Reference: Lykken, D. T. (1971). Multiple factor analysis and personality research. Journal of Experimental Research in Personality, 5, 161–170.

In 1971, David T. Lykken published a remarkable article titled Multiple Factor Analysis and Personality Research in the Journal of Experimental Research in Personality (now Journal of Research in Personality). More than half a century later, the article remains highly relevant to one of the fundamental problems in personality psychology:

What can we learn about the structure of personality from analyzing correlations among psychological measures with multiple factor analysis, now commonly called exploratory factor analysis (EFA)?

Factor analysis has played an enormous role in personality psychology. The Big Five, HEXACO, and various higher-order personality models are all closely connected to exploratory analyses of correlations among personality items, adjectives, or scales. Most instruments used to measure broad personality traits have been influenced by EFA results. The widespread use of these measures encourages the assumption that they measure specific traits. That is, scale scores are treated as measures of distinct underlying personality dispositions.

The basic procedure is familiar. Researchers collect a large number of personality measures, calculate their correlations, extract a smaller number of factors, rotate the solution, and assign names to the resulting dimensions. These factors often appear psychologically meaningful. They may be recognizable across datasets and may even predict important outcomes. But does this establish that the factors correspond to fundamental psychological dispositions?

Lykken thought not.

His article was written at the invitation of the journal’s editors to stimulate debate about the scientific role of factor analysis. He acknowledged Paul Meehl and Auke Tellegen for helpful comments, although he emphasized that they did not necessarily share his conclusions.

Indeed, Tellegen subsequently became an influential advocate of using EFA to identify the structure of affective experiences. Watson and Tellegen (1985) proposed that Positive Affect (PA) and Negative Affect (NA) are two largely independent dimensions of affect. Their model was based on the consistent appearance of two factors in orthogonally rotated exploratory factor analyses. This model subsequently guided the development of the widely used Positive and Negative Affect Schedule (PANAS; Watson et al., 1988).

This provides a particularly revealing example of Lykken’s concern. The independence of the two rotated factors is imposed by the orthogonal rotation procedure. It is not an empirical discovery that positive and negative emotional experiences are independent. Moreover, the PANAS scales were deliberately constructed using items with minimal secondary loadings, helping produce relatively independent scale scores.

The resulting scales may be useful measures of positive and negative activation, but their relatively low correlation does not establish that positive and negative emotional experiences are independent. The controversy illustrates how a statistical representation of a covariance matrix can be transformed into a substantive claim about psychological structure.

Thus, even within Lykken’s immediate intellectual circle, the fundamental issue he raised in 1971 remained unresolved: Does exploratory factor analysis reveal the underlying structure of psychological phenomena, or does it merely provide a convenient statistical representation of their correlations?

Lykken was not opposed to quantitative psychology or to the study of personality differences. Quite the contrary. He was an influential University of Minnesota psychologist whose research included temperament, psychophysiology, and behavioral genetics. He even coauthored a study with Tellegen showing that individual differences in subjective well-being are substantially influenced by genetic factors (Lykken & Tellegen, 1996).

Thus, Lykken’s criticism of EFA did not come from an outsider who questioned the existence or importance of personality traits. It came from a personality psychologist who regarded stable individual differences as important phenomena requiring scientific explanation but questioned whether EFA was an appropriate tool for discovering their underlying structure.

Three Uses of Factor Analysis

Lykken began by distinguishing three applications of factor analysis.

The first is data reduction. When researchers have hundreds of correlated measures, a smaller number of statistical dimensions may summarize much of the information. This can be useful, especially when the goal is to simplify a dataset without losing too much information.

The second is hypothesis testing. If a theory predicts that several measures reflect a common underlying disposition, researchers can examine whether their correlations are consistent with that prediction.

The third application is dimensional discovery: using exploratory factor analysis to discover the fundamental psychological variables responsible for individual differences.

