Age, Agency, and Epidemics

Linking psychological variation, behavioral heterogeneity, and transmission

John M. Drake

Odum School of Ecology & CEID, University of Georgia

Evolutionary Demography Society · EvoDemoS 11

Colorado State University, Fort Collins

June 17, 2026

Collaborators

Kerri-Ann Anderson

Pej Rohani

Mike Cacciatore

Glen Nowak

One recommendation, a whole range of responses

One vaccine, one recommendation — and uptake splits sharply by age.

Influenza vaccination coverage runs roughly twofold across age — about 36% in younger adults to about 70% at 65 and over — under the same recommendation. Source: US CDC FluVaxView.

The usual explanations predict a lot, but not all of it

Politics and partisanship
Trust in institutions and providers
Risk perception — of the disease, of the vaccine
Information and misinformation

Each of these is supported by our data (not shown). But the impulse behind them — the idea of an information-deficit, tell people more and tell it better — rarely changes behavior, because it fails to understand the problem.

None of these reaches the cognitive part: how a person prefers to make the decision in the first place.

The standard account is strong. Politics predicts vaccination. So does trust in institutions, so does how a person weighs the risk of the disease against the risk of the vaccine, and so does the information environment they live in. These are not strawmen, and I am not here to dislodge them.

But notice what they share. The implicit remedy, in almost every case, is more or better information — the deficit model: if people only understood the disease, the vaccine, the evidence, they would come around. Communication researchers have tested that assumption for decades, and it almost never works, because it misdiagnoses the problem. The gap is not mainly what people know; it is how they are disposed to decide before any fact arrives.

So the point is narrower. Each of these is about what a person knows, or trusts, or whose side they are on. None of them reaches the cognitive part — how a person prefers to make a health decision in the first place, before any particular fact or message arrives. That preference is stable, and it is what the usual determinants leave out. (~55s)

The framework comes from Your Medical Mind

Glen Nowak

Groopman J. & Hartzband P. (2011). Your Medical Mind: How to Decide What Is Right for You. Penguin Press. The book describes the three orientations we carry through the talk.

This did not start with us, and it did not start in epidemiology. Groopman and Hartzband are practicing physicians at Harvard, and they wrote this book for patients — a guide to noticing one’s own habits of mind before sitting across from a doctor. It is not an academic text, and it is not about vaccines.

Glen Nowak, on the right, is a social scientist who works on vaccine risk communication, and he is the one who read it. He came away with two thoughts. The first was that these orientations might explain the ideology people bring to vaccines. The second was that we could measure them — take what two clinicians noticed at the bedside and turn it into something measurable across a whole population. That second thought is what turned a patients’ guide into a research program, and it is where the rest of this work begins. (~30s)

Agency is how people decide, not what they know

The “medical mind”: three stable dimensions of how a person engages a health decision.

By agency I do not mean autonomy in the narrow bioethics sense, and I do not mean a belief. I mean the cognitive stance a person brings to a health decision, and the book gives it three axes: minimalist or maximalist, naturalist or technologist, doubter or believer.

Each axis has a natural orientation toward vaccination. The maximalist leans toward doing more, the technologist toward the engineered intervention, the believer toward trusting the recommendation — and the opposite poles lean the other way. Because the three point in a common direction, we can add them into a single score, and that sum is what I will call disposition: how readily a person leans toward a vaccine.

Measuring it took social scientists. Glen Nowak and Mike Cacciatore worked with the National Opinion Research Center to build a survey that places people on these axes — though others have measured some of these dimensions before us; what is new here is carrying all three together, on a representative sample.

The important word is disposition. These orientations are stable, they sit upstream of any particular vaccine, and they are not a deficit to be corrected with a pamphlet. A committed minimalist who declines a vaccine has applied a consistent way of deciding, not failed to understand it. That is the idea the rest of the talk builds on. (~60s)

Agency and Age

Disposition is measurable, and spread across the population

A nationally representative NORC AmeriSpeak survey places each adult on the three continua; the combined score spreads broadly across the population (mean ≈ 0.63). To our knowledge, the only representative measurement of vaccine decision-making spanning the entire US adult population.

