ArticlePDF Available

Some new methods for exploratory factor analysis of socioeconomic data

Authors:
  • Ulster Institute for Social Research

Abstract and Figures

Some new methods for factor analyzing socioeconomic data are presented, discussed and illustrated with analyses of new and old datasets. A general socioeconomic factor (S) was found in a dataset of 47 French-speaking Swiss provinces from 1888. It was strongly related (r’s .64 to .70) to cognitive ability as measured by an army examination. Fertility had a strong negative loading (r -.44 to -.67). Results were similar when using rank-transformed data. The S factor of international rankings data was found to have a split-half factor reliability of .93, that of the general factor of personality extracted from 25 OCEAN items .55, and that of the general cognitive ability factor .68 based on 16 items from the International Cognitive Ability Resource.
Content may be subject to copyright.
Submitted: 7th of September 2015
Published: 7th of November 2016
Some new methods for exploratory factor analysis of
socioeconomic data
Emil O. W. Kirkegaard*
Open Quantitative
Sociology & Political
Science
Abstract
Some new methods for factor analyzing socioeconomic data are presented, discussed and illustrated with analyses of new
and old datasets. A general socioeconomic factor (S) was found in a dataset of 47 French-speaking Swiss provinces from
1888. It was strongly related (r’s .64 to .70) to cognitive ability as measured by an army examination. Fertility had a strong
negative loading (r -.44 to -.67). Results were similar when using rank-transformed data. The S factor of international
rankings data was found to have a split-half factor reliability of .93, that of the general factor of personality extracted from
25 OCEAN items .55, and that of the general cognitive ability factor .68 based on 16 items from the International Cognitive
Ability Resource.
Keywords:
general socioeconomic factor, S factor, exploratory factor analysis, research methods, Switzerland,
reliability, intelligence, cognitive ability, IQ
1 Introduction
Exploratory factor analysis
1
is a method for finding
underlying dimensions, called factors, in datasets. Be-
cause factor analysis is so useful for many purposes,
it is widely used in many dierent sciences e.g. clima-
tology, chemistry, biology, geology, psychology, com-
puter science and sociology. The application of the
method is essentially limited only to those areas of sci-
ence where one needs to look for underlying patterns
among variables.
Factor analysis was first invented to be used in psy-
chometrics to analyze cognitive ability data (Cudeck
& MacCallum,2007;Spearman,1904). Cognitive abil-
ity data are easy to analyze in the sense that better
performance on one test is always correlated with bet-
ter performance on another test (called the positive
manifold) and the interest is mainly in the first factor
because this has most or all of the predictive validity
(Dalliard,2013;Jensen,1998;Ree et al.,2003). This
means that when a single factor is extracted, all the
*
Ulster Institute for Social Research, United Kingdom. E-mail:
emil@kirkegaard.dk
1
Principal components analysis is included here. Some readers
will bark and say that it is mathematically very dierent (Everitt
& Dunn,2001). That may be true, but in practice the results
are nearly the same as using regular factor analytic procedures
(Jensen & Weng,1994;Kirkegaard,2014b).
factor loadings are positive. This has certain simplify-
ing implications for methodology.2
In a previous publication (Kirkegaard,2014b), I used
factor analysis to analyze international country rank-
ings. I found that there is a strong general factor
that corresponds to the concept of general socioe-
conomic well-being or performance (which I called
S). Briefly put, this means that in most, but not all,
cases desirable outcomes have positive loadings and
undesirable outcomes have negative loadings. This
finding has since been replicated in a number of other
datasets covering intra-national regions and persons
grouped in various ways (Kirkegaard,2014a,2015g,d;
Kirkegaard & Tranberg,2015).
The purpose of this paper is to review methods pre-
sented in earlier papers and to introduce new ones
that were developed for the factor analysis of socioeco-
nomic data. Some of these were presented in various
earlier papers and some are new. The reason to have a
single review paper of the methods is that otherwise
2
Davide Pier pointed out the exception with elementary cog-
nitive tests (Jensen,2006). These have negative correlations
to the other tests because shorter response times mean better
performance. Sometimes researchers reverse the response times
(multiply by -1) to preverse the all positive correlation matrix.
These tests are rarely used and so the problems with the negative
correlations they cause are rarely present.
1
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
the reader would be forced to read a number of other
papers to learn about the respective methods. While
the methods were developed for the purpose of study-
ing socioeconomic data, they may be useful for other
types of data as well.
2 Jensen’s method with reversing
This method was first used in Kirkegaard (2014b).
Sometimes methods developed to be used for factor
analysis of data that has a positive manifold do not
work well when used on data where there can be gen-
uine negative loadings. Arthur Jensen invented what
he called the method of correlated vectors (I will refer
to this as Jensen’s method) to examine whether a given
criterion variable was related to the general cognitive
ability factor (g) or whether it was related to other
parts of the variance (Jensen,1985,1998;Jensen &
Reynolds,1982). The method consists of calculating
the correlation between the indicators’ g-loading and
the vector of the criterion variable’s correlations with
each indicator. The theory being that if the criterion
variable is related to general cognitive ability approx-
imated by the g factor, then the subtests that measure
this trait better should correlate more strongly with
the criterion variable as well. Likewise, if a criterion
variable is related to the g factor, but to other parts of
the variance, then the correlation should be negative.
A large and growing number of studies have used this
method to examine the relationship between the g
factor and variables such as brain size (Rushton &
Ankney,2009), group dierences (Frisby & Beaujean,
2015;Jensen,1985;McDaniel & Kepes,2014;te Ni-
jenhuis et al.,2015), the Flynn eect (te Nijenhuis &
van der Flier,2013), test training/re-taking gains (te
Nijenhuis et al.,2007), and education related gains
(te Nijenhuis et al.,2014). For criticism of the method,
see e.g. Ashton & Lee (2005).
The method is generally applicable. I have previ-
ously used it to examine the relationship between S
and other variables with strongly positive results, e.g.
(Kirkegaard,2014b). One problem with this is that
the Jensen coecient (the resulting correlation from
applying the method) is sensitive to whether there are
variables with negative loadings or not.
3
If there are,
then the Jensen coecient will be inflated towards
±
1
because the presence of the negative loadings greatly
increases the variance. However, the negative loading
of the variables depends on arbitrary choices made by
coders. For instance, one could use a variable called
percent of youth with at least a high school education
in which case one would get a positive loading. One
could also code the same data negatively and call it
3
Thanks to Marc Dalliard who was the first to notice this problem.
See the peer review thread for the first S factor paper at
http://
openpsych.net/forum/showthread.php?tid=77.
percent of youth without high school or better. The load-
ing of the indicator will depend on which way one
coded it, and this aects the application of Jensen’s
method. Thus, if one wanted, one could strategically
recode about half the variables in an S factor analyses
and so inflate the Jensen coecient to near
±
1. But
results should not depend on arbitrary coding choices
made by researchers, especially not when they can be
gamed.
A simple solution is to recode the variables such that
higher values correspond to desirable outcomes. This
works well in many cases, but not all. In some cases
(e.g. population density, fertility, economic inequal-
ity), there is no clear answer with regards to which
direction is the desirable one. Thus, subjective judge-
ment calls aect the results which is undesirable.
Instead another method was chosen to deal with the
problem: reverse all indicators with negative loadings
(called reversing). However, in the published studies
where this was done, comparison figures without re-
versing were not included. Thus, in order to illustrate
the method, a new S factor analysis is presented be-
low.
2.1 An S factor analysis of 47 French-
speaking provinces in 19th century
Switzerland
R
4
includes many datasets to use for testing and il-
lustrative purposes. One such dataset concerns 47
French-speaking provinces in Switzerland and dates
from around 1888. The dataset contains the following
variables:
Fertility: Ig, ‘common standardized fertility mea-
sure’
Agriculture: % of males involved in agriculture
as occupation
Examination: % draftees receiving highest mark
on army examination
Education: % education beyond primary school
for draftees.
