Health Insurance Literacy
Reframed health insurance enrollment from "better materials" to "better translation" after stats showed time and sources don't predict understanding.
Problem Space
Young adults selecting employer-sponsored health insurance for the first time encounter employer-provided resources that fail to establish meaningful insurance literacy. Across our research — 103 survey respondents, 8 moderated interviews, 16 AI-moderated interviews, and 14 evaluative sessions on a public cost-estimation tool — one pattern held in all four methods: the problem isn't access to information, it's translation. Most enter enrollment feeling healthy, default to the cheapest available option, and discover the coverage gaps later as unexpected bills.
"I don't know the difference, and it doesn't really matter to me."
My Role & Constraints
My Role
Survey analysis lead and synthesis writer — built the statistical analysis pipeline, ran survey recruitment across 30 distributed channels, moderated 2 of the 8 in-person interviews, wrote the cross-method insights synthesis tying every interview theme to its survey evidence, and authored the team's 3-page design brief. I used AI to teach myself non-parametric statistics and to write the analysis script.
Team
Natalia Avella, Eaghan Wright, Carly Zuklin. We developed our research plan and protocol as a team, coded the first two interviews independently, then built a shared codebook. Visualizations weren't mine.
Constraints
Ten-week quarter. Open enrollment had already closed before fieldwork, so we couldn't observe selection live — we opted for a portal walkthrough to study how people use insurance resources post-enrollment. Our original strict screener forced a mid-quarter revisit.
Research Methods
Moving to a panel out of necessity, then testing AI moderation on purpose
Our original plan was 8–12 human-moderated interviews under a tight screener, but recruitment stalled in the first weeks — strict criteria with no incentive made target participants hard to source. We pivoted to a paid panel (Maze/Prolific) to scale recruitment and unblock the study. The AI moderation choice was deliberate: many of the sponsors at UX360 were AI-moderation tools, and we wanted to test them in our own study.
We kept 8 human-moderated sessions for qualitative depth and ran 16 AI-moderated sessions through Maze, split into structured (8) and unstructured (8). Structured used our scripted guide; unstructured used general topics of interest. Structured hit the script but its probes were off-context, leading, or missing altogether. Unstructured probed better, but neither AI tier matched the nuance of a human follow-up.
Cost dominates selection
The interviews surfaced "cost first" as a near-universal pattern. The survey ranked it precisely: out-of-pocket costs, deductibles, and premiums sat at the top, ahead of every coverage-quality measure.
Top Selection Factors
% of survey respondents
"Definitely the cost. Like, I chose the lowest cost."
Time and Sources Don't Predict Understanding
We expected that participants who spent more time researching, or consulted more sources, would land in a more confident, more literate place. The data said otherwise. The survey gave us a clean null where we expected a positive correlation: time didn't predict confidence, number of sources didn't predict literacy on any item, and the most common source — employer HR materials — produced no measurable lift over not using HR at all.
The evaluative sessions then made the same pattern visible in real time: participants who had every piece of information in front of them on the Mayo Clinic cost calculator still couldn't translate it into an estimate they trusted. That ran directly against our "they should just try harder" assumption. I walked the team through the data, pairing each null and each evaluative observation against the matching interview theme, and we re-anchored our brief: the problem isn't engagement or effort, it's translation.
Where We Expected a Correlation, We Found None
We expected three things to predict insurance literacy — time spent learning, number of sources consulted, and using employer HR resources. None of them did.
More time doesn't mean more confidence
Confidence stays flat across every time bucket.
More sources doesn't mean more literacy
Every correlation is near zero, on every literacy concept.
HR users score no higher than non-users
Identical medians on every literacy item.
Outcome
The research reframed the design surface.
"Young adults need better employer materials" became "young adults need plans explained as real consequences" — because more material doesn't close the literacy gap. The synthesis became the design brief's three learnings: people only learn the aspects of insurance that bill them, cost is the lens but the total is invisible, and the tools meant to help fail at the highest-stakes moment. Asked to price a foot MRI on a public calculator, 6 of 14 evaluative participants abandoned insurance entirely.
The brief turned those into six How Might We statements and four design-exploration areas — research-backed direction without prescribing the form. The methods comparison produced its own conclusion: AI-moderated interviews are a viable scale lever for attitudinal data, but probe quality still belongs to human moderators.
Areas for Design Exploration
Four design moves, each tied to a moment where the research showed translation failing. None prescribe a specific form — they tell designers where to look.
Plain-language plan translator
A decision-time layer that reframes plan types and coverage as lived consequences — "with this plan, seeing a specialist works like…" rather than definitions. Targets the single best-supported literacy gap.
HMW 1Personalized cost estimator
A tool that takes a young adult's likely usage and returns a realistic annual cost, with insurance prefilled and deductible status folded into the result. One solution answers both cost opportunities: a real total and a tool that works on the first try.
HMW 3, 6Employer-delivered decision aid
A guided, plain-language alternative to the benefits packet, delivered through the channel young adults already trust and start with, and built to make a confident choice possible in under an hour rather than asking for more effort.
HMW 2, 4Post-enrollment support layer
Plain-language explanation of benefits, deductible tracking, and provider search built into the portal, designed for the two moments people actually log in: finding care and decoding a charge.
HMW 5Reflection
I'd rewrite the unstructured AI protocol to push deeper into the patterns our moderated interviews surfaced, rather than reusing the same guide — built as a hypothesis-testing tier, it could have produced genuinely new insight rather than a thinner version of what we already had.
I'd also add a portal-experience item to the survey, since the qualitative portal frustrations sit interview-only right now and a survey-level signal would let me defend them at scale.