Human Development and the Architecture of Everyday Worry: Latent Worry Profiles in the World Risk Poll

ABSTRACT

Whether public concern about risk rises or falls with national development has been debated for three decades, but the debate has compared countries hazard by hazard, asking how much people worry rather than how worry is organized. We provide a person‐centered analysis of everyday worry at global scale, applying latent class analysis to the six‐hazard worry battery of the Lloyd's Register Foundation World Risk Poll and linking profiles to the Human Development Index (HDI) in 140 countries ($n=139,111$; 2023 wave), with the 2021 wave ($n=118,703$; 119 countries) as an independent replication. Four worry profiles replicate almost exactly across waves, including a configurationally distinct secure‐basics profile (15% of respondents) combining low worry about food and water with high worry about crime, severe weather, and traffic. Profile prevalence is consistent with the ordered sequence predicted by a need‐priority account of worry allocation: pervasive worry falls from 33% to 13% between low‐ and very‐high‐HDI countries, low worry nearly doubles, and the secure‐basics profile holds steady across low‐ and medium‐HDI countries before declining at higher development. Traffic accidents are the most intense worry in every tier below very high HDI. These gradients persist after adjustment for self‐reported recent harm and demographic composition and are only partly accounted for by the objective hazard environment, though the cross‐sectional design cannot separate development from macroregional clustering. Development is associated with the composition of worry, not only its overall level; risk communication may be better matched to a country's worry architecture than to a single global template.

1. Introduction

What people worry about is not a simple function of what threatens them. Four decades of risk perception research have established that lay judgments of hazards are structured, multidimensional, and only loosely coupled to actuarial risk (Fischhoff et al. 1978; Slovic 1987). Yet, the comparative reach of this tradition has never matched its conceptual ambition. The cross‐national studies that followed the psychometric paradigm typically relied on small convenience samples in a handful of mostly industrialized countries, and reviews of that literature concluded that the observed country differences were substantial but undertheorized and difficult to compare (Boholm 1998; Sjöberg 2000). Even recent multicountry work spans 10 or so countries clustered at the upper end of the development distribution (Dryhurst et al. 2020). As a result, we know a great deal about how individuals within wealthy societies perceive hazards, and remarkably little about how the structure of everyday worry varies across the full range of societal development. The dominant explanatory currents in risk perception research—the role of affect and the availability of frightening instances (Slovic et al. 2007), of social trust in the institutions that manage hazards (Siegrist and Cvetkovich 2000), and of cultural worldviews that orient people toward or away from particular risks (Kahan et al. 2011)—were likewise developed and tested overwhelmingly within high‐income societies. Whether the very organization of risk concern differs across the development spectrum is a prior question that bears on all of them.

The question of how development relates to public concern about risk is itself unsettled. One tradition holds that concern about nonmaterial risks is a postmaterialist disposition that emerges once basic needs are secure (Inglehart 1995)—a scarcity hypothesis whose hierarchical core, we will argue, the worry data allow us to test more directly than the level comparisons on which the debate has run; others have shown that citizens of poorer nations often express equal or greater concern than citizens of rich ones (Dunlap and York 2008), while multilevel analyses of the same survey programs have recovered a positive wealth‐concern gradient (Franzen and Meyer 2010). Closer to our question, Lee et al. (2015) found that the predictors of climate change risk perception differ qualitatively between developed and developing countries, suggesting that development is associated not only with the level of concern but also with how it is organized. This debate has remained largely variable‐centered: it asks whether average concern about a hazard is higher or lower in richer countries. Framed this way, the debate conceals a different and arguably prior question—whether the configuration of concern across hazards, the way worry is allocated within persons, differs systematically across levels of development.

Two strands of evidence suggest that it should. First, research on economic risk preferences has documented a risk‐income paradox: within countries, richer individuals are more risk tolerant, yet across countries, populations of poorer countries are on average more risk tolerant than those of richer ones (L'Haridon and Vieider 2019; Bouchouicha and Vieider 2019; Mata et al. 2016; Falk et al. 2018). Since poorer populations also report more intense worry about everyday harms, perception and preference appear to diverge across the development gradient—a divergence that mean‐level comparisons cannot resolve. Second, theory predicts development‐linked differences in worry allocation rather than only in worry volume. If attention to risk is a finite resource (Weber 2006; Sisco et al. 2023), populations facing acute basic‐needs hazards should allocate concern differently—not merely more intensely—than populations whose food and water are safe. Person‐centered methods, which model respondents as holders of response profiles rather than as points on separate item‐by‐item scales, are the natural tool for this question, but they have so far been applied to risk perception only within single countries.

This article provides, to our knowledge, the first person‐centered analysis of everyday worry at a global scale. We use the Lloyd's Register Foundation World Risk Poll (Lloyd's Register Foundation 2024), which fields an identical multihazard worry battery—food, drinking water, violent crime, severe weather, road traffic, and mental health—to nationally representative samples in more than 140 countries. We analyze the 2023 wave ($n=139,111$ across 140 countries after exclusions) with the 2021 wave ($n=118,703$; 119 countries) as an independent replication, and link each country to its Human Development Index (HDI) from the United Nations Development Programme (United Nations Development Programme 2025). We ask three questions: (1) Do distinct multihazard worry profiles exist, and do they differ in configuration rather than only in level? (2) Does the prevalence of these profiles vary systematically with HDI? (3) Does any HDI gradient persist after adjustment for self‐reported recent harm—that is, is HDI associated with the allocation of worry across hazards beyond the plausible differences in exposure?

Anticipating the results, latent class analysis recovers four worry profiles that replicate almost exactly across the two waves, including a configurationally distinct profile—about 15% of respondents worldwide—that combines low worry about food and water with high worry about crime, severe weather, and traffic. Profile prevalence is strongly graded by HDI: the share of the pervasive‐worry profile falls by more than half from low‐ to very‐high‐HDI countries, while the low‐worry profile nearly doubles. These gradients are robust to adjustment for self‐reported recent harm and demographic composition, and they are only partly accounted for by the objective hazard environment, although they cannot be separated empirically from broad macroregional clustering—a scope condition we examine directly. The contribution is threefold: a global person‐centered typology of everyday worry; the use of HDI—a composite of health, education, and income—as the development measure, which connects the typology to mechanisms beyond national income; and a graduated identification analysis showing that HDI is associated with the allocation of concern over and above self‐reported harm and the measured objective hazard environment.

2. Theoretical Background and Hypotheses

2.1. Three Levels: Individual Worry, Development Context, and Objective Environment

Our argument distinguishes three levels that cross‐national work on development and risk perception often runs together. The first is the individual level: a respondent's pattern of worry across hazards, and in particular whether concern about food and water has receded relative to concern about physical‐safety hazards. The second is the country development context, indexed here by the HDI. The third is the objective hazard environment—the actual incidence of food insecurity, unsafe water, violent crime, road death, and disaster exposure—which varies both across countries and across individuals within them.

