basic physical constants were often closer to published values than would be expected given standard errors of measurement1. They realized that researchers were more likely to ‘confirm’ past results than refute them — results that did not conform to their expecta- tion were more often systematically discarded or revised. To minimize this problem, teams of particle physicists and cosmologists developed meth- ods of blind analysis: temporarily and judi- ciously removing data labels and altering data values to fight bias and error2. By the early 2000s, the technique had become widespread in areas of particle and nuclear physics. Since 2003, one of us (S.P.) has, with colleagues,
November 15, 2024 at 3:20 PM
btjo.bsky.social
NHS league tables are back: www.theguardian.com/society/2024...

Useless, and harmful, for many reasons, not least that rankings tell you nothing about actual performance. And because it redirects energy into gaming the numbers, often at the expense of patients:
europepmc.org/backend/ptpm...
Snapshot of the linked paper (from 2003): How to make a silk purse from a sow’s ear—a comprehensive review of strategies to optimise data for corrupt managers and incompetent clinicians David Pitches, Amanda Burls, Anne Fry-Smith The introduction of performance league tables for UK surgeons and hospitals has forced them to learn how to present data in the best possible light. Though there is an urgent need for guidance, official guidelines on how to optimise performance data are lacking
November 13, 2024 at 5:12 PM
btjo.bsky.social
FWIW the BPC's inquiry into the polling 'miss' in the 2015 UK general election (eprints.ncrm.ac.uk/id/eprint/37...) required member orgs to publish information on weightings, and changes in methods since their previous poll. Polling is very different in the US but it's an interesting read. #stats
Feom the press release here: https://www.britishpollingcouncil.org/bpc-inquiry-report/ "Specifically, the Council has agreed to implement immediately rule changes that will (i) require greater transparency about how polls have been weighted, (ii) specify what changes, if any, have been made since a company’s previous published poll in how the data have been weighted or otherwise adjusted, and (iii) place an obligation on members to supply to any inquiry or committee that has been established by the BPC the micro data set for any poll in which that inquiry or committee has an interest."
November 4, 2024 at 6:47 PM
btjo.bsky.social
This is why we do confirmatory trials. The prevalance of true diffs is much higher for the second trial precisely because we got an encouraging result from the first.

If you're using a frequentist framework, you can't trust a single test. It requires replication (and Bayes theorem shows why).
90% power, 95% confidence (alpha=0.05), 10% plausibility. So in 1000 trials we have 100 false nulls and we will find 90 of them. We will have 900 true nulls and we will incorrectly identify 45 as fals¬e. One in three p<0.05 will be false positives so PPV ~ 67%. Let’s confirm that result. 90% power, 95% confidence (alpha=0.05), 66.7% plausibility. So in 1000 trials we have ~667 false nulls and we will find ~600 of them. We will have 333 true nulls and we will incorrectly identify ~17 as false. 17/617 p<0.05 will be false positives so PPV ~ 97%.
October 23, 2024 at 9:27 AM
btjo.bsky.social
The pairing does a lot more than that. It allows you to analyse a single sample (of differences between pairs) rather than compare the means of two independent samples. If you have paired data, it's incredibly powerful by comparison (see plot in the lower third of this poster).
Screenshot of a poster, downloadable here: https://ora.ox.ac.uk/objects/uuid:a9abe009-1b43-4593-a601-bcde9ed08830 There's a plot at the bottom comparing the results from a study analysed as pairs vs independent groups, for a number of outcomes. The CIs for the paired analysis are tighter, much tighter for outcomes which are more variable at baseline, indicating an effective sample size for the paired analysis an order or two of magnitude greater than when analysed as independent groups. The data are from a systematic review of irradiated platelets, where vitro studies were included if they had split their samples (rather than randomised them into two independent groups). Only 14 of the 43 used a paired analysis and only one of the 14 reported enough information to reconstruct their paired analysis.
October 19, 2024 at 10:01 AM
btjo.bsky.social
Ah! It has an erratum! That corrects the error and resolves the confusion. I'll read the rest of it when I have some time.
Erratum Hernán MA. The hazards of hazards ratios [commentary]. Epidemiology. 2010;21:13–15. On page 15, column 1, second full paragraph: “… and the average HR is ensured to reach the value 1.” should instead say, “… and the risk ratio is ensured to reach the value 1.” Epidemiology. 22(1):134, January 2011.
October 16, 2024 at 9:28 AM
btjo.bsky.social
This is the closest example I have to hand where (almost) everyone is dead by the end of follow-up. From: openi.nlm.nih.gov/detailedresu...
A plot with two survival curves, one reaching ~zero survival at the end of 8 year follow-up, the other without ~5% alive at that time. The plot includes the HR, which is 0.49 (0.32, 0.73), roughly equal to ratio of the areas under the two curves (not the proportions alive at the end of follow-up). Note that this was not an RCT. The surgical group survived much longer but we can't conclude that this was due to surgery - it may just reflect that people offered surgery are considered fit enough to survive the surgery.
October 15, 2024 at 8:24 PM
btjo.bsky.social
It can also be tricky in the presence of competing events. I wrote a long-ish post on @f2harrell.bsky.social's forum about that, including some references: discourse.datamethods.org/t/background...

Pic is a poster on competing events AND in-hospital follow-up (quite common for Covid trials).
Image of a conference poster, can be downloaded here: https://ora.ox.ac.uk/objects/uuid:ed85eb18-ab98-499b-b14e-2e873f4a57bf
October 12, 2024 at 7:25 PM
btjo.bsky.social
Not difficult, just a bit more complicated than RR because you have to look at the survival curves to know what it means, and
because it can change with further follow-up. Pic is a slide I use to illustrate both points.

Any decent #stats textbook on the topic should make these points.
Slide shows three pairs of survival curves, based on the exponential distribution (black line) and a Weibull distribution (red line). The three curves show short, medium and long-term follow-up: With short follow-up the black line looks much better than the red line and the HR would strongly favour it. With longer follow-up, the two curves cross, with the red line having more long-term survivors. The HR would be ~1 (HR can be thought of as comparing the areas under two survival curves, like a relative risk averaged over time). With long-term follow-up, the early advantage to the black line looks trivial and the HR would strongly favour the red line.
October 12, 2024 at 7:19 PM
btjo.bsky.social
We did a poster related to this (competing events with in-hospital follow-up) using a toy dataset to illustrate some problems & solutions. Downloadable here: ora.ox.ac.uk/objects/uuid...

How you handle it depends on your specific context, and what question you want to answer.
Image of the poster which can be downloaded at the link. "Censoring in survival analysis implies that the event may occur with further follow-up. This is not true for progression-related events if deaths occur before progression. Nor for time to discharge if there are in-hospital deaths. If there is no post-discharge follow-up, it may also be a biased assumption for overall survival if discharge is due to clinical improvement."
September 17, 2024 at 1:08 PM
btjo.bsky.social
Not sleight-of-hand, it's how variance works. The two extremes are shown in this poster, where the data were paired (split samples). The figure at the bottom shows how much difference it makes if you get rid of population variance by analysing pairs instead of groups.
ora.ox.ac.uk/objects/uuid...
August 8, 2024 at 4:07 PM