American Journal of Epidemiology Advance Access originally published online on April 10, 2008
American Journal of Epidemiology 2008 167(9):1017-1026; doi:10.1093/aje/kwn005
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
How Much of the Data Published in Observational Studies of the Association between Diet and Prostate or Bladder Cancer Is Usable for Meta-Analysis?
1 Department of Social Medicine, University of Bristol, Bristol, United Kingdom
2 Division of Maxillofacial Surgery, University of Bristol, Bristol, United Kingdom
3 Department of Oral and Dental Science, University of Bristol, Bristol, United Kingdom
4 School of Population Health, University of Queensland, Queensland, Australia
Correspondence to Prof. Jonathan A. C. Sterne, Department of Social Medicine, University of Bristol, Canynge Hall, Whiteladies Road, Bristol BS8 2PR, United Kingdom (e-mail: jonathan.sterne{at}bristol.ac.uk).
Received for publication September 16, 2007. Accepted for publication January 3, 2008.
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
Epidemiologic investigations often report dose-response associations, which may be combined in meta-analyses. The authors examined how often the log odds, risk, or hazard ratio per unit increase in exposure, and its standard error, can be estimated from results reported from observational studies of diet and prostate or bladder cancer so that results are usable in meta-analyses estimating dose-response associations. Eight electronic databases were searched for studies reporting on the association of diet, nutrition, or physical activity with these cancers. A total of 767 papers reported 3,284 results; 1,999 (61%) results, reported in 545 (71%) papers, were usable in dose-response meta-analyses. The most important reason that results were not usable was the absence of sufficient information on exposure levels in the different groups. The proportion of results usable in "high-low" meta-analyses (comparisons of extreme categories) was similar (62%). Results that showed evidence of an association were more likely to be usable than results that found no such evidence. Insufficient detail in reporting of results of observational studies can lead to exclusion of these results from meta-analyses and is an important threat to the validity of systematic reviews of such research. Results providing evidence of associations may be overrepresented in meta-analyses of observational studies.
case-control studies; cohort studies; meta-analysis; publication bias
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
Systematic reviews and meta-analyses of observational studies are increasingly used to address etiologic research questions (1). Such meta-analyses often pose challenges additional to those encountered in reviews of randomized controlled trials, since exposure effect estimates from observational studies may be affected by confounding, selection bias, and error in measurement of exposure variables (2). The existence of a dose-response relation between exposure and outcome can help address uncertainties about misclassification effects and helps strengthen causal reasoning (3). It is therefore desirable that meta-analyses examine the evidence for dose-response effects rather than focus solely on simple binary ("high-low") comparisons (4–6).
Dose-response meta-analyses typically assume a linear association between exposure and the log of the odds, risk, or rate of disease. However, the data needed to estimate such an association are often incompletely reported, which may result in exclusion of results from meta-analyses. Failure to include all available evidence will reduce precision of summary estimates and may also lead to bias if propensity to report results in sufficient detail is associated with the magnitude and/or direction of associations. We quantified the proportion of results of observational studies of the association of diet, nutrition, and physical activity with prostate or bladder cancer that can be used in meta-analyses estimating dose-response associations and examined whether there was an association between the direction and precision of the result and the probability of the result being usable for meta-analysis.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
This work was based on two of a series of systematic reviews of the association of diet with different types of cancer (7). The statistical methods described here were developed to facilitate statistical analysis of data arising from these reviews.
We searched MEDLINE (National Library of Medicine, Bethesda, Maryland), EMBASE (Excerpta Medica Database; Elsevier, Amsterdam, the Netherlands), The Cochrane Library (The Cochrane Collaboration, Oxfordshire, United Kingdom), Pascal (Ovid Technologies, Inc., New York, New York), BIOSIS (Thomson Scientific, Stamford, Connecticut), ISI Web of Science (Thomson Scientific), LILACS (Latin American and Caribbean Health Sciences Literature; http://bvsmodelo.bvsalud.org/site/lilacs/I/ililacs.htm), and CAB Abstracts (Oxfordshire, United Kingdom; www.cabi.org/cababstracts) from inception to February 2004 for published studies that reported an association of diet, nutrition, and physical activity with the risk of cancer of the prostate or bladder. Details are available in an online supplement to this paper (posted on the Journal's website (http://aje.oupjournals.org/)). Papers were included if they reported original research: abstracts and "gray literature" were excluded. Here, we present results from cohort studies (including prospective and historical cohort studies, nested case-control and case-cohort studies) and case-control studies, which generated the majority of the data. Exposures eligible for inclusion were foods and beverages, dietary constituents, physical activity, and anthropometry. Journals were classified into three groups: epidemiology/public health, general medical (including urological), and cancer related.
