Estimating the Attributable Fraction for Cancer: A Meta-analysis of Nevi and Melanoma

Cutaneous melanoma is an important public health problem; it is both increasing in incidence and as a cause of mortality in most Caucasian populations throughout the world (1). Considerable progress has been made in quantifying the relative risks (RR) of disease associated with a broad range of environmental factors including UV radiation, and genetic/host factors including family history, number and type of melanocytic nevi, and pigmentary characteristics such as skin, eye, and hair color (2). The RRs derived from epidemiologic research alone, however, do not provide enough information to guide policy makers in directing prevention and screening efforts.

Primary prevention efforts aimed at reducing UV exposure, the major modifiable risk factor for melanoma, are complicated because the relationship between UV and melanoma is not straightforward and is highly dependent on phenotype (3). It may also be important to identify individuals at higher than average risk who can be targeted for prevention and screening efforts. This will require an understanding not only of the magnitude of the risk associated with each factor, but also of the public health effect of each factor. One widely used measure is the population attributable fraction (PAF), which considers both the strength of association between risk factor and outcome, as well as the prevalence of the factor in the community. Thus, the PAF estimates the proportion of all cases of a given disease in a population that is attributable to a particular factor.

Common melanocytic and atypical nevi have been shown repeatedly to be independent risk factors for melanoma (4, 5), and of all the host susceptibility factors identified to date, the presence of large numbers of melanocytic nevi confers the highest risk of melanoma (6, 7). The aim of this work therefore was to systematically evaluate the most recent epidemiologic evidence describing the relationship between the number and type of melanocytic nevi and melanoma, and then to use these data to estimate the fraction of melanomas attributable to these phenotypic characteristics.

A systematic review and meta-analysis was done in accordance with the Meta-analysis of Observational Studies in Epidemiology guidelines for reviews of observational studies (8).

We included case-control, nested case-control, or population-based cohort studies in the meta-analysis provided that they permitted quantitative assessment of the association between common or atypical nevi (whole-body counts) and histologically confirmed melanoma. Only studies published as original articles were included. We included studies reporting various measures of RR because melanoma is a rare disease, and in such instances, odds ratios (OR) and standardized incidence ratios provide a valid estimate of the RR.

Eligible studies up to September 2008 were identified by searching the following databases and by hand searching the reference lists of the retrieved articles.

• Medline 1950 (U.S. National Library of Medicine, Bethesda, MD) using PubMed software as the search interface

• Embase 1966 (Elsevier Science, Amsterdam, Holland) using the Embase search interface

• Conference Papers Index 1982 (CSA, Bethesda, MD) using the CSA Illumina search interface

• ISI Science Citation Index using the ISI Web of Science search interface

For computer searches, we used the following medical subject headings terms or text words (using both U.K. and U.S. spellings): melanoma, nevi, nevus, dysplastic nevus syndrome, pigmented nevus, pigmented lesion, skin lesion, melanocytic lesion, cutaneous factor, risk, etiologic, etiology, cohort studies, and case-control studies (Addendum). Studies that had been commonly cited in the literature were also included as citation search terms in the ISI Science Citation Index (1990 to present) to identify subsequent studies that had referenced them. Only studies of adult populations (>18 y) were included. The search was not limited to studies published in English. We read the abstracts of all identified studies to exclude those that were clearly not relevant. The full texts of the remaining articles were read to determine if they met the study inclusion criteria. Where multiple reports from one study were found, the most recent or most complete publication was used.

The primary computerized literature search identified 115 potentially eligible studies. After a review of the study abstracts, we retrieved 80 articles for further assessment, of which 2 reports of cohort studies (9, 10), and 47 from case-control studies (refs. 4, 5, 7, 11–54; total = 49) met the eligibility criteria (Fig. 1). The remaining 31 studies were excluded because they either did not present enough data to compute effect estimates of the association between nevus count and melanoma (n = 6), were not independent of other included studies (n = 14), or they represented an ineligible study design [e.g., case-series, nonpopulation-based cohort study (i.e., cohorts of people with atypical nevi; n = 11].

