# Instrument Psychometrics Dossier — Boojee Companion Care

**Prepared for:** Clinical Director, IRB, and payer clinical reviewers
**Scope:** psychometric profiles of every instrument the Companion Care system administers, with an
emphasis on (a) validity in **older adults**, (b) the caveats of **repeated / monitoring** use versus
one-time screening, and (c) the **mode-of-administration** question raised by delivering these items
**conversationally**.

**Honesty posture (load-bearing):** the four instruments below are well-validated *as self-administered
or standard-interview screeners*. That validation does **not** automatically transfer to (i) a
conversational-AI delivery mode or (ii) use as a longitudinal change-detection tool. Where we state a
number we cite it; where the evidence does not exist for our exact use case we say so plainly.

---

## 1. UCLA-3 — Three-Item Loneliness Scale

**Primary citation:** Hughes ME, Waite LJ, Hawkley LC, Cacioppo JT. *A Short Scale for Measuring
Loneliness in Large Surveys: Results From Two Population-Based Studies.* **Research on Aging.**
2004;26(6):655–672. doi:10.1177/0164027504268574. (PMC2394670 / PMID 18504506.)
**Parent instrument:** Russell DW. *UCLA Loneliness Scale (Version 3): Reliability, Validity, and
Factor Structure.* J Pers Assess. 1996;66(1):20–40.

### Derivation
Derived from the 20-item Revised UCLA Loneliness Scale (R-UCLA-20). Hughes et al. selected **three
items** — chosen to tap three facets of loneliness: relational connectedness ("lack companionship"),
social connectedness ("left out"), and self-perceived isolation ("isolated from others") — explicitly
so the scale could run on a **telephone survey**. The original R-UCLA wording and its 4-point (1–4)
frequency anchors were simplified to a **3-point** anchor for telephone/oral use.

### Items & scoring
- Item 1: "How often do you feel that you lack companionship?"
- Item 2: "How often do you feel left out?"
- Item 3: "How often do you feel isolated from others?"
- Anchors: **1 = hardly ever, 2 = some of the time, 3 = often.**
- **Total range 3–9** (higher = lonelier). A common threshold treats **total ≥ 6** as "probable /
  meaningful loneliness." NOTE: the ≥6 cut is a widely used field convention, **not** a diagnostic
  cutoff established by a criterion-standard interview; loneliness is a self-reported state, not a
  diagnosis, so "sensitivity/specificity" framing does not strictly apply.

### Reliability / validity in older adults
- Developed and tested **in older adults**: the original validation drew on the **Health and Retirement
  Study (HRS)** and a Chicago population sample; the telephone form was tested with a large sample of
  **middle-aged and older adults (~2,100)**.
- **Internal consistency: Cronbach α ≈ 0.72** in the telephone sample (Hughes 2004). Lower than the
  full R-UCLA-20 (α often ~0.89–0.94), which is expected for a 3-item scale, but "quite good" for three
  items per the authors.
- **Concurrent/convergent validity:** the 3-item score correlates strongly with the full R-UCLA and
  with related constructs (depressive symptoms, social support, health) — the authors report acceptable
  **discriminant and concurrent validity**. A later head-to-head psychometric comparison (Psychological
  Medicine) found the 3-item and 20-item versions perform comparably for many survey purposes while
  noting the 3-item version's narrower coverage.

### Administration validation
- **Telephone / oral administration is the mode the scale was BUILT for** — this is a genuine strength
  relative to the other three instruments (which were built as paper self-report). The item stems and
  the simplified 3-point anchors were designed for a live voice interviewer reading them aloud.
- Caveat: "telephone interviewer reads a fixed script" is **not identical** to "conversational AI
  weaves the item into dialogue." The oral-administration validation supports voice delivery of the
  *fixed stems*; it does not license paraphrasing or conversational reframing (see §5).

### Known limitations
- Three items = **narrow construct coverage**; cannot distinguish social vs emotional loneliness, and
  the word "lonely" is deliberately absent (reduces stigma but also face-content).
- **Coarse 3-point scale → limited granularity** and a real **floor/ceiling** concern for change
  detection (only 7 possible total scores). Small true changes may be invisible or, conversely, a
  one-anchor shift on one item moves the total by a full point.
- No criterion-standard "gold" for loneliness → cut points are conventions, not validated thresholds.