Keeping these levels distinct matters because the object we model is an individual worry profile, whereas development and the hazard environment are properties of the contexts in which such profiles are more or less common. We therefore do not claim that countries pass through stages in which basic needs are secured nationally; we ask how the prevalence of individual worry configurations varies across development contexts, and how far that variation tracks the objective environment rather than development as such. A worry configuration in which food and water concern is low can be common in a low‐HDI country without that country having secured food and water at the national level, provided that a substantial subgroup of its residents perceives those basic provisions as adequate while remaining concerned about safety hazards.

Several nonexclusive mechanisms could make worry configurations covary with development context. The most direct is exposure: residents of less developed contexts face objectively higher risks of unsafe food and water, road death, and weather‐related harm, and worry may track experience. Three further mechanisms would operate beyond exposure. Formal safety nets insure citizens against lifetime risks and can lower the subjective stakes of everyday hazards (Sinn 1995); informal insurance through dense kin and community networks plays the analogous role where welfare states are thin, and has been shown to alter risk judgments directly (Weber and Hsee 1998; Hsee and Weber 1999). Scarcity itself changes risk‐relevant cognition: poverty induces stress and negative affect and taxes attentional bandwidth, conditions under which threats loom larger (Haushofer and Fehr 2014). And sustained existential security shifts value priorities across generations, from survival toward self‐expression values (Inglehart and Baker 2000). These mechanisms differ in pathway but share an implication: across development contexts, the composition of concern should vary, not only its volume. They are not separately identified in our design, and we treat them as a family of accounts consistent with—rather than demonstrated by—the patterns we report.

2.2. A Finite Worry Budget Allocated by Need Priority

Two further ingredients turn that implication into an ordering prediction. The first is a constraint. Worry behaves as a limited resource: attention devoted to one class of risks crowds out concern for others, the “finite pool of worry” (Weber 2006; Sisco et al. 2023). Cross‐hazard worry must therefore be allocated, and the analytically interesting question is what governs the allocation.

The second ingredient is a priority rule. We draw on need‐hierarchy theory in its contemporary, evolutionarily grounded form, and present it as one theoretically motivated ordering of hazards to be evaluated against a simpler level‐shift account, not as an established stage theory. Maslow (1943) proposed that human motives are organized by prepotency: physiological and safety needs dominate attention until satisfied, after which higher order motives gain force. The modern renovation replaces the experiential pyramid with a set of evolved motivational systems—self‐protection, disease avoidance, affiliation, status, mate acquisition and retention, kin care—whose current activation depends on cues of threat and resource availability in the immediate environment (Kenrick et al. 2010; Neel et al. 2016). On this reading, the hierarchy is a priority schedule for threat vigilance: motivational systems tied to unmet or insecure needs capture the worry budget first. Global evidence is consistent with the ordering: across 123 countries, basic needs tend to be fulfilled in sequence with national wealth, even as each need contributes separately to well‐being (Tay and Diener 2011). The most influential account of value change in the development debate, Inglehart's scarcity hypothesis, was itself built explicitly on Maslow; returning to the need‐hierarchical source with person‐centered methods reconnects the levels debate to its theoretical foundation (Inglehart 1995; Inglehart and Baker 2000).

One measurement clarification is essential. The World Risk Poll measures worry about harms, not need fulfillment or aspiration. Our prediction therefore concerns threat‐vigilance allocation—which hazard domains occupy the finite worry budget—and not the satisfaction of needs as such. Hazards map onto the priority schedule by the needs they threaten: unsafe food and water threaten immediate physiological existence, while violent crime, severe weather, and traffic threaten physical safety. Worry about mental health does not map cleanly onto this physiological‐to‐safety ordering; we treat it as a cross‐cutting domain, plausibly reflecting psychological or existential security, and we do not use it to define the need‐priority sequence. Whether the configuration of interest depends on the mental‐health item is therefore an empirical question, which we examine with a five‐item robustness analysis (Section 4.2; Supporting Information, Section S2).

2.3. Hypotheses

The constraint‐plus‐priority architecture yields three hypotheses, stated at the levels just distinguished.

H1 (configuration; individual level). Cross‐national responses to the worry battery are not ordered by overall level alone: at least one well‐populated latent profile differs in shape, combining low worry about physiological‐need hazards (food, water) with high worry about safety‐need hazards (crime, severe weather, traffic). The hypothesis concerns the existence and replicability of this configuration within the six‐hazard battery, not the absolute number of worry profiles, which is necessarily bounded by the items available.

H2 (prevalence across contexts). The prevalence of these individual configurations varies systematically with development context. Because the secure‐basics configuration requires low worry about food and water alongside persistent safety worry, it should be most common where a substantial subgroup perceives basic provisions as adequate, while safety hazards remain salient—a condition met for sizeable subgroups across the lower‐to‐middle HDI range, not only at a particular national stage. We therefore predict that secure‐basics prevalence is elevated across low and middle HDI and declines at high and very‐high HDI, where low‐worry configurations predominate, while pervasive‐worry prevalence declines with HDI. This prevalence ordering is the distinctive prediction of the need‐priority account: a pure level‐shift account predicts uniform changes in worry, not an elevated intermediate configuration. We state the prediction as the prevalence of an individual profile across contexts, not as a claim that nations sequentially secure needs.

H3 (beyond reported harm). The association between development context and profile prevalence persists after adjustment for self‐reported recent harm in each hazard domain. We read a surviving association as consistent with the insurance, scarcity, and value mechanisms operating beyond reported harm, while emphasizing that the design neither distinguishes among these mechanisms nor identifies HDI as the operative cause independently of the objective hazard environment and covarying regional factors. We probe the objective environment and regional clustering directly (Section 4).

Two qualifications bound these hypotheses. First, the specific four‐profile structure and the intermediate position of the secure‐basics profile were identified in exploratory analysis of the 2023 wave; the 2021 wave provides an independent test of the refined predictions, and we report it as such. Second, because contemporary HDI differences partly reflect deep historical and cultural trajectories that cluster regionally (Becker et al. 2020), the design can establish that worry‐profile prevalence tracks the development gradient, but not that HDI rather than covarying macroregional or objective‐environmental factors is the operative cause; we examine both observed objective indicators and region‐adjusted specifications to map this boundary (Section 4.5).

