Projected new melanoma cases have doubled over a 23-year period.

Melanoma is a skin cancer, which usually develops on sun-exposed skin, such as the arms, back, face, and legs, but can also occur in the eyes or rarely in internal areas like the nose or throat. Most cases are linked to ultraviolet (UV) light from sunlight or tanning beds. In the U.S., sunburn prevalence is high, with 37.1% of adults and 55.8% of youth reporting at least one sunburn in the past year (1). Limiting UV exposure can help reduce the risk of melanoma (1–3).

Skin cancer also incurs remarkable direct and indirect costs of $8.9 billion annually (in addition to unmeasurable indirect costs). According to the CDC report on skin cancer in the United States, nearly 6 million people are treated for skin cancer each year (4, 5).

This study investigates the social and environmental factors contributing to melanoma incidence in the United States. Our research question is whether education can mitigate UV-related melanoma risk. The analysis explores how melanoma rates are influenced by medical literacy—proxied by population size and education levels—along with UV radiation and temporal trends. Using CDC data on non-Hispanic Caucasian males and females, the study applies quadratic Poisson regression and Two-Stage Least Squares (TSLS) over a 23-year period. The initial model uses the count of new melanoma cases as the dependent variable, with population size, state fixed effects (as proxies for UV exposure), and time as independent variables. To address overdispersion, extended models substitute the dependent variable with age-adjusted annual melanoma cases per 100,000 people and replace state dummies with direct UV index values and the share of the educated population. Building on prior research linking population size to knowledge spillovers and stronger health systems, the study finds that higher education levels—representing human capital—can mitigate environmental health risks like UV-related melanoma.

Results corroborate a strong connection between state dummies and UV radiation levels, showing that after accounting for these levels, projected new melanoma cases have doubled over the past 23 years. However, higher population and education levels are linked to fewer cases, even in high-UV states. Robustness tests reveal that the Poisson model may not be adequate due to overdispersion. Extended models confirm that a 12–16% rise in the educated population offsets long-term melanoma growth, and a 25–30% increase counters a 1% rise in UV exposure. These findings highlight the role of education in reducing melanoma risk.

The data for this study were generated, among other sources, from the CDC website and focus exclusively on non-Hispanic White males and females in the United States, as this demographic is the most susceptible to melanoma skin cancer. In 2022, the total population of non-Hispanic Whites in the United States was 197,639,521.

More specifically, the data were drawn from the following sources:

CDC: United States Cancer Statistics: Data Visualizations: Trends. Available at: https://gis.cdc.gov/Cancer/USCS/#/Trends/ (6)

The datasets and script for replications of the outcomes obtained are given in the following link: https://github.com/YuvalArbel1234/Frontiers_Melanoma_Human_Capital/releases

The combined dataset consists of 2,325 observations, each representing statewide-level data for a specific year, spanning from 1999 to 2022 (a total of 23 years). It is clear that the sample includes data from all 51 U.S. states. In 47 states, the sample contains 46 observations per state (one for Caucasian females and one for Caucasian males over 23 years). Two states have 42 observations each (spanning 21 years), one state (the District of Columbia) has 41 observations (covering 20–21 years), and one state (Mississippi) has 38 observations (covering 19 years). The total is calculated as 47 × 46 + 2 × 42 + 41 + 38 = 2 , 325 .

Table 1A reports the descriptive statistics of variables that are latter incorporated in the empirical model. Table 1B gives the covariance matrix between the variables. The average number of new melanoma cases per state is 631 per annum and the standard deviation is 702 (CaseCount). The 99% confidence interval is [594, 669]. The average population per state is 1,969,797 persons and the standard deviation is 1,737,117 persons (Population).

Referring to the correlation matrix, there is a strong positive Pearson correlation of 0.86 between the Case Count and Population variables. This indicates that the Population variable explains 86% of the standard deviation in the Case Count variable. Since the null hypothesis of zero correlation between Case Count and ( Year − 1999 ) cannot be rejected, there is also a positive time trend in the number of new melanoma cases. Lastly, as expected, there is a high Pearson correlation of 0.94 between the Population and Population squared variables.

Consider the following empirical model:

Model A: Basic Poisson Regression

In this specification, Case Count is the dependent variable. Case Count is the dependent variable, The independent variables include: Population _ sq (population squared), Population , and ( Year − 1999 ) , The parameters α 1 , α 2 , α 3 , α 4 are are to be estimated. A denotes a matrix of U.S. state dummy variables (with Alabama serving as the base category), and β T is the corresponding column vector of coefficients, and μ 1 is the random disturbance term.

Poisson regression is employed to model the counts of an event’s occurrences [e.g., (9), pp. 595–600]. Applications of the Poisson distribution include diverse events such as the pattern of hits by buzz bombs launched on London during World War II (10), the number of soldiers killed by horse kicks in the Prussian army (11), and telephone calls resulting in a wrong number (12). It has also been used to model disease incidence, most commonly over time but sometimes in spatial contexts.

