Sunday, April 26, 2020

The legend of the book of the film of the record of the poem of the graffiti of the urban myth

"Which came first, the egg or the chicken?" is the perennially annoying question that consultant philosophers use to impress naive clients.

A much more serious question is why people hate the film  of the book, or  love the film, but  hate the book  it came from, or love the film, and hate the book  that came from it.

To solve this problem, I think we need to carry out a large scale analysis based in causal inference  (why not) .  Clearly we have the equivalent of the adaptive clinical trial or the series of unfortunate natural experiments to choose between. We can start with some obvious candidates.

  1. The Godfather
  2. Paddington
  3. The Princess Bride
  4. The Hobbit
  5. 2001 or The Sentinal
For each of these we need to look at all the possible features that could lead  one to prefer a book or a film (length, sentence structure, plot, character, year published, alphabetic position of author directors name, jokes, box office takings, time in best seller charts, influences, sequels, spinoffs, live action, cartoon, comic book, illustrated, etc etc) and build a bayesian model of how one moves from one state of mind (Hobbits are boring) to another (Smaug is cool), or from one opinion (why did no-one ever actually read the princess bride) to another (You keep using that word, etc etc ).

Two dominant theories to date are

  • The Ordering Theory
  • The Gap Theory

Then we will finally know the truth.

Wednesday, April 22, 2020

cat > /dev/kb

I'd like to bring up a very important topic for all my inky-fingered friends (and I am not
referring to my experiments with spilling Quink on the ebony fretboard to see if I can play faster).

The cats. come and sit. on the keyboard. in front of the screen. while you're trying to work.

How can we fix this? 

Well, here's my patented invention that I think is going to work.

You know how the keyboard layout was invented by Benjamin Franklin for President Theodore Roosevelt when he was digging the Panama Canal? It is an oft told tale - the problem was that all the typists producing  reports to send back to Washington DC had to type an inordinate number of lower case 'a's every time just to get the headed paper to look right. For this reason, they moved the keys for 'l', 'n', 'm' and 'p' as far away from the 'a' as they could get them. Back in those days, most people were one or two fingered typists so this slowed them down, so that the 'a' key stopped breaking, leading to incredible misunderstandings between the  rmy, nvy and rfrce ('o' was a lesser,but not insignificnt prblm).

So the new keyboards worked a treat until they started working with a French company who still used the old Aztec name Pznzmz Cxnxl, which led to the second new keyboard, which nous aimions tres biens ces jours. and so on and so forth.

As code breakers amongst you will remember, (or possible Sherlock Holmes stick insect fans) the problem is to do with the popularity of different letters f the alphabet in different tongues. Why the alphabet is in alphabetical order is an interesting question which I'll leave for later  (just noting for now that it isn't, for example, in Arabic, Hebrew, Greek and Cyril Smith's languages, the third letter is 'g', not 'c' - go figure how the Romans got that wrong along with really poor ways of counting). And of course, Scrabble scores - which are in the opposite order to the popularity of letters ( a bit like Nathaniel Hawthorne's novels).

So "how does this solve a problem like a cat?", I hear you ask, Maria.

Easy peasy. we place bigger springs under the unpopular letters and so when the cat sits on the nice warm laptop keyboard, instead of getting a gentle purr like vive from the fan, it gets prodded uncomfortable in random places.

This will also let us revert the keyboard layout to being alphabetic, since it will just be harder to depress the unpopular letters which will slow us down between popular ones.

Of course we could confuse the cat further (as if such a thing were possible or even desirable) by choosing springs from a French or dare I  even suggest, a Chinese (Mandarin, not Catonese (sic), of course) layout of spring constants. By Hooke or by Crooke, we will have to solve the problem that the Chinese layout would require at least 5 springs of different strengths to operate really successfully, but we believe that yhe market for this in china will be as big as that for Dragon Nets.

I will be inviting investors to my alpha-beta-gamma product launches shortly, meanwhile I leave you with the experimental result that  you may wish to try as well, in these distracted days. You can teach your cat a foreign language easily. I have ours completely versed in French - when I say va't en or viens ici, she behaves in exacfly the same way as when I say "get the 'f off that keyboard" or "where are you didums". This does not work with dogs. In fact I know several dogs in the Dordogne that response to "Get off my leg fido" in exactly the same way as if you offer them a biscuit.