We identified four types of reported associations: 1) "dose-response": data reported as odds, risk, or hazard ratio per unit increase in exposure, or as regression coefficients; 2) "quantile based": data reported as ratios comparing groups as defined by quantiles of either total numbers or numbers of controls; 3) "categories": data reported as ratios comparing unequal-sized groups; and 4) "means": data reported as means or mean differences in exposure, comparing those with and without disease. Associations may be reported in more than one of these ways in the same paper; we used one association only, in the order listed. When associations were presented with different adjustments for confounding, we extracted the "best" adjusted model, defined as the one that controlled for the most factors not potentially on the causal pathway from the risk factor to the disease outcome and including at least age for prostate cancer and at least smoking for bladder cancer. Unadjusted results were included only when no others were given. Because prostate and bladder cancer are relatively rare in the populations studied, the risk ratio, odds ratio, and hazard ratio are approximately equal.
For each extracted finding, we attempted to estimate the log odds, risk, or hazard ratio per unit increase in exposure ("dose-response log ratio"), together with its standard error. Table 1 summarizes the data requirements and method of estimation for each of the four types of exposure-outcome associations. When it was not possible to estimate the dose-response log ratio, the reasons were documented.
|
For quantile-based or category data, the median or mean exposure in each group is required to estimate the dose-response log ratio. When these were not reported (the majority of papers), we estimated the mean in each group on the basis of the distribution of subjects across groups (8). For unbounded upper or lower categories, this method assumed a normal distribution of the exposure; for right-skewed distributions, the method was adapted to a lognormal distribution. When numbers of individuals per group were presented, the unadjusted log ratio was estimated by using logistic regression. For quantile-based data, if numbers of individuals were not reported, groups were assumed to be of equal sizes. We estimated the dose-response log ratio by using the method of Greenland and Longnecker (9), which allows for correlations between ratios related to the same reference group. When a comparison of only the highest versus lowest category was possible, the dose-response log ratio was estimated from this information, provided that the mean or median exposure in each group was reported or could be calculated.
Data reported as means were converted to an estimated dose-response log odds ratio by using the method of Chêne and Thompson (8). When results were presented without confidence intervals or standard errors, the p value was used to estimate the standard error via the z or t statistic. When only an upper bound was reported (e.g., "p < 0.05"), a conservative estimate of the standard error was calculated based on this value. If only medians were reported, they were assumed to be equal to the means. Table 2 shows the most frequently occurring problems in reporting and the solutions applied to make results usable in dose-response meta-analyses. Some studies reported results on subgroups only, for example, men and women. In this case, we estimated results for each subgroup, then derived the combined effect by using fixed-effect meta-analysis.
|
We also examined the number of results that were usable by comparing the highest versus the lowest category of the exposure. Results reported as comparisons of quantile-based data or categories were usable, provided that the standard error of the log ratio was reported or could be derived (from the confidence interval, p value, or
2 statistic). To assess whether the strength of evidence or direction of the association was related to the detail in which results were reported, we categorized all results into one of five types of associations: negative, suggestive negative, no evidence, suggestive positive, and positive (for detailed definitions, refer to table 3). Categories were also defined by ignoring the direction of association (none, suggestive, strong). Ordered logistic regression was used to examine whether the strength and direction of associations were related to the probability that results were usable.
|
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
The prostate and bladder cancer reviews included 466 and 301 papers, respectively, published in 156 different journals (18 epidemiologic, 101 general medical, and 36 cancer related). The median year of publication was 1998. A total of 3,284 results were reported: 274 cohort studies reported 1,276 results, and 349 case-control studies reported 2,008 results. Of these, 937 (28.5 percent) results were unadjusted. The median number of results per paper was 2 (interquartile range, 1–6).
