# Randomized Controlled Trial — Design Outline (Companion vs. Control) for an Efficacy Claim

**DRAFT protocol outline — for IRB and clinical-director review. Not effective until developed into a full protocol, reviewed, amended, and signed.**

- Program: Boojee Companion Care
- Document: `rct-design-outline.md`
- Version: DRAFT-1 / 2026-07-15
- Study type: **Randomized controlled trial** (parallel-group, superiority)
- Reporting standard: full protocol per **SPIRIT 2013** (33 items); results per **CONSORT 2010** (incl. the non-pharmacologic/behavioral-intervention extension). **Prospective registration on ClinicalTrials.gov is mandatory** (ICMJE requirement for publication).
- Status: **outline only.** This is the gold-standard study that payers and regulators eventually want for a genuine efficacy claim. It is deliberately not fully specified here — it is scoped after the fidelity study and pilot deliver their inputs (validated administration, effect-size estimate, SD of change, event rates).

---

## 0. Why an RCT, and why not yet

- Only a **randomized, controlled** comparison can support a causal claim ("the companion *reduces* loneliness/depression"). Single-arm change cannot — it is confounded by regression to the mean, natural course, seasonality, attention, and placebo.
- The RCT is expensive and slow, so it is sequenced **last**: fidelity (does the measurement hold?) → pilot (does anything move; what's the effect size and SD; is it feasible/safe?) → RCT (does it *cause* improvement vs. control?). Running the RCT before the pilot risks powering it on a guessed effect size and burning the one expensive study on wrong assumptions.

---

## 1. Objective & primary hypothesis

- **Objective:** Determine whether the companion, added to usual care, produces a greater reduction in loneliness (and depression) than the control condition over the trial period in isolated older adults.
- **Primary hypothesis (superiority):** Mean reduction in **UCLA loneliness** (primary instrument — see §5) at the primary endpoint is greater in the companion arm than control, by at least the minimally-important difference.

---

## 2. Design options (to be chosen with the clinical director & biostatistician)

- **Option A — Companion vs. Waitlist/Usual-Care.** Simplest, cleanest recruitment. Weakness: **no control for attention/expectancy** — a waitlist controls for time and natural course but not for "someone is paying attention to me," which for a *companionship* product may be a large part of the mechanism. Risk of overstating effect.
- **Option B — Companion vs. Attention/Placebo control.** Stronger internal validity: control gets an equal-dose non-companion contact (e.g., scripted check-in calls, a non-conversational content feed, or a "sham" limited bot). This isolates the companion's *specific* ingredient from generic attention. **But designing a credible attention control for companionship is genuinely hard** (see §3).
- **Option C — Three-arm** (Companion vs. Attention-control vs. Usual-care). Best inference (separates "any contact" from "this companion" from "nothing"), highest cost/N. Preferred if funded.
- **Option D — Stepped-wedge / waitlist-crossover.** Everyone eventually gets the companion (ethically attractive for a benefit population; helps recruitment/retention), with randomized timing. Adds analytic complexity and time-confounding that must be modeled.

**Recommendation:** if resourced, **three-arm (Option C)**; otherwise **Option B (attention control)** because the attention confound is the central threat for a companionship intervention. Waitlist-only (Option A) is acceptable only as a first efficacy signal, reported with the attention caveat foregrounded.

---

## 3. The attention/placebo problem (must be confronted, not hidden)

For a drug, placebo is a sugar pill. For **companionship**, the "active ingredient" and the "placebo" (attention, being contacted, routine) heavily overlap — a large share of any benefit may be *non-specific attention*, which is real but not unique to this product. Design responses:

- Use an **attention control matched on dose and contact frequency** but lacking the companion's hypothesized specific features (personalized memory, natural back-and-forth conversation, proactive companionship). E.g., scheduled scripted wellness calls or a non-conversational information service of equal contact time.
- Pre-specify that a companion-vs-usual-care effect is a **combined** (specific + non-specific) effect; only companion-vs-attention-control isolates the specific effect. Report both if three-arm.
- Acknowledge in the protocol and every downstream claim that, for this class of intervention, **the honest question is often "better than an equivalent amount of human/attention contact," not "better than nothing."**

---

## 4. Blinding — and its hard limits

- **Participants cannot be blinded** to whether they are talking to an AI companion vs. receiving check-in calls vs. a waitlist — the experience is obviously different. This is intrinsic to behavioral trials and is a stated limitation.
- Mitigations that *are* feasible:
  - **Blinded outcome assessment:** self-report instruments scored/handled by assessors blinded to arm; where an interviewer administers, they are blinded and arm-revealing cues are minimized.
  - **Blinded statisticians:** analysis on arm-masked data; SAP locked before unblinding.
  - **Objective/behavioral endpoints** (e.g., engagement, downstream healthcare utilization from claims) are less susceptible to expectancy than self-report.
- Report per CONSORT's items on blinding, including who was and was not blinded and why.