According to Chiang and Wainwright (13), the general form of a quadratic function is y = a x 2 + bx + c , where ( a ≠ 0 ) , and its second derivative is 2 a . Since this second derivative always has the same sign as the coefficient a , the function forms a U-shaped curve with a global minimum at ( − b 2 a , − b 2 + 4 ac 4 a ) when a > 0, and an inverted U-shaped curve with a global maximum at the same point when a < 0.

In this study, y represents the number of annual new melanoma cases in the state. ( Case Count ) in Equation 1; x 2 corresponds to Population _ sq , and x represents Population , with coefficients a = α 1 ; b = α 2 and c = α 3 (Equation 1). Compared to the linear model, the quadratic model offers flexibility by not assuming a strictly monotonic relationship between population and the number of annual new melanoma cases.

One limitation of the quadratic model is the high collinearity between Populatio n 2 and Population, with a Pearson correlation coefficient of 0.9439 (see Table 1B). As noted by Arbel et al. (14), such high collinearity can distort estimation results and lead to misleading conclusions. This issue can be addressed by excluding either Population or Population sq, leading to the following alternative specifications:

Model A1: Basic Poisson Regression (Partial Quadratic Model)

Finally, note that the variable ( Year − 1999 ) = 0 , 1 , 2 , ⋯ , 22 in Equations 1, 1a, and 1b captures the linear time trend. To relax the linearity assumption and permit non-linear relationships, an alternative version of the empirical models replaces ( Year − 1999 ) with dummy variables DUM _ Year 1999 , DUM _ Year 2000 ; ⋯ ; DUM _ Year 2023 where each dummy variable equals 1 if ( Year − 1999 ) = 0 , 1 , 2 , ⋯ , 22 and zero otherwise.

Table 2 reports the regression outcomes of the model defined by Equations 1a and 1b with and without linear time trend. Columns (1) and (2) give the structural coefficients and Columns (3) and (4)—the exponent of the structural coefficients. It may be readily verified that exp ( · ) − 1 yields the percentage of growth [e.g., (15)]. For example, a coefficient of 0.0317 results in exp ( 0.0317 ) − 1 = 3.2 % representing the percentage of growth.

The results show a clear time trend in melanoma cases. Over the 23-year period, the number of new melanoma cases in the United States is projected to double ( 1.032 23 = 2.0636 ) . However, as shown in Figure 1, projected new melanoma cases decrease as population size increases. For instance, there are around 800–835 projected new cases when the Caucasian population is 80,000 (the smallest state, the District of Columbia), compared to about 340–380 projected new cases when the population reaches approximately 8 million Caucasian people (the largest populated state—California). Given California’s sunny climate and high UV radiation, this outcome is particularly noteworthy. In fact, the UV index in the District of Columbia (7)—with 800–835 projected new cases per annum—is lower than that of California (10)—with only 350–370 new cases per annum.

The implication of this analysis is that the generic dummy variables for states capture variations in UV radiation levels, while population size reflects differences in medical literacy [for insights on population size and knowledge spillover effects, see, for example, O'Sullivan (16), pp. 61–62]. According to O’Sullivan: “The essential feature of knowledge spillover is that physical proximity facilitates the exchange of knowledge between people, leading to new ideas.” (p. 61).

In this context, Carlino and Hunt (17) examine, among other factors, the relationship between patent intensity and: (1) total employment, and (2) employment density (jobs per square mile). Their findings indicate that a 10% increase in total employment and employment density leads to an increase in patent intensity by 5.2% and 2.2%, respectively. Similarly, based on existing evidence, one could argue that a larger population contributes to knowledge spillover effects, which in turn foster medical literacy.

Moreover, densely populated areas are typically characterized by better health and education systems [e.g., (18)]. This, in turn, may facilitate knowledge spillover effects [e.g., (19)] and help reduce costs by utilizing the services of more skilled dermatologists and diagnosing problems at an earlier stage.

The relationship between UV radiation and the state dummy variables is illustrated in Figures 2, 3, which are based on the Poisson regression results from Table 2. While California ranks as the state with the highest projected number of new melanoma cases when controlling for population size (13.60 to 15.22 times higher than the prevalence in the base category, Alabama), Alaska ranks lowest in terms of projected new melanoma cases when population size is controlled (0.0761 times less than the prevalence in Alabama). These figures also correspond to the UV radiation index (see Appendix A), where California ranks at the top (UV radiation index = 10) and Alaska ranks at the bottom (UV radiation index = 1).

Figure 4 depicts the ratio of melanoma cases for each year in relation to the base year (1999). The data for this graph is taken from column 4 in Table 2 (the non-linear time trend). It illustrates the extent to which the number of melanoma cases in each year surpasses that of 1999 (the base category) by a factor greater than 1, while accounting for population size. This graph reinforces our earlier calculation, showing that the projected number of melanoma cases has more than doubled over the 23-year period.

As a robustness check, we include Appendix B, which presents key graphical features derived from various model specifications. The table displays features from the following three graphs: (1) the graph based on Model A1 (Equation 1a), which includes only the Populatio n 2 term and is shown in Figure 1; (2) the graph based on Model A (Equation 1), the full quadratic model that includes both Populatio n 2 and Population ; and (3) the graph derived from Model A2 (Equation 1b), which includes only the Population term.