Catholic: % ‘catholic’ (as opposed to ‘protes-
tant’).
Infant Mortality: live births who live less than 1
year.
4
R is a statistical computing language. See
https://www.r
-project.org/about.html.
2
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
Quoted from the dataset description. Use ?swiss in R to
see.
Four of these are clearly socioeconomic indicators: fer-
tility, workers in agriculture, secondary educational
attainment and infant mortality. The last two vari-
ables are demographic and cognitive, which we may
consider criterion variables in this analysis. Note
that the cognitive variable is a threshold measure of a
(presumably) normal distribution of cognitive ability.
This results in a non-linear transformation (La Grie
du Lion,2001,2007) that is expected to decrease the
correlations to some unknown degree.
The 4 S indicators were factor analyzed and Figure 1
shows the factor loadings.5
Note that the figure includes a note in the top left
corner that the indicators are reversed. This is be-
cause when factor analysis is performed, the factor
is turned such that most loadings are positive by de-
fault. This is generally what one wants. However,
in an S factor analysis, to align loadings and scores
with theory, undesirable indicators should have nega-
tive loadings and desirable positive. If one analyses
a dataset that includes mostly undesirable variables,
the result will probably be that these are assigned
positive loadings. To reverse this, one can multiply
the scores and loadings by -1.
This factor turning can result in problems if one ex-
amines dierent subsets of a larger dataset as we will
see later in this paper.
Table 1shows the intercorrelations between the vari-
ables.
Cognitive ability (Examination) correlates .70 with
S, a large correlation in line with results from many
other regional studies e.g. as found when analyz-
ing regions of India, Brazil and Italy (Kirkegaard,
2015h,c,f). There is a small negative correlation with
Catholic % for S, but a fairly large one for cognitive
ability. If we ignore the fact that there are too few vari-
ables here for Jensen’s method to work well (the pre-
cision is too low given only N
indicator
=4; (Cumming,
2012)), then we could use the method to examine
whether the observed correlations between the ex-
tracted factor scores and the criterion variables could
plausibly be attributed to the underlying trait.
Figures 2and 3show Jensen’s method when applied
to the S x Catholic % relationship with and without
reversing.
We see a drastic change depending on whether re-
versing is used or not. The strong negative correla-
tion seen in Figure 2was entirely due to the negative
5
Extracted with default settings. The scores were highly stable
across all method parameter choices, all r’s >.99.
loadings of the non-Education variables inflating the
variance. The standard deviation of loadings in the
first case is .78 and .37 in the second.
Figures 4and 5show Jensen’s method when applied
to the S x cognitive ability relationship.
For the relationship to cognitive ability, however, re-
versing made little dierence: the coecient changed
from 1.00 to .92. In other words, the strong posi-
tive relationship was not due to negative loadings
inflating the variance. Of course, because the num-
ber of indicators is only 4, not much certainty can
be ascribed to this finding (analytic 95 % confidence
interval: -.35 to 1.00).
To sum up, this study replicated the usual S factor
findings for a new dataset. The strong negative load-
ing of fertility in a dataset from the 19th century is
interesting, suggesting that dysgenic selection was
already in eect (Clark,2007;Lynn,1996). Caution
is advised for this interpretation because the data are
analyzed at the aggregate level. It is possible that
the individual level relationship is dierent from the
aggregate level (ecological fallacy).
3 Identifying structural outliers
This method was first used in Kirkegaard (2015b).
Exploratory factor analysis attempts to identify one
or more underlying dimensions in the data. Some
cases however might not exhibit the same structure
as the majority of the cases.
6
For instance, a case
might have high crime rates, high use of social ben-
efits as well as high income and high educational
attainment despite the loadings for these being nega-
tive and positive respectively. Such patterns are often
seen for cases that consist mostly of one large city
(Carl,2015;Kirkegaard,2015d). I previously called
this phenomenon mixedness because the indicators of
these cases give a decidedly mixed picture of the case,
but it seems more suitable to use the term structural
outlier (Kirkegaard,2015b).
Previously, I developed two metrics for measuring
the degree of structural outlierliness. The first metric,
the mean absolute residual, is calculated as follows:
1. Extract the general factor.
2. Extract the factor scores.
3. For each indicator:
1) Regress the indicator on the factor scores.
6
Another option is that the data are actually composed of two or
more sub-populations with markedly dierent factor structure.
This scenario is not considered further in the present paper but
warrants further study.
3
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
Figure 1: S factor loadings for the Swiss dataset.
Table 1: Intercorrelations in the Swiss dataset. S_noGe and S_rank are explained in Section 3.2.
Fertility Agriculture Examination Education Catholic Infant Mortality S S_noGe S_rank
Fertility 1.00 0.35 -0.65 -0.66 0.46 0.42 -0.67 -0.57 -0.49
Agriculture 0.35 1.00 -0.69 -0.64 0.40 -0.06 -0.64 -0.60 -0.66
Examination -0.65 -0.69 1.00 0.70 -0.57 -0.11 0.70 0.64 0.68
Education -0.66 -0.64 0.70 1.00 -0.15 -0.10 1.00 1.00 0.81
Catholic 0.46 0.40 -0.57 -0.15 1.00 0.18 -0.16 -0.21 -0.27
Infant Mortality 0.42 -0.06 -0.11 -0.10 0.18 1.00 -0.10 -0.05 -0.05
S -0.67 -0.64 0.70 1.00 -0.16 -0.10 1.00 1.00 0.81
S_noGe -0.57 -0.60 0.64 1.00 -0.21 -0.05 1.00 1.00 0.88
S_rank -0.49 -0.66 0.68 0.81 -0.27 -0.05 0.81 0.88 1.00
2)
Calculate and save the standardized residuals.
4.
Calculate the mean absolute residual by each
case.
The idea is that if a case follows the general structure
of the data, then we should be able to predict that
case’s scores on the indicator variables from the factor
score. The metric calculated above is a measure of
this predictability. A value of 0 means that indicator
scores are exactly predictable if we know the factor
scores.
The second metric, change in factor size, is calculated
as follows:
1.
Extract the general factor from the complete
dataset.
2. For every case:
1)
Create a subset of the dataset where this case
is excluded.
2) Extract the general factor from the subset.
3)
Extract the proportion of variance explained
and save it.
4)
Calculate the dierence in the proportion of
variance to the factor analysis using the com-
plete dataset and save it.
The idea here is that highly mixed cases decrease the
size of the general factor because they don’t fit the
pattern well. The opposite is also possible, namely
that they fit the structure ‘too well’ and so inflate the
factor size.
4
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
Figure 2: Jensen’s method applied to S x Catholic %. Without reversing.
Figure 3: Jensen’s method applied to S x Catholic %. With reversing.
5
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
Figure 4: Jensen’s method applied to S x cognitive ability. Without reversing.
Figure 5: Jensen’s method applied to S x cognitive ability. With reversing.
6
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
A variant of the second metric is the undirectional
version, where we are just interested in changes that
have a large eect on the recovered factor structure
and for this reason we use the absolute value.
3.1 Two variants of a new metric based on
changes in factor loadings
It is possible that exclusion of a given case results
in a large changes to the indicator loadings but in
such a way that the size of the general factor is not
changed. Such cases would be undetectable by the
change in factor size metric described above. The
extent to which it changes the structure to be similar
to itself will also determine the extent to which the
first metric will fail to capture the eect. Thus, a new
metric is proposed based on the change in loadings
themselves. The the two variants of the metric is
calculated as follows:
1.