3. Data and Methods

3.1. Data

We use the Lloyd's Register Foundation World Risk Poll (WRP), a cross‐national survey fielded by Gallup as part of its World Poll infrastructure (Lloyd's Register Foundation 2024). Each wave interviews nationally representative samples of the population aged 15 and over—typically around 1000 respondents per country, with larger samples in some populous countries—using probability sampling, by telephone or face to face depending on country. The same risk module has been fielded in 2019, 2021, and 2023. Because the worry battery used here is identical in wording and response format in the 2021 and 2023 waves, we use the 2023 wave (146,910 respondents in 142 countries) as the primary sample and the 2021 wave (125,911 respondents; 121 countries) as an independent replication. The 2019 wave is excluded: its worry battery did not include the traffic item—it fielded electrical power lines and household appliances instead—and its experience‐of‐harm module asked about personal experience only rather than personal or known‐other experience, so neither the profile measurement model nor the exposure adjustment is comparable.

The analysis sample retains respondents with valid responses on all six worry items and a matchable HDI value: $n=139,111$ (94.7% of the wave) in 140 countries for 2023, and $n=118,703$ (94.3%) in 119 countries for 2021. Item nonresponse is low (don't‐know and refusal rates of at most 1.4% per item). Taiwan and Kosovo are excluded because the UNDP does not publish HDI values for them. For the covariate‐adjusted models, additional exclusions for missing covariates leave $n=135,321$ (97.3% of the 2023 analysis sample).

3.2. Measures

3.2.1. Worry Profiles

The WRP asks, for each of several everyday hazards, how worried the respondent is that the hazard “could cause serious harm” to them: the food they eat, the water they drink, violent crime, severe weather events, traffic or roadside accidents, and mental health issues, each rated very worried, somewhat worried, or not worried. These six items form the profile indicators. A seventh item on workplace harm is asked only of respondents in work and is therefore excluded from profile estimation. The battery measures worry—an affective, self‐relevant state—rather than the full risk‐perception construct of the psychometric tradition, which also encompasses probability and severity judgments, dread, and perceived controllability (Fischhoff et al. 1978; Slovic 1987). Worry correlates with, but is not identical to, perceived risk (Slovic et al. 2007; Sjöberg 2000); we therefore speak of worry profiles throughout and treat their relation to the other components of risk perception as an open question taken up in Section 5.

3.2.2. Experience of Harm

A parallel battery asks whether, in the past two years, the respondent or someone they know experienced serious harm from each of the same six domains. We code one indicator per domain equal to one if the respondent reports any experience (personal, known other, or both) and zero otherwise.

3.2.3. Human Development Index

Country‐level HDI values are taken from the UNDP Human Development Report time series (United Nations Development Programme 2025) and matched to the survey year of each wave: the 2023 wave uses the 2023 index and the 2021 wave the 2021 index, which incorporates the UNDP's pandemic‐era revisions. In regression models, HDI is standardized so that one unit equals one standard deviation of the cross‐country HDI distribution. For descriptive comparisons, we use the UNDP tiers (low $<0.550$; medium 0.550–0.699; high 0.700–0.799; very high $\ge 0.800$), applied to the vintage‐matched value with boundary values assigned to the higher tier; they classify the 2023 analysis countries into 19 low, 27 medium, 33 high, and 61 very‐high‐HDI countries.

3.2.4. Covariates

Individual‐level controls are gender, age group (15–29, 30–49, 50, and over), education (elementary or less, secondary, tertiary), within‐country income quintile, and urbanicity (rural, small town, large city, suburb). A robustness specification adds indicators for the 15 WRP global regions.

3.3. Analytic Strategy

3.3.1. Latent Class Model

We estimate latent class models treating the six three‐category worry items as unordered categorical indicators with class‐specific multinomial response probabilities. Estimation uses expectation‐maximization weighted by the within‐country survey weights, with 10 random starts per specification and the best solution retained; we confirmed that the selected four‐class solution is not a local optimum by reestimating it from 200 random starts and by external validation against the stepmix package, and we screened item pairs for residual local dependence (Section 4.1; Supporting Information, Section S2). We estimate models with one to six classes. Following standard practice, we report the Bayesian information criterion (BIC), relative entropy, smallest class share, and mean posterior probability for each solution (Nylund‐Gibson and Choi 2018). At this sample size, the Bayesian information criterion declines monotonically with the number of classes and is therefore uninformative on its own; we select the number of classes on the substantive criterion of configurational distinctness—the smallest $K$ at which a class differs from the others in the shape of its response profile rather than only in level—which yields $K=4$, and we report $K=3$ and $K=5$ solutions as sensitivity analyses. Classification uncertainty is carried forward rather than discarded, in two ways. The descriptive and country‐level analyses (class profiles, tier prevalence, country shares, and maps) weight each respondent by their posterior class probabilities and assign no one to a single class. The structural models instead use the bias‐adjusted three‐step estimator described below, which takes the modal class assignment as an error‐prone indicator and corrects it with the misclassification rates implied by the measurement model.

3.3.2. Dimensional Alternative

Because a small set of ordinal items can be fitted equally well by a few correlated continuous dimensions as by discrete classes, we test the typological representation directly against dimensional alternatives: unidimensional and two‐factor graded response models, in exploratory and confirmatory (provision/safety) form, estimated by marginal maximum likelihood on the identical survey‐weighted sample and compared with the four‐class model on the same likelihood scale by BIC. As additional specification checks, we quantify residual local dependence for all 15 item pairs at K = 3–5 and reestimate the model (i) with the food–water pair replaced by its joint nine‐category indicator or a single collapsed provisions item, (ii) with items dichotomized as very worried versus not, removing the response midpoint to which cross‐culturally variable response styles attach (Baumgartner and Steenkamp 2001; Harzing 2006), and (iii) with survey weights rescaled so that every country contributes equal total weight to the measurement likelihood. Results of all comparisons are reported in Section 4 and the Supporting Information.

3.3.3. Cross‐Wave Replication and Measurement Equivalence

The full model is reestimated independently on the 2021 wave, and profiles are matched across waves by optimal assignment on Tucker's congruence coefficient. Because the 2021 wave predates the exploratory 2023 analysis, “replication” here denotes independent reestimation on an earlier, partly different sample—a strong robustness check—rather than out‐of‐sample prediction on new data. Because the surveyed country rosters also differ across waves, we repeat the congruence and gradient comparisons on the 118 countries present in both waves. To assess measurement equivalence across development levels, we compare multigroup specifications on the 2023 wave—complete homogeneity, partial homogeneity (invariant response probabilities with tier‐specific class proportions), and full heterogeneity—and complement the information‐criterion comparison with congruence between tier‐specific and pooled response profiles (Eid et al. 2003).

