Powered by glolg
Display Preferences Most Recent Entries Chatterbox Blog Links Site Statistics Category Tags About Me, Myself and Gilbert XML RSS Feed
Sunday, Nov 22, 2020 - 01:59 SGT
Posted By: Gilbert

By The Numbers

Events are moving very fast indeed, with the Hong Kong travel bubble already deferred for at least two weeks, after a rash of cases in the fellow former British colony, though we did level the playing field by mandating tests for those incoming from Hong Kong too. On the bright side, both Moderna and Pfizer have reported efficacies of over 90% for their vaccine offerings, and while there have been some caveats about the exceedingly low base rates involved, this still sounds pretty promising, if not quite the 100.0% protection vaunted by Sinopharm. Conversely, Remdesivir has sadly been unrecommended by the WHO, while the largest RCT on face masks known to date - over six thousand subjects by a group from Denmark - suggests that "...the recommendation to wear surgical masks to supplement other public health measures did not reduce the SARS-CoV-2 infection rate among wearers by more than 50%".

Again, however, one might consider the implications of setting a high level of decrease in infection/improvement as the hypothesis (50%, coincidentally also the level for the NEJM paper generally cited as evidence against early use of HCQ); in practice, this means that a smaller positive effect might in fact be observed, though under the threshold of significance - which, as it happens, is the case with face masks here. 1.8% of masked participants were infected, as opposed to 2.1% of control participants, which seems to suggest that face masks did help - but only by about 15%, relatively speaking.

Note that this does not contradict our reasoning from Feburary: masks, adequately-manufactured and worn properly, likely do help to some extent, but are obviously not foolproof (without getting into hazmat suit territory, anyway). It possibly becomes a bit of medical theatre when one accepts certain makeshift alternatives such as bandanas and neck gaiters as adequate, however, and the idea was hopefully always to stack multiple sources of prevention and protection - masking, voluntary social distancing, early treatment. On this, I'd gather there's been politicization to spare, on both sides of the American divide.


Negative example: Bandanas/gaiters, hole in mask
(Source: insider.com)


Returning to the reversal of the Hong Kong travel bubble plan. This, while disappointing, might hardly be worth a mention in the face of the United Kingdom bumping their defence spending significantly, and describing the situation as "the most perilous time since the Cold War". While our extremely-busy PM has claimed that Singapore would not join America's upcoming "Cold War-style coalition" - it's not like he was going to brandish an application form - I gather he's under few illusions about how the fracturing U.S.-China relationship's likely to develop, what with the Five Eyes cabal keeping the pressure up on Hong Kong disqualifying elected legislative council members.

That, by the way, came after Uncle Sam floated the idea of resurrecting the First Fleet with explicit and delicately-diplomatic references to "allies and partners like Singapore [and] India", not that the U.S. Army's waiting before purchasing user location data on Muslims en masse - including locals, apparently. Analysts are expecting Singapore to profess neutrality (officially, anyway) for as long as possible, but given how blocs appear to be forming whether our PM wishes it or not - and he did request for the Navy's presence, recall - there might soon not be much of a decision to be made. It's not as if hedging bets is working out that splendidly, seeing as how India's squeezing DBS on home ground, and contesting our proposal on food exports at the (fading) WTO as well, despite having an agreement; happily, as far as supplies go, we appear to be doing fine - so far.

It's slightly odd as to how the local media's jizzing themselves on a Biden transition, however, seeing as how he's assembling the same old collection of warhawks (honestly, did anyone expect any different?), and this in an environment where just about the only bipartisan initiative that looks like it can win broad support, is getting - and staying - tough on China. Notably, the first part may be a bit of an assumption in itself, with nearly half of American likely voters apparently of the belief that the election was stolen from the sitting POTUS... including 30% of Democrats. Facebook for one has quietly removed Biden's "President-elect" status for now, so it's probably not entirely unwarranted to go through how the drama might yet be settled, from what's been slung around the past fortnight.


Benford's Law Visited

We'll begin with a pattern mentioned back in May, from the piece on reviewing: Benford's law. It was presented as an example of a well-known forensic statistical test on the distribution of digits in natural collections of numbers then, but not further expounded on, and seeing as to how some of the earliest questioning of the election results' been based on it, this might be a fine opportunity to watch it applied in practice.

The original thesis, it seems, is that the distribution of Biden's numbers did not follow that as expected from Benford's law in some contested counties/cities, such as Milwaukee and Allegheny. A public analysis has been posted on Github, and gained a fair amount of traction. However, as it turns out, this is probably one of the weaker directions that has been explored, as far as uncovering possible duplicity goes. To begin with, analysis of the first digit may not be very appropriate with precinct-level data, especially where precincts are mostly of roughly-uniform size. As an example, if most precincts of a county have say six to eight hundred voters, and about 60% of them vote for some candidate, that candidate's leading digit should tend to be higher than one.

Now, this is not to say that Benford's law is entirely inapplicable to electoral data; one might realize that using the second digit would address those concerns to a large extent, which as it turns out seems to indicate nothing awry about Milwaukee or Allegheny's figures. Proper application would also involve comparison with a large set of known fair elections, and in the absence of such, it seems rational to set this angle aside for now - all the more as there's no lack of more fertile fields to till.


