Version 0.3 — Working Draft

Recovering Latent Structure from Scalar Affective Scores

A Retrospective Structural Analysis of Validated Instruments Using Three-Axis Geometric Decomposition

Brian Riggleman · Independent Researcher · March 2026

Working Draft Notice
This document is a working draft of a developing research thesis. It proposes a structural reinterpretation of existing validated instruments using publicly available datasets. No new data collection is involved. No clinical claims are made. All prior versions are publicly archived below.

AI tools were used for language refinement and structural editing. All ideas, analyses, and interpretations are the author’s own.

Thesis Statement

Existing validated affective instruments contain latent multidimensional structure that is not captured by scalar scoring. A three-axis geometric decomposition (valence, activation, intensity) applied retroactively to established datasets reveals predictive variation that the original scoring methods systematically discard.

Scalar equivalence does not imply state equivalence.

Chapter 1: Introduction

1.1 The Problem

Fifty years of affective science has produced validated, reliable instruments for measuring emotional experience. The PHQ-9 measures depression severity. The ACSI measures customer satisfaction. Employee engagement surveys measure workplace wellbeing. These instruments work. They have been tested, normed, and deployed at scale.

They all share one structural property: they reduce multidimensional experience to a single number.

This thesis does not argue that these instruments are wrong. It argues that their scoring methods compress information that is present in the data. Two respondents with the same score may occupy different positions in a multidimensional space. If those positions are associated with different outcomes, the compression has a cost. This thesis measures that cost.

1.2 The Claim

This thesis proposes that existing validated instruments may contain latent multidimensional structure that is not captured by scalar scoring. It demonstrates this claim by applying a three-axis geometric decomposition to three independent datasets and showing that the decomposition reveals outcome-relevant variation the scalar scores hide.

This work does not propose replacing established instruments. It provides an alternate representation of responses those instruments already generate. The questions do not change. The scoring changes.

1.3 What This Is

A retrospective structural analysis. Not a new survey. Not a clinical intervention. Not a critique of existing science. It is an observation that existing data behaves differently when viewed through a different representational lens, and that the differences are associated with real outcomes.

1.4 The Core Insight

We measure correctly, but interpret incompletely. Measurement and representation are not the same thing. The PHQ-9 measures nine constructs reliably. The scoring method represents those nine measurements as one number. This thesis asks what happens when you represent them as three numbers instead. The answer, across three datasets and 14,266 respondents: outcome-relevant variation appears that the single number cannot see.

1.5 Scope

The thesis makes one claim: three axes reveal more than one. It tests this claim on one primary dataset and confirms it on two additional datasets. It proposes a measurement instrument (the GAS) for prospective validation. It does not claim causation. It does not diagnose conditions. It does not prescribe treatment. It does not replace the PHQ-9 or the ACSI or the Likert scale. It provides an alternate representation of responses those instruments already generate.

Likely Reviewer Criticism

  • “You’re just doing factor analysis by another name.” — Response: Factor analysis identifies latent dimensions statistically. This thesis maps items to dimensions based on clinical content and tests whether the mapping predicts outcomes the total score does not. The item-to-axis assignment is theory-driven, not data-driven. The distinction matters for interpretability.
  • “If the structure is there, why hasn’t it been found?” — Response: It has been found. Kroenke et al. (2010) identified a two-factor structure in the PHQ-9. What has not been done is testing whether that structure predicts differential outcomes within the same severity band. This thesis does that.

Chapter 2: Background

2.1 Dimensional Models of Affect

Russell (1980) proposed two dimensions. Watson et al. (1988) proposed separate positive and negative dimensions. Bradley & Lang (1994) proposed three (valence, arousal, dominance). The field has known for decades that affect is multidimensional. Scalar instruments persist because they are simple, fast, and easy to aggregate.

2.2 The Compression Problem

A PHQ-9 score of 15 is computed by summing nine items. The sum discards the pattern. Two patients with the same sum can have different item profiles. If item profiles map to different clinical presentations, and different presentations require different treatment, then the sum is losing clinically relevant information.

This is not a new observation. It is a well-documented limitation. What is new is testing whether a specific three-axis decomposition recovers that information and whether the recovered information predicts outcomes.

2.3 The Three-Axis Framework

The geometric affective state space (Riggleman 2026j, 2026k, 2026l, 2026m) defines three axes: valence (positive-negative), activation (calm-activated), and intensity (magnitude of displacement from a baseline called sigma). This framework was developed for a synthetic agent architecture. This thesis tests whether the same three-axis structure applies to human-generated survey data.

