Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Nutritional and Socioeconomic Factors in Relation to Prostate Cancer Mortality: a Cross-National Study James R. Hebert, Thomas G. Hurley, Barbara C. Olendzki, Jane Teas, Yunsheng Ma, Jeffrey S. Hampl Background: Large international variations in rates of prostate cancer incidence and mortality suggest that environmental factors have a strong influence on the development of this disease. The purpose of this study was to identify predictive variables for prostate cancer mortality in data from 59 countries. Methods: Data on prostate cancer mortality, food consumption, tobacco use, socioeconomic factors, reproductive factors, and health indicators were obtained from United Nations sources. Linear regression models were fit to these data. The influence of each variable fit in the regression models was assessed by multiplying the regression coefficient b by the 75th (X75) and 25th (X25) percentile values of the variable. The difference, bX75 − bX25, is the estimated effect of the variable across its interquartile range on mortality rates measured as deaths per 100 000 males aged 45–74 years. Reported P values are two-sided. Results: Prostate cancer mortality was inversely associated with estimated consumption of cereals (bX75 − bX25 = −7.31 deaths; P = .001), nuts and oilseeds (bX75 − bX25 = −1.72 deaths; P = .003), and fish (bX75 − bX25 = −1.47 deaths; P = .001). In the 42 countries for which we had appropriate data, soy products were found to be significantly protective (P = .0001), with an effect size per kilocalorie at least four times as large as that of any other dietary factor. Besides variables related to diet, we observed an association between prostate cancer mortality rates and a composite of other health-related, sanitation, and economic variables (P = .003). Conclusions: The specific food-related results from this study are consistent with previous information and support the current dietary guidelines and hypothesis that grains, cereals, and nuts are protective against prostate cancer. The findings also provide a rationale for future study of soy products in prostate cancer prevention trials. [J Natl Cancer Inst 1998;90:1637–47] Prostate cancer is the most prevalent cancer among men and represents a large and growing health problem in the United States and other Western countries. Incidence and mortality rates vary widely across populations, with the highest rates in North America and northern Europe, intermediate rates in southern Europe and Latin America, and the lowest rates in Asia and Africa (1,2). Incidence increases markedly with age. Ecologic studies investigating the health experiences of migrants over time (3–6) and cross-national comparisons (7–10) suggest that environmental, rather than genetic, factors are responsible for large international variations in rates. Against the background of international variation, increases in prostate cancer mortality and incidence rates have been reported in the last three decades in the United States, Asia, South America, and central Europe. Prostate cancer incidence in the United States is currently increasing at a rate nearly three times as high as that of female breast cancer (11). The incidence of prostate cancer is extremely low in populations of western Africa, but in African-Americans, who originate from the same regions of Africa, the incidence is nearly 50% higher than that of U.S. whites (12), indicating a potentially strong environmental influence on the disease. Similar to epidemiologic studies of diet and breast cancer, studies of diet and prostate cancer have produced equivocal results (13). Ecologic studies and laboratory animal experiments tend to produce results strongly consistent with dietary hypotheses that propose a role for fat and animal product intake, whereas data from analytic epidemiologic studies are less consistent (8,9,14–19). Possible explanations include the fact that data on diet in humans are derived from self-reports that are subject to measurement error and a variety of biases (20) and that the range of dietary fat intake (e.g., grams per person per day or as a percent of dietary energy) within particular study populations tends to be very narrow (21), hampering efforts to detect dietary effects. Also, the effect of particular dietary constituents may be modified by other factors, such as the antagonism of the effect of fat by fiber (e.g., through its influence on enterohepatic circulation) (22). This study was conducted to identify major predictors of prostate cancer mortality by use of data from a total of 59 countries on which we had complete data on most covariates of interest. In addition to a wide range of nutritional predictors (including alcohol), we examined the significance of estimated per capita tobacco consumption, a variety of indicators of socioeconomic status (SES), and reproductive factors. MATERIALS AND METHODS Data Sources Data on a maximum of 59 countries were available from a number of United Nations sources (see Appendix Tables 1 and 2). Data were time-lagged so that the mortality information was from a period approximately 10 years later than data on other factors potentially affecting risk. Data on prostate cancer mortality (i.e., prostate cancer deaths per 100 000 male population for ages 45–74 years) were obtained from a data tape supplied by the World Health Organization (WHO) (23). Data were averaged for the 5 years from 1985 through 1989 and then age-truncated to focus on that portion of the distribution (ages 45–74 years) where cancer mortality rates and quality of reporting tend to be highest (9). The Affiliations of authors: J. R. Hebert, T. G. Hurley, B. C. Olendzki, J. Teas, Y. Ma, Division of Preventive and Behavioral Medicine, University of Massachusetts Medical School, Worcester; J. S. Hampl, Department of Nutrition, Simmons College, Boston, MA. Correspondence to: James R. Hebert, Sc.D., Division of Preventive and Behavioral Medicine, University of Massachusetts Medical School, 55 Lake Ave., North, Worcester, MA 01655. See ‘‘Note’’ following ‘‘References.’’ © Oxford University Press Journal of the National Cancer Institute, Vol. 90, No. 21, November 4, 1998 ARTICLES 1637 rates were then age-standardized to the 1987 world male population average. Mortality data provided by the WHO excluded countries with more than 15% of mortality due to senility or ill-defined causes. National nutrition data were obtained from the Food Balance Sheets—1979 Through 1981, Average, compiled and published by the Food and Agriculture Organization (FAO) of the United Nations (24). These data often are referred to as disappearance data. They represent a kind of inventory information that is based on the sum of imports and production minus exports, with adjustments made for non-food uses, losses during storage and transport, food fed to livestock, and food used for seeds and in manufacture for food and non-food uses. Data are prorated on a per capita basis as a surrogate for consumption. Values for energy intake plus protein and fat intake in grams are available from this source for overall vegetable and animal product consumption as well as for consumption of the major food groups (i.e., cereals, roots and tubers, sugars, pulses, nuts and oilseeds, fruits, vegetables, meat, eggs, fish and seafood, milk, oils and fats, vegetable oils and fats, animal oils and fats, spices, stimulants, and alcoholic beverages) and a list of about 300 commodities. Some of these items (e.g., soy products) were not available for every country in the dataset. Data on tobacco consumption are similar, with the exception that no allowance is made for food manufacture or livestock feed. SES data were obtained from a number of sources. Information regarding basic sanitation (e.g., proportion of the population