Lykken acknowledged that factor analysis is useful for the first two aims, but he was deeply skeptical about its ability to discover dimensions that correspond to actual psychological dispositions responsible for differences in behavior and life outcomes.

In other words, EFA can help describe individual differences, but identifying statistical dimensions is not the same as discovering the psychological mechanisms that explain those differences. A factor that summarizes correlations among several measures does not necessarily represent a common cause of those correlations.

Lykken began by distinguishing three applications of factor analysis.

The first is data reduction. When researchers have hundreds of correlated measures, a smaller number of statistical dimensions may summarize much of the information. This can be useful, especially when the goal is to simplify a dataset without losing too much information.

The second is hypothesis testing. If a theory predicts that several measures reflect a common underlying disposition, researchers can examine whether their correlations are consistent with that prediction.

The third application is dimensional discovery: using exploratory factor analysis to discover the fundamental psychological variables responsible for individual differences.

Lykken acknowledged that EFA is useful for the first two aims, but he was deeply skeptical about its ability to discover dimensions that map on actual personality dispositions that explain differences in behavior and life-outcomes. In other words, EFA can help with the description of personality differences, but it cannot be used to explain behavior.

The distinction is easy to overlook because the statistical computations are similar. However, the scientific claims are radically different. A statistical factor may efficiently summarize several correlated measures without representing a distinct psychological mechanism. Conversely, a psychological mechanism may be causally important without producing a strong factor in the available measurements.

Lykken’s central criticism concerned the unwarranted movement from the first claim to the second.

The Problem With Assuming That Nature Has a Simple Factor Structure

Traditional factor analysis is often motivated by the assumption that correlations among observed variables result from a relatively small number of underlying common factors.

A classic representation is a hierarchy in which narrow characteristics reflect broader dispositions, which may themselves reflect still more general dispositions.

This arrangement has an appealing simplicity. It suggests that a complex psychological system can be understood by identifying a relatively small number of fundamental dimensions.

Lykken questioned whether biological and psychological systems actually have this organization.

He compared personality research with the study of biological organs and complex machines.

An engineer trying to understand an unfamiliar computer would not simply measure its outputs, extract factors, and assume that the resulting dimensions correspond to its internal components.

Instead, the engineer would investigate how the system operates, identify its functional subsystems, and determine how those subsystems interact.

Similarly, a biologist studying the kidney would seek to understand its anatomical and physiological organization.

Lykken emphasized that the structure of biological systems is often hierarchical, but not necessarily a neat hierarchy in which a small number of broad causes produce progressively narrower effects.

Furthermore, important biological mechanisms do not necessarily generate the largest individual differences.

A process can be essential to survival and still show little variability across healthy individuals.

Consequently, extracting the dimensions that account for the most variance does not necessarily identify the mechanisms that are most important for understanding the system.

This is a fundamental limitation of using correlations among people to discover the organization of psychological processes.

The Problem of Mixing Different Levels of Abstraction

One of Lykken’s most insightful observations concerned the levels of abstraction represented by factors.

He described unpublished analyses of anxiety items that produced factors corresponding to several different kinds of fear.

Some concerned highly specific situations, such as roller coasters. Others concerned broader categories, such as heights or physical danger. Still another represented general fearfulness.

But these characteristics do not occupy the same conceptual level.

Fear of roller coasters is a special instance of fear of heights, which overlaps with a broader fear of physical danger, which may itself reflect general fearfulness.

A single exploratory analysis had placed these dimensions alongside one another as though they represented comparable factors.

This is not merely an issue of choosing better labels.

The factor analysis did not distinguish the hierarchy of psychological characteristics that the researcher was trying to discover.

Lykken pointed out that researchers might obtain a more orderly solution if they already understood the relevant hierarchy and selected their items accordingly.

But that creates an obvious problem: if the structure is already known, what exactly has the exploratory analysis discovered?

This observation is particularly relevant to contemporary personality research.

Personality questionnaires often combine broad domain indicators, narrow facet indicators, and highly specific behavioral descriptions.

