A review of the evidence for and against GFP (general factor of personality) and OCEAN (big five)
Note that much of the debate in the field is about whether or not GFP even exists. For example a 2013 paper opened “The overwhelmingly dominant view of the GFP is that it represents an artefact due either to evaluative bias or responding in a socially desirable manner.” (The general factor of personality: Substance or artefact?)
If it is an artifact, that would also invalidate the Big Five. GFP can be extracted from word data (via word vectors or surveys, the results are the same). Alternately, one can extract the first five PCs, and then use varimax rotation. In this rotation, significant portions of PC1 (the GFP) are distributed to PCs 3, 4, and 5 to make Conscientiousness, Neuroticism, and Openness. If GFP is an artifact, why do that?
Even more fundamentally, if PC1 is just an artifact, it makes for a strange version of the Lexical Hypothesis: "personality structure = language structure (minus the first PC)."
A set of personality factors should be said to be supported by factor analysis as far as it achieves simple structure rather than model fit. You can get equivalent fit for wildly different solutions [a], but the solutions which better display simple structure should be more interpretable and replicable in CFA (and I'd reckon, also more heritable).
I know varimax imposes an artificial orthogonality requirement and that many oblique rotation methods are unsatisfactory/subjective due to them relying on researcher hypotheses about how correlated a solution should be, but I don't think that it inherently has to be that oblique rotation has this flaw.
If we relax the requirement that factors be uncorrelated, a more modest criterion for simple structure might be that the correlations between factors can account for any examples of a variable correlating substantially with more than one factor.
Now, given the existence of some factor, how might we go about creating some variable for that factor which correlates with the other factors as closely as possible to what you'd predict from the variable's correlation with its 'parent' factor? The variable which does the best possible job of this is just a clone of the factor itself (or at least, a clone with some measurement error added in which doesn't correlate with any other factor or variable). As such, if we took all of the variables in a set that we think 'clumps' together, pretended that none of the other variables existed, and then took the first principal component of these variables, it's tautological to say that this first principal component is the linear combination of these variables which best facilitates simple structure.
If then, we looked at all the possible ways to group the variables and then select the grouping which minimizes the correlations between the principal components of its groups (or which minimizes the ratio of the variable-factor correlations to the factor-factor correlations; this is probably preferable but more cumbersome to calculate), it's safe to say that this would be among the best representations of the structure of our dataset, mirroring the graphical solutions that a researcher might come to, and letting the data "speak for itself".
Do this with likert self-ratings of a maximally-diverse set of actual behaviors rather than assessments of how similar people rate various adjectives to be, and I think you'd have a tough time coming up with a better model of personality.
As for testing for hierarchicalness,
A) It probably really depends that you get the rotations right
B) I'd be interested to see it looked at by an agglomerative approach rather than a divisive one. Posit next to as many oblique factors as could possibly be justified, and then when reduced to fewer factors, if it's always the case that a single factor in a given layer is responsible for all consolidation that happens relative to a lower one, then your data wasn't very hierarchical.
[a] NOTE: It's trivially the case that model fit can be improved in an exploratory context by just extracting more factors or principal components, but when people try to come up with rules about how many components to extract, this is never what they base the decision on. For a scree plot for example, it's not about which number of factors explains the most variance, it's about where (if anywhere) there's a sudden change of slope in the relationship between factor count and explained variance. With a set amount of factors to be extracted as a given however, all orthogonal rotations will be equivalent in a model fit sense even if they'd behave differently in a test for hierarchicalness or differ in replicability.
The optimal level of aggregation should be considered to be the one which performs the best when assessed for simple structure, with higher and lower levels also having validity if hierarchicalness is displayed. The only possible reason a test of model fit wouldn't conclude that more is better is if there's a punishment for complexity which concludes that at some point, model fit wasn't improved enough to justify the added complexity, and I'm dubious about how well anybody assesses complexity. Existing complexity punishments would, for example, conclude that hierarchical g models are worse than the oblique models they're derived from even when the difference in absolute fit between the two is statistically insignificant and practically negligible; if there's no difference in absolute fit but the hierarchical model is punished more for "complexity", then your measure of complexity is obviously retarded from a theoretical perspective even if it's sensible from an NHST perspective.
Kenneth Kendler has done some interesting work on the personality-psychopathology factor stuff you may be interested in although I'm assuming you're already familiar.
"Of course more neuroticism is bad" - that's simply not true. All personality factors have evolutionary advantages, otherwise they would have lasted millions of years. All of the Big Five will, at times, have the advantage at one or the other side of the spectrum - or in the middle. It depends on the SITUATION whether more neuroticism is good or bad.
Let me give you a real life example: four juvenile raccoons were born in my yard this spring, and I put out food for them every evening. One is very bold - he will run right up to me when I come out with the food dish. He is low in anxiety/neuroticism. Two are medium "normies", and one is high in neuroticism - he hides when I put out the dish, and only comes for the food when I'm back in the house.
At present, the low neuroticism raccon gets the lion's share of the food - but what if there were predators? We actually do have foxes and coyotes. What if they decided to eat a raccoon? The bolder one would likely be the one caught and eaten.
Likeswise, depression has evolutionary advantages. The saddest thing in any mammal's life - human or otherwise - is the death of a child. We will do ANYTHING to prevent our child dying. Imagine a situation where there was no depression associated with a child's death.....
Great, detailed summary. Thanks.
> Because of these correlations, you will get a more general factor which has positive loadings on extroversion, agreeableness, conscientiousness, openness, and a negative loading on neuroticism (thus, it is sometimes called emotional stability (ES) and reversed). It sure sounds like there is a weak tendency for everything good to go together, even if we are usually told that personality dimensions do not measure something better or worse.
Well the stereotype for high openness is a quasi-neurotic constantly chasing the newest trend. In fact, I sometimes suspect "openness" is just a nice way of saying gullibility.
You probably have seen this, but I thought I'd point it out: https://substack.com/inbox/post/136931675. I'd never heard about ego-resiliency, and it seems like an idea with some merit.
Well done, Emil! A solid review.
Ultimately the problem is that we're relying on self-report, and occasionally other-report data. However informative that may be, subjective ratings are still notoriously unreliable. Imagine what intelligence research would look like if we were left with self- and other-report. Eysenck had the right idea with salivary response, body sway, and similar physiological methods, but survey data is much cheaper, so here we are.
Ultimately the 7 factors of HEXACO+Psychoticism/Dissociation/Schizotypy is every useful, even as it's an incomplete description of personality variation. There are probably a dozen, if not a hundred, meaningful factors of personality, but those which influence behavior strongly enough are likely to have been revealed by the current research program. Most people, truly, won't even find any use for the majority of these dimensions, and will probably only care in knowing a person's 1. general intelligence, and 2. a combination of everything else (e.g. H+A+C-E-P) that tells whether they can be a reliable friend, spouse, or employee.