Extract the general factor from the complete
dataset and save the factor loadings.
2. For every case:
1)
Create a subset of the dataset where this case
is excluded.
2) Extract the general factor from the subset.
3)
Extract the factor loadings from the analysis
and save them.
4)
Calculate the mean and max absolute factor
loadings change compared with the full anal-
ysis and save these values.
The metric scores cases by their influence on the factor
loadings of the dataset, both on average and in their
strongest impact.
3.2 Structural outliers in the Swiss dataset
As an example, the three metrics described above
were applied to the Swiss dataset analyzed previously.
The Swiss data contain some relatively weak outliers.
There are no strong outliers on the mean absolute
residuals (MAR) metric. On the change in factor size
(CFS), V. De Geneve is an outlier, with a value of -.045.
In other words, this case increases the factor size by
4.5 %points which is substantial for a dataset with
47 cases. This is the city district of Geneva, so it is
not surprising it is an outlier. With respect to both
mean and max absolute loading change (MeanALC
and MaxALC), Geneva is also an outlier with Sierre
and Neuchatel also having fairly large eects on the
overall loadings.
Histograms of the structural outlierness metrics are
shown in Figures 6a-6e.
Since two dierent methods confirmed that Geneva
was a likely outlier, a parallel dataset without this
case was created for further analysis.
3.3 Which metric is to be preferred?
Conceptually, the metrics are somewhat distinct, so
depending on the goal of the analysis, a particular
metric may be the best suited for the task. But if the
goal is to identify structural outliers in general, it’s
not clear which one is to be preferred. My current
practice is using all of them and then comparing their
results. Sometimes, one indicator may give divergent
results and so one has to pay extra attention to see
if one can figure out why. A general approach is to
factor analyze the indicators to get a single structural
outlierness score for each case. When doing this, it’s
important to choose only one of the variants of a given
metric. E.g. do not use both mean and maximum
absolute loading change, pick one of them.
Lastly, a nice feature of the mean absolute residuals
method is that it allows one to see which indicators
cause a given case to be a structural outlier. The other
methods do not allow for this possibility.
4 Robust factor analysis
This method was first used in Kirkegaard (2015d).
A limitation of the methods presented above is that
they merely identify the outlier cases, if any. They
don’t oer a way to include them in an analysis with-
out disrupting the results. One way to do this is to
employ some kind of method that is less aected by
outliers. The rank-order correlation (Spearman’s rho)
is such a method when it comes to correlations. So
since factor analysis is based on the correlations be-
tween variables, one option is to convert the dataset
into rank-transformed data and then factor analyze
it.
Often it will be a good idea to conduct both stan-
dard factor analysis and ranked factor analysis so that
one may compare the results, e.g. as in Kirkegaard
(2015d).
Figure 7shows the factor loadings for standard factor
analysis, standard factor analysis without Geneva and
factor analysis on the rank-transformed data.
We can look back at Table 1and see that the S scores
correlated between .81 and 1.00, suggesting high but
7
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
(a) MAR in the Swiss dataset. (b) CFS in the Swiss dataset. (c) ACFS in the Swiss dataset.
(d) MaxALC in the Swiss dataset. (e) MeanALC in the Swiss dataset.
Figure 6: Histograms of metrics of structural outlierliness.
Figure 7: Factor loadings in the Swiss data using three factor analysis methods.
8
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
not great stability across extraction methods. The
correlation with cognitive ability, however, was robust
(.64 to .70).
Figures 8to 11 show the scatterplots of the data.
Both Figures 8and 9show a upside-down U-shaped
relationship, which is very curious.
We see that Geneva appeared to be inflating the corre-
lation, but when switching to rank-transformed data,
the overall pattern actually become more linear.
5 Redundant indicators
This method was first used in Kirkegaard (2015a).
One problem with extracting general factors from
datasets is that if the selection of indicators is not rep-
resentative of all possible indicators, the general fac-
tor will be ’colored’ or ’contaminated’ with variance
from one of more group factors that are represented
in the data (Carroll 1993, p. 596; Jensen 1998, p. 85).
Jensen states that the ’coloring’ of the general factor is
one reason to prefer hierarchical factor analysis over
direct factor extraction methods (Jensen,1998). He
cites no references or data for this however, so it is not
clear whether this is a good recommendation. A large
simulation study would probably be able to settle the
issue.
Meanwhile, we might try to avoid the problem by
trying to ensure that our sample of indicators is rep-
resentative. Because group factors arise when groups
of indicators are more strongly correlated than one
would expect based on their general factor loadings,
one can use this to try to avoid them. One simple idea
is to avoid including variables that are very highly
correlated.
In datasets of socioeconomic measures, such very
highly correlated pairs often represent one construct
measured in both genders (e.g. mean income by gen-
der), or measured in a negative and positive fashion
(e.g. percent of the population employed and per-
cent of the population receiving unemployment bene-
fits). Sometimes, two variables may simply be reverse
coded copies of each other (e.g. percent of population
with at least high school and percent of population
without high school). In the case of gender-split vari-
ables, they may or may not be highly correlated. If
they are, then averaging them to obtain one measure
is reasonable. If they are not, however, it is instructive
to include both as they may have dierential relation-
ships to the general factor or criterion variables which
could be of importance.
A similar fact applies to variables that measure the
same or very similar constructs positively and nega-
tively. For instance, whether employment rate and
use of unemployment benefits are very strongly corre-
lated or not depends on whether persons in dierent
regions have the same tendency to seek benefits when
not employed. They may or may not and this might
reveal an interesting pattern that could otherwise be
missed.
Because the number of intercorrelations between all
indicators increases quickly as a function of the num-
ber of indicators, it is practical to have an algorithm
that automatically excludes variables. An algorithm
was developed and works as follows:
1.
Calculate the correlation between every pair of
indicators.
2.
Sort the indicator pairs by their absolute correla-
tion.
3.
If any correlation between a pair of indicators
exceeds the threshold, the second indicator in
the pair is removed, then go to step (1).
The algorithm finishes when no pair of indicators
have a correlation that reach the threshold for re-
moval. Previous studies have used a threshold of .90
which seems to work well.
6 Bootstrapping indicators, reliability
and factor analysis
Since we can’t include every conceivable indicator of
S, there will be sampling problems with the indicators
as well. Worse, a single included indicator may have
a large eect on the results. This is often seen in
analyses with a small number of S indicators, such
as in the analysis of Swiss data above (N
indicator
= 4),
where one indicator is (nearly) perfectly aligned with
the extracted S factor (Kirkegaard,2015i;Kirkegaard
& Tranberg,2015). Jensen and others have called
this psychometric sampling error (Dragt,2010;Jensen,
1993;Kranzler & Jensen,1991), but a more general
term would be indicator sampling error.
Bootstrapping is a statistical method that involves cre-
ating new random samples from the obtained sample,
fitting a model to it, saving the model fitting param-
eters and finally calculating descriptive statistics for
the distribution of model parameters. This is an alter-
native way to finding confidence intervals for param-
eter values that does not involve the usual parametric
statistical assumptions. Unfortunately, we cannot use
bootstrapping for factor analysis for indicators, since
bootstrapping would result in duplicated indicators
which result in coloring of the general factor.
However, we can split the sample of indicators ran-
domly into two subsets and then extract the factor in
9
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
Figure 8: Scatterplot of S and Catholic % in the Swiss dataset.
Figure 9: Scatterplot of S_rank and Catholic % in the Swiss dataset.
10
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
Figure 10: Scatterplot of S and cognitive ability in the Swiss dataset.
Figure 11: Scatterplot of S_rank and cognitive ability in the Swiss dataset.
11
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
each sample independently, and finally correlate the
factor scores. In fact, in my first published S factor
study (Kirkegaard,2014b), I used sample splitting
to show that the S factor scores were fairly stable as
long as one had a reasonable number of indicators.