A word on measurement, because this audience will want the construct to hold up before I build on it. The survey runs through NORC’s AmeriSpeak panel: nationally representative of US adults, weighted to the population on age, race, education, and sex. To our knowledge it is the only representative measurement of how the whole adult population approaches vaccine decisions. And it is rich — in the same people we have demographics, the decision-making score, political orientation, risk perception, and vaccine confidence and intent; and in the 2023-to-2024 wave a follow-up on what they actually did, which links stated intent to behavior. That panel is the basis of another talk; today I use the cross-section.

A person’s answers place them on the three continua I just described, and we combine those into one score, which I rescale from zero to one so we can carry it through the talk. The distribution matters as much as the average. These scores are spread broadly across the whole range, not clustered into two camps or a small fringe against a compliant majority. The population fills the behavioral space, which is the condition under which within-group variation matters for the dynamics later. (~65s)

Disposition predicts vaccine confidence and hesitancy, beyond the usual determinants

In hierarchical regression on two nationally representative samples, decision-making preference is entered last — after demographics, perceived knowledge, physician recommendation, and other vaccine attitudes — and still adds independent signal:

Vaccine confidence — positive, p < 0.001 in both 2016 and 2018 (ΔR² ≈ 4–7%)
Vaccine hesitancy — negative and significant in both years
Recent vaccination and next-season intent — significant in 2016

Anderson, Nowak, Cacciatore, Rohani & Drake (2025), Vaccine 65:127804. OLS hierarchical regression; NORC AmeriSpeak national samples (2016, N ≈ 1000; 2018, N ≈ 1015).

Does disposition earn its place? Here is the published test — Kerri-Ann Anderson’s paper in Vaccine this year. She ran the standard hierarchical regression and entered decision-making preference last: after demographics, after perceived knowledge, after whether a physician recommended the shot, after the other vaccine attitudes. Entered last, against all of that, it still carried its own signal. It predicted higher vaccine confidence and lower hesitancy in two separate nationally representative samples, eight years apart, with every coefficient in the expected direction and significant. On confidence it added on the order of four to seven percent of explained variance on top of everything else — which, for a single construct entered last, is not a small thing.

A word for this audience on why confidence and hesitancy are two outcomes and not one. They are strongly correlated, but they are not complements. Confidence is epistemic — what a person believes about a vaccine, whether they judge it to work and to be safe. Hesitancy is procedural — how they approach the decision itself, how readily they act on it. Someone can think well of a vaccine and still drag their feet, or distrust it and still comply. Disposition predicts both, with the signs you would expect, which is why I read it as sitting upstream of either.

And that is why the four-to-seven percent is a floor, not a ceiling. It is the variance disposition adds on its own, after everything else is in the model — but disposition almost certainly also acts through confidence and hesitancy, not only beside them. A hierarchical regression hands the shared part to whatever was entered earlier, so that unique slice understates disposition’s real contribution. Measuring how much of its effect runs through those two mediators is a path-analysis question, and one we still owe.

For recent behavior and next-season intent the signal was clear in 2016 and weaker in 2018, so I do not lean on it. But the core claim is now on firm, published ground — how a person is disposed to decide predicts whether they trust and accept a vaccine, over and above what they know and who recommends it. (Backup, if pressed: the lasso variable importance and the theory mapping.) (~85s)

Disposition is correlated with age

Mean disposition is modestly higher in older age groups, while every age group spans nearly the full range — measured across people at one time.

Look across the groups: through middle age the average barely moves (the youngest are not even the lowest), and then in the oldest groups it climbs. Older groups carry more high-disposition people, and the shift is concentrated past fifty-five.