3.3.4. Profile Prevalence and HDI

Because HDI is a country‐level attribute, inference on the HDI‐profile association treats the country as the unit of analysis (Bryan and Jenkins 2015). We compute survey‐weighted, posterior‐probability‐weighted class shares per country and compare them across HDI tiers using Kruskal–Wallis tests with the epsilon‐squared effect size, and against continuous HDI using Spearman rank correlations, with Bonferroni correction across tests.

3.3.5. Experience‐Adjusted Models

To test whether the HDI gradient reflects more than differential exposure, we estimate multinomial logistic regressions of class membership in a three‐step framework: the measurement model is estimated first, and class membership then enters the structural model with its classification uncertainty corrected. Our primary estimator is the bias‐adjusted maximum‐likelihood three‐step method (Vermunt 2010; Bakk and Vermunt 2016), which treats the modal class assignment as an error‐prone indicator whose misclassification rates are fixed at the values implied by the measurement model. We also report the uncorrected proportional‐assignment estimator—each respondent entering as four pseudo‐observations weighted by posterior probability times survey weight (Vermunt 2003; Henry and Muthén 2010)—whose attenuation relative to the corrected estimates confirms that it is conservative. Four nested specifications add, in order: experience‐of‐harm indicators; demographic covariates; and global‐region fixed effects. We summarize each model by the average marginal effect of a one‐standard‐deviation difference in HDI on the probability of each class, with 95% confidence intervals from a cluster bootstrap that resamples countries (2,000 replications); each replication reestimates the structural multinomial model while holding the latent class measurement model, and hence the misclassification rates, fixed at the full‐sample solution. As a transparent complement, we regress country‐level class shares on standardized HDI and the country mean of the experience count using ordinary least squares with heteroskedasticity‐robust standard errors.

3.3.6. Objective Hazard Environment

To ask whether the gradient reflects the measurable hazard environment rather than development as such, we merge country‐level objective indicators corresponding to the worry domains—prevalence of undernourishment, intentional homicide rate, road‐traffic mortality, and population exposure to droughts, floods, and extreme temperatures (most recent available value per country, World Bank World Development Indicators)—onto the country‐level class shares. We regress each class share on standardized HDI alone, then adding the standardized objective indicators, and then adding world‐region fixed effects, with heteroskedasticity‐robust standard errors and the country as the unit throughout. A safely managed drinking‐water indicator with sparser coverage is added in a five‐indicator robustness specification reported in the Supporting Information.

3.3.7. Individual Provision (2021 Wave)

The most direct rival to a need‐priority reading of the secure‐basics profile is compositional: secure household provision combined with genuinely hazardous surroundings would produce the same configuration by exposure realism alone. The 2023 public file carries no individual provision measures, but the 2021 wave includes a deprivation module recording whether the respondent went without food, clean drinking water, electricity, medical care, or a telephone for more than a day in the past year. We therefore reestimate the covariate‐adjusted model on the 2021 wave adding the food and water provision items (and, in a further specification, all five module items), asking whether the HDI gradient on each profile survives adjustment for individual provision.

3.3.8. Weighting and Software

Within each country, the WRP poststratification weights are used as provided: latent class estimation weights every respondent's contribution by these weights, and country‐level class shares are the posterior‐probability‐weighted means of the same weights. Across countries, each country then enters every cross‐country analysis with unit weight regardless of its sample size, so that estimands refer to the average country rather than the average person in the world; no country‐size weighting enters the latent class model, whose likelihood is nonetheless dominated by the larger national samples and is reported as such. All analyses were conducted in Python 3.11 using a purpose‐written EM implementation and standard scientific libraries, except the dimensional comparison, in which the graded response models and the four‐class model refit on the same likelihood scale were estimated in R with the mirt package; analysis code and derived data are available in an anonymized repository for peer review (see the Data Availability Statement). The study is a secondary analysis of anonymized, publicly available survey data; ethical clearance for the original data collection, including interviews with respondents aged 15–17, rests with Gallup and the Lloyd's Register Foundation.

4. Results

Item‐level descriptives reproduce, with HDI, the pattern previously reported for income groups: the share of respondents very worried declines monotonically with HDI tier for every hazard. In low‐HDI countries, 34.0% are very worried about the safety of their food and 30.6% about their drinking water, against 15.0% and 12.5% in very‐high‐HDI countries; worry about traffic accidents falls from 54.7% to 27.7%, and worry about mental health from 42.1% to 16.8%. Taken alone, these level differences are compatible with a simple affluence account. The latent class analysis that follows shows that they conceal a more structured reallocation of worry across hazards.

4.1. Four Worry Profiles (H1)

Model fit statistics for one to six classes are reported in Table 1 (and, for both waves with the diagnostic plot, in Table S1 and Figure S1). As anticipated in Section 3, the BIC declines monotonically across this range at $n=139,111$ and cannot by itself select a solution; relative entropy is highest at $K=2$ (0.741) and declines gradually thereafter (0.689 at $K=4$), as does the survey‐weighted mean maximum posterior probability (0.828 at $K=4$). The two‐ and three‐class solutions separate respondents only by overall worry level. The four‐class solution is the first in which a class differs in profile shape—the substantive criterion set out in Section 3—and we retain it on that ground alone, not on any fit statistic; the three‐ and five‐class solutions, which bracket it, are examined as sensitivity analyses in Section 4.2 and the Supporting Information, and our conclusions are restricted to solutions of this battery rather than to the number of worry profiles in any absolute sense. The selected solution is stable across 200 random starts and under independent reestimation with the stepmix package, and the focal class does not depend on the residual association between the food and water items, the strongest local dependence in the battery: reestimating the model with that dependence freed, or with the two items collapsed into a single provisions indicator, preserves the secure‐basics configuration and its development gradient (Supporting Information, Section S2 and Tables S17– S18).

TABLE 1.

Latent class enumeration, World Risk Poll 2023 ($n=139,111$; six worry items, survey‐weighted).

$K$	Log‐likelihood	BIC	Relative entropy	Smallest class (%)	Mean posterior
1	$-886,164$	$1,772,471$	1.000	100.0	1.000
2	$-813,135$	$1,626,565$	0.741	44.1	0.925
3	$-790,827$	$1,582,103$	0.694	30.9	0.859
4	$-785,341$	$1,571,286$	0.689	15.0	0.828
5	$-780,539$	$1,561,837$	0.656	14.6	0.782
6	$-777,375$	$1,555,662$	0.645	11.7	0.763

Note: BIC declines monotonically, so the number of classes is selected on the substantive criterion of configurational distinctness (Section 3); $K=4$ is the smallest solution in which a class differs in profile shape rather than only in level. “Smallest class” is the smallest class share; “mean posterior” is the survey‐weighted mean of each respondent's largest posterior class probability.