with safe water and with excreta disposal facilities) from the late 1970s was obtained from the WHO (25). The following basic economic and health indicators were obtained from World Bank compilations (26,27): gross national product (GNP) per capita (1980), infant mortality rate (1980), crude birth rate (1970), total fertility rate (1970), annual percent of growth (natural increase) in population (average, 1970 through 1980), and life expectancy (at birth, 1980). Statistics on the number of physicians per unit population (physicians per 10 000 population) and hospital bed availability (beds per 10 000 population) were obtained from the WHO (28). Data from 59 countries were sufficiently complete to allow most multivariable analyses of these data. Statistical Analyses Preliminary analyses to examine the distributions and to identify outlier values were performed on all variables to assess the adequacy of the assumptions of linear regression (i.e., normality, linearity, and independence) (29). These statistics were computed for all 59 countries for which we had complete data in order to fit the final multivariable model. All P values are two-sided. Principal-components analysis (PCA) was performed to reduce the number of predictor variables with minimal loss of information. A complete description of this technique is beyond the scope of this article; however, excellent descriptions may be found elsewhere (30,31). The seven variables related to socioeconomic conditions included in computing the socioeconomic status principal components (SESPCs) were as follows: GNP, life expectancy at birth, infant mortality rate, proportion of the population without safe water or excreta disposal facilities, and the number of physicians or hospital beds per 10 000 population. To compute SESPCs, we used the PROC FACTOR procedure with the Varimax rotation option in SAS (SAS statistical software, version 6.11; SAS Institute, Cary, NC) (32). This approach allowed us to reduce a larger set of potentially correlated variables to a smaller set of variables (principal components). The principal components are optimally weighted combinations of the initial seven variables that maintain, as much as possible, the explanatory ability of the original variables. The Varimax rotation option creates principal components that are not correlated with each other. We selected the three SESPCs that account for the majority of the variability in the data and excluded the other factors that did not appreciably increase the explanatory ability. Similarly, PCA was used to examine the interrelationships between the large number of nutritional variables in the dataset. Variables were selected to minimize the impact of multicollinearity (i.e., intercorrelation among predictor variables) when fit in the multivariable models, without excluding potentially important predictors of prostate cancer mortality. For example, the intercorrelations between per capita animal energy (kcal/day—total energy derived from foods from all animal sources), animal fat energy (kcal/day), animal protein (g/day), milk energy (kcal/day), meat energy (kcal/day), total fat (g/day), and total energy (kcal/day) were very high, ranging from 0.72 to 0.99. Fitting such highly correlated variables simultaneously in a model would pose problems in terms of model specification and interpretation. We used PCA [PROC FACTOR in SAS (32) with the no-rotation option, which maintains the factor-loading weights unchanged] to identify nearly collinear sets of variables. We then selected a 1638 ARTICLES single variable from among each set of highly correlated variables by choosing the variable with the highest component weighting. In this way, we selected a set of largely uncorrelated nutrient variables to examine in a multivariable setting while we retained the inherent representation of the underlying PCA (i.e., the animal-related variables discussed above were represented by animal energy). Linear regression models were fit, regressing prostate cancer mortality on each predictor variable. All regression models were weighted by the malespecific population across countries for the age range 45–74 years. Model specification used stepwise regression techniques as a preliminary approach, as well as adding and removing candidate variables individually to produce the most parsimonious model while retaining a high level of explanatory ability. For this approach, we used a general linear model [the procedure PROC GLM in SAS (32)]. Diagnostic analytic procedures were used to examine the distribution of residuals and to identify influential cases and problems associated with multicollinearity (33). Also, we log-transformed both prostate cancer mortality data and nutrient data and reanalyzed our data. The results were not materially different from those obtained using untransformed data. We chose to present the untransformed results because the assumptions of the analysis were met (i.e., normality) and because use of untransformed data improves interpretability of results. To evaluate the actual effect of each variable fit in the final regression models, we multiplied the regression coefficient b by the 75th percentile value of the variable (X75) and the 25th percentile value of the variable (X25). The difference between bX75 and bX25 is the effect of the variable across its interquartile range (IQR). This manner of presenting the results combines the estimated magnitude of effect (b) and the range of the variable in the actual dataset in which the regression models were run. It provides a way to ‘‘standardize’’ the effect in terms of prostate cancer deaths. Also, we have presented the percentage change in the mean world rate of prostate cancer, which was calculated as [(bX75 − bX25)/world mean prostate cancer rate] × 100; this quantity gives a sense of the relative effect on prostate cancer based on a change across the IQR of the exposure variable. Regression model results for the 59-country models excluding tobacco consumption were similar on significant predictors to the 50-country models that included tobacco in the covariate list. Estimated tobacco consumption (from disappearance) was consistently unrelated to prostate cancer mortality. By excluding the tobacco consumption variable from the final model, we were able to include data from 59 countries in the multivariable analyses. Nuts and oilseeds subsume soybeans in the Food Balance Sheet dataset. We attempted to separate soybean consumption from this larger food group and were able to identify a subset of 42 countries on which we could obtain reliable data on soy disappearance (to estimate consumption). All analyses were repeated in this smaller dataset by use of the model-building approach described above. In addition, to assess the effect of total energy on prostate cancer mortality, we fit a model using energy-adjusted nutrients. Nutrient variables were regressed upon total calories, and the residuals were calculated as the difference between the actual value for the nutrient and the predicted value. The residual is that portion of the nutrient intake that is unexplained by the variation in total energy intake (34). The model specification process was the same as that described above. RESULTS In Table 1, summary statistics are presented for the dependent variables (sex-specific, age-truncated, and standardized prostate cancer mortality rates) and the independent variables (e.g., dietary and SES variables) used in the analyses. There was a slight rightward (a tendency for a few very high values) skewing for variables associated with poverty and a slight leftward skewing for variables associated with affluence. These statistics demonstrate a wide range of variability, as would be expected for such a diverse set of countries. Exploratory analyses indicated that data met