Why might this presentation be misleading?
(Source: theiowastandard.com)



Turnout/Time Series Studies

Continuing, a common complaint has been on unlikely turnouts being recorded, and while some of them seem to be based on faulty calculations (e.g. on Wisconsin turnout being near 90%), there have been no lack of slightly-strange ratios being pointed out. The most prominent of these were perhaps in Michigan, where a software settings error was for instance blamed for displaying a turnout of more than 100% in some precincts, for instance (on which more soon), which had Wayne County election board members rescind their certification of the results, after being denied an audit. One could think this kinda petty... until it's realized that a brief audit in Georgia turned up three thousand-plus uncounted ballots in three counties, all favouring the reigning POTUS in aggregate.

Expanding into the spatial domain, the fate of bellwether counties was a particularly large anamoly. Recall, bellwether counties are considered as historically representative of a country as a whole, and thus their winners are also almost always the winners of the entire election (at the state level, Ohio has often played this role). Well, it seems that the recognized top 24 bellwether counties since 1960 - which have maintained a predictive accuracy of at least 75% from then until 2016 - have suddenly managed an accuracy of just 33% this term. While this has been explained away as being due to polarization, it's only added grist to the rumour and doubt mill.

There's been analysis on Pennsylvania in particular, but perhaps the most innovative analysis I've come across thus far was attempted on time series data scraped off the New York Times website. The central principle is as follows: while batches of in-person votes can have wildly-differing distributions between the candidates, due to arriving from different precincts and even blocks that might have significantly varied political preferences, mail-in votes are shuffled in the post. As such, one would expect a messy cloud of points initially (corresponding to in-person votes) when vote ratios are plotted against elapsed counting time, but then a long tail of batches where the ratio is nearly perfectly-stable (corresponding to the mail-in votes). Indeed, this is supposedly what is generally seen, in most states.

Now, this fellow claims, this pattern specifically does not hold in Pennsylvania, Michigan and Georgia. In these states, there is a consistent and gradual (if slight) increase in the vote ratio towards Biden, as new batches of mail-in votes are reported:


Up, up and away!
(Source: threadreaderapp.com)


To be fair, one obvious explanation presents itself: a slight drift from D to R mail-ins had been said to be common, and was attributed to the batches with relatively more R votes arriving from outlying rural areas; one could, then, imagine it happening in reverse, with city votes (i.e. Philadelphia, Detroit, Atlanta?) being left for last in these states. The smoothness of the shift could happen if the mail-in batches were blended in alongside in-person votes according to some schedule, I guess, and there's also the issue of this NYT data supposedly already being preprocessed. Still, this appears in principle a pretty fun avenue to pursue, assuming the true raw vote counts and timings are available - which could be a teeny problem.


2019: They tested the machines by shaking them! Haha!
They can be hacked in a minute! Holy shit!
2020: Machines 100% secure, nothing to see here, citizen


The machinery behind the counting has come under some fire too, and in this respect, the biggest accusation by GEOTUS's legal team is probably that a percentage of his votes were switched to Biden by the software. The direct antecedent appears to be an analysis by a certain Dr. Shiva, who observed that TRUMP individual votes appeared to decrease linearly as the straight Republican vote percentage increased, in certain important counties. Admittedly, one would again expect a large-scale investigation as to the behaviour of other counties - as well as for Biden - to get a better idea of whether this is truly unusual, but the argument appears to have gained some admirers, including Michael Burry of Big Short fame.

This has seemingly been combined with the allegation that the election management system used can store votes as fractions, which frankly doesn't seem to have any legit reason for being, and to the best of my knowledge has not been debunked by the many eager fact-checkers around. Anyway, as it stands, the Dominion firm involved has backed out of a Pennsylvania hearing to lawyer up, and it's a little strange as to how the corporate media's so confident of their probity, given that Warren, Klobuchar and other Democrats had called for an investigation into just that company, less than a year ago.


The Routes Ahead

Having very briefly surveyed some of the election data analysis going around, it might be reaffirmed that the statistics were never going to be more than indicative. The real battles were always largely going to be procedural, especially on what ballots were or were not legal - whether they had arrived before some agreed cut-off timing (itself legally set), whether they could be verified, that sort of thing. The exceptional proportion of mail-in votes has arguably introduced greater-than-usual opportunity for partisanship - since one would expect a mailed ballot to be for one party far more often than others, there's an incentive to favour that party by minimizing the number of rejects - whether explicitly (and potentially illegally) or implicitly enforced.

Now, there are some plans that have TRUMP winning by running out the clock, so to speak, but that would be pretty bad, even given the terrible outcomes that remain. Even if Biden takes the seat, unity's looking a pipe dream at this point (although, as a consolation, I'd wager that there would be a lot less property damage than if it had gone the other way). The fundamental problems preventing an idealized coronavirus response would remain, to give but one example, from how every single county sheriff in the Los Angeles region has declined to enforce their Governor's statewide curfew.

And, in that case, one can hardly deny the draw that The Donald continues to represent within the G.O.P., whether among the common voter (record vote of over 73 million, remember) or the congress critters, and from how Fox News is slumping majorly after abandoning their audience. Given this, a four-year break to build up his own media empire could well be in the cards, before a comeback for the ages.



有一些叛逆, 还有一些疯狂!




comments (0) - email - share - print - direct link
trackbacks (0) - trackback url


Next: Data Binge (Keto)


Related Posts:
Déjà Vu
Being Accurate For Fun And Profit
Next Episode
Late Breaking News
Think Thunk
Back to top

Copyright © 2006-2025 GLYS. All Rights Reserved.