For instance, the average absolute correlation of S fac-
tors extracted from a random selection of 5 indicators
with the S factor extracted from all 54/42 indicators
was about .90 (see Section 6.1 for an explanation).
Similarly, if one chooses two sets of 5 indicators at
random, the S factors extracted from them correlated
.76-.80 on average (using absolute values to nullify
the fact that the factors are sometimes reversed).
The method described is an extension of Cronbach’s
alpha, which is the mean correlation of summed
scores for all possible split-halves. Because S fac-
tors can have negative loadings, using summed scores
would not work well as the positively and negatively
loaded variables would cancel each other out instead
of aggregating.
The original R code to run these analyses, however,
was not written well and would not easily be re-
useable for another dataset. It was probably also for
this reason that I have neglected to examine indicator
sampling stability in the later studies. To rectify this,
I have written a function that repeatedly splits the
dataset into two random subsets of the indicators, ex-
tracts the first factor, correlates the scores and saves
the results.
6.1 S score reliability in the international
dataset
To illustrate the method, the data used in the
international S factor paper were reanalyzed
(Kirkegaard,2014b). The data consist of 96 so-
cioeconomic indicators of which 54 come from
the Social Progress Index 2014 dataset (
http://
www.socialprogressimperative.org/data/spi
)
and 42 from the Democracy Ranking 2013 dataset
(
http://democracyranking.org/
). Complete data
is available for only 70 cases. A closer look at the
missing data reveals that many cases are missing only
a few (3 or less) datapoints. These datapoints were
imputed using deterministic imputation using irmi()
from the
VIM
package (Templ, Alfons, Kowarik, &
Prantner, 2015). This resulted in 105 cases with
complete data.
A quick reliability test is to extract the national S fac-
tor again and check the correlation with the published
scores which were calculated using a somewhat dif-
ferent method.
7
Bartlett’s scoring method was used
7
The published scores was an average of two S factor analyses,
one carried out on each dataset. This improves the coverage of
countries somewhat, but decreases the number of indicators in
each analysis.
Figure 12:
Histogram of split-half factor reliability runs
for S from 96 indicators. N=500. Mean = .93, median .95.
because it has been found to work well in datasets
with low case n
cases
/n
indicators
ratios, even ratios <1
(Kirkegaard,2015a). The correlation between the new
and previously published S scores was .997, so there
was negligible method variance.
The split-half factor analysis reliability algorithm was
run 500 times and a histogram of the results is shown
in Figure 12.
All runs produced strong correlations showing that
indicator sampling error is an unlikely error source
for the S factor in this dataset. In other words, there
is an indierence of the indicators used to measure it
(Jensen,1998).
6.2 General factor of personality score relia-
bility in a 25 item dataset
The general factor of personality (GFP) is another re-
cent child of factor analytic methodology (DeYoung,
2006;Digman,1997;Musek,2007). It is similar to
the S factor in that involves the use of negative factor
loadings. Briefly speaking, the general factor of per-
sonality is a proposed measure of ’good personality’
or social eectiveness (van der Linden et al.,2016)
and can be extracted from personality data in a simi-
lar fashion as the general cognitive ability factor can
be extract from ability data.
A dataset with 2800 cases with data for 25
OCEAN/big five items is included in the psych pack-
age (Revelle,2015). The items are from the Inter-
national Personality Item Pool and comes from the
Synthetic Aperture Personality Assessment project
(Revelle et al.,2010).
The same method as before was used, also with 500
runs. The histogram of split-half factor correlations
is shown in Figure 13.
12
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
Figure 13:
Histogram of split-half factor reliability runs
for GFP from 25 OCEAN items. N=500.
Figure 14:
Histogram of split-half factor reliability runs
for GFP from 25 OCEAN items. N=500. Absolute values.
Mean = .55, median .56.
Here we see very dierent results. The results are
split between two distributions. Why? It is because
sometimes, the split of the indicators is done such
that one sample has a majority of indicators with a
negative loading. The factor extraction method then
reverses them so that the factor has mostly positive
loadings. This means that the factor scores will also
be reversed resulting in negative correlations between
the factors.
One solution to this problem would be to use abso-
lute values as was done in the original S factor study.
This however will bias the reliability upwards in some
cases. If there genuinely is no reliability (our measure-
ment is pure noise), then then the factor correlations
will be normally distributed around 0. If we then
take the absolute values, they will all be positive and
indicate that there is some reliability when there isn’t.
More generally, this will happen whenever the re-
liability distributions overlap with 0. When this is
not the case, there is no bias. In Figure 13, we see
that the two reliability distributions do not overlap 0.
Figure 14 shows the histogram using absolute values.
In other words, for GFP in this dataset, we see
Figure 15:
Histogram of split-half factor reliability runs
for the GCA factor from the 16 ICAR items. N=500. Mean
= .68, median .68.
medium reliability across indicators.
6.3 General cognitive ability factor reliabil-
ity in a dataset of 16 items
The 16 items are part of the International Cognitive
Ability Resource test, a public domain test being de-
veloped (Condon & Revelle,2014;Kirkegaard & Nord-
bjerg,2015). The dataset is part of the psych package
for R and contains complete data for 1248 cases (Rev-
elle,2015). The 16 items includes 4 verbal reasoning
items, 4 alphanumeric series items, 4 matrix items
and 4 3D-rotation items.
Since all loadings are positive, the split-half method
is not predicted to be better than existing methods
for examining internal reliability. In general, it has
been found that it does not matter whether scores are
derived using unit weights, factor loadings or even
random numbers (Ree et al.,1998).
Figure 15 shows the histogram of the intercorrela-
tions. This was based on classical factor analysis be-
cause it has been found that item response theory
(IRT) based and classical scores correlate near 1, mak-
ing it unnecessary to use the more computationally
costly IRT method (Kirkegaard,2015e;Kirkegaard &
Nordbjerg,2015).
By comparison, Cronbach’s alpha is .83. Presumably
the decrease is due to the use of factor loadings.
7 Discussion and conclusion
A number of the methods reviewed in this paper are
still in the experimental stage and lack large simu-
lation studies that back up their eectiveness. Still,
the methods have been used in a number of analy-
ses seemingly without major issues, which does give
some confidence in them.
13
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
Many unresolved methodological problems with re-
gards to exploratory factor analysis of socioeconomic
data remain. First, how to meta-analysis S factor stud-
ies. Each study contains a unique batch of indica-
tors that do not overlap perfectly or sometimes at
all between studies, and it is unclear how such het-
erogeneous data can be aggregated in a sensible way.
Second, what is the best way to generalize the current
structural outlierliness methods to multi-dimensional
data. Third, it is unknown whether using hierarchi-
cal or bi-factor analysis can help alleviate the prob-
lem with group factor contamination of unbalanced
datasets. Fourth, it is not clear when one should use
rank-ordered data to include outlier cases or when
one should use interval data and exclude them. Other
options include winsorizing the outlying datapoints
and including them in the standard interval data anal-
ysis.
Supplementary material and
acknowledgements
Functions to calculate and plot Jensen’s method can
be found in my personal R package, kirkegaard. It is
found on GitHub:
https://github.com/Deleetdk/
kirkegaard.
The R source code and data files can be found at
the Open Science Framework repository:
https://
osf.io/3npj8/files/.
Thanks to L. J. Zigerell, Davide Pier and Noah Carl
for reviewing the paper.
The peer review thread can be found at:
https://
openpsych.net/forum/showthread.php?tid=248.