An abstraction form summarizing the study design, study population, and relevant raw and adjusted data were completed for each article by two independent reviewers, both investigators (CO, HC); inconsistencies were re-reviewed until consensus was achieved. The following information was recorded for each study: study design, location, years of data collection (case-control studies), number of cases and controls or person-year duration of follow-up (cohort studies), age of study population, variables for which statistical adjustment was done, definitions and categories of nevus variables, point estimates (RR, OR, or standardized incidence ratio), and 95% confidence intervals (95% CI). Where several risk estimates were presented, we abstracted those adjusted for the greatest number of potential confounders. Studies that reported results separately by gender, body site, or histologic subtype with no combined data were treated as independent data sets in the meta-analysis.

We did not assess the methodologic quality of the primary studies and hence did not exclude studies on the basis of quality score (8), but instead performed subgroup and sensitivity analyses according to study features that could potentially affect the strength of the association between nevus counts and melanoma.

To pool the RR estimates for the presence of atypical nevi, a weighted average of the log RR was estimated, taking into account the random effects using the method of DerSimonian and Laird (55). Because for this exposure measure we were interested in an “any” versus “none” comparison, and some studies presented affect estimates for several categories of nevus count, we used the method of Greenland and Longnecker (56) to derive the relative effect and precision estimates for the required comparison. Statistical heterogeneity among studies was evaluated using the Cochrane Q test and I² statistics. The Cochrane Q test, which is calculated as the weighted sum of squared differences between individual study effects and the pooled effect across studies, is widely known to have too much power to detect clinically unimportant heterogeneity if the number of studies is large (57). The I² statistic describes the percentage of variation across studies that is due to heterogeneity rather than chance (58), and does not inherently depend upon the number of studies considered (I² = 100% × (Q − df)/Q). We also conducted separate analyses by study type, geographic location, case family history of melanoma, definition of atypical nevi (standard or other), source of atypical nevus count (counted or self-report), year of publication (before and after 2000), and study size. Finally, we conducted sensitivity analyses, omitting each study in turn to determine whether the results could have been influenced excessively by a single study. We evaluated the publication bias by assessing the funnel-plot asymmetry (59, 60).

To pool RR estimates for common nevus counts, we conducted a dose-response meta-analysis using a method proposed by Greenland and Longnecker (56) and Orsini et al. (61), which estimated the study-specific slopes from the correlated natural log of the RR estimates across categories of nevus counts. Unlike other methods for pooling estimates that assume independence of the estimates, this method accounts for the correlation between estimates that depend on the same reference group. The value assigned to each category of nevus count was the midpoint for closed categories, and we assigned a midpoint for open-ended top categories (120 where ≤120, 150 where >120, and 180 where >150). We conducted sensitivity analyses around the assignment of these midpoint values for the open-ended top categories. Moderate changes (i.e., from 80-120 nevi) did not materially alter the pooled effect estimates nor improve model fit. We conducted further analyses using the “pool first” (61) method, which requires pooling data from all of the studies before estimating the dose-response model. In addition, we checked for nonlinearity of the dose-response relationship by estimating polynomial models. Overall, the best-fitting model for these data was a linear model pooling the study-specific slopes (Supplementary Table S1). We performed subgroup analyses to examine heterogeneity by study characteristics including study type, study size, year of publication, geographic location, case family history of melanoma, and source of common nevus count, and again conducted sensitivity analyses to assess the effect of each study on the summary risk estimates.