Table 2 displays the four profiles. Two classes are level classes: a pervasive high worry class (20.3% of the sample) with high intense worry across all six hazards, including food (69.1% very worried) and water (66.6%), and a low worry class (33.9%) with uniformly minimal concern. A moderate class (30.9%) sits between them, again with a roughly proportional profile. The fourth class is the configurational one predicted by H1: a secure‐basics class (15.0%) in which intense worry about food (14.0%) and water (7.4%) is nearly as low as in the low‐worry class, while worry about violent crime (64.0%), severe weather (72.2%), and traffic accidents (83.0%) approaches the levels of the pervasive class. The full three‐category probabilities (Table S6) make the shape explicit: in the secure‐basics class most respondents are “not worried” about food (64.2%) and water (79.1%) but “very worried” about crime, severe weather, and traffic, so the profile reflects movement into the lowest food and water category rather than a uniform downward shift. Worry, in this profile, is reallocated rather than reduced: the physiological‐need hazards have dropped out of the worry budget, while the safety‐need hazards remain high. This is the configuration that a need‐priority account predicts and that level accounts do not. The countries in which this profile is modal are identified in Figure 1, which places each country in the two‐dimensional worry space introduced in Section 4.2; all of them sit above the diagonal of equal provision and safety worry.

TABLE 2.

Worry profiles in the four‐class solution, World Risk Poll 2023 ($n=139,111$; 140 countries).

Class	Share	Food	Water	Crime	Weather	Traffic	Mental
Pervasive high worry	20.3	69.1	66.6	82.0	80.3	81.5	62.5
Secure‐basics	15.0	14.0	7.4	64.0	72.2	83.0	51.1
Moderate	30.9	12.3	11.0	26.3	24.2	26.1	10.3
Low worry	33.9	5.5	4.7	7.2	7.7	9.8	3.5

Note: Cells are the probability (%) of being very worried about each hazard given class membership; estimation is survey‐weighted. Class shares are model class proportions and may not sum to exactly 100% because of rounding.

FIGURE 1.

Countries in the two‐dimensional worry space implied by the six‐item battery, World Risk Poll 2023 (140 countries). Axes are survey‐weighted country means of factor scores from the confirmatory two‐factor graded response model (provision: food, water; safety: crime, severe weather, traffic, mental health), oriented so higher values indicate more worry; the dashed line marks equal worry on both dimensions. Colors give HDI tiers. Annotated countries (Botswana, Indonesia, Mauritius, Morocco, Namibia, Senegal, Tanzania) are those in which the secure‐basics profile is modal; all lie above the diagonal.

4.2. Replication and Robustness of the Typology

The same model estimated independently on the 2021 wave ($n=118,703$; 119 countries) recovers the four profiles almost exactly (Figure 2). Matching classes across waves by optimal assignment, Tucker congruence coefficients are 0.9999 (pervasive), 0.9995 (secure‐basics), 0.9990 (moderate), and 0.9939 (low worry)—all above the conventional 0.95 equivalence benchmark—with class shares within two percentage points and item‐level response probabilities differing by 1.1–1.5 points on average. The secure‐basics signature reproduces at an identical 15.0% share. The 2021 fieldwork, moreover, overlapped the COVID‐19 pandemic, which makes the near‐exact reproduction of the structure across waves more striking rather than less. Because the surveyed rosters differ across waves, we repeated the comparison on the 118 countries present in both, with the same result ($\varphi \ge 0.993$; Tables S20–S21).

FIGURE 2.

Cross‐wave replication of the four worry profiles. Each panel overlays the class‐conditional probabilities of being very worried estimated independently in the 2021 wave ($n=118,703$; 119 countries; dashed) and the 2023 wave ($n=139,111$; 140 countries; solid), annotated with the Tucker congruence coefficient $\varphi$ per matched class ($\varphi \ge 0.95$ denotes profile equivalence). The configurationally distinct secure‐basics profile reproduces at an identical 15.0% share in both waves.

The typology is also robust to specification; the checks are reported in full in Section S2 of the Supporting Information. The four shapes persist across three‐ to five‐class solutions, with the secure‐basics class surviving at $K=5$ and absorbed into the level classes only at $K=3$ (Table S4). Measurement is approximately equivalent across development levels: reestimating the model freely within each HDI tier recovers the same four shapes in every tier, with tier‐specific and pooled profiles congruent at $\varphi \ge 0.9875$ in 15 of 16 comparisons (Tables S2–S3), and the prevalence gradient reported below is preserved in sign and ordering when class shares are reestimated with tier‐specific response probabilities, so it is not an artifact of cross‐tier differences in response style. Finally, the secure‐basics configuration survives dropping the mental‐health item, which accompanies the profile but does not define it (Tables S7–S8), dichotomizing items as very worried versus not (Table S19), and rescaling weights so that every country contributes equally to the measurement likelihood (Table S22).

A deeper specification question is whether discrete classes are the right representation at all: with six ordinal items, a small number of correlated continuous dimensions could generate similar response patterns. We therefore fitted unidimensional and two‐factor graded response models—exploratory and confirmatory—on the identical survey‐weighted sample and compared them with the four‐class model on the same likelihood scale. The class model is preferred by BIC over every dimensional alternative (margin of roughly 3000 over the best two‐factor model; Tables S15–S16). The exploratory solution nonetheless recovers exactly the two correlated dimensions the configural class implies—a provision factor (food, water) and a safety factor (crime, severe weather, traffic, mental health), correlated at 0.62—and the two representations are complementary: Figure 1 locates each country on the two dimensions and shows both that development sorts countries along the joint diagonal and that the countries where the secure‐basics profile is modal sit distinctly above it, reporting more safety worry than their provision worry would predict.

4.3. Profile Prevalence Along the HDI Gradient (H2)

Figure 3 and Table 3 relate country‐level profile prevalence to HDI tier, with the country as the unit of inference. All four prevalences differ across tiers (Kruskal–Wallis $p<4\times {10}^{-5}$ throughout; all eight tests below survive Bonferroni correction), with large effect sizes (${\epsilon }^2$ between 0.15 and 0.32). The pervasive‐worry share declines monotonically from 32.7% of the population in the average low‐HDI country to 12.8% in the average very‐high‐HDI country (Spearman $\rho =-0.62$ with continuous HDI), while the low‐worry share rises from 22.2% to 42.5% ($\rho =+0.54$).

FIGURE 3.

Worry‐profile prevalence by HDI tier, World Risk Poll 2023. Bars show mean country‐level class shares with 95% confidence intervals of the tier mean (country as unit; 19, 27, 33, and 61 countries per tier). Pervasive worry declines and low worry rises monotonically with HDI; the secure‐basics profile remains at its maximum across low and medium HDI and falls only at very high HDI, the ordered pattern predicted by the need‐priority account (H2).

TABLE 3.