Science is a marvellous thing when used carefully.Electromagnets more so - in the new version (delta-key) of the boards above, we replace the springs by electromagnets and now can use this to train humans to type faster, and untrain cats to sit on the keys. Gnu Emacs key bindings will be available shortly.


qwerty or azerty

code breakers and scrabble


While waiting for the coffee to brew,

Monday, April 20, 2020

ethics, policy, regulation and contact tracing apps - babies and bathwater

0. How to get it right: Harvard ethics review/roadmap to pandemic resilience

1. if you are going to criticise the ethics of contact tracing app work, first establish a baseline - find out how manual contact tracing is done, what data is kept, what triggers it, what privacy risks there are. Who does the work? are they trained? is there a log (to avoid duplicating contact notifications by multiple staff and to record the test status of people). How does consent work from the pattient who's just tested positive and is probably stressed? How many false positive and negatives are there (people they falsely remembered they'd met, or encounters they forgot) etc etc- this is the standard an app has to meet at least.
2. know your medical ethics - it is standard that a new technology is introduced provided it is at least as good as existing "treatments" and no worse - see above. If contact apps tracing is also faster, and therefore reduces the number of people the virus spreads to before all possible infected people are found and isolated, then factor this in, as it is part of the care requirements.
3. Don't talk about stuff you don't understand - i've seen vague criticisms of the use of BLE (Bluetooth Low Energy) as it isn't "accurate enough" without a single citation on measurements that support this. There are multiple measurements and techniques that support that it is an ok proxy for encounters between people carrying capable smart phones with the relevant app running. The main criticisms are i) that might only be 60-75% of phones (depending on country/region/demographic) and ii) only around 75-80% of people have any sort of smart phone. See 1/ its additional, and faster/complimentary, not a replacement for manual contact tracing. Also see the care taken by Google/Apple (see prev blog) in terms of taking care of privacy- this is largely better at protecting people than manual tracing can be (there are modest exceptions - exercise for reader, think of one).
4. No-one's claimed you use contact tracing to replace testing (or more ludicrously, to replace the hunt for treatments, or vaccines or the actual provision of PPE for key workers who encounter a lot of potentially infectious people (including care homes, bus drivers, supermarket checkout staff etc). Don't claim people want to re-prioritise resources because they are techno-solutionists without actually finding out their motives and community. The idea of contact tracing apps came from (and is supported by) epidemiologists (e.g. from LSHTM in the UK). Tech people worked from what they asked, not from some bluesky fantasy. This includes the empirical testing of which is preferred proxy for encounters, and the fact that it is predicated on testing. Testing alone doesn't fix things fast enough either (unless you had a 15 minute test cheap enough to run on everyone nearly daily).
5. contact tracing doesn't have to be in more than a few percent of the population to be useful for its original purpose, which is to get more precision about the epidemic parameters to improve models, learn about asymptomatic carriers, infection rates between groups like children-to-adults, and the expiry date on immunity, whether acquired through surviving infection or from eventual vaccine deployment (many vaccines also have limited lifetime though usually better and longer than having had the disease, we still need to know).
6. Mission creep:-  discussed elsewhere - a mix of tech, regulatory and legal frameworks need to be clarified to minimise this risk- including (obviously) sunsetting.  This is not new.  People that work in clinical trials/medical ethics know this stuff - if you are a tech ethicist and you have not read up standard protocols in that space yet, please do so before criticising tech app writers who have. The goal of the privacy preserving/decentralised bluetooth API from Google/Apple is not to mess up earlier more centralised app designs, it is to offer a more ethical way forward and represents the way the tech sector has considered best practice ahead of some of the people criticising them

What's worse than techies who ignore ethics and context? ethicists who ignore the tech and context:

For the avoidance of doubt, Do No Harm.

Comprehensive list of tracer apps, initiatives, design docs etc

Friday, April 10, 2020

Some DP-3T & Apple/Google contact tracer abuse questions...

Contact tracing plus testing is a hope for getting out of lockdown, once we are well past the current peaks in the Covid-19 pandemic . Lots of apps have been proposed, some shipped. Most recently, privacy preserving apps have been designed in response to fears about misuse of the contact data. Apple&Google have specified an open API&Service for bluetooth low energy contact tracing with privacy. It looks like a good fit, technically, to some of the newer app designs. It does (a little) remind me of what adding privacy to WiFi AP scanning did (to prevent revelation of all the places someone had been by eavesdropping the list of prospective APs in their scan), but to a very different end and in a different way - see links to specifications below. Some comments added on NHS proposed app at the end now.