the assumption of normality, even for the full set of 59 countries. Table 2 displays the weighting of individual variables in the SES PCA. SESPC1 is a poverty-related factor, loading heavily with infant mortality and lack of sanitation or lack of safe drinking water. SESPC2 is related to access to medical services, with Journal of the National Cancer Institute, Vol. 90, No. 21, November 4, 1998 Table 1. Summary statistics for prostate cancer mortality, diet, and socioeconomic factors Percentile values Variable No. of countries Mean 23.27 Minimum Prostate cancer mortality rate* 59 Diet Total energy† Total fat, g/day Animal fat energy† Animal energy, including poultry† Animal protein, g/day Meat energy† Milk energy† Fish energy† Vegetable energy† Vegetable oil (extracted oils) energy† Vegetable fat (all fats), g/day Sugar energy† Root and tuber energy† Cereal energy† Pulse energy† Nut and oilseed energy† Soybean energy† Cabbage energy† Fruit energy† Alcohol energy† 0.53 59 59 59 59 59 59 59 59 59 59 59 59 59 59 59 59 42 59 59 59 3003 88.7 543 791 44 352 233 33 2215 52 28 375 125 1042 48 39 8 4 109 130 Socioeconomic factors Hospital beds per 10 000 population‡ Physicians per 10 000 population‡ Percentage without sanitation‡ Percentage without safe water‡ Per capita GNP (U.S. dollars)‡,§ Infant mortality rate‡,㛳 Life expectancy‡,¶ 59 59 59 59 59 59 59 64 12 16 13 4965 28 70 10 1 0 0 320 7 57 Other variables Total fertility rate# Annual percent growth in population** Crude birth rate†† Tobacco consumption (kg/person per year) 48 46 48 50 4 2 26 17 2 0 14 1 2018 14.1 46 95 9 9 9 2 1431 11 0.1 73 11 604 6 1 0 0 2 0 Maximum 69.50 3699 172.3 1173 1699 104 750 554 195 2966 146 70 644 446 2070 183 312 94 39 326 304 25th 75th 14.61 30.13 2712 52.9 268 406 27 157 116 15 2009 21 20 302 63 703 20 14 0.2 1 69 55 3393 128.6 868 1196 57 511 305 38 2384 70 37 446 187 1265 71 47 4 6 141 209 172 25 85 79 19 870 104 77 35 5 0 0 1430 11 66 101 17 20 20 7920 33 74 8 4 49 54 2 1 17 9 5 3 35 22 *Prostate cancer deaths per 100 000 male population 45–74 years old, age standardized to the 1987 world male population average. †Total energy based on disappearance data, prorated to kilocalories per person per day. ‡These variables are included in the socioeconomic principal-components analysis. The mean values across countries are derived by averaging the within-country percentages. §Per capita gross national product (1980). 㛳Rate per 1000 live births (1980). ¶Life expectancy, in years, at birth (1980). #Number of children that would be born to a woman if she were to live to the end of her childbearing years (15–49 years) and bear children in accordance with age-specific fertility rates (1970). **Average annual growth of population (percent) based on mid-year populations, 1970–1980. ††Number of live births per 1000 population (1970). large positive loadings on physicians and medical beds per unit population as well as life expectancy. SESPC3 is an affluencerelated factor, with a large loading on per capita GNP. The simple linear regression coefficient (b) and the 95% confidence interval along with the Pearson correlation coefficient (r) are presented in Table 3 for all candidate independent variables in the dataset. For the nutrition variables, positive associations were evident between prostate cancer mortality rates and variables associated with affluence, especially food-related variables including total fat, animal products including milk and dairy products, and energy from alcohol and sugar. Negative associations were observed with energy from cereals, vegetables, cabbage, soybeans, and fish. Among the SESPCs, there was a negative association with the poverty-related factor and a positive association with the affluence-related factor. The results of the regression model analysis are shown in Table 4. Overall, we observed protective effects from higher levels of intake of cereals, nuts and oilseeds (which include soy consumption), and fish. There was also a significant protective association for decreasing levels of the poverty-related factor (SESPC1). Animal energy was positively associated with prostate cancer mortality. Effect sizes, estimated by fitting the parameter estimate across the interquartile values of the variable (i.e., bX75 − bX25), are shown in Tables 4 and 5. For cereal energy, the difference in the estimated effect at the 75th and 25th percentile values was –7.31 prostate cancer deaths per 100 000 males aged 45–74 years, 31.4% of the world mean average prostate cancer mortality for males in this age group. Across the full range of data represented by these 59 countries, the difference between the lowest X0 and the highest X100 consumption of cereal energy when fitted to the regression coefficient equaled 19.1 deaths. In Journal of the National Cancer Institute, Vol. 90, No. 21, November 4, 1998 ARTICLES 1639 Table 4. Results of the general linear model based on prostate cancer mortality—including energy from soybeans as part of energy from nuts and oilseeds* Table 2. Results of principal-components analysis for variables representing socioeconomic status* Variable GNP㛳 Life expectancy Infant mortality Percent of households without safe water Percent of households without sanitation Physicians per capita Hospital beds per capita SESPC1† SESPC2‡ SESPC3§ −0.207 −0.495 0.746 0.869 0.914 −0.199 −0.301 0.306 0.568 −0.427 −0.213 −0.194 0.893 0.798 0.903 0.463 −0.355 −0.142 −0.145 0.176 0.324 95% confidence interval Independent variable *Principal-components analysis using PROC FACTOR in SAS with Varimax rotation. The principal components are optimally weighted combinations of the seven variables that maintain, as much as possible, the explanatory ability of the original variables. Data used are from the 59 countries on which we had complete data. The numbers in the last three columns indicate optimum factorloading weights for each combination. †SESPC1 ⳱ socioeconomic status principal component 1, computed as noted above and in the text. Poverty-related factor. ‡SESPC2 ⳱ socioeconomic status principal component 2, computed as noted above and in the text. Access to medical care factor. §SESPC3 ⳱ socioeconomic status principal component 3, computed as noted above in the text. Affluence-related factor. 㛳GNP ⳱ gross national product (U.S. dollars per person in 1980). contrast, the increase in mortality associated with an increase in animal energy across the IQR was 5.53 or 23.8% of the mean of prostate cancer mortality. The model focusing on soybean disappearance data for the subset of 42 countries where that data were available produced the results shown in Table 5. The same core group of variables observed to be associated with prostate cancer mortality in the Cereal energy Animal energy Nut and oilseed energy㛳 Fish energy SESPC1¶ b† Lower Upper bX75 − bX25‡ P§ −0.013 0.007 −0.052 −0.019 0.001 −0.085 −0.007 0.013 −0.019 −7.31 5.53 −1.72 .001 .02 .003 −0.064 −3.253 −0.101 −5.323 −0.027 −1.183 −1.47 −4.07 .001 .003 *In this model, the dependent variable—prostate cancer mortality rate per 100 000 population, truncated to ages 45–74 years and age standardized to the 1987 world male population—is regressed on all independent variables listed; 59 countries are included; model R2 ⳱ .80. †Regression coefficient for the independent variable. ‡bX75 − bX25 estimates the effect size of the variable listed by multiplying the regression coefficient b by the 25th percentile value of the variable and subtracting that value from the product of the regression coefficient b and the 75th percentile value of the variable. This is the effect of the variable across its interquartile range (in prostate cancer deaths per 100 000 males aged 45–74 years) in the 59 countries available for this analysis. §Two-sided P values were obtained from the F test of H0: ⳱ 0. 