References
Ashton, M. C., & Lee, K. (2005). Problems with
the method of correlated vectors. Intelligence,
33(4), 431–444. doi: https://doi.org/10.1016/
j.intell.2004.12.004
Carl, N. (2015). Iq and socioeconomic development
across regions of the uk. Journal of biosocial sci-
ence,48(3), 406–417. doi: https://doi.org/10.1017/
S002193201500019X
Carroll, J. B. (1993). Human cognitive abilities: A sur-
vey of factor-analytic studies. Cambridge University
Press.
Clark, G. (2007). A farewell to alms: a brief economic
history of the world. Princeton University Press.
Condon, D. M., & Revelle, W. (2014). The interna-
tional cognitive ability resource: Development and
initial validation of a public-domain measure. In-
telligence,43, 52–64. doi: https://doi.org/10.1016/
j.intell.2014.01.004
Cudeck, R., & MacCallum, R. C. (2007). Factor analy-
sis at 100: Historical developments and future direc-
tions. Mahwah, N.J: Lawrence Erlbaum Associates.
Cumming, G. (2012). Understanding the new statistics:
Eect sizes, confidence intervals, and meta-analysis.
Routledge.
Dalliard, M. (2013). Is psychometric g
a myth? Human Varieties. Retrieved
from
http://humanvarieties.org/2013/04/03/
is-psychometric-g-a-myth/
DeYoung, C. G. (2006). Higher-order factors of the
big five in a multi-informant sample. Journal of
personality and social psychology,91(6), 1138. doi:
https://doi.org/10.1037/0022-3514.91.6.1138
Digman, J. M. (1997). Higher-order factors of the
big five. Journal of personality and social psychology,
73(6), 1246.
Dragt, J. (2010). Causes of group dierences studied with
the method of correlated vectors: A psychometric meta-
analysis of spearman’s hypothesis (Master’s thesis,
University of Amsterdam). Retrieved from
http://
dare.uva.nl/cgi/arno/show.cgi?fid=176083
Everitt, B. S., & Dunn, G. (2001). Applied multivariate
data analysis (2nd ed.). Chichester: Wiley.
Frisby, C. L., & Beaujean, A. A. (2015). Testing Spear-
man’s hypotheses using a bi-factor model with
WAIS-IV/WMS-IV standardization data. Intelli-
gence,51, 79–97. doi: https://doi.org/10.1016/
j.intell.2015.04.007
Jensen, A. R. (1985). The nature of the Black–White
dierence on various psychometric tests: Spear-
man’s hypothesis. Behavioral and Brain Sciences,
8(02), 193–219. doi: https://doi.org/10.1017/
S0140525X00020392
Jensen, A. R. (1993). Psychometric g and achieve-
ment. In Policy perspectives on educational
testing (pp. 117–227). Springer. Retrieved from
http://link.springer.com/chapter/10.1007/
978-94-011-2226-9_4
Jensen, A. R. (1998). The g factor: The science of mental
ability. Praeger.
Jensen, A. R. (2006). Clocking the mind: Mental
chronometry and individual dierences. Elsevier.
Jensen, A. R., & Reynolds, C. R. (1982). Race, social
class and ability patterns on the WISC-R. Person-
ality and Individual Dierences,3(4), 423–438. doi:
https://doi.org/10.1016/0191-8869(82)90007-1
14
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
Jensen, A. R., & Weng, L.-J. (1994). What is a good g?
Intelligence,18(3), 231–258. doi: https://doi.org/
10.1016/0160-2896(94)90029-9
Kirkegaard, E. O. W. (2014a). Crime, income, edu-
cational attainment and employment among im-
migrant groups in Norway and Finland. Open
Dierential Psychology. Retrieved from
https://
openpsych.net/paper/29
Kirkegaard, E. O. W. (2014b). The international gen-
eral socioeconomic factor: Factor analyzing inter-
national rankings. Open Dierential Psychology. Re-
trieved from
https://openpsych.net/paper/28
Kirkegaard, E. O. W. (2015a). Examining the S
factor in Mexican states. The Winnower. Re-
trieved from
https://thewinnower.com/papers/
examining-the-s-factor-in-mexican-states
Kirkegaard, E. O. W. (2015b). Finding mixed cases
in exploratory factor analysis. The Winnower. Re-
trieved from
https://thewinnower.com/papers/
finding-mixed-cases-in-exploratory-factor
-analysis
Kirkegaard, E. O. W. (2015c). Indian states:
G and S factors. The Winnower. Retrieved
from
https://thewinnower.com/papers/indian
-states-g-and-s-factors
Kirkegaard, E. O. W. (2015d). IQ and socioe-
conomic development across regions of the
UK: a reanalysis. The Winnower. Retrieved
from
https://thewinnower.com/papers/
1419-iq-and-socioeconomic-development
-across-regions-of-the-uk-a-reanalysis
Kirkegaard, E. O. W. (2015e). Opinions about
nuclear energy and global warming, and
wordsum intelligence. The Winnower. Re-
trieved from
https://thewinnower.com/papers/
opinions-about-nuclear-energy-and-global
-warming-and-wordsum-intelligence
Kirkegaard, E. O. W. (2015f). S and G in
Italian regions: Re-analysis of Lynn’s data
and new data. The Winnower. Retrieved
from
https://thewinnower.com/papers/
s-and-g-in-italian-regions-re-analysis
-of-lynn-s-data-and-new-data
Kirkegaard, E. O. W. (2015g). An S factor among
census tracts of boston. The Winnower. Re-
trieved from
https://thewinnower.com/papers/
an-s-factor-among-census-tracts-of-boston
Kirkegaard, E. O. W. (2015h). The S fac-
tor in Brazilian states. The Winnower. Re-
trieved from
https://thewinnower.com/papers/
the-s-factor-in-brazilian-states
Kirkegaard, E. O. W. (2015i). The S fac-
tor in the British Isles: A reanalysis of
Lynn (1979). The Winnower. Retrieved
from
https://thewinnower.com/papers/
the-s-factor-in-the-british-isles-a
-reanalysis-of-lynn-1979
Kirkegaard, E. O. W., & Nordbjerg, O. (2015). Val-
idating a Danish translation of the International
Cognitive Ability Resource sample test and Cog-
nitive Reflection Test in a student sample. Open
Dierential Psychology. Retrieved from
https://
openpsych.net/paper/37
Kirkegaard, E. O. W., & Tranberg, B. (2015).
What is a good name? the S factor in Den-
mark at the name-level. The Winnower.
Retrieved from
https://thewinnower.com/
papers/what-is-a-good-name-the-s-factor
-in-denmark-at-the-name-level
Kranzler, J. H., & Jensen, A. R. (1991). Unitary g:
Unquestioned postulate or empirical fact? In-
telligence,15(4), 437–448. doi: https://doi.org/
10.1016/0160-2896(91)90005-X
La Grie du Lion. (2001, July). Pearbotham’s
law on the persistence of achievement gaps. Re-
trieved from
http://www.lagriffedulion.f2s
.com/adverse.htm
La Grie du Lion. (2007, January). Intelligence,
gender and race. Retrieved from
http://www
.lagriffedulion.f2s.com/g.htm
Lynn, R. (1996). Dysgenics: Genetic deterioration in
modern populations. Praeger.
McDaniel, M. A., & Kepes, S. (2014). An evalua-
tion of Spearman’s hypothesis by manipulating g
saturation. International Journal of Selection and
Assessment,22(4), 333–342. doi: https://doi.org/
10.1111/ijsa.12081
Musek, J. (2007). A general factor of personality:
Evidence for the big one in the five-factor model.
Journal of Research in Personality,41(6), 1213–1233.
doi: https://doi.org/10.1016/j.jrp.2007.02.003
Ree, M. J., Carretta, T. R., & Earles, J. A. (1998). In
top-down decisions, weighting variables does not
matter: A consequence of Wilks’ theorem. Orga-
nizational Research Methods,1(4), 407–420. doi:
https://doi.org/10.1177/109442819814003
Ree, M. J., Carretta, T. R., & Green, M. T. (2003). The
ubiquitous role of g in training. The scientific study
of general intelligence: Tribute to Arthur R. Jensen,
261–274.
