Dichlorodiphenyldichloroethane burden and breast cancer risk: a meta-analysis of the epidemiologic evidence.

The relationship of dichlorodiphenyltrichloroethane (DDT) exposure and breast cancer risk has received increasing attention since the beginning of the 1990s. Contradicting published results regarding the relationship between body burden levels of p,p'-dichlorodiphenyldichloroethane (p,p'-DDE)--the main DDT metabolite--and breast cancer, we argue that such differences stem from methodologic differences among those studies. We performed a meta-analysis of 22 articles using DerSimonian and Laird's method for random effects models. The Q-statistic was used to identify heterogeneity in the outcome variable across studies. The gradient of p,p'-DDE exposure in epidemiologic studies was homogenized to serum lipid bases (nanograms per gram). The potential for publication bias was examined by means of the Begg's test. We discuss methodologic features of the studies in an attempt to reconcile the findings. The summary odds ratio (OR) for selected studies was 0.97 (95% confidence interval, 0.87-1.09) and the gradient of exposure ranged from 84.37 to 12,948 ng/g. No overall heterogeneity in the OR was observed (chi-squared = 27.93; df = 23; p = 0.218). Neither the study design nor the lack of breast-feeding control or the type of biologic specimen used to measure p,p'-DDE levels were the causes of heterogeneity throughout the studies. Evidence for publication bias was not found (p = 0.253). Overall, these results should be regarded as a strong evidence to discard the putative relationship between p,p'-DDE and breast cancer risk. Nevertheless, the exposure to DDT during critical periods of human development--from conception to adolescence--and individual variations in metabolizing enzymes of DDT or its derivatives are still important areas to be researched in regard to breast cancer development in adulthood.

The insecticide dichlorodiphenyltrichloroethane (DDT) was first synthesized in 1874 and reached worldwide use during the early 1960s to control malaria and some agricultural pests. In 1972 the use of DDT was banned in the United States, and by the beginning of the 1980s this chemical was prohibited in most developed countries. However in India, Indonesia, and Italy, DDT was still produced until 1990, and in Mexico DDT was in use until 1997 to control malaria (Turusov et al. 2002). The long persistence and environmental spreading exhibited by DDT and its metabolites, along with their estrogenic potential, are the main concerns regarding its potential role in the etiology of estrogen-related malignant tumors (Snedeker 2001).
In 1993, Wolff et al. (1993) first reported on the presumed positive association between p,p´-dichlorodiphenyldichloroethane (p,p´-DDE)-the main metabolite of DDTand breast cancer, and subsequently the assessment of the body burden of DDT metabolites in relation to breast cancer risk received a lot of attention. During the last decade, results from several epidemiologic studies were published but most were unable to replicate the positive association between p,p´-DDE and breast cancer risk. Among other explanations for such negative results, some suggestedbut never proved-that single studies lacked an adequate gradient of exposure to p,p´-DDE among breast cancer cases and controls, and this flaw obscured the differences (Talbott et al. 1998). As demonstrated below, the gradient exposure in 22 published studies ranged from 84.37 to 12928.08 ng/g Dello Iacovo et al. 1999;Demers et al. 2000;Dorgan et al. 1999;Helzlsouer et al. 1999;Hoyer et al. 1998Hoyer et al. , 2000bHunter et al. 1997;Krieger et al. 1994;Laden et al. 2001b;López-Carrillo et al. 1997;Mendonca et al. 1999;Millikan et al. 2000;Moysich et al. 1998;Romieu et al. 2000;Stellman et al. 2000;van´t Veer et al. 1997;Wolff et al. 1993Wolff et al. , 2000aWolff et al. , 2000bZheng et al. 1999Zheng et al. , 2000, in sharp contrast to the narrower gradients achieved by most single studies.