There is no guarantee that an exploratory factor analysis will identify these levels correctly.

A prominent factor might represent a broad disposition, a particularly well-represented narrow facet, or some mixture of characteristics at different levels.

The factor’s position in an exploratory solution does not settle the question.

Five Electrical Circuits: An Ingenious Demonstration

Lykken’s most convincing argument involved artificial systems whose actual structure was known.

He constructed five different electrical circuits containing resistors and capacitors.

The circuits differed in the number of components, their arrangement, and the ways in which the components were connected.

For each circuit, he generated 50 simulated individuals by varying component values. He then obtained ten measurements describing the circuit’s response to inputs at different frequencies.

These measurements were subjected to factor analysis.

If factor analysis could discover the underlying structure of such systems, the different circuits should have produced solutions revealing their different internal components.

That did not happen.

All five circuits produced two-factor solutions, despite their different internal structures.

The extracted factors were not recognizable as the actual resistors or capacitors.

An engineer familiar with the circuits could have identified their internal structure from the original measurements more directly.

The factor solutions were orderly statistical descriptions, but they did not reveal how the circuits worked.

This demonstration remains powerful because it avoids the familiar difficulty of evaluating personality research: we do not know the true psychological mechanisms in advance.

With the electrical circuits, the mechanisms were known.

The results showed that a successful factor analysis could fail to recover them.

One can object that the circuits were not designed to satisfy the assumptions of factor analysis. But that is precisely the issue.

Why should we assume that psychological systems are organized in ways that satisfy those assumptions?

A factor-analytic method may perform very well when data are generated by a factor model. That does not establish that it can discover the structure of systems generated by other kinds of mechanisms.

Factor Analysis of Automobiles: Plausible Factors Without Useful Explanations

Lykken’s second example is equally illuminating.

He collected 17 measurements from road tests of approximately 100 automobiles. The variables included engine characteristics, vehicle dimensions, acceleration, fuel economy, and price.

Four factors accounted for approximately 86% of the common variance.

The rotated factors were readily interpretable.

One factor was associated with acceleration and engine size. Another was associated with braking and price. A third characterized smaller, economical automobiles. The fourth involved measures such as hill-climbing performance, price, and oil capacity.

A researcher could assign plausible labels to these dimensions.

But would the factors help an automotive engineer understand what determines automobile performance?

Lykken argued that they would not.

Consider the association between automobile weight and acceleration.

Across the sample, heavier vehicles tended to have more powerful engines and therefore could accelerate faster.

Yet the physical relationship within an automobile is different. Holding engine power constant, increasing vehicle weight reduces acceleration.

The observed correlation across automobiles reflected how manufacturers combined different components in particular vehicle designs.

It did not reveal the causal effect of vehicle weight.

This distinction is extremely important for psychology.

A correlation across individuals does not necessarily reveal how the same variables are causally related within an individual.

For example, people who report high motivation may also report high stress. But increasing motivation within a particular individual may not increase stress in the same way.

The between-person correlation can reflect many different combinations of underlying processes.

Lykken was warning against interpreting correlations among people as though they directly revealed the organization of the mechanisms operating within people.

That warning remains relevant to modern personality theories that interpret higher-order factors as underlying regulatory or motivational systems.

Why Interpretability Is Not Enough

The electrical-circuit and automobile examples illustrate another central problem.

Factor analysis almost always produces patterns that researchers can interpret.

Once a factor is defined by its strongest loadings, psychologists are skilled at finding a plausible common theme.

A factor involving kindness, generosity, sincerity, and courtesy might be called Agreeableness.

A factor involving energy, assertiveness, and sociability might be called Extraversion.

These interpretations may be useful for describing the measures.

But assigning a meaningful label is not evidence that the factor corresponds to a distinct causal process.

This is particularly problematic when the factors are subsequently reified.