This is a correlation with age, measured across people at a single time, and no more than that. Whether disposition actually rises as a person ages, or whether older and younger generations simply differ, I cannot tell you from these data — and I will come back to that. What I can say is that disposition is patterned by age, which makes it another age-associated trait for a society that studies age-associated traits.

Notice too that the within-age spread does not collapse as the average shifts. Every age group still holds the full range. Keep that in mind — it is the variation a compartmental model is about to throw away. (~70s)

A small shift in the mean, a large shift in who vaccinates

Disposition maps onto vaccination through a steep logistic curve. Holding age fixed, a person at the 10th percentile of disposition vaccinates with probability 28%, one at the 90th percentile with probability 71% — a 43-point swing within the population. The shift in group means across age, by contrast, moves it only about 9 points.

The gradient in the means looked small, but the variation underneath that average is not. Disposition feeds vaccination through this curve, and the curve is steep where most people sit. Holding age fixed and moving from the tenth percentile of disposition to the ninetieth, vaccination probability goes from twenty-eight percent to seventy-one percent — a forty-three-point swing, between people of the same age.

Set that against the age story from the previous slide. Sliding from the youngest group’s average disposition to the oldest group’s average moves vaccination probability by about nine points. The spread we carry inside every age group is four to five times larger than the shift in the averages across age. The curve itself is the link the model assumes, fit to survey uptake — but its steepness, and that asymmetry, do real work: a compartmental model keeps the age means and discards the spread, so it keeps the smaller effect and throws away the larger one. That is what the rest of the talk undoes. (~50s)

Behavior falls short of intent, more so as intent weakens

Flu vaccination by stated intent, from a 2023→2024 panel (intent measured in 2023, vaccination behavior in the same people in 2024): ~85% of those who say they will “definitely” get one do, versus ~55% of “probably,” ~15% of “probably not,” and ~4% of “definitely not.”

Now comes the bridge from psychology to dynamics, through a gap everyone in vaccine research knows: people do not do what they say they will. The shape of the gap is the useful part. This comes from a genuine panel — we asked the same people their intent in 2023 and then followed up on what they actually did in 2024, so this is stated intent against real behavior, not two separate cross-sections. Among those who said they would definitely get one, about eighty-five percent followed through. Among those who said probably, only about fifty-five percent did. Among the two unwilling groups, almost no one did — about fifteen percent, then four.

So vaccination tracks intent, but conversion falls off as intent weakens, and fastest at the hesitant end. The willing convert cheaply; the reluctant barely convert at all. That nonlinearity is what the next slide turns into a cost. (~55s)

The cost of reaching the reluctant depends on the vaccine

Implied effort to convert one person (the inverse of the conversion rate), by stated intent, for two vaccines. The familiar flu vaccine stays cheap from the willing (~1.2) to the merely probable (~1.8); a novel vaccine — the updated COVID shot — climbs about fivefold (~1.25 to ~6).

Turn that follow-through into a price. The effort to convert one person is the inverse of their conversion rate, and that lets me put two vaccines side by side. The navy line is the familiar flu vaccine: the willing cost about one and a quarter, and the merely probable still under two, so reaching past the eager costs almost nothing extra — even the hesitant mostly come along for a shot they already know. The gold line is a novel vaccine, the updated COVID shot, which is the closest data we have. It starts in the same place, the willing again near one and a quarter, but the merely probable cost about six (five times as much), because people are warier of an unfamiliar vaccine, and those who say only “probably” far more often never follow through.

And because willingness is graded by age, each of these costs is graded by age too. Coverage is cheaper to buy among older, higher-disposition adults and dearer among younger ones — a demographic statement about the price of an intervention, which is not a quantity we usually let ourselves think about.