㛳Includes energy from soy. ¶SESPC1 ⳱ socioeconomic status principal component 1, computed as described in the text. Poverty-related factor. previous analysis was similarly associated in this model, with the exception of fish energy. Here, however, energy from alcohol was significantly associated with increased mortality, while energy from soy was significantly associated with reduced mortality. Table 3. Regression coefficients (b) with the 95% confidence interval and Pearson correlation coefficients for candidate independent variables with age-standardized and truncated prostate cancer mortality rates* 95% confidence interval of b Variable Total energy㛳 Total fat, g/day Animal fat energy㛳 Animal energy, including poultry㛳 Animal protein, g/day Meat energy㛳 Milk energy㛳 Fish energy㛳 Vegetable energy㛳 Vegetable oil (extracted oils) energy㛳 Sugar energy㛳 Root and tuber energy㛳 Cereal energy㛳 Pulse energy㛳 Nut and oilseed energy㛳 Soybean energy㛳 Cabbage energy㛳 Fruit energy㛳 Alcohol energy㛳 SESPC1 (poverty-related factor)¶ SESPC2 (medical services-related factor)¶ SESPC3 (affluence-related factor)¶ No. of countries b† Lower Upper r‡ P§ 59 59 59 59 59 59 59 59 59 59 59 59 59 59 59 42 59 59 59 59 59 59 0.01 0.19 0.03 0.02 0.43 0.04 0.07 −0.10 −0.01 −0.08 0.06 0.07 −0.02 −0.07 −0.10 −0.24 −0.48 0.01 0.08 −5.06 6.64 3.30 0.004 0.147 0.024 0.016 0.322 0.032 0.054 −0.163 −0.020 −0.182 0.046 0.033 −0.026 −0.148 −0.157 −0.334 −0.858 −0.053 0.053 −8.296 3.702 0.632 0.016 0.233 0.036 0.024 0.538 0.048 0.086 −0.037 0.002 0.022 0.074 0.107 −0.014 0.008 −0.043 −0.146 −0.102 0.069 0.107 −1.824 9.578 5.968 .53 .74 .78 .79 .72 .76 .78 −.39 −.30 −.19 .71 .46 −.72 −.22 −.43 −.62 −.32 .03 .62 −.38 .51 .31 .0001 .0001 .0001 .0001 .0001 .0001 .0001 .002 .02 .14 .0001 .0002 .0001 .10 .0007 .0001 .02 .10 .0001 .003 .0001 .02 *Prostate cancer deaths per 100 000 male population 45–74 years old, age standardized to the 1987 world male population average. †Regression coefficient from regressing prostate cancer mortality on the independent variables. ‡r ⳱ Pearson correlation coefficient. §Two-sided P values were obtained from the F test of H0: ⳱ 0, which is identical to the t test of H0: ⳱ 0. 㛳Expressed as kilocalories per person per day. ¶Socioeconomic factors derived from principal-components analysis with Varimax rotation based on the seven socioeconomic status-related variables, as described in the text and in Table 2. 1640 ARTICLES Journal of the National Cancer Institute, Vol. 90, No. 21, November 4, 1998 Table 5. Results of the general linear model based on prostate cancer mortality—disaggregating energy from soybeans from that derived from nuts and oilseeds* 95% confidence interval Independent variable Cereal energy Animal energy Nut and oilseed energy㛳 Soy energy㛳 Alcohol energy SESPC1¶ b† Lower Upper bX75 − bX25‡ P§ −0.009 0.007 −0.035 −0.013 0.001 −0.062 −0.005 0.013 −0.008 −5.62 5.53 −1.16 .0003 .03 .02 −0.147 0.024 −1.961 −0.198 0.003 −3.803 −0.096 0.043 −0.119 −0.56 3.54 −2.46 .0001 .025 .045 *In this model, the dependent variable—prostate cancer mortality rate per 100 000 population, truncated to ages 45–74 years and age standardized to the 1987 world male population—is regressed on all independent variables listed; 42 countries are included; model R2 ⳱ .90. †Regression coefficient for the independent variable. ‡bX75 − bX25 estimates the effect size of the variable listed by multiplying the regression coefficient b by the 25th percentile value of the variable and subtracting that value from the product of the regression coefficient b and the 75th percentile value of the variable. This is the effect of the variable across its interquartile range (in prostate cancer deaths per 100 000 males aged 45–74 years) in the 42 countries available for this analysis. §Two-sided P values were obtained from the F test of H0: ⳱ 0. 㛳In this analysis, soy energy is disaggregated from the heading of nut and oilseed energy. ¶SESPC1 ⳱ socioeconomic status principal component 1, computed as described in the text. Poverty-related factor. A separate model was developed (data not shown) using energy-adjusted residuals for the nutrient variables and fitting total energy intake. This approach allows the examination of total energy intake as a predictor of mortality without being constrained by the problems associated with fitting total energy intake in a model with other energy-contributing nutrients. Once again, energy from cereals, fish, nuts and oilseeds, and the poverty-related factor (SESPC1) all demonstrated a protective effect, and the magnitude of these effects was not materially different from that reported in Table 4. However, in this model, total energy intake was independently and significantly associated with an increase in prostate cancer mortality (b ⳱ 0.005; bX75 − bX25 ⳱ 3.41) with a change across the IQR of 15%— calculated as [(bX75 − bX25)/world mean prostate cancer rate] × 100—of the mean world prostate cancer mortality rate. Animal energy, even when substituted for total energy intake, exhibited no association in this model. DISCUSSION In this hypothesis-generating study, our results showed that prostate cancer mortality is positively associated with dietary factors associated with affluence, including estimated intakes of energy, total fat, and animal products (specifically, milk, meat, and poultry). On the other hand, intakes of cereals, soybeans, other nuts and oilseeds, and fish were negatively associated with prostate cancer mortality and constituted the strongest effects in the data. Our results are consistent with the findings of previous researchers (14–16,35,36), who have shown that prostate cancer mortality is increased by intakes of animal fat but not of vegetable fat. Because the etiology of prostate cancer remains unknown, the mechanism by which fat (or other dietary constituents that cor- relate with fat) from animal foods may be associated with prostate cancer is still unclear. For prostate cancer and cancers of other sites that are sensitive to serum hormone levels, animal fat may influence the risk for cancer by raising adults’ sex hormone levels. This mechanism was postulated more than 20 years ago by Hutchison (37). Consistent with this hypothesis is the fact that, in the United States, African-American men have a rate of prostate cancer 40% higher than that in whites, whereas their testosterone levels are about 15% higher (38). Providing further evidence for the role of diet in control of hormone levels, Dorgan et al. (22) demonstrated, in a controlled feeding trial involving 43 healthy men, that urinary testosterone levels were lower after a low-fat, high-fiber diet than after a high-fat, low-fiber diet. Because cancers of the prostate and the breast are at least partially hormone mediated, they share several characteristics, and the mortality rates for prostate and breast cancers are fairly highly correlated (35). One disparity in the epidemiology of the two cancer sites is that breast cancer mortality is more strongly correlated with dietary fat than is prostate cancer mortality (35). In a previous study using international data (39), we showed that energy intake from animal foods was a strong predictor of breast cancer mortality. However, the results of the present study showed a considerably smaller effect due to energy from animal sources. One way in which energy-dense, high-fat foods can exert a biologic effect is through their effect on total energy intake. In the energy-adjusted model (fitting nutrients as the residual value after removing that part of the variability explained by total energy intake), we found an association for total energy intake. However, when fitting nutrients as residuals in the model, the effect of energy from animal foods that was evident in the unadjusted model disappeared. This result suggests that the effects of