Skip to main content
Companion Care
Clinical Research · Evidence Roadmap

From “uses validated instruments” to “has evidence.”

Using UCLA-3, PHQ-2/9, and GAD-7 is table stakes — the instruments are validated, but our conversational administration of them, and any outcome from the program, are not. This is the study program that closes that gap, staged so each step earns exactly one honest claim and nothing more: first prove the measurement holds, then look for a signal, then — only then — test whether it works.

SPIRIT 2013 protocols CONSORT 2010 trials STROBE observational Common Rule 45 CFR 46 ClinicalTrials.gov registration
01 — The Roadmap

Three stages. Each unlocks one claim.

Evidence is earned in order. Skipping a stage doesn't buy the claim — it just makes the claim wrong. The honesty rule: never state a claim a completed study cannot support.

1
Fidelity / Equivalence

Does conversational scoring equal the validated screen?

Within-subjects, counterbalanced. Same older adults take each screen two ways — standardized self-report vs. the companion in conversation — and we test agreement (ICC, Bland–Altman, weighted κ) and equivalence (TOST) against pre-set margins.

Unlocks: “The companion’s conversational scores are equivalent to the validated self-administered instruments.” Does NOT unlock any claim that the program helps anyone.
2
Single-Arm Outcomes Pilot

Does loneliness/mood move — and can we run this safely?

~40 isolated older adults, 12 weeks, pre/post UCLA-3 + PHQ-9 with the Reliable Change Index, plus real engagement and safety data. No control group.

Unlocks: “In an uncontrolled pilot, X% showed reliable improvement; the program was feasible and safe.” An association + feasibility claim. Does NOT unlock any claim that the companion CAUSED the change.
3
Randomized Controlled Trial

Does the companion actually cause improvement vs. control?

Registered, randomized, controlled — ideally against an attention-matched control, because for a companionship product the honest question is “better than an equal amount of contact,” not “better than nothing.” Blinded assessment; powered from the pilot’s effect size.

Unlocks: the real efficacy claim — “in a randomized controlled trial, the companion reduced loneliness/depression vs. control.” The evidence tier payers and regulators recognize. Still limited by a single trial, one population, and durability — replication strengthens it.
Where we are today: pilot-stage, zero enrolled participants, zero results. The category benchmarks we cite elsewhere (ElliQ, Papa, Pyx impact reports) are those vendors’ own program data, not peer-reviewed RCTs and not our results. These protocols are how we build our own evidence, honestly.
02 — The Protocols

Four IRB-grade draft documents.

Each is a full draft protocol written to the standard an IRB and a payer’s clinical team would expect — objectives, design, sample-size justification, analysis plan, safety, and honest limitations.

Study 1 · Equivalence

Instrument-Fidelity Study

Answers the #1 objection: a score obtained in conversation is not automatically the validated score. This proves (or disproves) equivalence.

  • Within-subjects, counterbalanced, 48h–7d washout
  • Margins: UCLA-3 ±1, PHQ-9/GAD-7 ±2 points
  • Agreement: ICC ≥0.75, Bland–Altman, weighted κ ≥0.60
  • Target N≈120 (ICC-precision driven)
  • PHQ-9 item-9 concordance gated as a safety analysis
Read protocol Study 2 · Pilot

Single-Arm Outcomes Pilot

The program’s first honest data point: does loneliness and mood move over 12 weeks, and can we operate safely? Uncontrolled by design.

  • ~N=40, 12 weeks, isolated older adults (UCLA-3 ≥6)
  • Primary: UCLA-3 + PHQ-9 via Reliable Change Index
  • Engagement + retention (≥70% target) + safety base rates
  • ClinicalTrials.gov registered before enrollment
  • Explicit: association only, not causation
Read protocol Study 3 · RCT

Randomized Controlled Trial — Outline

The eventual gold standard for a causal efficacy claim. Sequenced last, powered from the pilot. Confronts the hard attention/placebo problem head-on.

  • Companion vs. attention-control (or 3-arm) vs. usual care
  • Blinded outcome assessment (participants can’t be blinded)
  • Primary: UCLA loneliness; power from pilot effect size
  • CONSORT + prospective ClinicalTrials.gov registration
  • Full independent DSMB
Read outline Cross-cutting · Safety

Human-Subjects Safety Plan

Governs all three studies. Suicidality escalation, adverse-event reporting, elder-abuse disclosure, consent & capacity, and data safety.

  • PHQ-9 item-9 / risk language → 988 + human clinician, logged
  • AE/SAE definitions + IRB reporting timelines
  • Mini-DSMB (pilot) → full DSMB (RCT); stopping rules
  • Capacity-to-consent screen for older adults
  • Elder-abuse mandatory reporting; HIPAA data safety
Read plan
03 — Safety First

Crisis care always overrides research.

These studies enroll isolated older adults and screen for depression and suicidal ideation. Safety is the precondition for running at all. In every study — including the pure measurement study — a positive suicidality signal triggers the standing crisis pathway immediately, independent of any research procedure.

Standing rule. PHQ-9 item 9 > 0, or any suicidal statement in any mode, triggers immediate 988 Suicide & Crisis Lifeline resources (call or text 988), a real-time clinician alert, an audit-logged escalation, and — for imminent risk — emergency services per the clinical director’s SOP. No data point is worth a missed escalation.
04 — What This Program Does Not Claim

The honest ceiling of each stage.

  • No results exist yet. Nothing here has been run; we claim no outcomes.
  • Fidelity ≠ benefit. Proving conversational scores match validated ones says nothing about whether the program helps.
  • The pilot cannot prove causation. With no control arm, change may be regression to the mean, natural course, seasonality, attention, or measurement reactivity.
  • Companionship placebo is hard. Much of any benefit may be non-specific attention; only an attention-controlled RCT isolates the companion’s specific effect.
  • Participants can’t be blinded. Behavioral-trial limit; we blind assessment and analysis instead, and say so.
  • Findings are build-specific. Any change to the production model or prompt re-opens fidelity — results are pinned to a build hash.
  • Scope is bounded. English-language, capacity-intact older adults; other languages and advanced cognitive impairment require their own studies.

Trust & Evidence Center — full compliance & evidence status →