So the price of coverage depends on the vaccine a person is asked to take, not on the person alone. The model uses both curves: a familiar seasonal vaccine near the flat navy line, a novel pandemic strain — itself a new vaccine — near the steep gold one. We do not commit to one shape; we carry a family of convex forms spanning the two regimes and test every conclusion against all of them. At Oxford last year I listed this cost curve as future work; this is the future work. (~70s)

We cannot yet separate age from cohort

Everything so far is cross-sectional: disposition by age, measured across people at one time.

That cannot distinguish a life-course effect (people change as they age) from a cohort effect (generations differ). The two are confounded.

Separating them needs longitudinal data that follows people over time.

If it is cohort, the population’s disposition mix is non-stationary — uptake will drift as generations replace one another.

One might ask whether any of this is about aging at all. It is the right question. Everything I have shown is cross-sectional: I have disposition by age, pooled across survey years, but measured across different people. That design cannot separate two very different stories. One is a life-course story, in which a given person grows more disposed toward vaccination as they age. The other is a cohort story, in which older and younger generations simply differ and carry their dispositions with them. With cross-sectional data the two are confounded, and no amount of pooling fixes that — it is the old age-period-cohort problem. Telling them apart requires following the same people over time.

I flag it because the answer changes the long-run picture. If it is a life-course effect, the age pattern is stable and we can plan against it. If it is generational, the behavioral composition of the population drifts as cohorts replace one another, and so will the coverage we can expect. That is a demographic research program, and one I would like to do with people in this room. (~55s)

From persons to populations

Do small differences in agency scale up, or wash out?

Hypothesis: Small, age-graded differences in disposition → modest shifts in coverage → changes in epidemic timing, size, and mortality?

Following that chain requires a model that carries individuals, not group averages.

A compartmental model averages disposition within each age group

The age-structured SIRVD compartments split each age group only by vaccination status, so disposition enters as a single group-level rate — the within-age spread has nowhere to live.

Our first approach to this question was a compartmental model — an age-structured set of differential equations, the kind many of you have written: age groups, age-structured contact from the POLYMOD survey, vaccination, age-specific mortality. It reproduces an influenza season well, and it is the model behind our earlier work with Kerri-Ann Anderson.

The question it was built to answer is an old one: whom to protect. Vaccinate the people doing most of the transmitting — younger, high-contact — for an indirect benefit that shrinks the outbreak for everyone; or vaccinate the vulnerable — typically older — for a direct benefit against death. POLYMOD contact and age-specific mortality are what let a model weigh those two.

But one feature is built into its mathematics. Within each age group, everyone is identical — each group carries a single disposition, its average, and the within-age spread we just established has nowhere to live. That is not a failing of the modeler; it is what compartments are. And it may be a real problem: the spread that has nowhere to live is exactly what a targeted campaign reaches into. That is why we built the other kind of model, one that keeps every individual. (~65s)

An agent based model carries each person’s age and disposition

50,000 agents. Each carries age and disposition.

Contacts — age-structured mixing (POLYMOD)
Disease — age-specific infection fatality
Vaccination — a hazard driven by disposition and age
Effort — a fixed budget, allocated under the cost of coverage

The ODE had to average disposition within age. The agent model keeps age and disposition separate.

The model is deliberately ordinary in its epidemiology, so the behavioral part does the work. Fifty thousand agents, each with an age and a disposition score drawn to match the survey distributions we just saw. They mix according to an empirical age-structured contact matrix, so a teenager and an eighty-year-old meet the right people at the right rates. Infection and death follow age-specific rates. Vaccination is a hazard that depends on disposition and age.

The one piece that is not standard is effort. We give the campaign a fixed budget and make it pay the cost of coverage we just estimated, so reaching the disinclined draws the budget down faster. Every strategy I compare spends the same total effort; they differ only in where they spend it.

The difference is simple: the differential-equation version had to replace each person’s disposition with their age-group average. The agent model does not — and that single difference is what the next slide is about. (~60s)

The agent model

Each agent’s age and disposition feed contact, infection and death, and the vaccination decision; a fixed budget pays the cost of coverage. The ODE must average disposition within age; the agent model keeps them separate.