Serum and adipose tissue were the human biologic matrices used to estimate p,p´-DDE body burden and its potential relationship with breast cancer risk. Circulating lipids influence blood levels of DDT metabolites (Phillips et al. 1989), yet approaches to this condition varied greatly in the published scientific literature. Some studies reported p,p´-DDE serum levels in lipid bases (Demers et al. 2000;Dorgan et al. 1999;Helzlsouer et al. 1999;Hoyer et al. 1998Hoyer et al. , 2000aHoyer et al. , 2000bLaden et al. 2001;López-Carrillo et al. 1997;Millikan et al. 2000;Romieu et al. 2000;Ward et al. 2000;Wolff et al. 2000aWolff et al. , 2000bZheng et al. 2000), whereas others performed an indirect adjustment by fitting a cholesterol term in linear regression models (Dello Iacovo et al. 1999;Hunter et al. 1997;Moysich et al. 1998) and the rest only provided wet-based measurements (Krieger et al. 1994;Mendonca et al. 1999;Olaya-Contreras et al. 1998;Schecter et al. 1997;Wolff et al. 1993). This heterogeneity among biologic matrices and reported units of cumulative p,p´-DDE levels limited the ability to evaluate the gradient of p,p´-DDE body burden levels across the epidemiologic studies published so far.
In this article we aim to a) estimate the strength of the association between p,p´-DDE and breast cancer on the basis of published epidemiologic evidence; b) identify the gradient of exposure that was captured in the same epidemiologic studies; and c) discuss the consistency of published results in the context of their main methodologic features.

Materials and Methods
We searched for the epidemiologic evidence on p,p´-DDE and breast cancer risk in both the MEDLINE and PubMed databases (www.ncbi.nlm.nih.gov). A total of 35 analytic studies Bagga et al. 2000;Dello Iacovo et al. 1999;Demers et al. 2000;Dewailly et al. 1994;Dorgan et al. 1999;Duell et al. 2000;Falck et al. 1992;Güttes et al. 1998;Helzlsouer et al. 1999;Hoyer et al. 1998Hoyer et al. , 2000aHoyer et al. , 2000bHunter et al. 1997;Krieger et al. 1994;Laden et al. 2001a;Liljegren et al. 1998;López-Carrillo et al. 1997;Mendonca et al. 1999;Millikan et al. 2000;Moysich et al. 1998;Olaya-Contreras et al, 1998;Romieu et al. 2000;Schecter et al. 1997;Stellman et al. 2000;Unger et al. 1982Unger et al. , 1984van´t Veer et al. in English up to February 2001 were found using the following MeSH headings, key, and text words: breast cancer, organochlorines, pesticides. The articles identified were then reviewed to determine whether they met the following inclusion criteria for statistical analyses: to be epidemiologic cohort or case-control studies; to have enrolled at least 50 cases; to have reported measures of association and confidence intervals (CIs) for breast cancer risk; to have measured p,p´-DDE levels in biologic samples (serum or adipose tissue); and to have been published in journals listed by the Journal Citation Reports-Science Edition (JCR) (1999).
From each eligible report and using a predefined review form, two independent reviewers extracted the following methodologic information: name of the author, year and place of publication, epidemiologic design, type of controls and biologic specimens, confounding variables considered in the analysis, and the measure of association estimated for the highest versus lowest category of exposure along with the corresponding CI.
After their extraction, we entered relevant data into evidence tables. We then performed a meta-analysis using the method of the inverse of variance for fixed-effects models and the DerSimonian and Laird method for random-effects models. (DerSimonian and Laird 1986). Separate odds ratios (ORs) were used in the meta-analysis for one article that reported estimates from population-based and hospital controls , and the same was done with estimates from one study in which serum samples were taken and analyzed for two different periods of time . The results are displayed as summary ORs and 95% CIs for the effect of p,p´-DDE on breast cancer, corresponding to the contrast of the highest versus the lowest level of p,p´-DDE exposure.