2.4 Sigma as Individual Baseline

The key addition beyond Russell’s circumplex is sigma: a respondent-specific baseline. The same score means different things for different people depending on where they start. A naturally anxious person at a moderate anxiety score is closer to baseline than a naturally calm person at the same score. Sigma captures this. Total scores cannot.

Evidence Required

  • Literature review of dimensional affect models (Russell, Watson, Barrett)
  • Documentation of known limitations of scalar scoring
  • Review of factor analytic findings for each instrument used (PHQ-9, ACSI satisfaction scales)
  • Clear statement of what is new: the specific three-axis decomposition with sigma, tested against outcomes

Chapter 3: Method

3.1 The Decomposition

For each dataset, existing survey items are mapped to three geometric axes based on their content. The mapping is theory-driven: items measuring positive-negative evaluation map to valence, items measuring arousal or engagement map to activation, and where available, expectations or baseline measures provide sigma.

The single-axis score (total, mean, or sum) serves as the comparison. The question is always the same: do respondents with the same single-axis score occupy different positions in the three-axis space, and do those positions predict different outcomes?

3.2 The Transformation: How Scalar Becomes Geometric

The transformation is explicit and reproducible. It takes existing item-level responses and computes three axis scores instead of one total score. Here is the complete procedure for the primary dataset (PHQ-9).

Step 1: Group items by axis

The nine PHQ-9 items are assigned to three axes based on what they clinically measure. This assignment is theory-driven, not statistically derived:

AxisPHQ-9 ItemsWhat It Measures
ValenceQ1 (anhedonia), Q2 (depressed mood), Q6 (low self-esteem)How bad is the cognitive-affective state? The positive-negative dimension.
ActivationQ3 (sleep), Q4 (fatigue), Q8 (psychomotor)How stirred up or shut down is the person? The calm-agitated dimension.
IntensityQ5 (appetite), Q7 (concentration), Q9 (suicidal ideation)How strongly is the state registering? The severity-magnitude dimension.

Step 2: Compute axis scores

Each axis score is the mean of its assigned items (each item scored 0–3). This produces three numbers per respondent instead of one sum:

The traditional PHQ-9 total = sum of all nine items = (Valence score + Activation score + Intensity score) × 3. The total is recoverable from the three axes. The three axes are not recoverable from the total. Information flows one direction.

Step 3: Define subtypes

Split respondents by median valence and median activation within each severity band. This produces four geometric subtypes from two cuts. No clustering algorithm is required. No free parameters.

Step 4: Compare outcomes across subtypes within the same total score

Hold the PHQ-9 total constant. Ask: do the subtypes have different suicidal ideation rates? If yes, the total score is hiding variation that the geometry reveals.

Worked Example: Two Patients, Same Score

Patient A scores PHQ-9 = 12: Q1=2, Q2=2, Q6=2, Q3=1, Q4=1, Q8=1, Q5=1, Q7=1, Q9=1.
Valence = 2.0, Activation = 1.0, Intensity = 1.0. High valence-axis depression, low activation. Anhedonic subtype.

Patient B scores PHQ-9 = 12: Q1=1, Q2=1, Q6=1, Q3=2, Q4=2, Q8=2, Q5=1, Q7=1, Q9=1.
Valence = 1.0, Activation = 2.0, Intensity = 1.0. Low valence-axis depression, high activation. Agitated subtype.

Same total score. Different geometric positions. In the NHANES data, the Anhedonic subtype shows 27.5% suicidal ideation. The Agitated subtype shows 7.8%. The total score sees 12. The geometry sees two different clinical presentations.

3.3 Item-to-Axis Mapping: All Three Datasets

3.3.1 PHQ-9 (NHANES, n=5,455)

Detailed transformation in Section 3.2 above. Sigma: population median used as reference baseline.

3.3.2 ACSI (n=7,341)

AxisACSI ItemsConstruct
Valence (Quality)Overall quality, customization, reliabilityPositive-negative evaluation
Activation (Value)Price-given-quality, quality-given-priceEngagement / worth-the-effort
Sigma (Expectations)Overall expectations, customization expectations, reliability expectationsIndividual baseline — directly measured

3.3.3 IBM HR (n=1,470)

AxisHR ItemsConstruct
ValenceJob satisfaction, relationship satisfactionPositive-negative evaluation
ActivationInverse work-life balance, job involvementArousal / engagement
IntensityVariance across all satisfaction measures + environment satisfactionMagnitude amplifier

Sigma: inferred from item mapping (median baseline).