Revelle, W. (2015). psych: Procedures for psycholog-
ical, psychometric, and personality research (version
15
Published: 7th of November 2016 Open Quantitative Sociology & Political Science
1.5.4). Retrieved from
http://cran.r-project
.org/web/packages/psych/index.html
Revelle, W., Wilt, J., & Rosenthal, A. (2010). Indi-
vidual dierences in cognition: New methods for
examining the personality-cognition link. In Hand-
book of individual dierences in cognition (pp. 27–49).
Springer. Retrieved from
http://link.springer
.com/10.1007/978-1-4419-1210-7_2
Rushton, J. P., & Ankney, C. D. (2009). Whole brain
size and general mental ability: a review. Interna-
tional Journal of Neuroscience,119(5), 692–732. doi:
https://doi.org/10.1080/00207450802325843
Spearman, C. (1904). "general intelligence," objec-
tively determined and measured. American Journal
of Psychology,15(2), 201–292. doi: https://doi.org/
10.2307/1412107
te Nijenhuis, J., Jongeneel-Grimen, B., & Kirkegaard,
E. O. W. (2014). Are headstart gains on the g
factor? a meta-analysis. Intelligence,46, 209–215.
doi: https://doi.org/10.1016/j.intell.2014.07.001
te Nijenhuis, J., van den Hoek, M., & Armstrong, E. L.
(2015). Spearman’s hypothesis and Amerindians: A
meta-analysis. Intelligence,50, 87–92. doi: https://
doi.org/10.1016/j.intell.2015.02.006
te Nijenhuis, J., & van der Flier, H. (2013). Is the
Flynn eect on g?: A meta-analysis. Intelligence,
41(6), 802–807. doi: https://doi.org/10.1016/
j.intell.2013.03.001
te Nijenhuis, J., van Vianen, A. E., & van der Flier,
H. (2007). Score gains on g-loaded tests: No g.
Intelligence,35(3), 283–300. doi: https://doi.org/
10.1016/j.intell.2006.07.006
van der Linden, D., Dunkel, C. S., & Petrides, K.
(2016). The General Factor of Personality (GFP)
as social eectiveness: Review of the literature. Per-
sonality and Individual Dierences,101, 98–105. doi:
https://doi.org/10.1016/j.paid.2016.05.020
16
... Usually, factor analytic methods give near identical results, but in some rare cases they diverge substantially. To check whether this was the case, we factor analyzed the data using all possible combinations of method parameters (30 combinations) (Kirkegaard, 2016b). The median factor score correlation across these was .999 ...
... A general factor extracted from a dataset will be 'flavored' if the indicators in the sample are not representative of all possible indicators (Carroll, 1993;Jensen, 1998). One can measure the extent to which the indicator sampling impacts the scores by splitting the indicators into two subsets at random, extracting the factors in each subset and correlating the scores (Kirkegaard, 2016b). This was done 100 times. ...
... There was near-zero method variance because the scores from these 30 methods had a mean correlation of .99. Factor analytic indicator sampling error was examined using the split-half method, see Kirkegaard (2016b). The analysis indicated only a small amount of indicator sampling error (median absolute correlation between scores = .85). ...
Article
Full-text available
Analyses of the relationships between cognitive ability, socioeconomic outcomes, and European ancestry were carried out at multiple levels in Argentina: individual (max. n = 5,920), district (n = 437), municipal (n = 299), and provincial (n = 24). Socioeconomic outcomes correlated in expected ways such that there was a general socioeconomic factor (S factor). The structure of this factor replicated across four levels of analysis, with a mean congruence coefficient of .96. Cognitive ability and S were moderately to strongly correlated at the four levels of analyses: individual r=.55 (.44 before disattenuation), district r=.52, municipal r=.66, and provincial r=.88. European biogeographic ancestry (BGA) for the provinces was estimated from 25 genomics papers. These estimates were validated against European ancestry estimated from self-identified race/ethnicity (SIRE; r=.67) and interviewer-rated skin brightness (r=.33). On the provincial level, European BGA correlated strongly with scholastic achievement-based cognitive ability and composite S-factor scores (r's .48 and .54, respectively). These relationships were not due to confounding with latitude or mean temperature when analyzed in multivariate analyses. There were no BGA data for the other levels, so we relied on %White, skin brightness, and SIRE-based ancestry estimates instead, all of which were related to cognitive ability and S at all levels of analysis. At the individual level, skin brightness was related to both cognitive ability and S. Regression analyses showed that SIRE had little detectable predictive validity when skin brightness was included in models. Similarly, the correlations between skin brightness, cognitive ability, and S were also found inside SIRE groups. The results were similar when analyzed within provinces. In general, results were congruent with a familial model of individual and regional outcome differences.
... As a preliminary analysis, I carried out factor analyses at both the broad and detailed type-levels using both standard form and rank form data and correlated the extracted factor scores with the crime rate given found in the London Borough Profiles dataset, also in standard form and in rank form. The use of case ranks is to reduce effects of outliers (7). ...
... Before the factor analysis was run, I excluded variables that correlated at |r| > .9 with another variable to reduce the 'coloring' of the general factor by group factor variance (7). This resulted in the removal of two variables (Theft_Person, Wounding/GBH) in the case of the analysis using detailed types. ...
... To avoid 'coloring' the extracted general factor, a redundancy analysis was run (7). This resulted in the same two variables being excluded as in the preliminary analysis. ...
Article
Full-text available
A dataset of 30 diverse socioeconomic variables was collected covering 32 London boroughs. Factor analysis of the data revealed a general socioeconomic factor. This factor was strongly related to GCSE (General Certificate of Secondary Education) scores (r's .683 to .786) and and had weak to medium sized negative relationships to demographic variables related to immigrants (r's -.295 to -.558). Jensen's method indicated that these relationships were related to the underlying general factor, especially for GCSE (coefficients |.48| to |.69|). In multiple regression, about 60% of the variance in S outcomes could be accounted for using GCSE and one variable related to immigrants.
... Apart from these, some of the previous studies [7][8][9] suggest about the use for factor analysis for identification and classification of different type of indicators such as socioeconomic, and developmental indicators. Also, the factors are obtained at state and district level in these studies. ...
Article
Full-text available
It is widely known that fertility is determined by a complex set of biological, psychological, socioeconomic , cultural, political and legal factors. These factors affect the fertility dynamics directly or indirectly. Although the past studies have suggested about a lot of variables for the estimation of total fertility rate, and various fertility models have been developed for estimating the fertility parameters, yet it is difficult to include all the determining variables in a single estimation technique conveniently. In the present study, authors have used a multivariate technique i.e. factor analysis to examine the underlying patterns and relationships for a number of variables which affects fertility behaviour of a population. Also, authors have tried to summarize the information in smaller sets of distinct factors to estimate the children ever born. Present study is based on the data obtained from the fourth round of National Family Health Survey, India (NFHS-IV) dataset. The key finding of the present study is that the selected fertility parameters are classified into three distinct factors with a valid justification. On the basis of these estimated factors, the children ever born are estimated for 443 districts of India which is significantly good estimate of children ever born.
... General socioeconomic factor (S) -0.94 Figure 9. Jensen's method used on Muslim% and general socioeconomic status (S factor). Note that indicators with negative loadings were reversed to avoid variance inflation (Kirkegaard, 2016b), and are marked with "_r". This Jensen pattern 1 suggests that whatever it is Muslim% relates to, it is something that relates chiefly to the joint variation of a broad variety of indicators of well-being. ...