We plotted the outcomes for included studies for visual examination and performed meta-analysis regression using the Q-statistic to identify heterogeneity in the outcome variable across studies (Berkey et al. 1995;DerSimonian and Laird 1986). Potential sources of heterogeneity were evaluated, including the study design, control for breastfeeding and the kind of biologic specimen in which the DDT metabolites were measured. We assessed the potential for publication bias using a funnel plot in conjunction with the Begg's test, which is based on the fact that smaller studies tend to have larger effect size estimates and the publication bias induces a correlation between the effect estimates and their variances (Begg 1985(Begg , 1994.
To estimate the trend of p,p´-DDE body burden evaluated by the epidemiologic studies analyzed, we determined the crude mean p,p´-DDE levels among cases and controls reported by each study and homogenized Abbreviations: BMI, body mass index; ln, natural logarithm. a In design or analysis. b Date in which blood sample was returned, time of day that blood sample was obtained, fasting status at blood sampling and for postmenopausal homone use, BMI at blood collection, history of benign breast disease. c Number and dates of blood donations, day of menstrual cycle for premenopausal women, ln height, ln (BMI) -menopausal status at blood donation interaction. d Vital statistics at time of diagnosis and weight. e Year of blood draw and history of benign disease at the time of diagnosis. f Date of blood donation and day of menstrual cycle, race, BMI at age 20 or current. g Date of examination and vital status at the time of diagnosis, weight, height, alcohol consumption, smoking, physical activity, income, marital status, and education. h Serum lipids, month in which the blood sample was returned, time of day that the blood sample was obtained, fasting status at blood sampling, postmenopausal hormone use, history of benign breast disease, BMI. i Date of examination, length of follow-up after examination, race, date of joining the Kaiser Permanente Medical Care Program, year of multiphasic examination and BMI. j Number and date of blood donation, if premenopausal women: day of menstrual cycle at the time of the first blood drawing. them to serum p,p´-DDE levels in lipid bases (nanograms per gram) as follows: the arithmetic mean serum levels of p,p´-DDE in wet bases (nanograms per milliliter) were multiplied by a factor of 129.8 to convert them to the arithmetic mean of serum levels in lipid basis (nanograms per gram), and otherwise the arithmetic means of adipose tissue levels of p,p´-DDE were divided by a factor of 4.2 to estimate the corresponding serum levels in lipid basis (nanograms per gram) (López- . The percent of recovery of p,p´-DDE levels was not considered. Five articles did not provide mean values of p,p´-DDE and thus were not included (Dorgan et al 1999;Hoyer et al 1998Hoyer et al , 2000bLaden et al. 2001b;Stellman et al. 2000); also not included were two others that reported adjusted mean values of p,p´-DDE (Zheng et al. 1999(Zheng et al. , 2000 and three in which the p,p´-DDE levels were statistically modeled through the contents of triglycerides, serum, and total cholesterol (Dello Iacovo et al. 1999;Hunter et al. 1997;Moysich et al. 1998). Therefore, we included 12 studies in this step of the analysis. We estimated the trend of the mean p,p´-DDE body burden levels in nanograms per gram according to the year when the biologic samples were collected by linear regression.
To evaluate the gradient of p,p´-DDE body burden captured by studies of interest, we plotted the middle point of p,p´-DDE levels in nanograms per gram in serum (according to the methodology already described) for each category of exposure, against the corresponding ORs reported by 17 studies. We did not include two studies because no information on the magnitude of p,p´-DDE quartile distribution was provided (Hoyer et al. 1998(Hoyer et al. , 2000b and three studies in which p,p´-DDE was lipid-adjusted by regression methods (Dello Iacovo et al. 1999;Hunter et al. 1997;Moysich et al. 1998). All the statistical analyses were performed using the software Stata release 7.0 (Stata Corp., College Station, TX, USA).