The dose-response log odds, risk, or hazard ratio and its standard error could be estimated for 1,999 (61 percent) results, reported in 545 (71 percent) papers, which were therefore usable in dose-response meta-analyses. There were 817 (66 percent) usable results from cohort studies and 1,182 (60 percent) from case-control studies. Overall, 533 (27 percent) of these results could be used without further work or assumptions. Table 4 shows the different types of work or assumptions required to estimate the dose-response log ratio in the remaining 1,466 (73 percent) usable results and the numbers of results requiring these different types of work or assumptions. Only 189 (14 percent) of 1,379 usable results from quantile-based or category data required no further work or assumptions. For all other such results (n = 1,190), the mean exposure in each group was estimated by using the method of Chêne and Thompson (8). Two hundred fifty-two results required that the standard error be estimated by using raw numbers, with 104 also requiring the odds ratio to be estimated.
|
Results were most commonly reported as comparisons between categories (1,362; 41 percent) or quantile-based data (1,022; 31 percent), with only 264 (8 percent) reporting dose-response associations. Five hundred sixty-nine (17 percent) reported comparisons of mean exposure, and 67 (2 percent) gave qualitative comment only. Cohort studies were more likely than case-control studies to report dose-response data (12 percent vs. 6 percent) and less likely to report mean data (12 percent vs. 20 percent). Table 5 shows the proportions of usable results according to data and study type and the reasons that results were not usable (note that there could be more than one reason for the same result).
|
Unsurprisingly, nearly all dose-response results were usable. For quantile-based data, 76 percent were usable, as were 44 percent for categories and 64 percent for exposure means. The most common problem was the absence of sufficient information on the range of exposure in quantile-based data and categories, followed by not reporting either numbers of individuals in exposure groups or units or frequency of measurement. Where data appeared as exposure means, the most frequent omission was a measure of precision. A total of 562 reported results were based on dichotomized exposures (e.g., ever vs. never drinking coffee). Dose-response log ratios could be derived for only 11 (2 percent) of these, primarily because the range of exposure in one or both of the groups was almost always reported too vaguely to enable the mean or median exposure to be estimated.
For 545 (71 percent) papers, a dose-response log ratio and its standard error could be estimated for at least one result. The median number of results per paper was 3 (interquartile range, 1–9) for cancer journals, 2 (interquartile range, 1–5) for epidemiology journals, and 3 (interquartile range, 1–5) for general medical journals. The median number of results in the 222 (29 percent) papers for which no dose-response log ratio could be estimated was 2 (interquartile range, 1–4).
Table 6 shows the relation between the total number of reported results and the proportion of usable results in each paper. Papers reporting more than 10 results tended to have the highest proportions of usable results. Of the 32 papers reporting more than 20 results, 24 were from cancer journals, five from epidemiology journals, and three from general medical journals. The two papers reporting most results were a case-control study on diet and prostate cancer (10) (57 results, 97 percent usable) and a report of the Health Professionals Follow-up study (11) on fruit and vegetable intake and bladder cancer (46 results, 37 percent usable).
|
Table 7 shows the proportion of usable results according to study design, journal type, and publication period, categorized as pre-1992, 1992–1997, 1998–2000, and 2001 onward. The proportions of usable results were 64 percent and 59 percent for cohort and case-control studies, respectively, and they varied from 20 percent in general medical journals in the period 1992–1997 to 78 percent in epidemiology journals in the same period. Results published in cancer-related and epidemiology journals were more likely to be usable than results published in general medical journals. In general, the proportion of usable results was higher in more recently published papers, although the proportion of usable results from epidemiology journals has been high (69 percent or greater) since 1992.
|
Representative examples of why published results were not usable are given in table 8, which shows that even those results presented in accordance with accepted practice (odds ratios with 95 percent confidence intervals related to a baseline group or quantile) may not be usable in dose-response meta-analyses if information such as range of exposure or units of measurement is missing. Table 9 shows representative examples of results that were usable for dose-response meta-analyses and the assumptions that were made, if needed.
|
|
Meta-analyses comparing the highest with the lowest category of exposure (high-low meta-analyses) could have included 2,032 (62 percent) of these results. Thus, there would be no substantial increase in usable results compared with a dose-response meta-analysis. In such meta-analyses, the type of data determines whether results are usable: quantile-based and category data are always usable, but dose-response data and mean comparisons are not.