With heterogeneity turned off, the agent model and the equations agree

Seasonal and pandemic share one virus — the same transmissibility (R₀ = 2.0) — and differ only in the population’s immunity: the seasonal epidemic meets prior immunity and a pre-vaccinated public (effective R ≈ 1.4), the pandemic a fully susceptible one (effective R = 2.0). The pandemic is worse not because the pathogen changed but because the population is susceptible.

With disposition averaged within age and no targeting, the compartmental trajectory falls inside the agent model’s stochastic envelope. Agent model: 50,000 agents, 51 replicates per scenario; R₀ = 2.0, exponential infectious period (mean 3 days).

Three ways to spend the same vaccination effort

The experiment crosses two factors — how we represent people, and how we target the campaign. Three targeting rules, each spending the same total effort:

Broad — a small uniform lift across all adults
Priority — extra effort to the age groups with the lowest baseline uptake
Behavioral — extra effort to the people most disposed to vaccinate

They differ only in where the effort goes, not how much.

Averaging by age understates the deaths a campaign can avert

Under pandemic conditions with a steep cost of coverage, the same vaccination effort averts about three times as many deaths when individual disposition is represented (median 14 per 50,000) as when it is averaged within age (median 5).

This is the payoff of representing the variation, measured in lives. Same population, same fixed budget, the same mean behavior in every age group — the only difference is whether we carry each person’s disposition or replace it with their age-group average. The number of deaths a campaign averts should not depend on a modeling choice that subtle. It does.

Under pandemic conditions, with a realistically steep cost of coverage, the same budget averts about three times as many deaths when we keep individual disposition — a median of fourteen per fifty thousand against five. The mechanism is the cost curve: finding the cheap, willing people one at a time stretches the budget further. Under ordinary seasonal conditions the gap is small, so this is not a universal effect — it shows up where epidemics are severe and people are expensive to reach, which is when it matters most.

And the lesson generalizes: a compartment is an average, and collapsing either age or disposition into it throws away the within-group spread where the consequences live. One last, humbling result — the figure pools the three targeting strategies because the differences between them were small. Which rule one chooses matters far less than whether the variation was represented at all. (~65s)

Heterogeneity’s ruses, in behavioral form

Frailty selection makes a cohort’s mortality hazard bend below every member’s still-rising hazard (Vaupel & Yashin 1985); averaging disposition within an age group hides behavioral variation in exactly the same way.

The title of this slide is not mine. It is the title of a 1985 paper by James Vaupel and Anatoli Yashin — “Heterogeneity’s Ruses” — and I borrow it on purpose, to set what I have shown this morning in a tradition that is yours, not epidemiology’s. Their phrase means something specific.

Their lesson was that a population with unobserved variation behaves in ways no individual in it does. Their case was frailty, and the figure shows its shape. Every thin line is one person’s mortality hazard, and every one of them rises. The frail die first, so the survivors are selected, and the bold line — the cohort average we actually observe — bends below its own members and levels off, even though no individual’s hazard ever does. The “ruse” is that the aggregate misleads precisely because the variation is real and unmodeled.

What I have shown this morning is the same phenomenon, with behavior in place of frailty. Averaging disposition within an age group erases the spread — and the spread is what a targeted campaign reaches into and what the cost curve prices. Smooth it away and both the lever and the cost go with it. This is a cousin of Simpson’s paradox, not the same thing (Simpson’s is an association that reverses under conditioning, while here the aggregate takes a shape no member has), but it is the same family: aggregation hiding what holds within the groups.

One guardrail, so that I am not overselling the behavior: disposition is not what drives the epidemic — age-structured contact does that, the way who-meets-whom always does — and disposition is what lets a campaign steer within that structure. Both are age stories, and the unobserved-heterogeneity lesson runs through both. (~75s)

Conclusion: Agency is a demographic variable

Agency is an age-structured demographic trait

Two age-structured layers run through this talk: behavior — who contacts whom, measured by POLYMOD, and who gets vaccinated, measured by the NORC survey — and cognition, how a person is disposed to decide. Cognition is upstream of behavior, and it is where our interventions land.