---

## 2. PHQ-2 and PHQ-9 — Patient Health Questionnaire (depression)

**Primary citations:**
- Kroenke K, Spitzer RL, Williams JBW. *The PHQ-9: Validity of a Brief Depression Severity Measure.*
  **J Gen Intern Med.** 2001;16(9):606–613.
- Kroenke K, Spitzer RL, Williams JBW. *The Patient Health Questionnaire-2: Validity of a Two-Item
  Depression Screener.* **Med Care.** 2003;41(11):1284–1292.
- Löwe B et al. *Monitoring depression treatment outcomes with the PHQ-9.* Med Care. 2004;42(12):
  1194–1201. (PMID 15550799) — change/monitoring properties.

### Items & scoring
- **PHQ-9:** 9 items mapped to the DSM depression criteria; each 0–3 ("not at all" → "nearly every
  day") over the last 2 weeks. **Total 0–27.**
- **PHQ-2:** the first two PHQ-9 items (depressed mood + anhedonia). **Total 0–6.**
- **Severity bands (PHQ-9):** 1–4 minimal, 5–9 mild, 10–14 moderate, 15–19 moderately severe, 20–27
  severe (Kroenke 2001).

### Diagnostic accuracy (general/primary care)
- **PHQ-9 ≥ 10:** sensitivity **88%**, specificity **88%** for major depression vs a criterion-standard
  interview (Kroenke 2001). ≥10 is the standard screen-positive threshold.
- **PHQ-2 ≥ 3:** the author-recommended screen cut (sensitivity ~83%, specificity ~92% for MDD at the
  ≥3 threshold in validation work; some studies report a ≥2 cut trades sensitivity up for specificity
  down). Our system uses **PHQ-2 ≥ 3 → administer PHQ-9**, consistent with the 2003 paper.

### Older-adult / geriatric validation (this is the population that matters for us)
- The PHQ-9 is validated in older adults, but the **optimal cut point tends to be LOWER** than 10 in
  geriatric samples:
  - Swedish older-adult study: **PHQ-9 ≥ 5** gave sensitivity **100%**, specificity **81%** (Youden 0.81)
    for a major depressive episode; **≥10** gave excellent specificity **95%** but sensitivity fell to
    **71%** (PMC8559588). Implication: the standard ≥10 cut **under-detects** depression in elders.
  - Taiwanese elderly sample: optimal **PHQ-9 ≥ 6** (sensitivity 1.00, specificity 0.85).
- **Clinical implication for us:** in older adults we should treat **PHQ-9 ≥ 10** as the *severity/escalation*
  anchor but recognize that a **5–9 "mild" band is NOT negligible** in this population and warrants human
  review, not reassurance. This should be encoded in the escalation SOP.
- **Cognitive impairment caveat:** PHQ-2/PHQ-9 performance degrades with cognitive impairment in aging-
  services clients (PMC3930057) — self-report reliability drops. Users with suspected MCI/dementia need a
  clinician-administered or informant-supplemented pathway, not automated self-report alone.

### Change / monitoring properties (central to a monitoring product)
- The PHQ-9 **is** sensitive to change and is validated for treatment monitoring (Löwe 2004).
- **Minimal important difference (MID) ≈ 5 points** (estimated as ~2 SEM on the 0–27 scale).
- **Reliable change (Jacobson–Truax) ≈ a reduction of ≥ 6 points** — the more conservative threshold
  many services use to call an *individual* change "reliable" rather than measurement noise.
- These thresholds are for **standard self-administration**; they are our best available anchors but were
  not derived for conversational-AI delivery.

### Item-9 / suicide caveat (must be handled correctly)
- PHQ-9 **item 9** ("thoughts that you would be better off dead, or of hurting yourself") is a **safety
  trigger, NOT a suicide-risk assessment.** Its positive predictive value for actual suicidal behavior is
  low and it will produce false positives; conversely a "0" does **not** rule out risk. Any endorsement
  (item 9 > 0) must route to a **human + crisis resources (988)**, and a structured tool (e.g., C-SSRS)
  administered by a clinician remains the standard. Automated systems must never treat item 9 as
  sufficient risk stratification.