3.4 Analysis Approach

For each dataset:

  1. Compute single-axis score (total or mean)
  2. Compute three-axis coordinates from item mapping
  3. Define geometric subtypes (quadrants or sigma-relative groups)
  4. Compare outcome rates across subtypes within the same single-axis score band
  5. Report the magnitude of hidden variation (ratio of highest to lowest outcome rate)

Likely Reviewer Criticism

  • “The item-to-axis mapping is subjective.” — Response: Acknowledged. The mapping is theory-driven (items grouped by clinical content), not data-driven (no exploratory factor analysis). Two different mappings on two different datasets produce the same structural finding. The ACSI mapping uses directly measured expectations as sigma, eliminating subjectivity for that axis.
  • “Why these three axes and not two or four?” — Response: Two axes (Russell’s circumplex) are a special case of this framework. The third axis (intensity) captures magnitude of displacement that the other two do not. Four or more axes may be viable but have not been tested. Three is the minimum that produces the observed separation.

Chapter 4: Results

4.1 Primary Analysis: NHANES PHQ-9 (n=5,455)

The PHQ-9 is the anchor dataset. It is the most widely used depression screening instrument in the world. It has nine items. It produces one number. The transformation described in Section 3.2 produces three numbers from the same nine items. This section shows what the three numbers reveal that the one number cannot.

4.1.1 Depression Subtypes Within the Moderate Band

Among 455 respondents scoring in the moderate depression band (PHQ-9 10–14), four geometric subtypes emerge with different suicidal ideation rates:

SubtypenSuicidal Ideation Rate
Anhedonic (high valence-axis, low activation)16727.5%
Classic Depressed5619.6%
Somatic-Dominant (high activation, lower valence)11715.4%
Agitated (high activation, lower valence-axis depression)1157.8%

Variation: 3.5x (27.5% vs. 7.8%) within the same PHQ-9 severity band.

At PHQ-9 = 10 exactly: 132 respondents occupy four distinct geometric positions from the same nine items.

4.2 Confirmatory Analysis 1: ACSI (n=7,341)

If the finding is real, it should not be specific to clinical data. The ACSI dataset tests whether the same structural problem — scalar scores hiding geometric variation — appears in a completely different domain with a different instrument, different items, and a different outcome variable.

4.2.1 Customer Satisfaction Subtypes

Among 1,034 respondents rating satisfaction exactly 7/10, complaint rates vary by geometric subtype:

SubtypenComplaint Rate
Exceeded Both (quality and value above expectations)32315.8%
Quality Up, Value Down29612.5%
Quality Down, Value Up8527.1%
Fell Short Both33028.5%

Variation: 2.3x (28.5% vs. 12.5%) within the same satisfaction score.

Cross-industry: The pattern holds in all four industries (food, airlines, ISPs, banks).

Sigma finding: “Fell Short Both” respondents had higher expectations (sigma 8.80) than “Exceeded Both” (sigma 7.46). Same score, different baseline, different complaint behavior.

4.3 Confirmatory Analysis 2: IBM HR (n=1,470)

A third domain, a third instrument, a third outcome variable. If the finding appears here too, the structural problem is general.

4.3.1 Employee Attrition Subtypes

Geometric quadrants defined by median valence and activation splits:

QuadrantnAttrition Rate
High Valence + High Activation74711.8%
High Valence + Low Activation32617.5%
Low Valence + High Activation27421.9%
Low Valence + Low Activation12326.0%

Variation: 2.2x (26.0% vs. 11.8%) within overlapping single-axis score ranges.

4.4 Cross-Dataset Convergence

FindingPHQ-9 (n=5,455)ACSI (n=7,341)IBM HR (n=1,470)EES (n=715, 4 waves)
Hidden variation within same score3.5x2.3x2.2x2.9x
Outcome predictedSuicidal ideationComplaint behaviorEmployee attritionAttrition intent & advocacy
DomainClinicalCustomer satisfactionEmployee retentionEmployee engagement
Sigma sourcePopulation medianDirectly measuredInferredPopulation median
LongitudinalNoNoNoYes (4 waves)
Total n14,981 respondents

Four datasets. Four domains. Four outcome variables. The same structural finding: scalar equivalence does not imply state equivalence. The EES adds the first longitudinal evidence that the decomposition structure persists across time.