Article
Full-text available
We examined regional inequality in Belgium, both in the 19 communes of Brussels and in the country as a whole (n = 589 communes). We find very strong relationships between Muslim% of the population and a variety of social outcomes such as crime rate, educational attainment, and median income. For the 19 communes of Brussels, we find a correlation of-.94 between Muslim% and a general factor of socioeconomic variables (S factor) based on 22 diverse indicators. The slope for this relationship is-7.52, meaning that a change in S going from 0% to 100% Muslim corresponds to a worsening of overall social well-being by 7.52 (commune-level) standard deviations. For the entire country, we have data for 8 measures of social inequality. Analysis of the indicators shows an S factor which is very similar to the one from the Brussels data only based on the full set of indicators (r's = .98). In the full dataset, the correlation between S and Muslim% is-.52, with a slope of-8.05. Adding covariates for age, population density, and spatial autocorrelation changes this slope to-8.77. Thus, the expected change going from a 0% to 100% Muslim population is-8.77 standard deviations in general social well-being. We discuss our findings in relation to other research on immigration and social inequality, with a focus on the causal influence of intelligence on life outcomes in general.
... 2 Second, based on Pesta et al. (2010) we implemented a two-stage / higher-order grouping of variables by type into the categories proposed by Lynn et al. (2018;economics, education, health, crime, fertility) except that we did not include crime. We then aggregated the categories into a single, global index of S. The advantage of this procedure is that it forces equal importance to each of the different subscales in the analysis (see also Kirkegaard, 2016c). That is, the aim is to ensure that data sets with an imbalance in the S indicators do not produce S scores that unduly weight one aspect (e.g., crime versus health) of social inequality. ...
Article
Full-text available
We compiled cognitive, ethnic, and socioeconomic data for the 63 provinces of Vietnam. The cognitive data came from math and reading achievement tests administered to 70,000 fifth-graders in 2001 (World Bank, 2004). Ethnic and socioeconomic data were coded from various official sources (e.g., The General Statistics Office of Vietnam). Analysis of the socioeconomic data revealed a general factor (S) that was robust to variations in extraction method and controls. The average cognitive ability of the provinces correlated .47 with the S factor. The strongest predictor of S, however, was ethnicity. Specifically, the percent of Vietnamese (Kinh) within each province correlated .74 with S. Moreover, this effect was not mediated by cognitive ability. The lack of mediation is inconsistent with results from earlier studies that examined relations between ethnicity, cognitive ability, and socioeconomic outcomes (see, e.g., Fuerst & Kirkegaard, 2016). Also inconsistent with prior studies, although latitude correlated positively with cognitive ability, it did so inversely with the S factor. We discuss several potential hypotheses for why these discrepant effects occurred.
... A positive association, however, was not found for Germany (Roivainen, 2012), India (Lynn & Yadav, 2015) or the UK (Carl, 2016a). Associations were not reported for China ; Denmark (Teasdale et al., 1988); Finland (Dutton & Lynn, 2014); France (Lynn, 1980); Portugal (Almeida et al., 2011); Switzerland (Kirkegaard, 2016c); or Vietnam (Holsinger, 2007). Since several models predict a fairly consistent association between latitude and cognitive or socioeconomic outcomes (e.g., Lynn's evolutionary cold winters hypothesis, and León's contemporaneous UV hypothesis) it would be useful if future researchers included geoclimatic variables, so the consistency of these associations can be assessed. ...
Article
Differences in intelligence have previously been found to be related to a wide range of inter-individual and international social outcomes. There is evidence indicating that intelligence differences are also related to different regional outcomes within nations. A quantitative and narrative review is provided for twenty-two countries (number of regions in parentheses): Argentina (24 to 437), Brazil (27 to 31), British Isles (12 to 392), to 79), Spain (15 to 48), Switzerland (47), Turkey (12), the USA (30 to 3100), and Vietnam (61). Between regions, intelligence is significantly associated with a wide range of economic, social, and demographic phenomena, including income (r unweighted = .56), educational attainment (r unweighted = .59), health (r unweighted = .49), general socioeconomic status (r unweighted = .55), and negatively with fertility (r unweighted = −.51) and crime (r unweighted = −.20). Proposed causal models for these differences are noted. It is concluded that regional differences in intelligence within nations warrant further focus; methodological concerns that need to be addressed in future research are detailed.
Article
Full-text available
A dataset of socioeconomic, demographic and geographic data for US counties (N≈3,100) was created by merging data from several sources. A suitable subset of 28 socioeconomic indicators was chosen for analysis. Factor analysis revealed a clear general socioeconomic factor (S factor) which was stable across extraction methods and different samples of indicators (absolute split-half sampling reliability = .85). Self-identified race/ethnicity (SIRE) population percentages were strongly, but non-linearly, related to cognitive ability and S. In general, the effect of White% and Asian% were positive, while those for Black%, Hispanic% and Amerindian% were negative. The effect was unclear for Other/mixed%. The best model consisted of White%, Black%, Asian% and Amerindian% and explained 41/43% of the variance in cognitive ability/S among counties. SIRE homogeneity had a non-linear relationship to S, both with and without taking into account the effects of SIRE variables. Overall, the effect was slightly negative due to low S, high White% areas. Geospatial (latitude, longitude, and elevation) and climatological (temperature, precipitation) predictors were tested in models. In linear regression, they had little incremental validity. However, there was evidence of non-linear relationships. When models were fitted that allowed for non-linear effects of the environmental predictors, they were able to add a moderate amount of incremental validity. LASSO regression, however, suggested that much of this predictive validity was due to overfitting. Furthermore, it was difficult to make causal sense of the results. Spatial patterns in the data were examined using multiple methods, all of which indicated strong spatial autocorrelation for cognitive ability, S and SIRE (k nearest spatial neighbor regression [KNSNR] correlations of .62 to .89). Model residuals were also spatially autocorrelated, and for this reason the models were re-fit controlling for spatial autocorrelation using KNSNR-based residuals and spatial local regression. The results indicated that the effects of SIREs were not due to spatially autocorrelated confounds except possibly for Black% which was about 50% weaker in the controlled analyses. Pseudo-multilevel analyses of both the factor structure of S and the SIRE predictive model showed results consistent with the main analyses. Specifically, the factor structure was similar across levels of analysis (states and counties) and within states. Furthermore, the SIRE predictors had similar betas when examined within each state compared to when analyzed across all states. It was tested whether the relationship between SIREs and S was mediated by cognitive ability. Several methods were used to examine this question and the results were mixed, but generally in line with a partial mediation model. Jensen's method (method of correlated vectors) was used to examine whether the observed relationship between cognitive ability and S scores was plausibly due to the latent S factor. This was strongly supported (r = .91, Nindicators=28). Similarly, it was examined whether the relationship between SIREs and S scores was plausibly due to the latent S factor. This did not appear to be the case.
Article
Full-text available
Cultural practices resulting in lasting modifications of the body can be a productive focus of investigation. Artificial cranial modification involves the alteration of cranial vault shape by cultural means, and is performed during infancy or early childhood while the cranial bones remain malleable. Our results indicate that tumpline deformation and cradle deformation should be added to the suite of cranial deformations that were detected in Armenia. It has been determined that the occipital modification is most likely unintentional. Post-coronal depression also appears to be unintentional. Cradle deformation is heavily influenced by infant sleep position, and constant supine positioning is a frequent cause of deformation during infancy. In the Armenian burials 32 skulls showing signs of cradle deformation are identified. Post-coronal depression, positioned on both parietals and slightly posterior to the coronal suture, was detected on 44 of the skulls. The deformation was a correlate of economic activities of the population and was an inadvertent consequence of carrying loads with a band across the parietal bones. In 12 cases a combination of the two types of deformation (tumpline and cradle), is observed. The only modification that has been classified as intentional is one case of parieto-temporal modification.