### Limitations for our use
- **2-week reference window** vs our (potentially more frequent) monitoring cadence: re-administering
  PHQ-9 weekly asks about overlapping 2-week windows — statistically autocorrelated, and users may
  anchor/round to prior answers (**practice/response-set effects**). Cadence should respect the reference
  window; over-frequent re-administration inflates apparent stability.
- **Floor effects** near remission and **ceiling** at severe end limit change resolution at extremes.
- Somatic items (sleep, appetite, energy, concentration) overlap with **normal aging and comorbid
  medical illness** → risk of inflated scores in elders from non-mood causes.

---

## 3. GAD-7 — Generalized Anxiety Disorder 7-item scale

**Primary citation:** Spitzer RL, Kroenke K, Williams JBW, Löwe B. *A Brief Measure for Assessing
Generalized Anxiety Disorder: The GAD-7.* **Arch Intern Med.** 2006;166(10):1092–1097.

### Items & scoring
- 7 items, each 0–3 over the last 2 weeks. **Total 0–21.**
- **Severity bands:** ≥5 mild, ≥10 moderate, ≥15 severe.
- **Screen-positive cut ≈ 10.**

### Diagnostic accuracy
- At **GAD-7 ≥ 10:** sensitivity **89%**, specificity **82%** for GAD (Spitzer 2006, n=965 with
  criterion interviews). Also performs reasonably as a screen for panic, social anxiety, and PTSD at the
  same cut (lower specificity).
- **Factor structure:** predominantly **unidimensional** (single anxiety factor); good internal
  consistency (α ~0.89–0.92 across samples, incl. heterogeneous psychiatric samples, PMC6691128 /
  PMC5333929).

### Older-adult validation
- The GAD-7 performs acceptably in older adults and is widely used in that population, though it has
  **less geriatric-specific validation than the PHQ-9**, and anxiety presentation in elders is often
  entangled with depression and medical illness (lower discriminant precision). Use the standard ≥10
  screen-positive cut, but interpret in older adults with the depression/medical-comorbidity overlap in
  mind and treat it as a screen that routes to a human, not a diagnosis.

### Limitations for our use
- Same **2-week window / repeated-measure autocorrelation** caveat as PHQ-9.
- No established GAD-7 MID/reliable-change value as clean as the PHQ-9's 5/6-point anchors; change
  interpretation is weaker. Treat GAD-7 trends as **directional signal**, not calibrated change.
- Somatic anxiety items overlap with cardiac/respiratory disease common in elders.

---

## 4. Cross-cutting caveat — MONITORING vs one-time SCREENING

All four instruments were validated primarily as **one-time screeners** (or, for PHQ-9, treatment
monitoring under standard administration). Using them as a **repeated longitudinal signal** in a
consumer product introduces measurement issues that must be stated to the IRB and payers:

- **Practice / familiarity effects:** repeated exposure to identical items changes responding
  (memory of prior answers, response sets, "just say what I said last time").
- **Regression to the mean:** users who enroll while distressed will tend to improve on re-test
  regardless of any intervention — a *major* confound for any pre/post "it works" claim.
- **Floor/ceiling & granularity:** especially UCLA-3 (7 possible scores) — limited resolution to detect
  real change; PHQ-9/GAD-7 better but bounded at extremes.
- **Reference-window overlap:** re-administering a "last 2 weeks" item weekly produces
  autocorrelated, non-independent observations.
- **Only PHQ-9 has clean individual-change anchors** (MID ≈5, reliable change ≈6). UCLA-3 and GAD-7
  lack equally established per-person change thresholds — we must not manufacture them.

**Consequence:** these instruments can support *honest monitoring and triage*. They cannot, by
themselves, support a *causal efficacy claim* for the product (see `measurement-plan.md`).

---

## 5. THE CENTRAL FIDELITY CONCERN — conversational vs standardized administration

This is the most important caveat in this dossier and we state it without hedging.

**What we do:** the companion delivers each instrument **conversationally**, one item at a time, in
dialogue, using the **exact validated item stems and response anchors**.

