PVQ-RR: Revised Portrait Value Questionnaire

• PVQ-RR is a psychometric instrument designed to measure basic and higher-order human value dimensions using 57 systematically constructed vignette items.
• It operationalizes Schwartz’s value theory through hierarchical scoring and multidimensional scaling, validating a circular circumplex structure of human values.
• Recent advancements integrate neural network embeddings via the SQuID pipeline to recover latent value structures with high congruence to human-assessed data.

The Revised Portrait Value Questionnaire (PVQ-RR) is a psychometric instrument designed to measure the structure of human values in accordance with Schwartz’s value theory. The PVQ-RR operationalizes value measurement via a set of systematically constructed items, each describing hypothetical persons with specific motivational emphases. Scores from the PVQ-RR permit hierarchical aggregation from fine-grained value domains to broader higher-order value dimensions, with the empirical structure validated by multidimensional scaling (MDS) revealing a circular circumplex consistent with theoretical predictions. Recent developments have demonstrated that neural network embeddings, processed according to specific methodologies, can recover the latent structure of these values as effectively as human rater data, offering scalable alternatives for psychometric evaluation (Pellert et al., 29 Sep 2025).

1. Instrument Design and Theoretical Foundations

The PVQ-RR comprises 57 items, each a concise vignette depicting a person exemplifying a specific motivational orientation (e.g., “It is important to him/her to form his/her views independently.”). Respondents indicate the correspondence of each hypothetical person to themselves on a 6-point Likert-type scale ranging from “not like me at all” to “very much like me.”

Each of the 19 basic value domains (for example, Self-Direction-Thought, Power-Dominance, Universalism-Concern) is measured by exactly 3 items ($57 = 19 \times 3$). These 19 value domains are theory-guided and map hierarchically to 4 higher-order dimensions—Openness to Change, Conservation, Self-Enhancement, and Self-Transcendence—arranged in a well-validated circumplex structure reflecting Schwartz’s motivational continuum.

The scoring procedure computes the mean response over the 3 items for each value domain:

$y_d = \frac{1}{3} \sum_{i \in items(d)} r_i$

Aggregated scores for the higher-order value dimensions are calculated by averaging the means of their respective basic domains.

Schwartz’s value theory posits that values form a circular, motivationally structured space in which adjacent value domains are positively correlated (indicating compatibility), whereas diametrically opposed domains exhibit negative correlations (indicating motivational conflict). Empirically, this structure is evaluated by MDS of the inter-value correlation matrix, with recovery of such patterns—particularly negative correlations—serving as a stringent test of the theoretical model.

2. Scoring, Validity, and Internal Consistency

Internal consistency for each value domain is typically assessed using Cronbach’s $\alpha$, given $k = 3$ items per domain:

$$\alpha = \frac{k}{k-1} \left(1 - \frac{\sum_i \text{Var}(y_i)}{\text{Var}(\sum_i y_i)}\right)$$

Empirical benchmarks indicate that for respondent ratings, mean $\alpha$ values are approximately 0.70, supporting sufficient reliability.

The circumplex arrangement of values is validated by MDS applied to the inter-domain correlation matrix, with recovery of positive and negative inter-value relations serving as key evidence. Achieving accurate negative correlations (e.g., between Conformity and Self-Direction) remains critical for discriminant validity as required by the theoretical framework.

3. Neural Network Embedding Approaches

Survey and Questionnaire Item Embeddings Differentials (SQuID) is a methodological pipeline facilitating neural embedding-based reconstruction of the PVQ-RR’s latent value structure (Pellert et al., 29 Sep 2025). SQuID operates independently of respondent data, utilizing text embeddings generated by LLMs or domain-specific transformers.

Several embedding architectures have been evaluated:

• Linq-Embed-Mistral (7.11B, open, prompt-based)
• gemini-embedding-exp-03-07 (closed API)
• jina-embeddings-v3 (572M, multilingual)
• KaLM-mini-instruct-v1.5 (494M, prompt-based)
• mpnet-personality (109M, domain-finetuned)

Typical embedding dimensionalities span 384 to 4096. For empirical baselines, 4096-dimensional random Gaussian vectors are used.