Decision-making disposition — the overlooked, cognitive layer — is itself patterned by age: a trait to set beside the mortality and fertility schedules demographers study.

What it points to:

communication matched to how people decide, not only to what they know
a measurable, age-graded price for moving cognition — without the trials that measuring behavior would demand
whether disposition is age or cohort

The whole argument reduces to one claim: decision-making agency is patterned by age. I will not call it a life-course schedule, because I cannot yet separate aging from cohort — but it is an age-associated trait, one to set beside the mortality and fertility schedules you study, and it propagates into population dynamics through exactly the heterogeneity arguments this society has made about frailty for forty years.

One distinction matters most. The contact behavior in this model — POLYMOD, who meets whom — is behavior, and behavior is hard to move. Cognition, how a person is disposed to decide, sits upstream of it, and cognition is what our interventions actually target: a doctor’s recommendation, a public-health message. Two things follow. Surveys like ours measure the cost of moving cognition directly, where measuring an intervention’s effect on behavior would demand randomized trials we rarely get to run; and understanding the path from cognition to behavior lets a campaign act more surgically. Age here is the organizing variable, not a nuisance proxy, and the agent model is what let me stop averaging it away.

It points three ways: communication matched to how people decide, not only to what they know; a measurable, age-graded price for moving cognition; and the age-versus-cohort question as the next real study. (~65s)

Thank you

John M. Drake · Odum School of Ecology & CEID, University of Georgia

Supported by the U.S. National Science Foundation — PIPP Phase I: Heterogeneous Model Integration for Infectious Disease Intelligence (DEB-2200158). With thanks to my collaborators, the AmeriSpeak / NORC survey teams, and the survey participants whose answers made the empirical work possible.

Backup

Replicate variability falls with N but persists at 50,000

Replicate-to-replicate variability falls as population size grows but remains substantial at N = 50,000.

The survey items behind the three continua

Two items anchor each continuum: Naturalist–Technologist, Doubter–Believer, Minimalist–Maximalist.

A lasso retains disposition items, distinct from the usual theories

Left: lasso variable importance for stated vaccination intent (overall R² = 0.62) — retained predictors draw from all four frameworks, including the Medical Mind disposition items (Q16). Right: survey items mapped to the Health Belief Model, Theory of Planned Behavior, and COM-B, with the Medical Mind items forming a construct of their own. Survey analysis: Anderson et al.

This is the evidence behind the claim that disposition earns its place, for when someone presses on it. We mapped the survey items onto four behavioral frameworks — the Health Belief Model, the Theory of Planned Behavior, COM-B, and the Medical Mind — and the mapping itself was part of the work, because many of these items had never been classified this way. Then we ran a lasso on stated vaccination intent: it penalizes complexity and keeps only what predicts.

Two things to take from it. No single theory wins — the retained predictors draw from all four, so health decision-making is genuinely multidimensional. And the Medical Mind items survive that selection. They are not the largest coefficients; the threat and recommendation items lead. But they are retained, and they sit in their own theoretical circle rather than being folded into the others. That is exactly what I mean on the main slide when I say disposition adds signal of its own. Overall R-squared is 0.62.

Why the cost of coverage is steeper for a novel vaccine

Implied effort to convert one person (the inverse of the conversion rate), by stated intent, for a familiar (flu) and a novel (COVID) vaccine. Across the willing, cost rises ~1.5x for flu but ~4.8x for COVID; the familiar curve is flat, then jumps — the piecewise shape the model carries.