People are concerned about how this might lead to privacy invasive apps in the future, but first, why do we want this now:

Aside, to keep an epidemic in "virtual lockdown" you need to able to trace and isolate cases before they infect further people and restart the epidemic exponential growth ahead of your trace rate capability. This means there's a relationship between the reproduction rate (R0) of the epidemic in normal population behaviour (contacts that might lead to infection) and the fraction of people likely to be able to give fast accurate contact information - with nominal R0 around 2, this is estimated in the range 40%+ of people out and about. If people wear masks and observe social distancing, the baseline R0 might be somewhat lower. With proactive testing (random or periodic) you also trigger things earlier for people testing positive so the effective R0 is then even lower - the goal is to keep it always effectively well below 1. But note the number  40% of UK population (or even just households) is 20M (10M) roughly.

Could you build an app to "round up all the co-conspirators"?
or all people that were at this protest at this time with this person?

1. agency (replace healthcare with bad cops:) coerce person to equivalent of test +ve: sends notifications: 
2. agency coerce people to reveal whether notified or not 

Could latter be required by, say, employers (e.g. good ones like healthcare, or bad ones like xxx)?
How is that new compared to current Real World contact trace/notify done through interviews/phone visit

Firstly, service doesn't give precision time, nor is their geo-location as part of it.

Phones may already potentially separately run geo-location, so not clear this adds a lot apart from additional evidence of co-location, and spatial precision. So if any of the phones in a co-lo event are also reporting position, you "infect" contacts with a possible inference, if someone can coerce ALL possible contacts to reveal presence or lack of notifications...obviously people at protests could turn off service, and later not ask for notifications. Would that then be evidence too? This seems like a pretty complicated and far fetched scenario...Not very good evidence that some people out of 20M might be co-conspirators. Not clear how the coercion scales without becoming somewhat visible.

Explainer/proper use case/reference:
Google/Apple BLE explainer
Tech spec:

Pandemic mission creep "best intention" temptations:-

1. "Self-report" and Test certification verification.

Given the trigger for the upload of crypted contact info is a positive test with authorisation by the health authority, there's a strong temptation to bundle test certificates 
+ve/-ve/timestamp/ virus v antibody, into an app...

This is orthogonal to the contact side. but employers (especially healthcare employers) might require verifiable clear tests for staff (like CRBs for teachers etc). Is
failure to do something about being notified also a breach of some employment agreement? Is commerce going to coerce?

I suspect people who work in jobs that you care will actually want to respond to notifications and  get tested too, so can tell to self isolate/get treated/get better and back to work in safe knowledge, so
incentives are aligned, no?

In the NHS app case, there are two separate triggers for using contact history to send notifications: 1/ is a self report (yellow alert), 2/ is a positive test result (red alert). A colleague suggests that there should be an intermediate trigger where a call to the UK's 111 service that results in suggestion to self-isolate, could be accompanied (like the positive test result) with an authorization code to the app (given over the phone to the subject) so that like the test, this trigger (say amber alert) would be much harder to troll with fake self-diagnoses and might act as a deterrent to such behaviour since the 111 caller would be identifiable. re-linking with the patient is no more risk than it was in the test case, either.

2. Isolation/lockdown location compliance

Since we don't have absolute geoloc at all, is there a way to find if notified people were in contact with a person who was infected and in breach of isolation/lockdown rules, more than current Real World contact tracing would reveal...? This seems not to be made easier by these 
contact tracer approaches. See above. 

Other concerns include false positive rates in self-reporting - this applies whether the data is centralised (NHSX current app design as of 12.4.2020) or decentralised as with the Google/Apple/DP-3T.