PAFs for the presence of atypical nevi and common nevi in categories (0-10, 11-24, 25-49, and 50+) were estimated using Levin's formula (62), which requires only the RR estimate and the prevalence of the risk factor. CIs for the PAFs were derived using the substitution method described by Daly (63). For common nevi, we used the pooled RR for the number of nevi at the midpoint of each category for the PAF calculations (i.e., 5 for the 0-10 category, 18 for the 10-25 category, 37 for the 25-49 category, and 80 for the 50+ category). Because reporting of nevus counts varied across studies, we estimated the prevalence of the nevus exposures by calculating an average of the prevalence reported in study controls, which was weighted by the size of the control group. For common nevi, this was restricted to those studies that presented their nevus counts in similar categories. This approach to calculating PAFs has inherent limitations due to the underlying heterogeneity in prevalence estimates; however, reliable population-based data on nevus prevalence are scarce. We therefore conducted sensitivity analyses to evaluate the effect of different ranges of common and atypical nevus prevalence on the PAF estimates. For the 0 to 10 common nevi category, we calculated the PAF for a change in prevalence in increments of 0.025 from 0.3 to 0.5; the difference was distributed over the remaining three categories in a weighted fashion based on the original estimates. Similarly, for the 11 to 25 and 25 to 49 nevi categories, we calculated the PAF for prevalence ranging from 0.1 to 0.3, and calculated PAF for prevalence for the 50+ nevus category from 0 to 0.2. We modeled the PAF of atypical nevi based on prevalence ranging from 0.025 to 0.2, again in increments of 0.025.

All analyses were conducted using Stata 10.

Measurement of nevus burden varied across studies; however, the nevus variables collected by the greatest number of studies were whole-body atypical nevus counts and whole-body common nevus counts, and thus, we used these two variables in our meta-analyses.

A total of 25 studies presented data on the relationship between atypical nevi (whole-body counts) and melanoma (refs. 4, 5, 7, 11, 17, 18, 20–22, 25, 26, 29, 33, 36–38, 43, 45–49, 52–54; Fig. 1); these were all case-control studies (Table 1). One study presented data stratified by sex only (20); for that study, we included the data for men and women in the meta-analyses as two independent data sets. A single cohort study reported atypical nevus counts from the legs only (10), and this study was not included in the meta-analysis. For all studies, the pooled RR of melanoma for people with one or more atypical nevi compared with those with none was 3.63 (95% CI, 2.85-4.62) with evidence of significant heterogeneity (P ≤ 0.001; Fig. 2). The results of subgroup analysis to examine the heterogeneity are provided in Table 2. Only two of the studies were population based; the pooled RR for these two studies was lower than for the nonpopulation-based case-control studies (2.63 versus 3.74; Table 2). No heterogeneity was observed for the two population-based case-control studies. When stratified by geographic location, the highest pooled RR was observed for the five studies conducted in North America (pooled RR, 5.86; 95% CI, 3.50-9.82), and the lowest was observed for those conducted in Central Europe (pooled RR 2.83; 95% CI, 1.41-5.69). There did not seem to be significant differences in the pooled RRs according to the definition of atypical nevus used, and only one study relied upon self-reported atypical nevus counts. For studies that included cases with a family history of melanoma (n = 12), the pooled RR was 3.92 (95% CI, 2.89-5.31) compared with 3.27 (95% CI, 1.77-6.04) for studies that did not (n = 5); the remaining eight studies did not state whether they included cases with a family history of melanoma. Studies published since 2000 were more homogeneous, and our most homogeneous estimate was derived by excluding the one study that relied on self-reported atypical nevus counts from this group; the pooled RR was 3.70 (95% CI, 2.76-4.96), which was very close to the pooled RR for all studies (3.63). Finally, the summary statistics were not influenced by sensitivity analysis excluding one study at a time, with the pooled RR ranging from 3.46 (2.74-4.38) with the omission of Rhodes et al. (11), to 3.79 (3.00-4.80) with the omission of Grulich et al. (33) The funnel plot of the effect estimates for the risk of melanoma related to atypical nevi was close to symmetrical and there was no evidence of publication bias using the Egger weighted regression method (P_bias = 0.148) or the Begg rank correlation method (P_bias = 0.076).

