The consequence of ignoring measurement (and sampling) error(s) in subject matter. An example with a race-denialism argument.

When I come to read the same fallacy over and over again, I become tired. As shown in the right upper corner on my blog, an isolated anecdote cannot be equated with data. The reason is due to both measurement error and sampling error. To evaluate the pattern of an event, enough individuals is needed, if possible randomly sampled and representative of the population we want to generalize, and for each person sampled we would need multiple measurement/assessment. I suppose it is easier to rely on anecdote(s) because those are events that have been experienced personally and have some part of emotion. They are easily kept in memory. Data are just an endless list of numbers. Too cold. Too abstract. And people don't accustom with that. Someone like Kanazawa (2004a, 2004b) can even say that his savanna principle can account for it. Why not. But that is not the subject I need to discuss.

To this matter, I will illustrate with an example I have seen in certain forum polluted by clowns. The topic of discussion was about the validity of the concept of race. The discussants agreed that if we cannot tell the exact number of races, i.e., if there is no perfect agreement, the word itself has no meaning, no validity. I will not provide a link toward the said forum, because it is polluted with hate comments by haters who don't have civilized behavior. These morons don't deserve a link. I don't give jam to pigs. I just eat them. But enough with the off-topic. The focus was the following :

These people use that picture in order to dismiss the concept of race. It was asked to name the races of those people. The argument on which the procedure is based is dubious. When we look at the colors in a rainbow, they don't consist of discrete bands but instead they are just like a continuum. All of these variations are not easily classified. On the other hand, we can still distinguish red, blue, yellow, green, purple, violet, etc. We can distinguish how close these variations are, relative to red, blue, etc. With respect to the spectrum, red color itself shades imperceptibly into orange and orange into yellow. Another way to look at the problem is to examine IQ scores. The IQ distribution is a perfect continuum. And yet we are able to distinguish and conceive racial differences in IQ. We are able to group people by low-, middle-, high-scoring groups, and yet without the same criteria of score-group(ing). The categories defined in intelligence studies are purely arbitrary, but need no universal agreement, and yet the concept of group differences is not invalidated. One can appreciate how Baker (1974, pp. 100, 110) defended the concept of race. If there are no intermediates, there are no races. Just as there are no group differences in IQ without continuum. Some other useful insights are provided by Jensen (1998, p. 425), Sarich & Miele (2004, pp. 209-210), Sesardic (2013), Fuerst & Boetel (2014).

The same goes for the racial groups. A subgroup among the asian group will not be classified as belonging to the sub-saharan african group, in most cases. Obviously, the more we divide the racial groups by subgroups, the more likely the errors. That does not invalid the concept. Genome clustering studies support this perspective.

Artifactual errors are what cause a diminishment in correlations. One obvious example is beauty evaluated with photos given to some evaluators, e.g., those "interviewers" who give the questionnaires in survey data (sometimes longitudinal). The common argument is that because beauty is subjective, we cannot evaluate a "beauty score" so to speak. I have extensively addressed this problem in an earlier (modified) post. Measurement unreliability causes agreement to be lower, thus when the same pictures would be given to other people, the rank-ordering in beauty score is not expected to be in agreement among the different evaluations. This necessarily results in low(er) correlations. Most of the empirical data shows however that the correlation is not trivial. Subjectivity does not preclude agreement.

Concerning the photos, measurement errors can be more pervasive than expected. People may look very presentable on the picture, depending on the angles. And on some other angles, they look less presentable. For the question of determining with precision which race which person belongs, it is even more complicated. Under some angles, some physical traits that are more commonly acknowledged among one own's (or other) population can be markedly hidden (or pronounced). This renders classification even more hazardous. Some people may look better (or think they are) when they are pictured on the side rather than taken in front. Some people may even feel they look better on the right (or left) side on their face. Because of some beauty spot, or some other specific features, who knows. But some features at some location can confuse people when asked to name the racial groups.

Now, back to the core of the matter. I remember the guy who posted the above picture was the same who gave the answers. To be fair, I have read the answer given by that guy. Only stupid people play stupid games. But I suppose I would need to say it anyway. E is Angelina Jolie. I doubt everyone would recognize her. F is indian. Well done again. Indian people are quite varied, surely due to admixture. G is Sri Lankan. Who is going to guess that ? Even though these three are celebrities. C is negrito. I is australian. The most interesting faces, perhaps, are B and J. Both of these players are swiss. I bet everything no one would be able to classify them as belonging to the exact same stock. Like I said above, the pictures are selected so as to disseminate confusion.

Regardless, the difficulty is disproportionately high because it wasn't asked to name the major races (e.g., africans, caucasians, mongoloids) to which these people belong. No, it was something much more precise. But how can someone give a detailed response without being provided with accurate descriptions ? And how can someone categorize anyone of these people if they belong to one (or more) race(s) that the respondents had never heard or learned about ? Members of population X who live day to day with members of population X certainly have much more ability to distinguish between members of X and Y or Z than distinguishing between members of Y and Z because members of X have much more contact with members of X than with the others. Hence, all of the aforementioned conditions will not be fulfilled. That's why the question is ridiculous to begin with. Thus, one side of the problem is that if someone hit the correct response, it would most likely be due to chance, owing to informational problems. Any attempt to answer that question has no meaning if the respondents rely on guessing. If such phenomenon occurs, then, in the field of psychometrics, it is called measurement bias or, alternatively, differential item response functioning. Scores between respondents are not fully comparable because they would have achieved the same total score but with different response structures. Now, the other side of the problem, as said above, is sampling error. One single picture to represent one person is not necessarily accurate. We don't have the average. Also, one person cannot accurately represent one racial (sub)group for the same reason one person cannot speak for everyone. Collections of pictures allow averaging of the outliers or the extreme values in any given characteristics. A representative sample must have equal mean and variance than what is seen in the population estimates. That is the purpose of the averaging method.

A similar story emerges regarding the so-called Lewontin's Fallacy (Mitton, 1977, 1978; Smouse et al., 1982; Edwards, 2003; Sesardic, 2010). When too little information is sampled and compared across groups, the differences won't emerge so easily, but when the informations accumulate, the differences emerge because of the larger reliability of the relevant information. One illustration is given by Witherspoon et al. (2007, Figure 1). We see that the overlap diminishes as the information increases.

Surely, this post isn’t funny to write up.