Profile prevalence by HDI tier, World Risk Poll 2023.

	HDI tier
Class	Low	Medium	High	Very high	${\epsilon }^2$	${\rho }_{\text{HDI}}$
Pervasive high worry	32.7 (3.2)	25.0 (1.8)	23.3 (2.1)	12.8 (1.3)	0.30	−.62
Secure‐basics	19.7 (2.0)	20.2 (1.1)	16.0 (1.2)	10.7 (0.7)	0.32	−.59
Moderate	25.4 (2.0)	26.1 (1.3)	30.9 (1.5)	34.0 (1.0)	0.15	+.37
Low worry	22.2 (3.0)	28.6 (2.2)	29.8 (2.5)	42.5 (2.0)	0.22	+.54

Note: Entries are mean country‐level class shares (%) with standard errors in parentheses; shares are survey‐weighted within country and posterior‐probability‐weighted across classes. Countries per tier: low 19, medium 27, high 33, very high 61. The final columns report Kruskal–Wallis tests across tiers and Spearman correlations with continuous HDI ($n=140$ countries). Every test is significant at $p<4\times {10}^{-5}$ and survives Bonferroni correction across the eight tests; exact test statistics for both waves are reported in Table S11.

The sequencing prediction of H2 concerns the secure‐basics profile, and the data follow it: prevalence is at its maximum and nearly identical across low‐ and medium‐HDI tiers (19.7% and 20.2%), declines modestly at high HDI (16.0%), and drops sharply only in the very‐high tier (10.7%; $\rho =-0.59$ overall). The profile that combines low food and water worry with active safety concern is thus most common across the lower‐to‐middle HDI range—where a substantial subgroup perceives basic provisions as adequate while safety hazards remain salient—rather than accumulating monotonically with development, as a simple affluence account of “worry relief” would suggest. The 2021 wave reproduces every gradient with the same signs and somewhat larger magnitudes (pervasive worry: 39.3% to 11.6%, $\rho =-0.72$; low worry: 15.7% to 40.9%, $\rho =+0.61$; secure‐basics $\rho =-0.69$; Table S5), providing the independent test of the sequencing prediction announced in Section 2. Part of the cross‐wave difference in magnitude reflects the different country rosters (the 2021 low‐HDI tier contains 10 countries against 19 in 2023); on the common 118‐country roster the gradients agree closely across waves (Table S21).

The geography of these gradients is mapped in the Supporting Information (Figure S2), which colors each country by its dominant (plurality) profile: pervasive worry is modal across much of Latin America and sub‐Saharan Africa, the moderate profile across a band spanning Eastern Europe and Central, South, and Southeast Asia, and low worry throughout North America, Western Europe, East Asia, and Oceania, while the secure‐basics profile is modal in seven countries spanning all four HDI tiers—Senegal (39.5% of respondents), Indonesia, Mauritius, Botswana, Namibia, Tanzania, and Morocco (29.6%)—four of which (Senegal, Indonesia, Mauritius, Morocco) are also modal in 2021 (Table S14). The list leans toward Africa, a regional concentration we return to when weighing compositional explanations in Section 4.4 and Section 5. Because profiles are within‐country mixtures rather than country types (median dominant share 41%), the substantive importance of the secure‐basics profile lies in its within‐country share across the low and medium tiers (Table 3) rather than in where it ranks first.

4.4. Beyond Self‐Reported Harm (H3)

Do these gradients merely restate that people in less developed countries experience more harm? Table 4 reports average marginal effects of a one‐standard‐deviation difference in HDI on the probability of each profile, from the three‐step multinomial models ($n=135,321$; ML‐corrected estimates, with the conservative proportional‐assignment estimates in Table S23). With no controls, one standard deviation of HDI is associated with an 8.0‐percentage‐point lower probability of pervasive worry. Adjusting for self‐reported harm in all six domains attenuates this effect by roughly a third (to $-5.2$ points), and adding demographic composition leaves it at $-5.6$ points (95% country‐cluster bootstrap CI $[-7.7,-3.4]$); the intervals for all four classes exclude zero in the adjusted model. Experience matters in its own right—in country‐level regressions, each additional harm domain experienced is associated with a 7.4‐point higher pervasive‐worry share ($p<.001$)—but HDI retains an independent association of $-4.6$ points per standard deviation ($p<.001$) alongside it. The secure‐basics profile is the least exposure‐dependent of the four: its HDI effect is essentially unchanged by experience adjustment ($-5.6$ to $-5.2$ points across M0–M2), and country‐level experience does not predict its prevalence ($p=0.27$), consistent with its interpretation as a configuration of worry rather than an exposure residue.

TABLE 4.

Average marginal effect (percentage points) of a one‐standard‐deviation difference in HDI on the probability of each worry profile, bias‐adjusted (ML) three‐step multinomial models, World Risk Poll 2023 ($n=135,321$; 140 countries).

	M0	M1	M2 (main)	M3
	(HDI only)	(+ experience)	(+ demographics)	(+ region FE)
Pervasive high worry	$-8.0$	$-5.2$	$-5.6$ $[-7.7,-3.4]$	$+0.3$
Secure‐basics	$-5.6$	$-6.2$	$-5.2$ $[-6.7,-3.9]$	$+0.2$
Moderate	$+3.3$	$+4.5$	$+2.5$ $[+0.2,+4.8]$	$-3.9$
Low worry	$+10.3$	$+6.9$	$+8.3$ $[+5.1,+11.6]$	$+3.4$

Note: M1 adds six experience‐of‐harm indicators; M2 adds gender, age group, education, income quintile, and urbanicity; M3 adds world‐region fixed effects. Brackets give 95% percentile confidence intervals from a country‐cluster bootstrap (2000 replications) for the main specification (M2). The uncorrected proportional‐assignment estimates, uniformly closer to zero, are reported in Table S23.

Self‐reported harm does not, however, capture whether basic provision is actually secured, which the compositional rival to H3 requires (Section 3). The 2021 deprivation module supplies that test. Adding indicators for having gone without food (weighted prevalence 17.0%) or clean drinking water (29.2%) to the 2021 covariate model leaves the secure‐basics gradient essentially unchanged—$-6.0$ points per standard deviation of HDI, 95% bootstrap CI $[-7.7,-4.5]$, against $-6.0$ without the provision terms—while the pervasive‐worry gradient attenuates from $-7.6$ to $-5.7$ (and to $-5.3$ with all five deprivation items). The provision items themselves behave exactly as an allocation account expects: going without food or water sharply raises the probability of the pervasive profile ($+8.5$ and $+7.8$ points) but leaves secure‐basics membership essentially unchanged ($+0.5$ and $-1.5$ points). Deprivation raises overall worry, as an allocation account expects; the measured deprivation indicators, however, do not explain away the configural profile or its HDI gradient (Tables S25–S27). Because these items are informative but imperfect measures of household provision and exposure, this weakens the compositional account without excluding it.