A statistical dimension called Extraversion becomes a psychological disposition called Extraversion. Researchers then explain correlated behaviors by claiming that they are caused by Extraversion.

The explanatory argument risks becoming circular.

We observe that people who describe themselves as sociable also tend to describe themselves as outgoing. We extract a common factor called Extraversion. We then explain the correlation between sociability and outgoingness by invoking Extraversion.

That explanation becomes scientifically informative only when the latent disposition has implications beyond the correlations used to define it.

This is why independent validation matters.

A personality construct becomes more credible when it predicts new observations, relates systematically to independently measured biological or behavioral processes, and survives demanding tests of competing explanations.

Factor analysis can contribute to that process, but it cannot substitute for it.

Discovery Versus Verification

Lykken’s distinction between the context of discovery and the context of verification is one of the most enduring lessons of the article.

Exploratory factor analysis can be useful for generating hypotheses.

Researchers can inspect different dimensional solutions, try alternative rotations, examine clusters of variables, and search for patterns worth investigating.

There is nothing wrong with using statistical methods creatively to identify potentially interesting relationships.

The problem arises when the exploratory procedure is treated as though it had already verified the resulting psychological theory.

Lykken described a study by Meehl, Lykken, Schofield, and Tellegen that used factor analysis to identify patterns among indicators of schizophrenia-related characteristics.

They also conducted a separate exercise in which judges attempted to reproduce factor labels from subsets of high-loading items.

This demonstrated that the factor interpretations could be reproduced without simply relying on the original researcher’s intuition.

But Lykken emphasized that such agreement did not establish that the factors represented distinct psychological mechanisms.

A cluster of related clinical ratings might reflect observer expectations, a genuine common cause, or several processes producing similar symptoms.

The labels could be reliable without the proposed explanation being correct.

This distinction is as relevant today as it was in 1971.

Replicating an exploratory factor structure across datasets is valuable evidence of a stable statistical pattern.

But it is not necessarily a decisive test of the substantive theory used to interpret that pattern.

Common Method Variance: An Especially Modern Problem

Another remarkable passage concerns a study of stress tolerance using several measurement methods.

The researchers included rating scales, perceptual-motor tests, MMPI scores, and physiological measures.

Factor analysis identified five major factors.

However, these factors were largely associated with the different methods used to collect the data.

In other words, the statistical analysis primarily recovered shared method variance rather than the psychological constructs that the researchers hoped to identify.

Lykken recognized that this was a problem for interpreting the factors.

The broader lesson is that observed covariance can arise from several sources.

Measures may correlate because they assess the same substantive characteristic. But they may also correlate because they share a method, response process, item format, or source of systematic error.

A factor analysis does not automatically distinguish these possibilities.

This problem is especially important in personality psychology, where much research relies on self-report questionnaires.

People differ not only in their substantive personality characteristics but also in how they describe and evaluate themselves.

A generally favorable self-description can produce positive correlations among desirable characteristics from different personality domains.

A corresponding tendency to reject undesirable descriptions can create additional covariance.

Such correlations may be reproduced across samples without establishing a substantive general personality disposition.

The relevant scientific task is therefore not simply to recover the covariance matrix. It is to determine which processes generate its different components.

What Lykken’s Critique Means for the Big Five

Lykken published his critique before the Big Five became the dominant taxonomy in personality psychology.

Nevertheless, his argument applies directly to contemporary lexical factor analyses.

There is substantial evidence that familiar broad dimensions recur in personality ratings. The Big Five are useful descriptive summaries, and measures based on them predict many important outcomes.

These accomplishments should not be dismissed.

But they establish less than is sometimes claimed.

The existence of five prominent dimensions in a correlation matrix does not establish that personality consists of five fundamental causal dispositions.

A five-factor solution may reflect a useful compression of a much more complicated structure.

The distinction becomes particularly important when moving from broad factors to higher-order constructs such as Stability, Plasticity, or a General Factor of Personality.