Results
Tables 1, 2 and 3 describe the 22 studies that were included in the meta-analysis. All were case-control studies, and of these nine were prospective (nested case-control) (Dorgan et al. 1999;Helzlsouer et al. 1999;Hoyer et al. 1998Hoyer et al. , 2000bHunter et al. 1997;Krieger et al. 1994;Laden et al. 2001b;Wolff et al. 1993Wolff et al. , 2000b and 13 retrospective Dello Iacovo et al. 1999;Demers et al. 2000;López-Carrillo et al. 1997;Mendonca et al. 1999;Millikan et al. 2000;Moysich et al. 1998;Romieu et al. 2000;Stellman et al. 2000;van´t Veer et al. 1997;Wolff et al. 2000a;Zheng et al. 1999Zheng et al. , 2000. Among the retrospective studies, four were populationbased case-control studies (Dello Iacovo et al. 1999;Millikan et al. 2000;Moysich et al. 1998;van´t Veer et al. 1997) and seven were clinic-based case-control studies López-Carrillo et al. 1997Stellman et al. 2000;Wolff et al. 2000a;Zheng et al. 1999Zheng et al. , 2000; in one study only a subsample of a populationbased case-control study population was analyzed ; and another study included two types of controls: population and clinical . All the studies are presented in the tables according to decreasing date of publication and design features.
The inclusion criteria of the referent groups varied among studies. Individuals with skin cancer were accepted in the control group by some studies, (Dorgan et al. 1999;Helzlsouer et al. 1999;Wolff et al. 2000a) whereas in others controls had been diagnosed with benign breast disease (Aronson et al. 2000;Stellman et al. 2000;Wolff et al. 2000a), and in the remaining studies only healthy individuals and/or subjects with no cancer diagnosis made up the comparison group (Dello Iacovo et al. 1999;Demers et al. 2000;Hoyer et al. 1998Hoyer et al. , 2000bHunter et al. 1997;Krieger et al. 1994;Laden et al. 2001b;López-Carrillo et al. 1997;Mendonca et al. 1999;Millikan et al. 2000;Moysich et al. 1998;Romieu et al. 2000;van´t Veer et al. 1997;Wolff et al. 1993Wolff et al. , 2000b.
Overall, the data provided by the published studies do not support an association between p,p´-DDE body burden levels and breast cancer risk, because the summary OR was 0.97 (95% CI, 0.87-1.09) (Figure 1). We found no evidence for significant overall heterogeneity in the OR [χ 2 = 27.93; degrees of freedom (df) = 23; p = 0.218].
In Figure 3 we depict a significantly decreasing trend (β = -130.59; p = 0.001) of the mean levels of p,p´-DDE that were evaluated by the studies analyzed, according to the date when the biologic sample was collected. All those levels were converted to the corresponding equivalent nanograms per gram in lipid serum bases. Significantly, p,p´-DDE levels as reported by the study carried out in Mexico City by Romieu et al. (2000), were at great variation with the other studies performed at about the same time, including the one performed in that same city (López-Carrillo et al. 1997).
In Figure 4 we present the ORs for the effect of p,p´-DDE on breast cancer risk from each study, according to the gradient of the p,p´-DDE body burden levels expressed as nanograms per gram in serum lipid bases (middle point of p,p´-DDE levels for each category of exposure). The range of that gradient varied from 84.37 ng/g in the study performed by Mendonca et al. (1999), to 12928.08 ng/g in the study by Krieger et al. (1994). As shown in Figure 4, most studies reported p,p´-DDE body burden levels in the range of 84.37-9,000 ng/g and yielded negative results, with two exceptions Wolff et al. 1993). The studies that had the highest levels of p,p´-DDE body burden levels (9,001-12928.08 ng/g) did not show an increasing risk of breast cancer due to p,p´-DDE body burden levels (Dorgan et al. 1999;Krieger et al. 1994;van´t Veer et al. 1997).