**Why it is a genuine deviation:** every published psychometric property above was established under a
**standardized administration mode** — paper/computer self-report for PHQ/GAD, and a **fixed telephone
script** for UCLA-3. Conversational-AI delivery differs in ways that can move scores:
- surrounding dialogue and rapport can create **social-desirability / demand** effects (users may
  under-report distress to a "friendly" companion, or over-report for attention);
- any **paraphrase, reordering, or empathic reframing** of a stem breaks standardization;
- voice/turn-taking pacing differs from silent self-completion;
- an AI that has been conversing warmly is not a neutral administrator.

**What the current literature says (honestly, both directions):**
- *Encouraging but preliminary:* a GPT-4o voice-interactive chatbot administering the PHQ-9 showed
  **high concordance with self-administration (ICC ≈ 0.91; median absolute difference ≈ 1 point), no
  systematic severity bias** (HopeBot, arXiv 2507.05984, 2025). Virtual-human questionnaire
  administration studies and paper-vs-computer-vs-mobile equivalence work (e.g., PMC5607438) also
  generally find **rough equivalence** across modes.
- *But:* those studies are **small, recent, not in older adults, not our product, and not peer-reviewed
  at the level our other citations are.** Face-to-face/interview modes have shown **lower reliability
  (α ≈ 0.80) than self-administration (α ≈ 0.87)** in some comparisons — i.e., mode *can* matter.
- **No study validates OUR conversational companion administering these instruments to OLDER ADULTS.**

**Position:** conversational administration is **plausibly near-equivalent** based on early chatbot/IVR
equivalence data, but it is **not validated for our population and mode**, and we will not claim
equivalence until we run it. Concrete mitigations, already recommended in the Clinical Brief:
1. Preserve **verbatim** item stems and anchors; no paraphrase, no reordering (enforced in code).
2. Offer a **standardized self-administration mode** as an option/comparator.
3. Run a **within-subject concordance/fidelity study** (conversational vs standardized, counterbalanced)
   in the target older-adult population, reporting ICC, mean bias, and limits of agreement (Bland–Altman)
   before any efficacy claim.

---

## Sources
- Hughes, Waite, Hawkley, Cacioppo (2004), *Research on Aging* 26(6):655–672 — https://journals.sagepub.com/doi/10.1177/0164027504268574 ; PMC https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2394670/
- Russell (1996), *J Pers Assess* — UCLA-3/R-UCLA parent — https://www.researchgate.net/publication/14623374
- Kroenke, Spitzer, Williams (2001), *J Gen Intern Med* 16(9):606–613 — https://onlinelibrary.wiley.com/doi/abs/10.1046/j.1525-1497.2001.016009606.x
- Kroenke, Spitzer, Williams (2003), PHQ-2, *Med Care* 41(11):1284–1292
- Löwe et al. (2004), PHQ-9 monitoring, *Med Care* — https://pubmed.ncbi.nlm.nih.gov/15550799/
- PHQ-9 in older adults (Swedish GDS-15/PHQ-9) — https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8559588/
- PHQ-2/PHQ-9 in aging-services clients with cognitive impairment — https://ncbi.nlm.nih.gov/pmc/articles/PMC3930057
- PHQ validation in Taiwanese elderly — https://www.sciencedirect.com/science/article/pii/S187395981630117X
- Spitzer, Kroenke, Williams, Löwe (2006), GAD-7, *Arch Intern Med* 166(10):1092–1097 — https://ctc-ri.org/sites/default/files/uploads/gad7_ref.pdf
- GAD-7 psychometrics (heterogeneous psychiatric sample) — https://pmc.ncbi.nlm.nih.gov/articles/PMC6691128/ ; https://pmc.ncbi.nlm.nih.gov/articles/PMC5333929/
- Reliable Change Index / MID for PHQ-9 — https://novopsych.com/support/user-guide/reliable-change-index-rci/ ; Jacobson–Truax https://pubmed.ncbi.nlm.nih.gov/15637779/
- HopeBot LLM PHQ-9 administration concordance — https://arxiv.org/pdf/2507.05984
- Paper/mobile/tablet depression-item equivalence — https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5607438/