As shown in the table, all the derived graphs exhibit a decreasing trend with respect to the population variable. However, the graph in Figure 1 is concave—representing the decreasing segment of an inverted U-shaped curve—and ranges from approximately 800 projected new cases at the lowest population level to about 375 at the highest. In contrast, the graphs based on Model A (Equation 1) and Model A1 (Equation 1a) are convex—representing the decreasing segment of a U-shaped curve—and range from approximately 1,650 to 7,500 projected new cases at the lowest population level to around 0 to 375 at the highest.

One concern associated with the Poisson model is the problem of overdispersion. The Cameron and Trivedi (20) overdispersion test evaluates whether variance exceeds the mean in Poisson regression. Under the Poisson assumption: Var ( y i ) = E ( y i ) = μ i (the population mean of the dependent variable). They propose an alternative: Var ( y i ) = μ i + α μ i 2 , where α is the overdispersion parameter. Rejection of the null hypothesis α = 0 , supports the presence of overdispersion; non-rejection suggests the Poisson model is adequate. The test is based on an auxiliary regression using squared Pearson residuals: r i 2 = ( Y i − u ̂ i ) 2 − u ̂ i u ̂ i regressed on u ̂ i (the residuals from estimating Model A1) excluding the constant term, If the coefficient on u ̂ i is statistically different from zero, this indicates overdispersion [e.g., (21)].

As indicated by the descriptive statistics for the Count Case variable in Table 1, the mean (631) is close to the variance (702). However, the formal Cameron–Trivedi test for overdispersion (Table 3) rejects the null hypothesis—that the mean of the random disturbance term is zero—at the 1% significance level [e.g., (20, 21)]. This result suggests that the Poisson model may be inappropriate for addressing the research question.

Additional concerns include the use of total state population as an indirect proxy for medical or health literacy, which may not effectively represent it. Furthermore, state dummy variables may not sufficiently capture differences in UV radiation across states. Lastly, age is a known factor that plays an important role in melanoma prevalence.

To address these concerns, this section introduces an extended empirical model that replaces state dummy variables with a direct UV index measure from World Population. Additional control variables include a time indicator and the proportion of the population with a bachelor’s degree or higher, serving as a measure of educational attainment. The goal is to demonstrate that population size functions as a dependable proxy for medical literacy. Finally, to account for the influence of age on melanoma prevalence, the model replaces the count of melanoma cases with the annual incidence of age-adjusted melanoma prevalence.

Figure 5 shows the share of educated population within each U.S. State’s Total Population. This histogram shows the distribution of the Pop _ BA -to- Population ratio, where Pop _ BA denotes the number of individuals with a bachelor’s degree or higher, and Population refers to the total state population. According to the descriptive statistics of this variable, the median proportion of educated population is 33.81% and the mean is 34.86%. The minimum share of educated population is 24.12% (West Virginia) and the maximum is 63.05% (District of Columbia).

The previous section presents robust empirical evidence supporting the claim that larger populations are associated with higher levels of medical knowledge. This is done via the following:

Extension of the Empirical Model—We augment the model to include the percentage of the educated population (defined as individuals with at least a bachelor’s degree) in each state and apply the Two-Stage Least Squares (TSLS) methodology to address potential endogeneity concerns.

Melanoma treatment costs have risen significantly, largely due to sun-induced skin damage. This study investigates how medical literacy—approximated by population size and education—affects melanoma incidence. Prior research links medical literacy, education, health systems, and population size (16–19).

Using CDC data on non-Hispanic white males and females—the most at-risk group—a quadratic Poisson regression was employed. The model incorporates annual melanoma case counts, squared population size, state dummy variables (for UV radiation), and a time trend. Results indicate a doubling of melanoma cases over the past 23 years, closely tied to UV exposure.

To address overdispersion, the extended models replaced raw case counts with age-adjusted melanoma rates per 100,000 people and substituted state-level dummy variables with direct measures of UV index and the proportion of the population with higher education. These improvements reveal that education—serving as a proxy for human capital—plays a significant role in reducing melanoma risk. Even in regions with high UV exposure, states with larger and more educated populations exhibit lower melanoma incidence rates.

The extended models indicate that a 12–16% increase in the proportion of educated individuals can offset long-term melanoma growth, while a 25–30% increase can counteract a 1% rise in UV exposure. An increase in BA-level education by 16–30 percentage points could prevent ~22,000 to ~41,000 melanoma cases in 20 years.¹

Literature highlights disparities in early detection, especially among rural, minority, and high-risk groups, where delayed diagnoses worsen outcomes (31, 32). Access to dermatologic care remains uneven (33), and clinical training often overlooks melanoma in darker skin, particularly acral lentiginous melanoma, a subtype more common among non-white patients (34, 35). Addressing these gaps is essential for equitable outcomes.

Public policy should prioritize sun safety education and promote regular skin exams—especially in high-UV, underserved areas. Investing in health literacy today is far more cost-effective than treating preventable melanoma in the future.