energy from animal foods and total energy intake may be confounded. When the energy-adjusted animal food residual was substituted for total energy intake in this model, we found that energy from animal foods remained unrelated to prostate cancer, indicating that the original association may be due to the association of animal energy with total energy intake. These results must be interpreted with caution because energy adjustment on group level variables may not function the same as energy adjustment on the individual level. Control for confounding on the group level may not be complete and may introduce additional bias (40). The evidence from case–control and cohort studies on the impact of total energy intake is conflicting; it is not clear if energy intake per se or if the intake of energy-dense foods, such as animal products, is associated with prostate cancer. Unlike meat and poultry, fish appears to be protective against prostate cancer; this effect is perhaps due to -3 fatty acids, which appear to inhibit the growth of tumors in the prostate as well as in other sites (41,42). Our results showed a moderate, inverse association between prostate cancer mortality and energy from fish. In contrast, a study among Seventh-day Adventist men (18) showed several nonsignificant trends of increasing risk for prostate cancer due to consumption of fish, but only one statistic was significantly different from unity (relative risk [RR] ⳱ 1.68; 95% confidence interval [CI] ⳱ 1.16–2.43) for men Journal of the National Cancer Institute, Vol. 90, No. 21, November 4, 1998 ARTICLES 1641 who ate fish less frequently than once per week compared with men who never ate fish. These results need to be interpreted cautiously because the authors do not provide details regarding their dietary assessment methodology; moreover, their results for intakes of other foods (e.g., beef, poultry, and milk) do not agree with the majority of the prostate cancer literature. Also, fish was not categorized as being from fresh or salt water or by temperature of the water. All of these factors are known to influence the -3 fatty acid content of fish. The results of that study may be unique because their sample comprised a homogeneous group of predominantly white middle-class vegetarian men. In our initial model, which did not separate soybeans from other nuts and oilseeds, energy from fish was associated with prostate cancer mortality. However, when soybeans were entered separately from nuts and oilseeds, the effect of fish was not significant. Our results showed a strong correlation between energy from fish and soy products (r ⳱ .85) because countries in our dataset that had a high soy intake also had a high fish intake. Future studies assessing the role of diet in prostate cancer should ensure that a protective effect from fish is not actually due to the consumption of other foods, such as soy products. Apart from some countries where reported death rates are not reliable, the lowest death rates from prostate cancer are in the Far East (i.e., Japan and Hong Kong) (2). Although many hypotheses have been offered for the lower rates of prostate cancer mortality among Asians, the one that possibly has received the greatest amount of attention and for which the data are most consistent concerns the dietary differences between populations of Western and Eastern nations. In particular, Asians have a greater per capita consumption of soy products, consumed primarily as tofu, soymilk, tempeh, and miso. Our results showed that energy from soybeans was inversely associated with prostate cancer mortality. On a per unit energy (kilocalories) basis, the soy effect was the largest in the dataset, although its effect across the IQR of the 42 countries was relatively small, owing to the low per capita disappearance of soy products. Although the soybeans in our dataset were used for human consumption, we do not have data indicating what types of soy products were consumed (e.g., fermented products like tempeh). However, our results are similar to those of Severson et al. (17), who showed that an increased consumption of tofu was associated with a decreased risk for prostate cancer, and to those of another study (43) showing that soy products are protective against cancer at other sites. It has been proposed that certain compounds in soy, isoflavonoids and lignans, have a metabolism similar to that of biologic steroids but have a profoundly different effect on hormone-sensitive target cells (44,45). A protective effect of these compounds may be the result of a competitive binding to hormone receptors, with the weaker phytoestrogen activity displacing the harmful stimulatory effect of endogenous hormones. Isoflavones and lignans also exhibit antioxidative activity and appear to increase serum levels of sex hormonebinding globulin (SHBG) production in the liver, which in turn results in a lowering of free testosterone, thus reducing hormonal activity (45). Higher levels of SHBG and decreased concentra1642 ARTICLES tions of free testosterone have been reported both in Asian men and in men who are vegetarian (46,47). Vegetarian and Asian diets typically contain other foods, such as lentils, that have phytoestrogenic activity and other beneficial properties (48,49). However, these foods are not itemized in the Food Balance Sheets (24) because they are of minor economic importance. Besides isoflavonoids and lignans that affect hormone metabolism, soybeans (and other legumes) contain compounds that could protect through other mechanisms. These other compounds include protease inhibitors, phytic acid, and plant sterols (50). We could not examine all categories of vegetables and fruits in these data, even though we know that some of them contain isoflavonoids and other potentially protective factors (51). However, our univariate results did show a moderate, inverse association between prostate cancer mortality and cabbage consumption (Brassica sp.). The Brassica genus comprises other nutrient-dense vegetables such as kale, cauliflower, Brussels sprouts, kohlrabi, and broccoli. Although not rich in phytoestrogens, the Brassica vegetables do have a relatively high content of glucosinolates, which are hydrolyzed to produce indoles and isothiocyanates (52,53). Sulforaphane, a potent isothiocyanate that has been extracted from broccoli, blocks chemical carcinogenesis, indicating that these vegetables could have anticarcinogenic activity (54). Perhaps because they are not a plentiful source of phytoestrogens, the Brassica vegetables have not before been found to consistently protect against cancers of the prostate, ovary, and endometrium, which are hormonedependent sites; however, their chemopreventive role against cancers of the lung and the digestive system is fairly well established (52). Similar to consumption of vegetables, consumption of cereals typically has been associated with a reduced risk for cancers of the digestive system (55). Evidence is also growing that consumption of cereals is inversely related to prostate cancer mortality. Our results showed a strong negative association between cereal energy and prostate cancer (b ⳱ –0.013 and b ⳱ −0.009 for the overall and soy-specific models, respectively). Our findings support those of Rose et al. (35), who showed that cereal intake was more strongly related to prostate cancer mortality than was intake of meat or milk. The protection afforded by cereals is likely related to the consumption of traditional breads, which frequently contain whole flaxseeds, rye, and buckwheat flour (47). These foods are rich in lignans that are weaker than isoflavonoids but may have an independent effect against cancer at hormone-sensitive sites (47,51). Plant lignans are located next to the aleurone layer of the grain and are often removed in the milling processes