Why Do Cognitive Abilities Correlate?

One of the most important discoveries in intelligence research is
also one of its simplest. Give many people a varied set of cognitive tests and the scores tend
to move together. People who do relatively well on one cognitive task
are, on average, somewhat more likely to do well on others. Verbal
ability correlates with reasoning. Spatial performance correlates with
memory. Speed correlates with broader cognitive performance. The correlations are not perfect. But they are predominantly
positive. This positive manifold has survived more than a
century of measurement, new tests and new statistical methods. What has not survived is any easy agreement about why it exists.

The
discovery of g solved one problem and created another

Factor analysis offers a powerful way to summarize the common
pattern. A general latent dimension can be estimated from covariance
among diverse cognitive measures. That dimension is conventionally
called g. But a statistical summary is not automatically a causal
explanation. This distinction is easy to lose because g is so useful. It
predicts important outcomes and captures a robust feature of
cognitive-test data. Yet extracting a general factor does not tell us
what biological or psychological process makes the tests correlate. The positive manifold is the observation. g is one way of
representing its common structure. The cause of that structure is a
further question. Much of intelligence theory begins there.

Perhaps diverse
tasks share something general

The traditional intuition is that many cognitive tasks depend partly
on a broadly general capacity. If reasoning, comprehension, memory and spatial problem solving all
draw on some common cognitive resource or highly general set of
mechanisms, people who differ in that common machinery should differ
across many tasks at once. Modern versions need not imagine a literal reservoir of mental
energy. Executive attention, processing efficiency and broad neural
properties have all been proposed as candidates. The challenge is to identify a mechanism that predicts the covariance
rather than simply renaming it.

Or perhaps
g emerges because processes overlap

Process Overlap Theory takes a different route. Complex tasks recruit
combinations of domain-general and domain-specific processes, with
executive processes particularly likely to recur across many tasks. Tests then correlate because their process requirements overlap. On this account, a general factor can be psychometrically real
without corresponding one-to-one with a single general psychological
capacity. That is a subtle but important possibility: the statistical structure
can be robust even if its causal architecture is distributed.

Development may help
create the pattern

Another family of theories asks whether the correlations themselves
develop. Mutualism proposes that cognitive abilities can reinforce one another
over time. Better reasoning can support learning; greater knowledge can
make later reasoning easier; memory can facilitate acquisition; acquired
skills can create new opportunities for cognitive growth. Abilities that began less strongly correlated could therefore become
mutually coupled through development. Network approaches extend this idea by modeling intelligence as an
emergent property of interacting cognitive components rather than the
effect of one common cause. The direction of explanation has changed. Instead of asking what
single thing sits behind all the abilities, these theories ask how the
abilities might help create the general pattern together.

An old
alternative sounds surprisingly modern

Godfrey Thomson was making a related conceptual point more than a
century ago. His sampling theory proposed that tests draw on many elementary
processes or “bonds.” Different tests correlate because they sample
overlapping sets of them. Under some conditions, a hierarchy resembling
a general factor can emerge without one single general process causing
every task. The details of Thomson’s model belong to an earlier psychometric era.
The warning does not: a general factor does not uniquely identify a general
cause.

The answer may not be
singular

These theories are often presented as rivals, and in important
respects they are. But the positive manifold could have more than one
source. Tasks can overlap in their process demands. Cognitive systems can
influence one another during development. Broad biological properties
can affect multiple systems. Test design can alter the correlations we
observe. The scientific task is to determine how much explanatory work each
mechanism can do. That leaves intelligence research in an unusual position. The
phenomenon itself is extraordinarily secure; its deepest explanation is
not. We know that diverse cognitive abilities tend to rise and fall
together across people. We can model that shared variation very
effectively. More than a century after Spearman, why they correlate
remains one of the questions that makes intelligence a science rather
than merely a scoring system.