Evidence Status

  • DEMONSTRATED: Geometric subtypes with different outcome rates exist within the same scalar score bands in all four datasets
  • DEMONSTRATED: The finding replicates across domains (clinical, commercial, HR, engagement)
  • SUPPORTED: Sigma (where directly measured) is associated with differential outcomes
  • SUPPORTED: Three-axis decomposition is temporally stable across four longitudinal waves (EES, Appendix A)
  • SUPPORTED: Axis-specific psychometric coherence confirmed (Cronbach’s α 0.79–0.89 per axis, EES)
  • NOT YET TESTED: Whether native three-axis collection outperforms retroactive decomposition
  • PARTIALLY ADDRESSED: Longitudinal stability demonstrated for EES; individual-level trajectory tracking still required

Chapter 5: The Geometric Affective Survey (GAS)

5.1 Purpose

If the retroactive decomposition reveals useful structure, a native instrument designed to collect three-axis data directly should capture that structure with less ambiguity than item-to-axis mapping of existing instruments.

5.2 Design

Three simultaneous bipolar sliders per item (valence, activation, intensity). Sigma established via baseline questions at intake. Output is a coordinate triple, not a number. Detailed in Riggleman (2026r).

5.3 Status

The GAS is proposed and designed. It has not been administered to human respondents. Prospective validation requires IRB approval and is identified as the primary next step.

Chapter 6: Limitations

6.1 Methodological Limitations

6.2 Instrument Limitations

6.3 Clinical Limitations

6.4 The Critical Boundary

This thesis cannot say: “This would have prevented X.” It can say: “This analysis suggests that individuals with identical scores may occupy different regions of affective space associated with different outcomes.” That is the claim. Nothing more.

Likely Reviewer Criticism

  • “You’re claiming the PHQ-9 is broken.” — Response: No. The PHQ-9 measures what it was designed to measure, reliably. The thesis observes that the scoring method discards dimensional information present in the items. The instrument is validated. The scoring could be augmented.
  • “The suicidal ideation finding is irresponsible to publish without clinical validation.” — Response: The finding is explicitly presented as a cross-sectional observation in a general population sample with multiple caveats. It is framed as motivation for clinical replication, not as a clinical recommendation.

Chapter 7: IRB Considerations

No human subjects were involved. All datasets are publicly available and de-identified:

Prospective GAS validation and clinical trajectory tracking will require IRB approval. These are identified as future work.

Chapter 8: Future Work

Chapter 9: Conclusion

We do not need new questions yet. We need a better representation of answers we already have.

The PHQ-9 asks nine questions. The ACSI asks fifteen. Employee engagement surveys ask twenty to forty. All of them compress the answers to a single number. This thesis shows that the compression has a measurable cost: 2.2x to 3.5x variation in outcome rates hidden within the same scalar score, across four independent datasets in four different domains totaling 14,981 respondents — now including longitudinal evidence across four survey waves.

The variation is not noise. It is associated with real outcomes: suicidal ideation, complaint behavior, employee attrition. It is present in the data these instruments already collect. It is discarded by the scoring method, not by the instrument.

This thesis does not propose replacing established instruments. It proposes that their interpretive structure can be augmented. The same items, scored differently, reveal structure that the total score cannot see. Whether that structure improves clinical decisions, business outcomes, or scientific understanding is the question this thesis is designed to motivate — and that prospective validation is designed to answer.

Supporting Papers

PaperRole in This Thesis
2026j — ValenceDefines the valence axis
2026k — ActivationDefines the activation axis + accumulator
2026l — IntensityDefines the intensity axis + decay framework
2026m — Emotional GeometryUnifies the three axes + sigma
2026r — GAS SurveyInstrument design + IBM HR and ACSI analyses
2026s — Clinical MetricNHANES PHQ-9 analysis + clinical metric proposal
Appendix A — EES Replication4th-dataset longitudinal replication + psychometric axis validation

Appendix A: Employee Engagement Survey — Longitudinal Replication

A.1 Purpose

This appendix extends the thesis with a fourth dataset and, critically, the first longitudinal test. If the geometric decomposition holds across four survey waves in a structurally different domain (organizational engagement), two open questions are addressed: (1) does the method generalize beyond the original three datasets? and (2) is the decomposition temporally stable?

A.2 Data

Source. Four-wave Employee Engagement Survey (EES) from a single organization: Oct 2022, May 2023, Oct 2023, May 2024. 39 Likert items (1–5) plus department. Total n=715 respondents pooled across waves (151 / 170 / 188 / 206).