Full hierarchical regression — 2016

Predictor	Confidence	Hesitancy	History	Intent
Block 1 — Demographics
Gender (M=1; F=2)	0.04	0.24**	-0.09	-0.15*
Age	0.00	0.00	0.02***	0.02***
Race (Non-Wht=0; Wht=1)	-0.17***	-0.28**	-0.10	-0.16*
Household Income	-0.01	-0.01	0.00	-0.01
Education	-0.01	-0.13***	-0.03	0.00
R² (%)	2.7***	7.5***	14.5***	14.5***
Block 2 — Information
Perceived Flu Knowledge	0.22***	-0.05	0.23***	0.18***
Healthcare Rec. (Y=1; N=2)	-0.20**	0.27**	-0.58***	-0.55***
Inc. R² (%)	9.4***	5.6***	17.1***	15.6***
Block 3 — Flu Impacts
Flu Seriousness	0.13**	0.06	0.16***	0.25***
Flu Spread	0.04	-0.14**	0.15***	0.11*
Inc. R² (%)	2.8***	1.8***	4.6***	5.3***
Block 4 — Vaccine Attitudes
Vaccine Confidence	N.A.	-0.52***	0.23***	0.31***
Vaccine Hesitancy	-0.30***	N.A.	-0.22***	-0.26***
Access Difficulty	0.16**	0.00	-0.01	-0.01
Inc. R² (%)	15.9***	15.4***	11.5***	15.2***
Block 5 — Health Decision-Making Preferences
Decision-Making Preferences	0.27***	-0.13*	0.13**	0.17***
Inc. R² (%)	3.5***	0.45*	0.49**	0.65***
Total R² (%)	34.3	30.8	48.2	51.2

Anderson, Nowak, Cacciatore, Rohani & Drake (2025), Vaccine 65:127804. OLS hierarchical regression, NORC AmeriSpeak 2016; standardized coefficients, blocks entered in order. N = 1001 (Confidence, Intent), 998 (Hesitancy), 978 (History). N.A. = outcome itself enters as a predictor. p<0.05, p<0.01, p<0.001.

Full hierarchical regression — 2018

Predictor	Confidence	Hesitancy	History	Intent
Block 1 — Demographics
Gender (M=1; F=2)	-0.03	0.02	-0.02	0.06
Age	-0.01***	-0.01*	0.01***	0.01***
Race (Non-Wht=0; Wht=1)	-0.11*	-0.22**	-0.03	-0.06
Household Income	0.00	0.00	0.00	0.00
Education	-0.02	-0.04	0.05*	0.04
R² (%)	0.70	3.3***	8.1***	7.3***
Block 2 — Information
Perceived Flu Knowledge	0.14**	-0.01	0.19**	0.18**
Healthcare Rec. (Y=1; N=2)	-0.20***	0.11	-0.64***	-0.59***
Inc. R² (%)	9.7***	6.0***	16.8***	15.6***
Block 3 — Flu Impacts
Flu Seriousness	0.09*	-0.06	0.17***	0.21***
Flu Spread	0.04	-0.07	0.16***	0.16***
Inc. R² (%)	3.6***	3.2***	5.4***	6.3***
Block 4 — Vaccine Attitudes
Vaccine Confidence	N.A.	-0.73***	0.24***	0.34***
Vaccine Hesitancy	-0.39***	N.A.	-0.24***	-0.32***
Access Difficulty	0.10*	0.02	-0.06	-0.15*
Inc. R² (%)	31.8***	32.1***	13.7***	20.5***
Block 5 — Health Decision-Making Preferences
Decision-Making Preferences	0.41***	-0.14**	0.00	0.04
Inc. R² (%)	6.6***	0.4**	0.0	0.0
Total R² (%)	52.4	45.0	44.0	49.7

Anderson, Nowak, Cacciatore, Rohani & Drake (2025), Vaccine 65:127804. OLS hierarchical regression, NORC AmeriSpeak 2018; standardized coefficients, blocks entered in order. N = 1015 (all outcomes). N.A. = outcome itself enters as a predictor. p<0.05, p<0.01, p<0.001.