Article
Full-text available
Previously, we looked at the association between overall state-level biogeographic ancestry (BGA) and overall state-level outcomes. It was found that European BGA relative to African and Amerindian BGA was associated with better outcomes. In this paper, the analysis is extended by looking at the state-level ancestry-outcome associations individually for black and Hispanic self-identified race-ethnicity (SIRE) groups. General socioeconomic factor (S) scores were calculated for US states by SIRE groups based on three indicators. The S factor loadings were generally stable across subgroup analyses and the factor scores were stable across factor analytic extraction methods (for the latter, almost all r's ≈ 1). For Whites, Blacks and Hispanics, there were strong correlations between cognitive ability scores and S factor scores across states (r = .55 to .78; N = 28-50). This pattern also held when all data were analyzed together (r = .86, N = 115). Furthermore, the size of the Hispanic-White and Black-White S and cognitive ability gaps strongly correlated across states (r = .62 to .69; N = 36-37). Lastly, parasite prevalence did not plausibly explain SIRE gaps in cognitive ability because gaps were smaller in more parasite-rich states (combined analysis r = -.17, N = 91). We found that climatic and geospatial variables did not correlate strongly with cognitive ability and S scores when scores were decomposed by SIRE group, but did so at the total state level, even after statistically controlling for SIRE composition. Key words: Inequality, General socioeconomic factor, S factor, USA, States, Cognitive ability, Intelligence, NAEP, Race, SIRE, Biogeographic ancestry, Ecology
Article
Full-text available
Two datasets of Japanese socioeconomic data for Japanese prefectures (N=47) were obtained and merged. After quality control, there were 44 variables for use in a factor analysis. Indicator sampling reliability analysis revealed poor reliability (54% of the correlations were |r| > .50). Inspection of the factor loadings revealed no clear S factor with many indicators loading in opposite than expected directions. A cognitive ability measure was constructed from three scholastic ability measures (all loadings > .90). On first analysis, cognitive ability was not strongly related to 'S' factor scores, r = -.19 [CI95: -.45 to .19; N=47]. Jensen's method did not support the interpretation that the relationship is between latent 'S' and cognitive ability (r = -.15; N=44). Cognitive ability was nevertheless related to some socioeconomic indicators in expected ways. A reviewer suggested controlling for population size or population density. When this was done, a relatively clear S factor emerged. Using the best control method (log population density), indicator sampling reliability was high (93% |r|>.50). The scores were strongly related to cognitive ability r = .67 [CI95: .48 to .80]. Jensen's method supported the interpretation that cognitive ability was related to the S factor (r = .78) and not just to the non-general factor variance.
Article
Full-text available
We translated the International Cognitive Ability Resource sample test (ICAR16) and the Cognitive Reflection Test (CRT) into Danish. We administered the test online to a student sample (N=72, mean age 17.4). Factor analysis revealed a general factor. The summed score of all test items correlated .42 with GPA. Item difficulties correlated .85 with those reported in the Internet norming sample. Method of correlated vectors analysis showed positive relationships between g-loading of items/subtests and their correlation with GPA (r=.53/.85). Model comparisons revealed that for predicting GPA the CRT did not have incremental validity over the ICAR16, but the evidence was not strong.
Article
Full-text available
A reanalysis of (Carl, 2015) revealed that the inclusion of London had a strong effect on the S loading of crime and poverty variables. S factor scores from a dataset without London and redundant variables was strongly related to IQ scores, r = .87. The Jensen coefficient for this relationship was .86.
Article
Full-text available
Two methods are presented that allow for identification of mixed cases in the extraction of general factors. Simulated data is used to illustrate them.
Article
Full-text available
A factor analysis was carried out on 6 socioeconomic variables for 506 census tracts of Boston. An S factor was found with positive loadings for median value of owner-occupied homes and average number of rooms in these; negative loadings for crime rate, pupil-teacher ratio, NOx pollution, and the proportion of the population of ‘lower status’. The S factor scores were negatively correlated with the estimated proportion of African Americans in the tracts r = -.36 [CI95 -0.43; -0.28]. This estimate was biased downwards due to data error that could not be corrected for.
Article
The General Factor of Personality (GFP) is a higher-order factor causing lower-order personality traits to show consistent correlations in a socially desirable direction. The literature on the GFP reveals that there are various scientific interpretations of this construct. One interpretation is that it is a substantive factor reflecting general social effectiveness and exerting a broad influence on behavior. Another interpretation is that it merely reflects methodological or statistical artifacts and has no further relevance for personality research. We review the empirical literature on the nature of the GFP, its possible links to evolutionary processes, and its relation to other constructs overlapping with social effectiveness. We conclude that the substantive interpretation of the GFP is the most plausible, whereas the notion that it is a psychologically meaningless methodological artifact would be rather difficult to uphold.
Chapter
Education’s traditional value of enhancing the quality of life for the individual need not be eclipsed by the growing recognition of its importance to the national welfare. A well-educated population is now deemed crucial in this technological era. The cultivation of excellence in the kinds of achievement that depend on an educated work force is an undisputed goal in all industrial societies. Regardless of differences in coutries’ political and economic systems, we see implicit agreement with Adam Smith’s dictum that the wealth of nations depends on the abilities of their people. Virtually every head of state appoints a minister of education. The government of Venezuela, in addition, even appointed a “Minister for the Development of Intelligence.” Obviously, the modern world perceives the supply of educated intelligence as vitally related to the general welfare.
Article
Cross-regional correlations between average IQ and socioeconomic development have been documented in many different countries. This paper presents new IQ estimates for the twelve regions of the UK. These are weakly correlated ( r =0.24) with the regional IQs assembled by Lynn (1979). Assuming the two sets of estimates are accurate and comparable, this finding suggests that the relative IQs of different UK regions have changed since the 1950s, most likely due to differentials in the magnitude of the Flynn effect, the selectivity of external migration, the selectivity of internal migration or the strength of the relationship between IQ and fertility. The paper provides evidence for the validity of the regional IQs by showing that IQ estimates for UK nations (England, Scotland, Wales and Northern Ireland) derived from the same data are strongly correlated with national PISA scores ( r =0.99). It finds that regional IQ is positively related to income, longevity and technological accomplishment; and is negatively related to poverty, deprivation and unemployment. A general factor of socioeconomic development is correlated with regional IQ at r =0.72.
Article
Spearman's hypothesis (SH) is a phrase coined by Arthur Jensen, which posits that the size of Black–White mean differences across a group of diverse mental tests is a positive function of each test's loading onto the general intelligence (g) factor. Initially, a correlated vector (CV) approach was used to examine SH, where the results typically confirmed that the magnitude of g loadings were positively correlated with the size of mean group differences in the observed test scores. The CV approach has been heavily criticized by scholars who have argued that a more precise method for examining SH can be better investigated using a multi-group confirmatory factor analysis (MG-CFA). Studies of SH using MG-CFA have been much more equivocal, with results not clearly confirming nor disconfirming SH.
Article
Historical Understanding of the Problem Natural Selection in Preindustrial Societies The Breakdown of Natural Selection The Genetic Deterioration of Health Intelligence and Fertility Sibling Studies Intelligence and Fertility in the United States Intelligence and Fertility in Europe Resolving the Paradox of the Secular Rise of Intelligence Education and Fertility Socioeconomic Status and Fertility Socioeconomic Status Differences in Intelligence Socioeconomic Status Differences in Conscientiousness The Genetic Basis of Socioeconomic Status Differences in Conscientiousness Dysgenic Fertility for Conscientiousness Dysgenic Fertility in Economically Developing National Counterarguments and Rejoinders.