SQuID proceeds as follows:

1. Extract $e_i \in \mathbb{R}^D$ for each PVQ-RR item $i$.
2. Compute the mean embedding $\overline{e} = \frac{1}{57} \sum_{i=1}^{57} e_i$.
3. Calculate centered differentials $\Delta_i = e_i - \overline{e}$ to amplify negative semantic signals and subtract shared “language-baseline” components.
4. Aggregate value-domain embeddings by averaging the 3 $y_d = \frac{1}{3} \sum_{i \in items(d)} r_i$0 for each domain.

Pairwise domain similarity is quantified as $y_d = \frac{1}{3} \sum_{i \in items(d)} r_i$1, with Pearson’s $y_d = \frac{1}{3} \sum_{i \in items(d)} r_i$2 computed over embedding dimensions. These similarities are transformed into dissimilarities $y_d = \frac{1}{3} \sum_{i \in items(d)} r_i$3 for MDS analysis.

4. Embedding-Based Psychometric Structure Recovery

Empirical results demonstrate that SQuID-treated embeddings substantially replicate psychometric properties as observed in human data. For Linq-Embed-Mistral:

• Mean $y_d = \frac{1}{3} \sum_{i \in items(d)} r_i$4 for value domains is ≈0.77, slightly above human data (mean $y_d = \frac{1}{3} \sum_{i \in items(d)} r_i$5 ≈ 0.70), and greatly exceeding a random baseline ($y_d = \frac{1}{3} \sum_{i \in items(d)} r_i$6 ≈ 0).
• Correlations between embedding-based and human-derived domain-domain similarities reach $y_d = \frac{1}{3} \sum_{i \in items(d)} r_i$7 (95% CI: [0.66, 0.80]), corresponding to $y_d = \frac{1}{3} \sum_{i \in items(d)} r_i$8 of explained variance.
• Ordinal MDS analysis yields a two-dimensional circular configuration consistent with theory. Procrustes-aligned comparisons between human and embedding MDS recover factor congruence coefficients $y_d = \frac{1}{3} \sum_{i \in items(d)} r_i$9 and $\alpha$0 for the first and second MDS axes, implying substantial configural overlap.
• The SQuID procedure enables recovery of negative and positive inter-domain correlations, an advance over prior embedding methods which failed to resolve negative relations without domain-specific fine-tuning.

The following table summarizes core comparative metrics:

Data/Model	Mean $\alpha$1 (per value domain)	$\alpha$2 (embedding vs human)	Factor Congruence $\alpha$3 (1st/2nd MDS axis)
Human	≈ 0.70	—	—
Linq-Embed-Mistral + SQuID	≈ 0.77	0.55	0.88 / 0.82
Random Embeddings	≈ 0	≪ 0.55	Near zero

5. Methodological and Applied Implications

The embedding-based reconstruction of PVQ-RR value dimensions introduces methodological efficiencies to psychometrics:

• The SQuID pipeline is model-agnostic and requires no fine-tuning or human-in-the-loop annotation.
• All 57 PVQ-RR items can be processed in under 20 minutes on a standard CPU.
• Pre-validation of novel items, item translations, or adaptations is feasible via embedding similarity without recourse to human rater panels.
• Country- or language-specific MDS comparisons enable profiling of “value orientations” latent in LLMs, suggesting avenues for cross-cultural computational value analysis.
• Integration of SQuID outputs into automated “pseudo-factor analysis” pipelines is possible, potentially generalizing to domains beyond value measurement, such as attitudes or personality constructs.

6. Limitations and Open Questions

Despite these advances, critical limitations persist:

• Possible overlap between LLM pretraining data and item texts introduces the potential for indirect data leakage, though no explicit prompt steering has been applied.
• The process for “reverse scoring” embeddings to match psychometric norming standards remains undetermined.
• Embedding-driven similarity metrics and value structures may reflect the cultural and linguistic biases of model pretraining corpora, challenging claims to universal validity.
• SQuID methodology requires further field validation against outcome measures and applied contexts to establish construct generalizability.

The embedding-based approach represents a scalable adjunct to human survey-based psychometrics. However, its use must be contextualized within broader validity and representativeness concerns intrinsic to foundation model outputs and the latent structure of human values (Pellert et al., 29 Sep 2025).