The final specification adds fixed effects for the 15 world regions and marks the boundary of what this design can identify: the HDI effects on the pervasive and secure‐basics profiles shrink to near zero. HDI varies primarily between regions, so region fixed effects absorb the development gradient together with all other region‐level differences—including the deep cultural and historical trajectories that covary with development. We read this as a scope condition rather than a refutation: the data establish a robust, harm‐adjusted association between development level and the organization of worry, but cannot apportion it between HDI and the macroregional history in which HDI is embedded. We return to this in Section 5.

4.5. The Objective Hazard Environment (H3, Continued)

Self‐reported harm is itself perception‐laden. To ask whether the gradient instead reflects the measurable hazard environment, we merged country‐level indicators of the objective conditions behind the worry domains—prevalence of undernourishment (food), intentional homicide (crime), road‐traffic mortality (traffic), and population exposure to droughts, floods, and extreme temperatures (weather)—from World Bank sources, and reestimated the gradient with the country as the unit of analysis ($n=116$ countries with complete indicators). Here, the country‐level class‐share regressions are the primary estimand; the respondent‐level models of Table 4 are the individual‐level complement, and the two agree in sign and ordering.

Two patterns sharpen the interpretation. First, the objective safety environment carries the expected signal: a one‐standard‐deviation higher homicide rate is associated with a 3.1‐point higher prevalence of pervasive worry ($p<0.001$) and a 2.6‐point lower prevalence of low worry ($p=0.01$), and higher road‐traffic mortality is associated with a 2.0‐point higher secure‐basics prevalence ($p=0.02$). Second, and in contrast, objective food security does not predict profile prevalence once HDI is held constant (undernourishment $p=0.41$ for secure‐basics, $p=0.80$ for pervasive worry). The absence of a country‐level food signal is consistent with the three‐level reading of Section 2: the secure‐basics profile is an individual configuration whose prevalence varies across contexts, not a marker of nationally secured food and water.

Table 5 nests the development gradient in three steps. Adjusting for the objective hazard environment attenuates the HDI association but does not remove it: the gradient for pervasive worry falls from $-8.2$ to $-5.8$ points per standard deviation (29% of the association accounted for) and that for low worry from $+9.2$ to $+8.0$ (13%), both still distinguishable from zero ($p<0.001$). Adding region fixed effects instead absorbs most to all of the gradient—the remaining point estimates range from roughly a quarter of the unadjusted association (pervasive worry, $-2.6$; low worry, $+2.2$) to a slight sign reversal for secure basics—and renders every profile's gradient statistically indistinguishable from zero, reproducing at the country level the boundary already seen in Table 4. This dependence on between‐region variation is not driven by any single region (leave‐one‐region‐out and within‐region slopes, Table S9), and a five‐indicator specification adding safely managed drinking water, available for fewer countries, yields the same pattern (Table S10). Development is thus associated with the architecture of worry over and above the measured objective hazard environment, while remaining empirically inseparable from macroregional clustering.

TABLE 5.

Association (percentage points of class share per one standard deviation of HDI) between development and worry‐profile prevalence, nested in three country‐level specifications, World Risk Poll 2023 ($n=116$ countries with complete objective indicators).

	HDI only	+ objective env.	+ region FE
Pervasive high worry	$-8.2$	$-5.8^{\ast \ast \ast }$	$-2.6$
Secure‐basics	$-4.7$	$-2.5^{†}$	$+0.4$
Moderate	$+3.7$	$+0.3$	$+0.0$
Low worry	$+9.2$	$+8.0^{\ast \ast \ast }$	$+2.2$

Note: “+ objective environment” adds undernourishment, homicide, road‐traffic mortality, and disaster exposure; “+ region FE” adds fixed effects for world region. Because of the complete‐case restriction, the HDI‐only coefficients differ slightly from the 140‐country slopes reported in Table S9. Significance of the HDI coefficient: , .0

5. Discussion

This article asked whether human development is associated with the organization of everyday worry rather than only its amount. Across two independent global samples, the answer has three parts. Worry responses are well captured by four profiles that include a configurationally distinct secure‐basics pattern, in which concern about food and water has receded while concern about crime, severe weather, and traffic remains high (H1). Profile prevalence is consistent with the ordering predicted by a need‐priority account: pervasive worry declines with HDI, the secure‐basics profile is most common at low and medium HDI and least common in the very‐high tier, and low worry becomes the modal profile only at high levels of development (H2). And the development gradient is not fully accounted for by differential exposure: it persists after adjustment for self‐reported recent harm in all six domains, with the secure‐basics profile the least harm‐dependent of the four, and is only partly accounted for by the measured objective hazard environment (H3).

The long‐running disagreement over whether concern about risk is a rich‐country or a poor‐country phenomenon (Inglehart 1995; Dunlap and York 2008; Franzen and Meyer 2010) looks different from a person‐centered vantage point. Both sides are right about levels—intense worry about almost every hazard is more common in less developed countries, yet substantial concern exists everywhere—because levels are not where development leaves its clearest mark. Development is associated with a reallocation of a finite worry budget (Weber 2006): hazards exit the budget in roughly the order in which the needs they threaten are secured (Maslow 1943; Kenrick et al. 2010; Tay and Diener 2011). The intermediate secure‐basics configuration, invisible to item‐by‐item comparisons, is the empirical signature of that sequence. It is worth noting that Inglehart's scarcity hypothesis was built on precisely this hierarchical intuition; our results return the debate to that foundation with data its original protagonists did not have.

The results also sharpen a puzzle inherited from the economics of risk. Behavioral and survey measures of risk preference find populations of poorer countries more risk tolerant (L'Haridon and Vieider 2019; Bouchouicha and Vieider 2019; Mata et al. 2016), while we find them more worried. Read jointly, the two literatures suggest that worry and risk‐taking are governed by different machinery: worry appears to track unsecured needs, whereas preference measures may reflect hardship‐calibrated valuation of gains and losses (Falk et al. 2018). A person‐centered design that elicits both perception and preference from the same respondents would test this directly; the World Risk Poll and the Global Preference Survey, both fielded on Gallup infrastructure, make such an integration feasible.