A total of 31 studies provided data on the association between common nevi (whole-body count) and melanoma. An additional eight case-control studies (14, 24, 32, 34, 35, 39, 40, 49) and one cohort study (9) did not report on whole-body nevus counts, but rather specific body sites including the arms or legs. A further six studies presented data only for palpable and/or large common nevi (13, 19, 28, 30, 31, and 44) and one study did not use quantitative common nevus counts (27); these were also excluded from the analyses (Fig. 1).

For the dose-response meta-analysis, 23 case-control studies were eligible to be included in the meta-analyses of whole-body common nevus counts (Table 3; refs. 5, 7, 15, 17, 18, 21–23, 25, 29, 33, 36–38, 41, 43, 45–47, 50–53); seven studies that presented data in dichotomous categories only (4, 12, 16, 20, 26, 42, 54) were excluded because the analyses required three or more exposure categories. One study presented data stratified by sex only (50), and the data for men and women were included in the meta-analyses as two independent data sets.

Using a random effects model, the pooled RR for common nevi (whole body) for single-unit increase in nevus count for all studies was 1.017 (1.014-1.020; Fig. 3). Although there was evidence of significant heterogeneity (P ≤ 0.001; Table 4), all RRs were in the same direction and the heterogeneity was due to differences in the magnitude of the association. The results of subgroup analysis to examine the heterogeneity are provided in Table 5. Two of the studies were population based; the pooled RR for these studies was lower than that for the nonpopulation-based case-control studies (1.008 versus 1.018). As expected, the pooled RR for studies that included cases with a family history of melanoma was higher than for studies that did not (1.014 versus 1.012), and there was no evidence of heterogeneity among the four latter studies (P = 0.41). However, a large number of studies (n = 10) did not report information on inclusion/exclusion of cases with a family history. The pooled RR for studies conducted in Northern and Central Europe was higher than for studies from North America or Mediterranean Europe. Sensitivity analyses excluding one study at a time resulted in pooled RR ranging from 1.016 (95% CI, 1.013-1.019) when the study by Swerdlow et al. (15) was excluded, to 1.017 (95% CI, 1.015-1.020) when the study by Augustsson et al. (22) was excluded. Again, there was no evidence of publication bias using the Egger weighted regression method (P_bias = 0.330) or the Begg rank correlation method (P_bias = 0.239).

Estimates of the PAF associated with the presence of atypical nevi and common nevus counts, calculated using the meta-analysis–derived summary RRs and the weighted average of the prevalence estimates from all studies, are presented in Table 6. There was a large degree of variation in prevalence of both atypical nevi and common nevus counts across studies, particularly for the lowest category of common nevi (0-10). The estimated PAF of melanoma associated with the presence of one or more atypical nevi was 0.25. Level-specific PAFs for common nevus counts ranged from 0.04 for 0-10 nevi to 0.27 for 50 or more nevi. Figure 4 presents an evaluation of the effect of different prevalence assumptions on the PAF of common nevi. We modeled the PAF under different scenarios of nevus prevalence. Varying the prevalence in the bottom two categories of nevus counts (0-10, 11-24) did not have a substantial effect on the total PAF of common nevi with the PAF associated with these counts remaining under 0.14 under all scenarios. In contrast, under all prevalence scenarios examined, the highest burden was always associated with higher nevus counts (25-49, 50+), especially the 50+ category. We also modeled the PAF of atypical nevi based on prevalence ranging from 0.025 to 0.2; the PAF ranged from 0.06 to 0.34.

