[3/6]
It is important to note a question: "While some commentators provided relatively partisan analysis, others merely expressed surprise at the near-vertical leaps in some of these vote updates," the Substack analysis↱ observes in its characteristically credulous way, the authors go on to ask: "Is it likely this phenomenon would arise organically?" The answer is actually, yes.
Nonetheless, having justified their pretense of further scrutiny, the rickety setup teeters into its explanation of Concept, Intuition, and Measurement. The section opens with a nearly benign paragraph:
Data analysis relies on recognizing and evaluating patterns in data. When we find anomalous data, that is often an indication of underlying differences. This is why in this report we focus on these four vote updates.
The first two sentences aren't absolutely horrible, but in the larger context of concern about vote integrity, what is the relationship between the underlying differences and the purpose of investigation? So far, it's a matter of everything wrong with the setup. The underlying differences are often apparent; the challenge is understanding how to formulate them compared to vagary, ambiguity, and a mysterious and unexpressed idyll.
For instance, something about their sense of intuition comes off as far too simplistic:
There are also a number of general intuitions upon which we draw to direct our research. In general, the larger the sample size, the smaller we expect the deviation from the population average to be. While anomalous vote ratios may occur, the statistical chance of anomalous margins goes down as the size of the sample (or vote update) goes up.
The basic intuition is: big margins are one thing, and so are super-skewed results, but it's weird to have them both at the same time, as they generally become inversely related as either value increases.
The basic intuition is: big margins are one thing, and so are super-skewed results, but it's weird to have them both at the same time, as they generally become inversely related as either value increases.
And we're right back to the wordplay about what is surprising. Still, look at how the logic works: "In general, the larger the sample size, the smaller we expect the deviation from the population average to be." This requires some particular consistency within the sample range that is not necessarily in effect. Check that basic intuition, "as they generally become inversely related". Yes, generally. But think of Milwaukee and environs; we already expected, from voter reg and mail-in request data, an overwhelming Biden trend in those ballots. We already had an idea what those sums would look like, and the idea that those numbers posted about when we expected they would is not surprising. As analytical practice goes, the prospect of insisting on generalization in lieu of data known to indicate otherwise is unsound.
Inasmuch as the report seeks to "demonstrate … that four key vote updates … cut against this intuition", the question of general and particular persists. Given what the sample range actually represents, and acknowledging that four, or even seven, out of over eighty-five hundred, is unusual or uncommon, the question remains whether we should be surprised at the fact of deviation from so general an intuition, or to what degree a lack of deviation from so general a "mathematical property" is a realistic expectation.
Thus, if the authors purport to "show the existence of a very strong inverse relationship within vote updates, across all states and times, between the difference of votes for Joe Biden and Donald Trump … and the the ratio of Joe Biden's votes to Donald Trump's votes", what actually stands out is the range of all states and times. Yes, this is what the data from 8,954 updates shows, but, again, we consider the general and particular. "At any geographical level," the authors explain, "we can test the assumption of an inverse relationship between vote update size and the extremity of the ratio between the candidates' votes, and, as we will see here, the relationship is extremely strong." Again, note the generality, i.e., "any geographical level". Because, while, "Across states red and blue, where turnout is high and low, there is an obvious inverse relationship between the two", an abiding question remains: So what?
Any geographical level. All states and times. Now, then: What do we expect of the mere fact that deviations occur? Is the idea of a statistical outlier according to generalized expectation really so surprising?
The paragraphs spent explaining their metrics are shot through with wondering, "So what?" or, "And?"
Here, think of it this way:
Let us now attempt to quantify the nature of the inverse relationship in the context of a particular state. First we take our data set of running vote totals for each state, and, for each state, calculate the vote differential for each candidate between updates. This produces a sequence of vote differences, the sum of which, within any given state, is the total.
Something about what happens in Milwaukee and environs compared to the rest of Wisconsin, or Detroit compared to the rest of Michigan, or Atlanta compared to the rest of Georgia, seems grossly understated in the basic intuition and mathematical property the analysis describes and requires. There is a reason this argument to cast doubt on the legitimacy of the election depends on abstraction in lieu of substance.
The math is set up so that—
given X for Biden and Y for Trump, either metric will produce a score which is the opposite of what it would produce if the update instead had Y votes for Biden and X for Trump. This property is extremely useful, and will come in handy during the statistical analysis.
—the question becomes whether it is applicable: One can formulate mathematics to do nearly anything they want; it's just a matter of convincing other people. To wit, such as it is—
Readers might ask: Why are you measuring the ratio? Why not measure the difference between the vote proportions (or, equivalently, their percentages). The answer to this lies in what we are looking for, i.e. evidence of fraud or foul play which manifests in extremely unusual outcomes. In particular, ratios are almost never used in expressing vote counts (one typically hears of percentages or, when a race is close, numbers) and so anyone committing fraud and looking to "cover their tracks" is more likely to be "gaming" the metrics they're used to, and much more likely to leave tells in metrics they're not considering.
—they have gone to where circumstance indicates expectation of deviation from arbitrary statistic and complained that a deviant outcome has occurred. The obvious counterpoint remains that the outcome is not deviant unto itself; per prior information, we could expect a Biden flood in those votes, and, per law and circumstance, yes, the overnight reports were the likely time those votes would start to post.
Here is a question of application:
1. Ratios demonstrate an important property: the farther ahead a candidate is, the harder it is to move the next 1 percent ahead. They reflect the relative difficulty of each marginal vote as the pool of remaining votes decreases.As a candidate approaches 0% or 100% of the vote, the rates at which the ratio of that candidate’s votes to the other candidate’s votes converge to zero or infinity are very different.
2. Ratios allow us to spot a potential sign of fraud: unusually low ratios between the losing (major) candidate and other, less well-known candidates. Because those who watch and participate in elections tend not to think in these terms, if there is fraud, they’re much less likely to have covered their tracks in this respect. A tin-pot-dictator style election where the favored candidate gets 99% of the vote is obviously suspect, but less attention is often paid to details like whether the ratio between the most popular losing candidate and long-shot third-party candidates actually makes sense. Looking at metrics which are less popular in practical use will be tremendously helpful here, as we will see.
2. Ratios allow us to spot a potential sign of fraud: unusually low ratios between the losing (major) candidate and other, less well-known candidates. Because those who watch and participate in elections tend not to think in these terms, if there is fraud, they’re much less likely to have covered their tracks in this respect. A tin-pot-dictator style election where the favored candidate gets 99% of the vote is obviously suspect, but less attention is often paid to details like whether the ratio between the most popular losing candidate and long-shot third-party candidates actually makes sense. Looking at metrics which are less popular in practical use will be tremendously helpful here, as we will see.
These are the "critical differences" between the ratio described in their "mathematical property" and the difference between vote percentages. As true as the essence of the first is, its applicability remains a question. The second is of dubious value to Biden-Trump. But what follows are two pained examples of dubious applicability, intended to illustrate the first of these. And if the second of those differences includes self-gratification, so does the bit about San Francisco:
which, despite being one of the bluest cities in America numerically and culturally, is one where Democratic Presidential candidates consistently get about 90% of the vote but never seem to crack 95%. There are Republicans in San Francisco, however few of them, and converting half of them is a tall order. This makes ratios a useful tool in our arsenal for answering questions of the form “how much is too much”?. This allows us to assess the data in a way which we believe is qualitatively different — and qualitatively superior — to the common forms of assessment used by average individuals and the news media.
But it really is a strange appeal to behavior, because consideration of behavioral economy is what the analysis lacks. Everything about the setup relies on generalization and abstraction.
[(cont.)]