We can assume that there will be fairly high levels of people stressed in the current lockdown, and potentially experiencing some symptoms (e.g. coughing at the slighted thing). We're currently heading out of the period of seasonal flu, so people having genuine symptoms, but caused by something less risky, will be in smaller numbers perhaps? Nevertheless, this is going to contribute a significant "false positive" rate. However, given the goal of all this tech (coupled with more wide scale testing) is to be able to leave lockdown, the effect would be to have some larger number of people self-isolating than expected, but a much much smaller number than the current 65M people stuck indoors. It remains to be seen what that rate would be, but even if 5 times the rate of real symptoms, this would (after the current peak is over - say early May) be quite a modest number. And it is "failsafe"

Another threat sometimes claimed to these systems is trolling. This I don't buy. The whole point of the bluetooth scanning algorithm (since we did ours 11 years ago in Fluphone) is that someone would have to stand next to you (less than 2 meters away) for 15 minutes (or so) to trigger adding you as a contact. You'd probably notice people doing that in the supermarket, on the pavement, etc. Fleeting encounters are not triggers. That's part of the design.

The third criticism I've seen of these contact tracer apps is that they need a significant fraction of the population to run them for them to "work" - actually, this is not strictly true - they need a significant fraction of an infected person's social group (friends, family, colleagues) to run the app to help. This is true for the contact tracing side. but all contact tracing is partial - it is an attempt to reduce the reproduction rate of the epidemic below 1 - any contribution to that reduction helps us avoid a second wave.
The app is also useful (as discussed above) for gathering details to build a more precise model of the epidemic, mathematically, so things like pre- and asymptomatic carrier infection is characterised, and the rate of child-to-adult is understood better, and even the expiry of immunity. For that to work, any reasonable number of people running the app will help. Given other apps (eg. Zoe/Kings app, or the Covid-Sound app have seen  thousand of downloads a day, it is clear that reaching a decent target for that purpose is achievable, whereas to get to herd-levels of contact tracing coverage many be harder.

Question: apple are saying they will mandate the use of the new privacy/decentralised bluetooth scanning API for IOS devices to run scanning in background - is this already in place, or is it after they (and google) release the new scanning code? Would a centralised-store app like the NHS one be blocked (either from release through the Apple App store, or further, actually unable to run (in background) on IoS devices right now? Will update this soon as someone upates me:-)

Baseline: how does manually tracing contacts (extracting addresses/phone numbers from a tested person, and subject to imperfect recall, possibly including people they didn't actually see and forgetting ones they did)- how is that better than digital contact tracing in safety&security?

meanwhile, also, some data on manual versus app based tracing impact on reducing R0 - from lancet paper based on data from china

Tuesday, April 07, 2020

covid-19 & interventions - very very speculative

looking at Mark Handley's graphs of many countries evolution of the pandemic, and the interventions, i'm going to engage in some idle speculation - please don't take this seriously or as a prediction - its just thinking out loud...

people see family multiple times a day
people make daily trips home->work/school
people make weekly trips (home->shops or trips to country)
people make monthly trips (business -> other countries)

this pattern of self-similar journeys underlies this study of cinter-ontact intervals and duration.

If you look at clustering starting in Wuhan, and then to rest of china, then to other country, it really looks like that. Rhyhm and randomness in human movements as also explored in this ref paper.

It seems that in a trip you might meet 1-100 people but only infect 1-2 so its quite hard for most people to catch, but somewhat easy for some
(severity is a whole other question, maybe, but maybe its related too)

35% daily increase corresponds to a doubling time of 2.5 days
assume basic R0 is 2-3 -i.e. as above, so each person adds 1-2 people a day
but they aren't necessarily infectious for 3-5 days so you get double every 2-3 days

Social Distancing
22% daily increase corresponds to a doubling time of 3.5 days
distancing works weakly - i.e. your infectious person travels
more carefully in their daily trips but not carefully enough. see this LSHTM paper for more info which looks consistent?

Speculate - if Covid-19 can be spread by touch, this might be
evidence that 2 meter is fine if everyone washes every time they were near where an infected person touched. so maybe combination of masks and washing would be as good as lockdown if 100% observant, but if 50% obvservant, only roughly halves the spreading rate.
13.5% daily increase corresponds to a doubling time of 5.5 days
lockdown week 1 < works, but you've only removed half the people in the first week

8% daily increase corresponds to a doubling time of 9 days
lockdown week 2, remove other half....but still have tail of who was infected 2 weeks back, showing symptoms this week

lockdown week 3?

Blog Archive

About Me

My photo
misery me, there is a floccipaucinihilipilification (*) of chronsynclastic infundibuli in these parts and I must therefore refer you to frank zappa instead, and go home