If the first-order factors are themselves mixtures of substantive dispositions and systematic measurement variance, their correlations cannot be interpreted straightforwardly as evidence for additional substantive common causes.

Extracting another factor from those correlations does not resolve the underlying measurement problem.

It simply summarizes the correlations at another level.

Lykken’s objection to interpreting statistical hierarchies as biological hierarchies is therefore especially relevant to claims about fundamental personality organization.

What Has Changed Since 1971?

Modern structural equation modeling (SEM) allows researchers to specify and compare more differentiated measurement models.

Instead of assuming that five common factors explain all shared covariance among hundreds of personality items, researchers can distinguish multiple sources of variation.

For example, a hierarchical measurement model can include:

• Narrow personality facets measured by multiple items.
• Broader factors representing covariance among facets.
• Secondary loadings where items reflect more than one disposition.
• Facet-specific residual covariance.
• Systematic measurement factors, such as acquiescence or general evaluative halo.

This makes it possible to examine whether a proposed personality structure reproduces the observed covariance matrix after accounting for narrower substantive dimensions and measurement effects.

Structural equation modeling does not change the fact that patterns of correlations alone cannot reveal the causal mechanisms that produce them. However, it offers important advantages over conventional EFA.

EFA is essentially a restrictive measurement model that relates observed variables to unobserved factors. It separates common variance from item-specific variance, including random measurement error. However, it generally assumes that residual covariances are zero and does not distinguish substantive common variance from systematic measurement variance. Consequently, an EFA factor may reflect a mixture of substantive personality characteristics and shared measurement biases.

SEM can be used to specify more differentiated measurement models that include method factors as well as substantive factors. This makes it possible to estimate substantive trait variance separately from identifiable sources of systematic measurement error and potentially obtain more valid measures of the intended constructs.

However, these factors are still statistical representations of covariance among the observed variables. Better measurement models do not automatically reveal the causal mechanisms underlying personality differences.

The most important additional advantage of SEM is that the correlations among factors in the measurement model can be examined using structural models derived from substantive theories. This makes it possible to test whether different theories predict the observed patterns of relationships among latent constructs.

SEM is sometimes conflated with confirmatory factor analysis (CFA), and CFA is often presented as a method for confirming factor structures previously identified by EFA. This is one possible application, particularly when the goal is to develop and validate a measure of a single construct.

However, SEM is not limited to confirming exploratory factor structures. When multiple constructs are modeled simultaneously, it offers opportunities to evaluate both measurement and substantive theories.

First, researchers can test whether a theoretically specified measurement model adequately reproduces the observed covariance matrix. A model that fails this test provides a questionable basis for interpreting its factors as measures of the intended constructs.

Second, researchers can examine whether a theoretically specified structural model accounts for the correlations among these factors without producing a substantial loss of model fit.

For example, a theory may predict that several personality facets are related because they share a broader underlying disposition. Another theory may predict that the correlations arise from multiple narrower dispositions or shared measurement biases. SEM allows researchers to specify these explanations and evaluate whether their implied covariance structures are consistent with the data.

Good model fit does not prove that a causal theory is correct, because different causal models can generate the same correlations. Nevertheless, poor fit can provide evidence against a theory’s statistical implications, and comparisons among competing models can help determine which explanations remain viable.

These applications of SEM go well beyond extracting exploratory factors or confirming simple factor structures. They provide an opportunity to move from describing correlations toward testing the empirical implications of explanatory theories.

One important development is the availability of structural equation modeling.

Modern SEM allows researchers to specify and compare more differentiated measurement models.

Instead of assuming that five common factors explain all shared covariance among hundreds of personality items, researchers can distinguish multiple sources of variation.

For example, a hierarchical measurement model can include:

• Narrow personality facets measured by multiple items.
• Broader factors representing covariance among facets.
• Secondary loadings where items reflect more than one disposition.
• Facet-specific residual covariance.
• Systematic measurement factors, such as acquiescence or general evaluative halo.