A.3 Axis Construction

17 items allocated to three axes following standard organizational-behavior theory. Two outcome items held out from all axes for a clean hidden-variation test.

Valence (affective commitment, 6 items; α = 0.892)

Activation (engagement energy, 6 items; α = 0.792)

Intensity (commitment depth, 5 items; α = 0.882)

Cronbach’s α values (0.79–0.89) confirm the theory-driven partition is psychometrically sound. This is the strongest axis-level psychometric validation across all four datasets.

Held-out outcomes

A.4 Results: Hidden Variation Within Engagement Bands

Scalar total (sum of 17 axis items, range 17–85) partitioned into quartiles. Octants assigned via median splits on 3 axes.

Band Q2 (total 60–67, n=163): the critical middle band

Octant (V-A-I)nMean total% plan to leave% wouldn’t recommend
LLL8864.512.5%18.2%
LHL3266.76.2%18.8%
HLL2366.70.0%0.0%
LLH1166.418.2%27.3%

Hidden ratio: 2.91× (attrition). Respondents at identical engagement totals (~66) show an 18-point attrition gap depending on octant. HLL employees — high valence, low activation, low intensity — are the “stable-but-underdeveloped” group: they feel good about the company but are neither energized nor advancing. Their zero attrition rate despite middling total scores is the clearest demonstration of the thesis in this dataset.

Permutation inference

BandOutcomenRatioPerm. p
Q1_LowLOW_RECOMMEND1992.900.008
Q2LOW_STAY1632.910.403
Q3LOW_STAY1802.790.602

One of six comparisons reaches p<0.05 (Q1_Low advocacy, p=0.008). Ratios are directionally consistent with the thesis (2.8–2.9×) but small cells limit permutation power. The within-band outcome-table patterns are qualitatively aligned; formal significance requires a larger sample.

A.5 Longitudinal Stability

Octant composition across four waves (%)

WaveHHHHHLHLHHLLLHHLHLLLHLLL
Oct 202231.111.37.36.62.06.01.334.4
May 202338.87.65.95.34.18.24.725.3
Oct 202320.27.45.95.34.36.95.944.1
May 202423.85.82.45.82.910.22.446.6

Coefficient of variation across waves: mean 0.31. The decomposition is moderately stable. HLL is the most stable octant (CV 0.11); the structure persists even as engagement levels shift.

Wave-level axis means

WavenTotalValenceActivationIntensity% planning to leave
Oct 202215169.426.024.119.214%
May 202317070.326.224.419.814%
Oct 202318865.824.622.918.421%
May 202420666.524.623.418.515%

Between May 2023 and Oct 2023, total engagement fell 4.5 points. Attrition intent spiked to 21%, then receded to 15% by May 2024 while the total remained low. The scalar recovered partially; attrition risk decoupled from it. This is precisely the pattern a scalar metric cannot resolve — and precisely what the three-axis decomposition is designed to see.

A.6 Interpretation

The EES data provides three contributions that the original three datasets could not:

  1. Psychometric axis validation. Cronbach’s α of 0.79–0.89 per axis confirms the theory-driven partition is internally coherent. The items assigned to each axis hang together, which no prior dataset formally tested.
  2. Temporal robustness. The decomposition does not dissolve across four waves. Octant proportions shift with real organizational events, but the axis structure persists.
  3. Non-uniform drift. When engagement drops, it drops differently across axes. A scalar reports “engagement went down.” The decomposition says which kind of engagement went down. This is the operational case for multi-axis scoring in HR contexts.

A.7 Limitations

Version Archive

Every version of this thesis is publicly archived. Nothing is hidden. Nothing is deleted.

VersionDateChangesArchive
0.3 2026-04-06 Added Appendix A: Employee Engagement Survey (EES) — 4th dataset, 715 respondents across 4 longitudinal waves. Adds psychometric axis validation (α 0.79–0.89), first longitudinal stability evidence, and hidden-variation ratios of 2.8–2.9×. Updated cross-dataset convergence table to 4 datasets / 14,981 respondents. Updated evidence status and conclusion language. View
0.2 2026-03-28 Three fixes. (1) Language: “alternate representation” not “augmentation,” added “measurement ≠ representation” insight as Section 1.4. (2) Added complete transformation walkthrough in Section 3.2 showing step-by-step how PHQ-9 items become geometric coordinates, including worked example with two patients at same score. (3) NHANES promoted to primary analysis, ACSI and IBM HR demoted to confirmatory analyses with framing text. View
0.1 2026-03-28 Initial draft. 9 chapters. Three datasets. View