Critical thinking
Article: Eaton, Krueger, Docherty, and Sponheim (2013)
Toward a Model-Based Approach to the Clinical Assessment of Personality Psychopathology

This paper illustrates how new statistical methods can inform conceptualization of personality psychopathology and therefore its assessment.

The relationship between structure and assessment

Structural assumptions about personality variables are inextricably linked to personality assessment.

• reliable assessment of normal-range personality traits, and personality disorder categories, frequently takes different forms, given that the constructs of interest are presumed to have different structures.
• when assessing personality traits, the assessor needs to measure the full range of the trait dimension to determine where an individual falls in it.
• then assessing the presence or absence of a DSM-V personality disorder, the assessor needs to evaluate the presence of absence of the binary categorical diagnosis.
• given the polythetic nature of criterion sets, the purpose of the assessment is to determine which criteria are present, calculate the number of present criteria, and note whether this sum meets or exceeds a diagnostic threshold.

The nature of the personality assessment instrument reflect assumptions about the distributional characteristics of the construct of interest.

• items on DSM-oriented inventories are usually intended to gather converging pieces of information about each criterion to determine whether or not it is present.

Distributional assumptions of personality constructs

Historically, many assumptions about the distributions of data reflecting personality constructs resulted form expert opinion or theory.
Both ‘type’ theories and dimensional theories have been proposed.
Assessment instruments have reflected this bifurcation in conceptualization.

• The resulting implications for assessment are far from trivial
  The structure of a personality test designed to determine whether an individual is one or two personality types, needs only to assess the two characteristics, as opposed to assessing characteristics that are more indicative or mid-range.
  • There is no mid-ground in type theory, so items covering middle-ground are not relevant.

Because the structure of personality assessment is reflective of the underlying distributional assumptions of the personality constructs of interest, reliance solely on expert opinion about these distributions is potentially problematic.

Model-based tests of distributional assumptions

It is critical for personality theory and assessment that underlying distributional assumptions of symptomatology be correct and justifiable.

• different distributions impact the way clinical and research constructs are conceptualized, measured, and applied to individuals.
• characterizing these latent constructs properly is a prerequisite for efforts to asses them.
  • it is of limited value to assess an improperly conceived construct with high reliability.

Latent trait models

Latent trait models: posit the presence of one or more underlying continuous distributions.
Individuals are dispersed along this trait dimension in a particular way (following a normal distribution) such that there are no zones of rarity.
Zones of rarity: locations along the dimension that are unoccupied by some individuals.

• The latent trait is continuous.
• Latent trait models form the basis of traditional exploratory factor analysis (EFA), confirmatory factor analysis and IRT.
  • Items are associated with a particular location along the span of the latent trait.
  • Item discrimination can be modelled as well.

Discrimination: the measure of how strongly the item taps into the latent trait.
Quasi-continuous: the construct would be bounded at the low end by zero, a complete absence of the quality corresponding with the construct.

Latent class models

Latent class models: based on the supposition of a latent group (class) structure for a construct’s distribution.

• Typically evaluated via latent class analysis (LCA)
  • A finite number of latent classes are posited.
    They are mutually exclusive and nominal, separated by zones of rarity of observations.
  • Every individual is assigned to his or her most probable class of membership given his or her responses to items on the assessment measure.
    These classes are defined by patterns of item endorsement across individuals, assuming conditional independence.
    Conditional independence: that inter-item correlations solely reflect class membership.

Conditional independence has important implications for how the construct itself is conceptualized and assessed.

• Because inter-item correlations are accounted for solely by class membership, there are assumed to be no inter-item correlations within each class.
  • This way, assessment measures that reflect a latent class model yield a nominal result for each individual.
• There is assumed to be no variation within classes, so all individuals assigned to a particular class are expected to be of the same type and not to differ among one another on the construct of interest.

Hybrid models

Hybrid models (of factor mixture models): combine the continuous aspects of latent trait models with the discrete aspects of latent class models.

• Multiple forms of factor mixture models exists.
  • Some mixture models identify classes of individuals within a particular factor structure.
  • others identify classes of individuals, within which a dimensional factor structure differentiates individuals.

EFMA: exploratory factor mixture analysis.
A set of latent classes are fit to the data and, within each class, a latent trait solution captures additional variance.

Model comparison

It is critical to characterize latent structures first, then develop appropriate measurement devices, and finally apply these assessment tools for clinical and research purposes.

Discussion

Distributional assumptions about personality psychopathology are critical to how it is conceptualized and assessed.

• Measures of personality psychopathology that reflect trait, class, and hybrid disorder distributions will necessarily differ markedly, reflecting the fundamentally different ways that these competing models characterize individual differences.

Concerning model selection

The decision of which fit indices to use to adjudicate between models can have large impact on interpretation.

• If multiple fit indices converge on the same model, this can strengthen inference about this model’s optimality.
• If they diverge, one is left in something of a statistical quinary.