After controlling for all covariates, we found county-level estimates of solar UV exposure to be positively associated with rates of early-stage melanoma among men aged 35 or older and among women aged 65 or older, but negatively associated with late-stage melanoma rates among women aged 15 or older and not significantly associated with rates of late-stage melanoma among men aged 15 or older. We have no explanation for the different relationships observed between AVGLO exposure and melanoma incidence in the various sex and age categories.
Our results indicate the importance of conducting separate analyses for early-stage and late-stage melanoma cases. For example, the model for early-stage cases had a sex-age-AVGLO exposure interaction, whereas the model for late-stage cases did not. Similarly, the state of residence at time of diagnosis was a significant effect in Poisson models for early-stage melanoma in about half of the registries used in our study, but it was a significant effect for late-stage melanoma in only 2 of the 42 registries. Possible reasons for the differences include differences by state in residents' sun protection behaviors and genetics, the thoroughness of melanoma case reporting, and the prevalence of geographic features (such as beaches or mountains) that could affect residents' UV exposure. We were unable to explore these possibilities.
Our adjusted findings that county-level estimates of solar UV exposure were positively associated with the rates of early-stage melanoma results in older adults but not among younger adults suggests that artificial sources of UV exposure or other factors might be mitigating the melanoma risk difference between younger adults in high AVGLO counties and those in low AVGLO counties.21–22
Our finding that the incidence rate of early-stage melanoma was positively associated with county level SES and physician density could be explained by higher SES individuals being more likely to go on vacations resulting in sunburn or to frequent tanning salons, as well as to live in areas with relatively high physician density.23–25
Our finding that the incidence rate for late-stage melanoma was not associated with county-level SES was in conflict with results of a previous study showing that education level was an important predictor of melanoma prognosis, defined as the melanoma mortality ratio to incidence ratio and measured using census tract level data from 1988 to 1993 in 9 SEER registries.26
The conflict might be accounted for by factors such as differences in: the outcome (incidence rates versus ratio of mortality to incidence ratios), the size of the geographic unit of analysis (county versus census tract), the number of registries (42 versus 9), the study populations (Hawaii excluded versus Hawaii included), and the melanoma incidence years of diagnosis (2004–2006 versus 1988–1993).
Our study used different time periods for solar UV exposure and melanoma incidence. Use of the 2004–2006 time period for melanoma incidence enabled an expanded US geographic area to be included in our study, because a larger number of counties released county-level incidence data that met the high-quality data criteria for United States Cancer Statistics
reporting in 2004–2006 than in earlier time periods.15
Analysis of the relationship between solar UV exposure from earlier time periods and melanoma incidence in more recent time periods is of potential interest because melanoma incidence is associated with the accumulated effects of lifetime UV exposure.13
Use of a 30-year period (1961 to 1990) for exposure also provides more accurate solar UV exposure estimates for US counties than possible with shorter time periods.12
Furthermore, adding another decade of solar UV data would not add to the quality of our estimates of potential solar UV exposure, because the annual average measures of solar UV exposure did not vary significantly over the three decades considered.12
Our study has five notable limitations. First, the accuracy of our findings for early-stage melanoma could have been affected by incomplete reporting of thinner melanomas or by regional differences in diagnostic scrutiny for melanoma.9–10,15,27
Second, although the 42 cancer registries whose data we used cover close to three-fourths of the U.S. population, several missing registries with relatively large populations were in locations classified as having either high or low AVGLO exposure, and their exclusion may have affected our findings.2,8
Third, although AVGLO exposure values can be estimated for census tract or even smaller geographic units, our analysis had to be based on less precise county-level data because melanoma incidence data were only available at the county level. Our analyses also did not account for individual variations in sun exposure protection behaviors or use of tanning beds or other artificial UV sources or for subjects' residential history prior to the time of their melanoma diagnosis. We analyzed relationships between incidence rates that were geographically aggregated to the county-level and solar UV exposures that were geographically aggregated to the county level. Caution is needed when drawing inferences about individuals based on the aggregate data for a group.28
Disaggregating data may reveal statistical relationships that are different from those at the aggregated level; geographers refer to this situation as the ecological fallacy
In our county-level analysis, there is no guarantee that the individuals with the highest risk of melanoma are also those with the highest solar UV exposures, or vice-versa. If our analysis had been based on individual life-time solar UV estimates32
that accounted for the timing, amount, and nature of the exposure for each case, then we might have observed different results.
Fourth, our Poisson regression results may have been affected by small melanoma case counts for some of the county-level analysis cells and by correlations between county location, county solar UV exposure values, and state random effects.
Finally, our study did not include details on the various ethnic and ancestry categories included within non-Hispanic whites in the United States, and geographic clustering of individuals with higher levels of genetic risk for melanoma could have affected our results.29–31