particular to Western societies. Lignan end products are excreted when subjects have ingested the whole grain. Lignans appear to act as protease inhibitors by a mechanism similar to that of the pharmaceutical Finasteride used to treat prostate cancer (47). The combined action of both isoflavonoids and lignans, from their dietary sources of soy, legumes, cereals, and vegetables, may be protective against prostate cancer. Our results showed that consumption of milk was strongly and positively associated with prostate cancer mortality, a finding consistent with the results of other studies that have found Journal of the National Cancer Institute, Vol. 90, No. 21, November 4, 1998 that high intakes of milk and other dairy products are associated with increased risk of prostate cancer (35,36). La Vecchia et al. (16) showed that two or more servings of milk per day significantly increased risk (RR ⳱ 5.1; 95% CI ⳱ 1.5–16.6), although intakes of high-fat dairy foods like cheese and butter did not. A case–control investigation from the Roswell Park Memorial Institute (56) found a strong association of prostate cancer incidence with consumption of whole milk but not of skim milk, supporting the general hypothesis that prostate cancer is related to intake of animal fats. Although data regarding skim milk were available from the FAO dataset, we did not include this information in analyses because quantities were not specified for most countries. Also, unlike soybeans, for which the putative biologically active constituents are not represented in the summary statistics in the dataset, displacement of high-fat dairy products by skim milk products is reflected in summary measures including total fat, animal fat, and milk energy. The results of the PCA of the socioeconomic variables can be interpreted in the context of developed versus undeveloped nations. Each of the three principal components had a moderate correlation with prostate cancer mortality. Unlike the other principal components, SESPC1, which was associated with poverty factors, was negatively correlated with prostate cancer mortality. SESPC1 also was negatively associated with prostate cancer mortality in the regression models (b ⳱ −3.253 and b ⳱ −1.961 for the overall and soy-specific models, respectively). Hence, poverty is associated with lower prostate cancer mortality. This effect might be achieved because of a lower life expectancy or greater number of deaths due to other causes among populations of undeveloped nations. However, because all of our data were age adjusted and excluded countries with obvious problems in reporting vital statistics, it is unlikely that this was a major problem. Alternatively, the principal components may be acting as a proxy for the plethora of sociodemographic factors that characterize a population, such as total wealth and access to health-related resources. Further evidence for correlates of affluence exerting an effect to increase prostate cancer mortality comes from the univariate analysis of SESPC2 and SESPC3. These variables, which are related to access to medical services and affluence, respectively, were directly associated with prostate cancer mortality (r ⳱ .51 and r ⳱ .31, respectively). The datasets from the FAO and WHO are a rich source of information, but they are not without limitations. Criticisms can be made that countries differ in how accurately they report mortality data, although none of the countries included in these analyses reported having more than 15% of mortality due to senility and ill-defined causes. In addition, although the food disappearance data measure total food that is produced and imported, minus food that is exported, fed to animals, or unavailable for human use, they do not account completely for food wastage, selective avoidance of foods, or food fed to domestic animals and pets. Therefore, the data apply to populations; they cannot be interpreted as having a cause–effect relationship, and they should not be used to make dietary guidelines for individuals. Nevertheless, each method for determining the dietary intake of individuals or groups has its unique limitations; e.g., we are not convinced that the food-frequency questionnaires used in many prostate cancer studies were able to better describe the food intakes of their samples (20,57). In this study, we have used prostate cancer mortality rather than prostate cancer incidence because of the availability and the quality of the data. Clearly, these two measures are not the same because mortality is a combination of incidence, stage and age at diagnosis, and survival. Although incidence rates of prostate cancer have increased sharply in recent years among some of the more developed countries (due in part to the implementation of widespread prostate-specific antigen screening), mortality rates for this cancer have not changed appreciably (58–60). Apparently, this phenomenon is a result of the identification of earlier stage, more slowly growing neoplasms and the generally indolent nature of the disease. Moreover, the lack of a decrease in mortality rates may be due to a lag in the effect of screening (61). Analysis of ecologic data can present a problem with the misinterpretation of results analyzed at the group level as indicative of relationships on the individual level (i.e., the ecologic fallacy) (62). Aggregating data by country can obscure interesting relationships within a country, and it limits control of potential confounders and effect modifiers. In our analyses, we have attempted to account for factors related to SES and level of development by examining and controlling for the effects of the principal component factors (poverty, access to medical facilities, and affluence) in our multivariable models, while acknowledging that control in this manner cannot completely remove these biases (40). In our analyses, we have used age-standardized and truncated mortality rates to reduce the impact of reporting errors. However, there is potential for bias when regressing age-adjusted rates on crude measures of exposure (63). The extent of this bias in our study should be small because the within-country variation in dietary intake is generally much smaller than the between-country variation in dietary intake, suggesting that crude estimates and age-adjusted estimates of aggregate measures of diet would be relatively similar. Other kinds of studies would be needed to relate specific, individual exposures to prostate cancer-related outcomes. In Table 3, we have presented Pearson correlation coefficients to facilitate comparisons of our results with those from previous studies. However, caution must be used in the interpretation of these results because the associations may be overestimated as a result of the effects of aggregation bias (64). Prostate cancer is considered to be a disease of the affluent because the incidence is greater in developed nations. For example, Armstrong and Doll (9) showed that the GNP was strongly and positively correlated with prostate cancer. Among wealthy nations, the effects of occupation, environment, and lack of physical activity should not be ignored (65,66). However, even with consideration of these factors, estimates of dietary intake consistently have been shown to be associated with prostate cancer incidence and mortality. Western diets typically are rich in total energy, animal fat, and meat but low in vegetables, soy products, and other nuts and oilseeds. On the basis of the results of this study and other studies, it appears that the Western diet may contribute to risk for prostate cancer mortality. Journal of the National Cancer Institute, Vol. 90, No. 21, November 4, 1998 ARTICLES 1643 Appendix Table 1. Country-specific values of prostate cancer mortality and selected dietary variables for the 59 countries used in