Sunday, July 29, 2018

Numbers, We Got Numbers.

Or not.

Biology is sorely deficient in math when compared to physics or chemistry, and is therefore a sort of Junior Varsity member of Team Science -- at least from the 18th century, Scientific Revolutionary perspective. You may recall that one of the Pillars of that Revolution was the privileging of Mathematics as the Language of Discourse for Science. See Descartes for details. (It is possible to quibble with those details -- after all, if the only tool you allow yourself is a hammer, all you will ever see is a nail, and by restricting themselves to what was mathematically "visible," the Scientists blinded themselves to that which was not measurable.) And so we have equations for Newton's Theory, Maxwell's Theory, Boyle's Theory, Einstein's Theory, and sundry others. But we have, alas, no equations expressing Darwin's Theory.

Fair is fair. Math does show up here and there in biology, usually in biophysics and biochemistry. It appears also in genetics, which was pioneered by Br. Gregor Mendel, trained as a physicist be it noted, who conducted a real, no-foolin' designed experiment, the first in all biology. One also sees biostatistics. But statistics is not mathematics and a regression equation, though it bears a passing resemblance to a mathematical equation, is not actually of the same species. Think of it as mimicry. This distinction is a topic for another day, except that the softer the science, the worse its statistical praxis. And when the science is social the praxis gets downright mushy, usually learned via cookbooks in "Stats 101" courses.

Which brings us to today's topic: correlation. Ever since David Hume, correlation has held pride of place over causation due to the inability of inductive reasoning ever to establish causes. It establishes only a co-relation between two (or more) variables measured on the same unit. Thus, it requires:
  • measurements
  • on the same units
Neither is as simple as it sounds.

"There can be no science of any hard empirical variety when the very act of identifying one’s object of study is already an act of interpretation, contingent on a collection of purely arbitrary reductions, dubious categorizations, and biased observations. There can be no meaningful application of experimental method. There can be no correlation established between biological and cultural data."
-- David Bentley Hart, "Daniel Dennett Hunts the Snark"
An example recently promoted by "The Gift that Keeps on Giving," a.k.a. Jerry Coyne, is that "the happiest countries are the least religious." It is itself a happy¹ illustration of Thucydides' dictum that people will swallow anything if it accords with their prior beliefs.² Also that experts who wander off their reservation are no more insightful than the rest of us fools. Herewith, the evidence in chief as cited by the True Coyne:

We note first of all that the author of the graph (who was not Coyne) is unfamiliar with Edward Tufte's classic work, The Visual Display of Quantitative Information. Fair enough, since the display is not, strictly speaking of quantitative information at all.
Notes:
1. happy. Lucky, fortunate. See also mayhap, happen, happening, happenstance; also hapless.
2. Thucydudes, History of the Peloponnesian War, Book IV, 108

Happy, Happy! Joy, Joy!

The Coynester has committed the scientific gaffe of reifying abstractions: to wit: both happiness and religious. Countries cannot be "happy" (or "religious"), only human beings can be either. Perhaps Coyne meant that "the majority of the people in country X" are happy, but this raises the question of how this can be known. How were they contacted? By sample? Was the sample randomized? Stratified? In some countries, folks do not have telephones; indeed, in some countries they do not have street names or addresses. In some, they do not have streets. How do you randomly sample in Burkina Faso? Is it as straightforward as in Singapore? San Jose, Costa Rica, has streets, but the houses are not numbered. How does one identify a sampling frame from which to select random units? Did the data collector even bother to do so? And, if not, in what manner could the results be generalized from the sample to the target population? Allow TOF to express Profound Doubts on these points. To learn that these data were collected by Gallup is encouraging, although that they were gathered under UN auspices is not.