Risk factors for melanoma have been quantified in terms of RRs. From the perspectives of public health and health policy, however, the effect of any given risk factor depends not only on the magnitude of the association, but also on the distribution of the factor in the population. We have evaluated the most up-to-date available epidemiologic evidence about the magnitude of the relationship between atypical nevi and common nevus counts and melanoma risk by conducting a systematic review and meta-analysis. Our findings suggest that ∼25% of melanoma cases are attributable to the presence of one or more atypical nevi. High common nevus counts (50 or more common nevi) account for ∼27% of melanoma cases, whereas individuals with few common nevi (0-10) account for only ∼4% of melanoma cases. To our knowledge, this is the first study to systematically evaluate the PAF of melanoma associated with atypical nevi and common nevus counts using estimates of RR derived through systematic review and meta-analysis.

Several limitations must be acknowledged when interpreting these findings, relating to the pooled RR estimates and to the nevus prevalence estimates. First, the studies contributing to the pooled RR estimates are prone to the usual biases associated with case-control studies, including selection bias, recall bias, and measurement error. It is unlikely that recall bias has unduly influenced the pooled RR estimates because in most studies, nevi were counted by independent personnel rather than by patients themselves. Nondifferential exposure misclassification would likely bias OR estimates toward the null, which would result in an underestimate of the true PAF (64). Selection bias due to control recruitment from dermatology clinics or hospitals in the majority of studies would also result in an attenuated effect if controls recruited in this way are more likely to have higher nevus counts than those recruited randomly from a population-based source. Second, there is uncertainty surrounding our calculations of the prevalence of common nevi in the study population due to a lack of data available from population-based sources. The studies contributing to the pooled estimates of RR represent diverse populations, and there was a large amount of between-study variation in the prevalence of common nevus counts. Our sensitivity analyses to examine the effect of changing prevalence of common nevi on the PAF were reassuring in this regard; the lowest total PAF was 0.31 based on a prevalence of 0 in the 50+ nevus category. However, even if the PAF estimates are based on imprecise assessments of the distribution of nevus counts in the population, our estimates are nevertheless useful in quantifying the potential for prevention by directing the level of screening surveillance at people with >25 common nevi and/or atypical nevi.

Another limitation to acknowledge in our estimation of the PAF associated with common and atypical nevi is our single risk factor approach, which does not take into account the potential interaction or confounding by other risk factors. For example, we know that common nevus counts are highly correlated with atypical nevus counts (65); however, from these analyses, we were unable to determine the degree of overlap between people with high nevus counts and atypical nevi. Multifactorial attributable fractions can only be estimated reliably when the data are presented and published in this way. We would encourage future publications to present the risk estimates for combinations of common risk factors so that additive effects can be explored. There are also challenges associated with the interpretation of PAFs, which have been discussed by others (66, 67).

An earlier systematic review and meta-analysis (68) included studies conducted up to 2002; however, a large number of relevant studies have entered the literature since that review was published in 2005. Indeed, we included 10 new studies in our meta-analysis (45–54) that were not published in time for the earlier analysis. Our study differs from that earlier review in that we have analyzed the effect of our pooled effect estimates on the disease at a population level by estimating the PAFs.

Although there is insufficient justification to recommend routine screening of all patients for melanoma (69, 70), there is some evidence that screening of patients at high risk of melanoma is cost effective (71) and likely to be associated with improved survival (72). The results of these analyses suggest that patients with 25 or more common nevi and/or one or more atypical nevi are a high-risk group for which identification, screening, and education may achieve disease prevention or early detection at low cost. A PAF of 0.42 for melanomas associated with 25 or more common nevi equates to 67,200 patients newly diagnosed with melanoma from an annual worldwide total of 160,000 new cases (73). Further work incorporating our findings with considerations of the cost and viability of alternative prevention strategies is required.

Search strategy to identify observational studies on the association of common or atypical nevi and cutaneous melanoma

No potential conflicts of interest were disclosed.

Grant Support: This research was supported by a grant from the Xstrata Community Partnership Program Queensland. The researchers are independent of the funding source. D. Whiteman is a Principal Research Fellow of the National Health and Medical Research Council of Australia.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.