---

## 5. Primary endpoint, effect size, power

- **Primary instrument:** loneliness. Consider the **full UCLA Loneliness Scale (v3, 20-item)** as the primary efficacy instrument rather than UCLA-3, because the 20-item has higher reliability (α .89–.94; 1-yr test–retest r=.73) and more range to detect change; UCLA-3 remains the *screening* instrument. Final choice with biostatistician.
- **Co-primary / key secondary:** PHQ-9 depression change.
- **Primary endpoint timing:** e.g., 12 weeks (aligned to pilot) with a follow-up (e.g., 24 weeks) to test **durability** — a common weakness of companionship interventions is fade-out when novelty passes.
- **Effect size:** taken from the **pilot's observed change + SD** and the literature MID (a between-arm difference of ~0.3–0.5 SD is a reasonable target range; the exact MID/anchor is pre-specified). Do **not** hard-code a number here — it is set from pilot data.
- **Power / N (illustrative, to be finalized):** for a two-arm superiority test at α=0.05 two-sided, power 0.90, detecting a between-group standardized difference d≈0.4 → roughly **~130–140 per arm** analyzable; inflate for ~20–30% attrition and (if three-arm) multiplicity. **Final N comes from the SAP using pilot-derived SD**, not from this outline.
- **Multiplicity:** pre-specified handling for co-primary endpoints and (if three-arm) pairwise comparisons.

---

## 6. Population, randomization, arms

- **Population:** isolated older adults (≥60), UCLA-3 screen-positive, capacity to consent, not in active crisis (as in the pilot).
- **Randomization:** central, concealed allocation, stratified by key prognostic factors (e.g., baseline depression severity, age band, living-alone status); block or minimization to keep arms balanced.
- **Arms:** per chosen design option (§2). Usual care continues in all arms (no one is denied indicated treatment; the companion is *additive*, never a substitute for care).

---

## 7. Safety
Full **DSMB** (independent), pre-specified stopping rules, expedited SAE reporting, suicidality monitoring and 988 escalation identical to the pilot but with DSMB oversight. See `safety-monitoring-plan.md`. Ethical duty: a waitlist/attention arm must still have crisis escalation available — safety pathways are never withheld for study design.

---

## 8. Analysis
Intention-to-treat primary; pre-specified per-protocol sensitivity; mixed models for repeated measures handling dropout under stated assumptions; CONSORT flow diagram; pre-registered SAP; report effect sizes with CIs (not p-values alone).

---

## 9. What the RCT unlocks (and its remaining limits)

- **If positive:** a genuine, causal **efficacy claim** — "in a registered RCT, the companion reduced loneliness/depression vs. control" — the evidence tier payers and regulators recognize, suitable for peer-reviewed publication and a payer clinical-evidence dossier. Strength depends on the control (companion-vs-attention-control is the credible efficacy claim; companion-vs-waitlist is "better than nothing," weaker).
- **Even if positive, still limited by:** single trial (needs replication for a strong claim), specific population/setting/language, durability question (hence long follow-up), and the un-blindable nature of the intervention.
- **A single-arm result, no matter how large, never substitutes for this.**

---

## 10. Registration & reporting commitments
- **Prospective ClinicalTrials.gov registration before first enrollment** (with primary outcome pre-declared to prevent outcome-switching).
- SPIRIT-compliant full protocol; CONSORT-compliant results (including the behavioral-intervention extension); publication regardless of direction of result (no file-drawer).

---

### Sign-off (required before this outline becomes a protocol / before enrollment)
- Principal Investigator: ______________________
- Licensed Clinical Director: ______________________
- Independent Biostatistician: ______________________
- DSMB Chair: ______________________
- IRB approval # / date: ______________________
- ClinicalTrials.gov ID: ______________________

*Version DRAFT-1 / 2026-07-15. Design OUTLINE for IRB and clinical-director review. To be developed into a full SPIRIT-compliant protocol once fidelity and pilot inputs are available. Not effective until signed.*