About 1000 people were surveyed in each country in each year. The reported happiness results are the combined results for all the years from 2014-2016. (Although not all countries were surveyed in every year.) Results from multiple years were combined in order to tighten the confidence interval via larger nominal sample sizes. A confidence interval is a mythic figure regarding a parameter of a statistical distribution. Multiple surveys can be combined in this manner only if there has been no substantive change in the population during the time when the various samples were taken, nor any change in the manner of data collection.³ Otherwise, you may be averaging apples and oranges.
Notes.
3. combining surveys. For example, suppose in the prior year there had been much anticipation that candidate X would be elected and usher in the eschaton because the Opposing Party had been cozened into nominating a Bull Goose Loser. But then in the following year, the Bull Goose Loser unaccountably has won the election, causing much weeping and tooth-gnashing. There might be a sea change in happiness, at least within some strata of the population. 
Next, how do you measure "happiness"? With a hap-o-meter? (Preferably one calibrated to a standard certified by NIST). This is perhaps more evidently a problem to a physicist than to a biologist or a social "scientist." The latter in particular is conditioned to accept a questionnaire as an "instrument" and to confuse the answers to a suite of questions with a "measurement." And indeed, so it happened.

The Happiness Scale was measured [sic] using the Cantril life ladder. The English Language version of the key question runs as follows.
“Please imagine a ladder, with steps numbered from 0 at the bottom to 10 at the top. The top of the ladder represents the best possible life for you and the bottom of the ladder represents the worst possible life for you. On which step of the ladder would you say you personally feel you stand at this time?


That's right, it's all about feelz.

Note that the life ladder begs the question of what the "best possible life" means to different people, especially in different cultures. One easily imagines countries in which the "best possible life" is not very good at all; but one also imagines situations in which access to television has introduced formerly happy people to the life of Western Europeans and the sin of envy, which is defined as unhappiness at the good fortune of others.

The original survey also asked about other factors, such as: GDP, life expectancy, social support ("If you were in trouble, do you have relatives or friends you can count on to help you whenever you need them, or not?"), generosity ("Have you donated money to a charity in the past month?"), corruption perception ("Is corruption widespread throughout the government [or throughout business] or not") etc. Collectively, the UN Report claims, these factors account for nearly all the variation in happiness in the world.
Looking ahead, none of these factors included "religiosity." Hence, there was no "room" in the model for religion to account for happiness positively or negatively. Naturally, this stopped Coyne and his religious enthusiasts in their tracks.
Ho, ho. TOF jests. One meaningless social science exercise does not forestall another.

Ordinal in the Court! The Happiness Scale is an ordinal scale, not a ratio scale. A response of "10" on the ladder does not mean that the respondent is twice as happy as someone answering "5". Not like 10 cm. is twice as long as 5 cm., or 10Ω is twice as resistant as 5Ω. This is an illusion created by the imagery of the "ladder" and its "rungs." Without a ratio scale, the whole procedure of averaging and correlating is bogus from the get-go.⁴

It is not clear what the steps on the "ladder scale" mean. A happiness of "8" is eight whats? Not volts, TOF is sure. Happiness is said to "light up a room," so perhaps eight lumens? But TOF doubts it.

Did happiness ratings within a country form a single statistical population? If not, there may be no average for that country. Consider the paste weights of battery grids which were made at different times using different batches of paste, milled fortuitously at different densities. The resulting weights had two distinct averages, not one: those grids pasted prior to 4:00 PM and those pasted after the batch change. 
 If you stick your head in the oven and your feet in the freezer, on the average you're comfortable.
An average is a measure of central tendency, and not all processes have a central tendency. In that sense, there is no average, even if you can calculate one.  What is the central tendency of the number of testicles possessed by human beings? But we digress.⁵
Notes
4. ratio scale. See Deming, The Statistical Adjustment of Data for discussion.
5. testicles. The average is just less than one. Half of all humans have zero. The other half usually have two, but there are occasional accidents and elective surgeries.
Holy Mackerel!
Most of the same points apply to the religiosity scale. What exactly is "importance of religion"? What is meant by a "religion"? Is shamanism even the same kind of thing as Buddhism? In those countries possessing established churches, citizens are enrolled in the approved churches for tax purposes, whether they attend church or not. Does this count? Does pro forma attendance? Does the devoutness of the attendees? Who measures such subjective attitudes? How do they do so? With what precision? Does "religiosity" mean the same thing in China as it does in Bolivia?⁶ Some folks like to point out that atheistic countries like the quondam Soviet Union are "really" religious because communism is "really" a religion. (See Midgley, Evolution as a Religion for a discussion.)

Notice that an individual may regard "religion" (broadly speaking) as "important" (broadly speaking) even if he himself is not a believer.

It is not clear what the scale measures, except that it appears to be a proportion running from 0.00 to 1.00. If so, it is at least a ratio scale.