Discussion
The results of the meta-analysis of 22 studies showed no evidence for an association between p,p´-DDE body burden levels and breast cancer risk. The summary OR reported in this manuscript was 0.97 (95% CI, 0.87-1.09), very similar to the one recently estimated from a pooled analysis of five studies (OR = 0.99; 95% CI 0.77-1.27) performed in the United States (Laden et al. 2001a) and had the same covariates to be included in the logistic models. Some studies (Hunter 1997, Laden 2001a, Wolff 1993, 2000b apparently made repeated use of some subjects as part of their study populations in subsequent papers. Hence, one additional check was to remove in subsequent steps those studies for which we believed that such condition could be met, and the estimates of summary ORs remained almost unaltered (data not shown).
An intrinsic flaw in many environmental epidemiologic studies is the lack of an adequate gradient of exposure both within and throughout the different populations studied. On average, the difference between the highest and the lowest levels of p,p´-DDE in the 22 studies was 6928.92 ± 6414.5 ng/g. The studies that had the widest internal gradient of exposure were negative (Dorgan et al. 1999;Krieger et al. 1994) as were the studies that reported the highest levels of p,p´-DDE. (Dorgan et al. 1999;Krieger et al. 1994). In this regard, evidence from occupational studies, which evaluated much higher levels of p,p´-DDE exposure, does not suggest a high risk for breast cancer (Austin et al. 1989;Fleming et al. 1999). Hence, we believe we can rule out the possibility that contradictory results among the 22 studies are caused by differences in p,p´-DDE levels.
Methodologic features among the studies that may explain the contradictory results include differences in the temporal relationship between the measurement of p,p´-DDE Article | A meta-analysis of DDT and breast cancer risk Environmental Health Perspectives • VOLUME 112 |  Confounding is a potential explanation for inconsistent epidemiologic results. Among the confounders that might distort the relationship between p,p´-DDE and breast cancer risk are breast-feeding and diet. Lactation is a way of eliminating the body burden levels of p,p´-DDE (López- Carrillo et al. 2001) and has been found to decrease the risk of breast cancer in several studies (Romieu et al. 1996). However, we found no heterogeneity in our meta-analysis to assign explanatory relevance to the lack of control by this variable in some of the published studies. Yet this analysis is not enough to rule out the possibility that equivocal results might be partially explained by differences in the ranges of values for the adjustment variables across the studies; moreover, measurement error in the confounding variables is likely to result in unpredictably biased estimates of effect for the main variable of interest when adjustments are performed.
Residues of p,p´-DDE were reportedly found in several foods (fish, dairy products, meat) (Galván- Portillo et al. 2002), and the consumption of some of them may be related to breast cancer risk. For example, meat intake is related to an excess risk for breast cancer, whereas fish intake, presumably because of the presence of omega-3 fatty acids, seems to be inversely related to breast cancer incidence (World Cancer Research Fund 1997). A study performed by Verma et al. (1997) showed that genistein, an isoflavonoid present in soybeans, and curcumin, a component of turmeric powder and also a widely used spice, can inhibit the action of pesticides with estrogenic activity. The great variation in breast cancer risk raises the possibility that dietary factors are related to its etiology. In this regard, dietary factors and particularly specific compounds such as phytoestrogen were scarcely or not at all taken into account as covariates in studies of p,p´-DDE and breast cancer risk; thus, the lack of adjustment by these variables might partially explain the equivocal results so far available.
Another methodologic issue of concern is the type of controls that were enrolled-i.e., hospital or population based-in that one should expect that p,p´-DDE body burden levels were not related to the diseases identified among clinical controls, and also that population controls actually constitute a representative sample of the p,p´-DDE body burden levels present in the target population (Rothman and Greenland 1998). We were not able to find heterogeneity according to the study design, except for a borderline significant result within retrospective population-based case-control studies (χ 2 = 11.23; df = 5; p = 0.047). And this variation mainly arose from the study by van't Veer et al. (1997), which assembled a referent group that combined hospital-and populationbased controls and provided no explanation for such an unusual combination.   (2000) contributed data for clinical and population controls, which were considered separately. That was also the case for Helzlsouer et al. (1999), in which serum samples were reported for two different moments in time: 1974 and 1989. Thus the sample size for the analysis became 24.