regression analyses Country Argentina Australia Austria Bahamas Barbados Belgium Belize Bulgaria Canada Chile Costa Rica Cuba Czechoslovakia Dominica Dominican Republic Ecuador Egypt England/Wales Federal Republic of Germany France German Democratic Republic Greece Guatemala Honduras Hong Kong Hungary Iceland Ireland Israel Italy Japan Kuwait Malta Mauritius Mexico The Netherlands Netherlands Antilles New Zealand Northern Ireland Panama Paraguay Peru Poland Portugal Republic of Korea Romania Scotland Singapore Spain Sri Lanka St. Lucia Suriname Syrian Arab Republic Thailand Trinidad/Tobago United States Uruguay Venezuela Yugoslavia Prostate cancer mortality* Total energy† Fish energy† Nut and oilseed energy† Soy energy† Alcohol energy† Animal energy† Cereal energy† 25.24 29.59 31.23 63.49 51.61 33.78 25.03 17.12 29.20 23.48 24.17 29.86 30.13 69.50 22.47 15.80 2.75 28.38 32.21 31.43 25.88 15.83 5.77 1.12 5.44 33.72 31.72 28.16 16.10 22.57 5.87 15.21 24.36 7.09 14.61 29.98 28.63 35.09 25.48 24.22 13.84 12.59 22.46 25.60 0.90 17.98 27.40 7.47 24.86 0.70 12.27 32.22 0.67 0.53 46.88 32.19 34.94 24.14 22.16 3380 3055 3575 2200 3020 3639 2714 3619 3340 2759 2653 2796 3393 2018 2130 2114 3175 3249 3351 3529 3689 3668 2138 2135 2771 3484 3087 3699 3060 3688 2852 3344 2843 2766 2890 3617 2712 3573 3249 2338 2839 2195 3479 3204 3056 3346 3249 3165 3294 2251 2390 2590 3005 2330 2837 3641 2886 2646 3550 10 22 15 19 59 28 10 12 29 38 13 34 17 38 15 35 9 22 24 34 27 29 2 3 89 8 184 27 26 25 195 20 41 39 19 17 36 13 22 14 2 49 35 47 67 15 22 63 53 31 53 46 4 37 24 23 11 21 8 14 33 43 7 51 16 48 29 68 3 21 15 24 47 27 10 20 34 39 22 16 86 10 2 81 16 7 5 88 43 115 45 37 13 24 29 7 15 34 30 50 13 1 24 113 18 34 35 46 312 20 30 93 65 56 75 9 8 11 5 0 0 — — 1 — 0 0 — 4 9 0 — 3 4 4 0 1 0 0 0 3 — 36 0 0 0 1 0 93 — — — — 2 — — 0 — 32 1 0 0 94 0 0 29 1 1 — 3 — 18 — 1 — 2 0 216 179 238 160 134 255 80 232 129 122 71 37 263 63 48 47 1 209 304 269 280 111 32 31 52 252 54 150 61 254 142 0 66 46 55 177 153 158 209 103 94 111 135 210 161 183 209 34 236 5 115 93 9 21 88 186 98 91 150 1107 1019 1233 768 850 1483 705 756 1297 447 500 650 1157 364 287 403 209 1196 1200 1343 1307 810 211 227 834 1210 1432 1315 705 913 590 870 794 311 406 1512 836 1699 1196 405 573 287 1151 609 269 823 1196 689 843 95 518 325 405 152 516 1316 1019 543 804 987 798 701 640 869 697 958 1582 662 1343 911 1050 1034 607 703 655 2032 685 696 784 907 1165 1240 1153 956 1131 604 960 1082 1282 1239 1265 980 1382 1443 631 808 762 685 909 851 960 1224 1258 2070 1455 685 1443 853 1275 661 1306 1514 1540 1140 661 938 971 1668 *Prostate cancer mortality per 100 000 male population aged 45–74 years and age standardized to the 1987 world male population. †Values are based on food disappearance data, as described in text, and prorated to an energy equivalent per capita (kilocalories per person per day). — ⳱ missing data. 1644 ARTICLES Journal of the National Cancer Institute, Vol. 90, No. 21, November 4, 1998 Appendix Table 2. Country-specific values of prostate cancer mortality and socioeconomic variables for the 59 countries used in regression analyses Country Argentina Australia Austria Bahamas Barbados Belgium Belize Bulgaria Canada Chile Costa Rica Cuba Czechoslovakia Dominica Dominican Republic Ecuador Egypt England/Wales Federal Republic of Germany France German Democratic Republic Greece Guatemala Honduras Hong Kong Hungary Iceland Ireland Israel Italy Japan Kuwait Malta Mauritius Mexico The Netherlands Netherlands Antilles New Zealand Northern Ireland Panama Paraguay Peru Poland Portugal Republic of Korea Romania Scotland Singapore Spain Sri Lanka St. Lucia Suriname Syrian Arab Republic Thailand Trinidad/Tobago United States Uruguay Venezuela Yugoslavia Prostate cancer mortality* Per capita GNP† Infant mortality‡ Hospital beds§ 25.24 29.59 31.23 63.49 51.61 33.78 25.03 17.12 29.20 23.48 24.17 29.86 30.13 69.50 22.47 15.80 2.75 28.38 32.21 31.43 25.88 15.83 5.77 1.12 5.44 33.72 31.72 28.16 16.10 22.57 5.87 15.21 24.36 7.09 14.61 29.98 28.63 35.09 25.48 24.22 13.84 12.59 22.46 25.60 0.90 17.98 27.40 7.47 24.86 0.70 12.27 32.22 0.67 0.53 46.88 32.19 34.94 24.14 22.16 2520 11 140 11 140 4650 3040 10 760 1000 2625 10 193 2210 1430 840 5800 460 1330 1350 690 9660 12 460 11 680 5340 4290 1130 660 5340 2270 9000 5150 5090 6840 10 080 19 870 2036 1052 2270 10 930 10 930 7920 5150 2120 1610 1310 4670 2450 1187 2560 9660 5910 5430 320 698 2600 1680 790 6840 13 160 2650 4140 2800 44 10 10 32 28 12 28 20 15 27 18 17 16 21 65 68 104 11 12 10 12 14 66 83 10 20 10 11 16 14 7 32 15 33 53 8 8 12 11 33 45 83 20 26 32 29 11 11 10 32 22 30 58 51 26 11 34 39 34 56 124 113 40 87 89 45 87 87 35 35 41 123 43 29 20 21 86 118 106 107 64 22 14 43 88 172 105 56 104 106 39 106 35 12 101 92 102 112 39 14 18 76 53 16 92 116 38 54 29 51 54 10 12 45 63 43 31 60 Physicians§ Percentage without sanitation Percentage without safe water Life expectancy㛳 19 15 23 7 7 23 3 23 18 6 7 9 25 2 5 6 9 15 20 16 19 22 2 3 9 23 17 12 5 21 12 13 13 4 8 17 6 14 15 8 5 6 17 14 5 14 17 8 18 2 4 5 4 1 5 17 14 11 13 21 0 0 0 0 0 0 0 0 20 13 0 0 0 85 75 74 0 0 0 0 0 71 65 7 0 74 0 0 0 0 0 4 6 63 0 0 0 0 55 9 63 0 0 0 0 0 20 0 32 0 0 55 54 7 0 41 15 0 43 0 0 0 0 0 0 0 0 16 16 0 0 0 41 55 25 0 0 0 0 0 55 56 0 0 62 0 0 0 0 0 0 1 44 0 0 0 0 18 79 49 0 0 22 0 0 0 0 67 0 0 29 37 2 0 20 19 0 70 74 74 66 71 73 60 72 73 70 74 75 72 58 62 63 57 74 73 75 73 74 60 60 75 71 76 73 74 74 77 71 71 69 65 76 76 73 73 71 65 58 72 71 64 71 74 72 74 69 68 66 66 63 68 75 73 68 71 *Prostate cancer mortality per 100 000 male population aged 45–74 years and age standardized to the 1987 world male population. †Per capita gross national product in U.S. dollars (1980). ‡Rate per 1000 live births (1980). §Per 10 000 population. 㛳Life expectancy, in years, at birth (1980). Journal of the National Cancer Institute, Vol. 90, No. 21, November 4, 1998 ARTICLES 1645 REFERENCES (1) Parkin DM, Muir CS, Whelan SL, Gao YT, Ferlay J, Powell J. Cancer incidence in five continents. Lyon: International Agency for Research on Cancer; 1992. (2) Levi F, Lucchini F, LaVecchia C. Worldwide patterns of cancer mortality, 1985–89. Eur J Cancer Prev 1994;3:109–43. (3) Buell P. Changing incidence of breast cancer in Japanese-American women. J Natl Cancer Inst 1973;51:1479–83. (4) Haenszel W, Kurihara M. Studies of Japanese migrants. I. Mortality from cancer and other diseases among Japanese in the United States. J Natl Cancer Inst 1968;40:43–68. (5) King H, Li JY, Locke FB, Pollack ES, Tu JT. Pattern of site-specific displacement in cancer mortality among migrants: the Chinese in the United States. Am J Public Health 1985;75:237–42. (6) Staszewski J, Haenszel W. Cancer mortality among the Polish-born in the United States. J Natl Cancer Inst 1965;35:291–7. (7) Prentice RL, Sheppard L. Dietary fat and cancer: consistency of the epidemiologic data, and disease prevention that may follow from a practical reduction in fat consumption [published erratum appears in Cancer Causes Control 1990;1:253]. Cancer Causes Control 1990;1:81–97. (8) Hursting SD, Thornquist M, Henderson MM. Types of dietary fat and the incidence of cancer at five sites. Prev Med 1990;19:242–53. (9) Armstrong B, Doll R. Environmental factors and cancer incidence and mortality in different countries, with special reference to dietary practices. Int J Cancer 1975;15:617–31. (10) Correa P. Epidemiological correlations between diet and cancer frequency. Cancer Res 1981;41:3685–90. (11) Kosary CL, Ries LA, Miller BA, Hankey BF, Harras A, Edwards BK. SEER cancer statistics review, 1973–1992. Bethesda (MD): National Institutes of Health, National Cancer Institute; 1995. (12) Ries L, Hankey B, Miller B, Hartman A, Edwards B. Cancer statistics review, 1973–88. Bethesda (MD): National Institutes of Health, National Cancer Institute; 1991 Report No.: DHHS