This makes it possible to examine whether a proposed personality structure reproduces the observed covariance matrix after accounting for narrower substantive dimensions and measurement effects.

Structural equation modeling does not change the fact that patterns of correlations do not reveal causal mechanisms. However, it offers important advantages over conventional EFA. EFA is essentially a restrictive measurement model that relates observed variables to unobserved factors. By separating common from unique variance, it can reduce the influence of random measurement error on estimates of common factors. However, common variance may also reflect systematic measurement error, which conventional EFA does not distinguish from substantive trait variance. SEM can be used to specify measurement models with method factors to obtain better measures of substantive factors. However, these factors are still descriptive and depend on the observed variables.

The real advantage of SEM is that the correlations among the factors in the measurement model can be modeled using structural theories that predict specific patterns of relationships among them. This makes it possible to test the empirical implications of competing causal theories.

SEM is sometimes equated with Confirmatory Factor Analysis (CFA), and it is often assumed that its only role is to confirm the results of an EFA. Confirming a previously identified factor structure is one application of CFA, but SEM can do considerably more, especially when multiple constructs are modeled simultaneously.

Here, SEM offers two important opportunities. First, it is possible to test whether a theoretically specified measurement model adequately fits the data. Second, it is possible to examine whether a structural model can explain the pattern of correlations among the latent factors without substantially reducing model fit.

These applications of SEM go well beyond EFA and the confirmation of simple factor structures. They offer an opportunity to move from description to testing the predictions of explanatory theories.

SEM was still developing as a statistical method when Lykken published his criticism of EFA in 1971. However, it subsequently became a highly influential tool in twin research for estimating genetic and environmental contributions to individual differences. Lykken used biometric structural models in his own twin research.

These analyses typically relied on personality scale scores rather than item-level measurement models that simultaneously distinguished substantive trait variance from systematic measurement variance. It would have been difficult, although not impossible, to conduct these studies using more elaborate measurement models.

For analyses of single personality traits, the additional complexity might not always have changed the main conclusions concerning genetic influences. However, accounting explicitly for measurement error and systematic method variance can change estimates of genetic and environmental contributions.

Multivariate twin models that simultaneously represented numerous personality facets, broad factors, and method effects would have been particularly challenging to estimate with the statistical software and computing resources available at the time. Advances in statistical software and computer hardware have substantially reduced these technical obstacles.

Given his skepticism about exploratory dimensional discovery in 1971, it is reasonable to ask whether Lykken would have supported contemporary efforts to replace EFA-derived sum scores with more comprehensive SEM measurement models that adequately fit the observed data.

Whether Lykken would have embraced modern SEM as a way to advance personality research remains speculative. However, his criticism of using EFA to discover the underlying structure of personality leaves little doubt that he would have been skeptical of subsequent attempts to interpret EFA-derived personality hierarchies as representations of biological systems. In particular, he explicitly rejected the assumption that biological systems have the simple hierarchical structure often imposed by factor-analytic models:

“Now the structure of organisms, while hierarchical, certainly does not have the simple-minded Christmas-tree shape envisaged by MFA” (Lykken, 1971, p. 163).

Conclusion

Why did Lykken’s criticism have so little impact on personality research, even among close colleagues with whom he collaborated?

One possible explanation is that identifying the limitations of an established method is not sufficient to change scientific practice. As Thomas Kuhn emphasized, scientific paradigms are rarely abandoned simply because their problems become apparent. Researchers also need a viable alternative that allows them to make progress.

In 1971, such an alternative was not readily available for investigating complex personality structures. Today, advances in structural equation modeling, statistical software, and computing power have made it possible to specify and test measurement models that would have been prohibitively difficult to estimate at the time.

Lykken’s criticism therefore deserves renewed attention, not because its arguments have suddenly become valid, but because researchers now have better tools to address the problems he identified. The opportunity is to move beyond extracting and naming factors toward developing and testing explicit theories of personality structure.