These religiosity "data" were not collected in the same survey as the happiness "data." That is, the X and Y "measurements" were not performed on the same units. And yet a correlation was performed! 
Notes
6. cross-country comparisons should be made with great care because different countries often use different operational definitions of the variables. Infant mortality is a well-known example. Even within a country, definitions sometimes change. See the discussion sections in the Historical Abstracts of the United States for examples.

Rules of Engagement

A vital issue: Were the happiness and the religiosity "measured" on the same units (people) or only within the same geographical region (country)?

Imagine trying to determine the relationship between the nitrogen content and tensile strength of steel if the two properties were not measured on the same heats of steel. Even a biologist might hesitate to rely on such results -- even if he did not notice the metallurgist rolling on the floor laughing his guts out. As a thought experiment, imagine Coyne's reaction to a correlation between cancer rates in US census districts versus the usage of lawn services in those same districts, with no attempt to discover whether the households experiencing the cancers were the same households employing the lawn services!

Another issue to be considered is whether one should treat all countries as equal units when they vary widely in size: Singapore is small and compact; Brazil is not.

The reported correlation coefficient is r= ‒0.58. No self-respecting engineer would entertain such a value or r for a New York minute, although TOF has been told that soft "scientists" put much stock in any r greater than zero, provided they have wee p-values. This can be ascribed to their training in "cookbook statistics". But let it be said that you can have a very high "confidence" around a very wrong value. A confidence interval is a statement about the precision of an estimate, not about its accuracy, let alone its appropriateness. An r= ‒0.58 means an r²= 0.34. This means that only about one-third of the variation-among-countries in happiness is "explained" by its association with religiosity of those countries. (Whatever that means.)

Hot Dogs and Hamburgers

A correlation coefficient, so beloved by soft "scientists" is almost meaningless. Among the assumptions built into the woodwork is the assumption that the data come from a statistical population. That is, that the data represent a constant system of common causes. But out in the wilds, we most often encounter mixtures of populations: units produced at different times, from different material lots, under different operational conditions. Sometimes these differences have no practical effect on the measurement of interest, but sometimes they do. You cannot leave that to assumption. TOF has seen cases where it mattered who the operator was who ran the machine, or made the measurement, or even what time of day. And if these muck up manufactured product, where the output is supposed to be uniform and to specifications, how much more so wild data which is under no such domesticated obligation?




It is not clear that the scatterplot is a hot dog or a hamburger with a tail. That is, the weak appearance of a correlation is due to multiple clusters of points. The vast majority of points form an amorphous ball on the right. A second cluster in the northwest consists of Western Europe and a third cluster in the southwest consists of Eastern Europe and East Asia. This is a common pattern on scatterplots.

Apparent correlations between X and Y can indeed come about when:
  1. X is a cause of Y
  2. Y is a cause of X
  3. Z is a lurking cause⁵ of both X and Y
  4. coincidence

The correlation in the left hand plot apparently shows that errors decrease with increasing workload on the clerks. The managers were delighted. To reduce errors we will give the clerks more work! But wait. There were two clerks: Adam and Betsy. Betsy was more experienced. She got more work done and made fewer errors than Adam. There was a causal relationship, but it was not between X and Y! It was between Z and X and between Z and Y. (In the actual case, there were four clerks. The case has been simplified for presentation purposes.)

These are all technical issues associated with the use of the statistics; but there are also substantive issues associated with the hypothesis supposedly being tested.

Everyone gives lip service to the fact that correlation is not causation, but then turns around and acts as if it were. The Coynester is no exception to this rule and chortles over the "fact" that religion does not result in happiness for its practitioners. (Notice the leap in logic here. That is not even what the data is supposed to show. These are countries, not people.)  But why should anyone suppose that "religiosity" however defined should be expected to entail "happiness" however defined? It may be the opposite case: unhappiness may entail religiosity, at least of certain types. Recall


Notes
5. lurking cause. A nice article on the subject is Brian Joiner. "Lurking Variables: Some Examples." The American Statistician 35(4): 227-233 (Nov 1981)

Wonder and Anticipation, the Likes of Which We Have Never Seen

  Hello family, friends and fans of Michael F. Flynn.   It is with sorrow and regret that I inform you that my father passed away yesterday,...