Our design cannot adjudicate among the insurance, scarcity, and value‐change mechanisms of Section 2; all three are consistent with a harm‐adjusted development gradient. Two boundaries deserve emphasis. First, self‐reported harm is not a rival to be dismissed: it independently predicts pervasive worry at both the individual and country level, and the objective hazard environment accounts for part of the gradient in turn. The claim supported by the data is that HDI is associated with worry allocation over and above self‐reported harm and the measured objective environment, not instead of them. The most parsimonious rival for the secure‐basics profile itself deserves separate naming: it is compositional. A respondent with piped water and adequate food in a high‐crime, high‐road‐mortality city—a common situation in much of urban Africa, Latin America, and Southeast Asia, and consistent with the Africa‐leaning list of countries where the profile is modal—would produce this configuration by exposure realism alone, with no allocation psychology required. The 2021 deprivation module provides a direct check: adjusting for whether the respondent actually went without food or clean drinking water leaves the HDI gradient on the secure‐basics profile essentially unchanged, while the pervasive‐worry gradient attenuates (Section 4.4). The measured provision items therefore account for part of overall worry, as they should, but do not explain away the development patterning of the secure‐basics configuration—the outcome a purely compositional account predicts they would explain first. Second, the region‐adjusted models mark the limit of cross‐sectional identification. Because HDI varies mainly between world regions, region fixed effects absorb the development gradient together with everything else that distinguishes regions, including the deep historical trajectories known to shape preferences (Becker et al. 2020). The data therefore establish a robust association between development level and the architecture of worry, but cannot apportion it between contemporary development and the regional history in which development is embedded. Distinguishing the two requires within‐country change: as further WRP waves accumulate, countries moving across HDI thresholds will provide that test.

Five limitations qualify the findings. First, the data are repeated cross‐sections; no individual is followed over time, and profile‐sequence claims are strictly between‐country. Second, cross‐cultural measurement deserves particular caution. Comparing subjective reports across more than a hundred countries assumes that “worried” survives translation, that willingness to report worry does not itself vary with development, and that referents such as “severe weather events” mean comparable things in different climates (Deaton 2008; Boholm 1998). Midpoint and extreme response styles are known to vary across cultures and with education—and hence potentially with HDI (Baumgartner and Steenkamp 2001; Harzing 2006). The three‐item response format compounds the concern by limiting within‐profile granularity. Three checks bound, without eliminating, these threats: emission profiles reestimated freely within each HDI tier remain congruent with the pooled solution (Table S3); the profile structure and its gradients persist when items are dichotomized as very worried versus not, removing the midpoint on which midpoint response style operates; and the configural secure‐basics profile is defined by a contrast between domains within respondent, which a uniform reporting shift cannot produce. Third, experience of harm is self‐reported by the same respondents who report worry, so the exposure adjustment shares method variance with the outcome; it likely overadjusts rather than under‐adjusts, which would make H3 conservative. Fourth, the number of classes is selected on substantive rather than purely statistical grounds, because the BIC is uninformative at these sample sizes; the typology and all gradients are, however, stable across three‐ to five‐class solutions and across waves. Relatedly, because the model pools all countries, a worry pattern specific to a single country would surface only as noise or reduced congruence, not as a profile of its own; the profiles describe the dominant cross‐national configurations, not an exhaustive inventory. Fifth, classification is probabilistic (relative entropy 0.69); all analyses propagate that uncertainty, and the structural estimates use a bias‐adjusted three‐step estimator whose corrected gradients are, as expected, somewhat larger than the proportional‐assignment estimates also reported—but assignment quality differs across classes, and residual attenuation cannot be excluded.

The profile structure carries a practical message: the prevailing worry configuration differs by development context, and the map in the Supporting Information (Figure S2) shows where each profile is most common. The data describe worry, not efficacy, trust, institutional capacity, or protective behavior, so the implication concerns the targeting of risk communication rather than its effect; high worry can motivate protective action but can also produce fatalism or avoidance. In countries where pervasive worry is modal, concern is already abundant, and raising threat salience may add little relative to emphasizing efficacy, trust, and feasible protective action. In the low‐and‐middle‐HDI band where the secure‐basics profile is concentrated, attention has already shifted to violent crime, severe weather, and above all traffic—the hazard with the highest intense worry in every tier below very high HDI—so domain‐specific safety messages may resonate more than messages about food and water. In very‐high‐HDI countries, where low worry is modal, the first communication challenge is instead salience for hazards that have left the worry budget. Matching the message to the prevailing worry architecture, rather than to a single global template, is the practical suggestion these results support; whether such matching changes behavior is a question the present data cannot answer.

6. Conclusion

Across 140 countries and two independent survey waves, responses to the six‐hazard worry battery are well captured by a small set of replicable profiles whose prevalence tracks the HDI in the order that a need‐priority account predicts, and this association persists after adjustment for self‐reported recent harm. What differs across the development gradient is less how much people worry than what worry is about—its architecture more than its volume.

The contribution is threefold. First, the article reframes a three‐decade debate that has asked whether concern about risk rises or falls with national development: by moving from how much people worry to how worry is organized, it shows that the two camps were arguing about levels while development leaves its clearest mark on allocation. Second, it identifies and replicates a configurationally distinct secure‐basics profile—low concern about food and water alongside persistent concern about crime, severe weather, and traffic—that is invisible to the item‐by‐item comparisons on which the original debate was built, yet is consistent with a finite worry budget reallocated as material hazards become less salient. Third, to our knowledge, it supplies the first cross‐national evidence that this intermediate configuration is systematically graded by development: a four‐profile typology that is approximately measurement‐invariant across human‐development tiers and reproduces across the 2021 and 2023 waves, with a prevalence gradient that persists after adjustment for self‐reported recent harm and demographic composition and is only partly accounted for by the objective hazard environment.

Two directions follow. For theory, worry and risk preference appear to diverge across the development gradient, and the World Risk Poll and the Global Preference Survey—fielded on shared infrastructure—make a joint person‐centered design feasible for testing why. For identification, cross‐sectional data establish the association but cannot apportion it between contemporary development and the regional history in which development is embedded; as further waves accumulate, countries crossing HDI thresholds will allow the within‐country test the present design cannot provide. For practice, the profile map locates where the binding constraint on safety is the security of basic provisions, where it has already shifted to crime, weather, and traffic, and where the task is instead to create salience for hazards that have left the worry budget—an argument for matching risk communication to the prevailing architecture of worry rather than to a single global template.

Conflicts of Interest

The author declares no conflicts of interest.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not‐for‐profit sectors.

Supporting information

Supporting File: risa70393‐supp‐0001‐SuppMat.pdf

Data Availability Statement

The World Risk Poll microdata are publicly available from the Lloyd's Register Foundation (https://wrp.lrfoundation.org.uk). Human Development Index data are publicly available from the United Nations Development Programme (https://hdr.undp.org/data‐center), and the objective hazard indicators from the World Bank World Development Indicators (https://data.worldbank.org). Analysis code and derived country‐level datasets are publicly available in the Open Science Framework repository at https://doi.org/10.17605/OSF.IO/QA94P.