Publ No. (NIH)91-2789. (13) Mettlin C. Recent developments in the epidemiology of prostate cancer. Eur J Cancer 1997;33:340–7. (14) Le Marchand L, Kolonel LN, Wilkens LR, Myers BC, Hirohata T. Animal fat consumption and prostate cancer: a prospective study in Hawaii. Epidemiology 1994;5:276–82. (15) Giovannucci E, Rimm EB, Colditz GA, Stampfer MJ, Ascherio A, Chute CC, et al. A prospective study of dietary fat and risk of prostate cancer. J Natl Cancer Inst 1993;85:1571–9. (16) La Vecchia C, Negri E, D’Avanzo B, Franceschi S, Boyle P. Dairy products and the risk of prostatic cancer. Oncology 1991:406–10. (17) Severson RK, Nomura AM, Grove JS, Stemmermann GN. A prospective study of demographics, diet, and prostate cancer among men of Japanese ancestry in Hawaii. Cancer Res 1989;49:1857–60. (18) Mills PK, Beeson WL, Phillips RL, Fraser GE. Cohort study of diet, lifestyle, and prostate cancer in Adventist men. Cancer 1989;64:598–604. (19) Graham S, Haughey B, Marshall J, Priore R, Byers T, Rzepka T, et al. Diet in the epidemiology of carcinoma of the prostate gland. J Natl Cancer Inst 1983;70:687–92. (20) Hebert JR, Miller DR. Methodologic considerations for investigating the diet–cancer link. Am J Clin Nutr 1988;47:1068–77. (21) Wynder EL, Hebert JR. Homogeneity in nutritional exposure: an impediment in cancer epidemiology [letter]. J Natl Cancer Inst 1987;79:605–7. (22) Dorgan JF, Judd JT, Longcope C, Brown C, Schatzkin A, Celvidence BA, et al. Effects of dietary fat and fiber on plasma and urine androgens and estrogens in men: a controlled feeding study. Am J Clin Nutr 1996;64: 850–5. (23) Health statistical data collections—data tape on: mortality by cause of death by sex and age group and population by sex and age group. 1985–1990. Lyon: World Health Organization (International Agency for Research on Cancer), 1991. (24) Food Balance Sheets 1979–81, Average. Rome: Food and Agricultural Organization, 1981. (25) The international drinking water supply and sanitation decade: review of national baseline data. Geneva: World Health Organization; 1984. (26) World development report. New York and Oxford: World Bank; 1984. (27) World development report. New York and Oxford: World Bank; 1993. 1646 ARTICLES (28) World Health Statistics Quarterly. Vol 33. Geneva: World Health Organization; 1980. (29) Tukey JW. Exploratory data analysis. Reading (MA): Addison-Wesley; 1977. (30) Kleinbaum DG, Kupper LL. Factor analysis. Applied regression analysis and other multivariable methods. 1st ed. North Scituate (MA): Duxbury Press; 1978. p. 376–413. (31) Hebert JR, Stoddard AM, Harris DR, Sorensen G, Hunt MK, Morris DH, et al. Measuring the effect of a worksite-based nutrition intervention on food consumption. Ann Epidemiol 1993;3:629–35. (32) SAS. SAS User’s Guide. Cary (NC): SAS Institute; 1997. (33) Kleinbaum DG, Kupper LL, Muller K. Applied regression analysis and other multivariable methods. 2nd ed. Boston (MA): PWS–KENT Publishing Co.; 1988. (34) Willett WC. Nutritional epidemiology. Oxford: Oxford University Press; 1990. (35) Rose DP, Boyar AP, Wynder EL. International comparisons of mortality rates for cancer of the breast, ovary, prostate, and colon, and per capita food consumption. Cancer 1986;58:2363–71. (36) Serra-Majem L, La Vecchia C, Ribas-Barba L, Prieto-Ramos F, Lucchini F, Ramon JM, et al. Changes in diet and mortality from selected cancers in southern Mediterranean countries, 1960–1989. Eur J Clin Nutr 1993;47 Suppl 1:S25–34. (37) Hutchison GB. Epidemiology of prostate cancer. Semin Oncol 1976;3: 151–9. (38) Pienta KJ, Goodson JA, Esper PS. Epidemiology of prostate cancer: molecular and environmental clues. Urology 1996;48:676–83. (39) Hebert JR, Rosen A. Nutritional, socioeconomic, and reproductive factors in relation to female breast cancer mortality: findings from a cross-national study. Cancer Detect Prev 1996;20:234–44. (40) Greenland S, Morgenstern H. Ecological bias, confounding, and effect modification [published erratum appears in Int J Epidemiol 1991;20:824]. Int J Epidemiol 1989;18:269–74. (41) Cave WT Jr. Dietary n-3 (-3) polyunsaturated fatty acid effects on animal tumorigenesis. FASEB J 1991;5:2160–6. (42) Steinmetz KA, Potter JD. Food-group consumption and colon cancer in the Adelaide case–control study. II. Meat, poultry, seafood, dairy foods, and eggs. Int J Cancer 1993;53:730–7. (43) Persky V, van Horn L. Epidemiology of soy and cancer: perspectives and directions. J Nutr 1995;125(3 Suppl):709S–712S. (44) Rose DP. Diet, hormones, and cancer. Annu Rev Public Health 1993;14: 1–17. (45) Adlercreutz H. Phytoestrogens: epidemiology and a possible role in cancer protection. Environ Health Perspect 1995;103 Suppl 7:103–12. (46) Adlercreutz H. Evolution, nutrition, intestinal microflora, and prevention of cancer: a hypothesis. Proc Soc Exp Biol Med 1998;217:241–6. (47) Morton MS, Griffiths K, Blacklock N. The preventive role of diet in prostatic cancer. Br J Urol 1995;77:481–93. (48) Thompson LU, Robb P, Serraino M, Cheung F. Mammalian lignan production from various foods. Nutr Cancer 1991;16:43–52. (49) Reinli K, Block G. Phytoestrogen content of foods—a compendium of literature values. Nutr Cancer 1996;26:123–48. (50) Kennedy AR. The evidence for soybean products as cancer preventive agents. J Nutr 1995;125(3 Suppl):733S–743S. (51) Hutchins AM, Lampe JW, Martini MC, Campbell DR, Slavin JL. Vegetables, fruits, and legumes: effect of urinary isoflavonoid phytoestrogens and lignan excretion. J Am Diet Assoc 1995;95:769–74. (52) Verhoeven DT, Goldbohm RA, van Poppel G, Verhagen H, van den Brandt PA. Epidemiological studies on brassica vegetables and cancer risk. Cancer Epidemiol Biomarkers Prev 1996;5:733–48. (53) Witte JS, Longnecker MP, Bird CL, Lee ER, Frankl HD, Haile RW. Relation of vegetable, fruit, and grain consumption to colorectal adenomatous polyps. Am J Epidemiol 1996;144:1015–25. (54) Zhang Y, Talalay P. Anticarcinogenic activities of organic isothiocyanates: chemistry and mechanisms. Cancer Res 1994;54(7 Suppl):1976s–1981s. (55) Jacobs DR Jr, Slavin J, Marquart L. Whole grain intake and cancer: a review of the literature. Nutr Cancer 1995;24:221–9. (56) Mettlin C, Selenskas S, Natarajan N, Huben R. Beta-carotene and animal fats and their relationship to prostate cancer risk. A case–control study. Cancer 1989;64:605–12. Journal of the National Cancer Institute, Vol. 90, No. 21, November 4, 1998 (57) Hegsted DM. Errors of measurement [editorial]. Nutr Cancer 1989;12: 105–7. (58) Merrill RM, Brawley OW. Prostate cancer incidence and mortality rates among white and black men. Epidemiology 1997;8:126–31. (59) Nakata S, Sato J, Imai K, Yamanada H, Ichinose Y. Epidemiological characteristics of prostate cancer in Gunma Prefecture, Japan. Gunma University Urological Oncology Study Group. Int J Urol 1995;2:191–7. (60) McCaul KA, Luke CG, Roder DM. Trends in prostate cancer incidence and mortality rates in South Australia, 1993–1997. Med J Aust 1995;162: 520–2. (61) Gann PH. Interpreting recent trends in prostate cancer incidence and mortality [editorial]. Epidemiology 1997;8:117–20. (62) Morgenstern H. Uses of ecologic analysis in epidemiolgic research. Am J Public Health 1982;72:1136–44. (63) Rosenbaum PR, Rubin DB. Difficulties with regression analyses of ageadjusted rates. Biometrics 1984;40:437–43. (64) Langbein LI, Lichtman AJ. Ecological inference. Beverly Hills (CA): Sage Publications; 1978. (65) Boyle P, Zaridze DG. Risk factors for prostate and testicular cancer. Eur J Cancer 1993;29A:1048–55. (66) Lee IM, Paffenbarger RS Jr, Hsieh CC. Physical activity and risk of prostatic cancer among college alumni. Am J Epidemiol 1992;135:169–79. NOTE Manuscript received July 21, 1997; revised August 14, 1998; accepted August 27, 1998. Journal of the National Cancer Institute, Vol. 90, No. 21, November 4, 1998 ARTICLES 1647