The relation of the precision and direction of results with the proportion of results that are usable is shown in table 10. When the direction of associations was taken into account (five categories of association), we found only weak evidence of association between type of association and the probability that results were usable (odds ratio per category = 0.96, 95 percent confidence interval: 0.90, 1.02 for dose-response meta-analyses and odds ratio = 1.05, 95 percent confidence interval: 0.99, 1.12 for high-low meta-analyses). However, when the direction of results was ignored (three categories: no, suggestive, and strong evidence), there was clear evidence that stronger associations were more likely to be usable (odds ratio per category = 1.61, 95 percent confidence interval: 1.45, 1.78 for dose-response meta-analyses and odds ratio = 1.92, 95 percent confidence interval: 1.73, 2.14 for high-low meta-analyses.
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
In a series of 767 papers providing 3,284 results on the association of food and nutrition with cancer of the prostate and the bladder, 61 percent of associations were reported in a form usable for meta-analyses estimating dose-response associations. This proportion would increase only slightly if meta-analyses were based on the highest and lowest categories of exposure. The proportion of usable results was higher in more recent publications and lower in general medical journals than in public health or cancer journals. Results that showed evidence of associations were more likely to be presented in a form usable for dose-response meta-analyses.
We conducted comprehensive literature searches, and we checked reference lists of included studies. We excluded abstracts, but results from abstracts are less likely to be usable for dose-response meta-analyses because they are reported very concisely. We did not contact authors of the papers for additional information, although doing so presumably could increase the proportion of useful results substantially. In a review of randomized trials, contacting authors resulted in usable data for 13 of 39 studies for which data were requested (12). Papers reporting observational research often include complex analyses of associations of many exposures with multiple outcomes. Therefore, it may be more difficult to obtain additional data from authors of observational research.
We made only limited statistical assumptions to derive dose-response odds, risk, or hazard ratios usable in meta-analyses. For example, for quantile-based data, we estimated the number of controls when this information was not reported. Stronger assumptions would increase the proportion of usable results, but we judged that the decisions required would be too arbitrary. To scale log ratios to a common unit of measurement, we had to make assumptions regarding portion sizes in about 25 percent of results, which may well have introduced (random) misclassification; for example, "serving size" for milk will depend on whether it is consumed as a beverage or added to tea or coffee.
Several authors have recommended how to report results of observational studies (13, 14). As early as 1981, suggested guidelines for documentation of epidemiologic studies were published (15). However, they focused on design and data analysis and made few recommendations on how to present results. More recently, two studies have focused on the quality of reporting of observational studies. Tooth et al. (16) found that, on average, 17 of the 33 criteria related to the validity of observational research were reported, and they suggest that a structured approach to presenting results from such studies is required. Pocock et al. (17) expressed concerns regarding inadequacies in the analysis and reporting of epidemiologic research based on a survey of published studies in high-impact epidemiologic and general medical journals. Although these papers did not assess whether published results were usable for meta-analysis, their findings that reporting of observational studies needs to be improved are consistent with the results presented here.
Other authors have considered difficulties in extracting data for meta-analysis of different types of studies. Chan et al. (18) reported data required for meta-analyses of randomized controlled trials, whereas Riley et al. (19) focused on prognostic studies. Consistent with the findings reported here, Chan et al. (18) found that the odds of an outcome being reported were more than doubled if the treatment effect for an outcome was statistically significant. Selective reporting of outcomes thus has the potential to cause serious bias in meta-analyses of randomized controlled trials.
The majority of usable results in our study (n = 2,384; 74 percent) were reported as odds, risk, or hazard ratios comparing one or more exposure categories with a baseline category. Reporting results in this way has advantages: the reader can check informally for nonlinear exposure effects, and the magnitude of the association is readily interpretable. However, estimates of the dose-response association derived from categorized exposure data are dependent on the exposure value assigned to each category. Il'yasova et al. (20) found that use of midrange scores for the exposure level in each category led to reasonably good approximations to the results from the individual-level analysis, provided that the value assigned to the uppermost, open-ended category was at least as high as the lower bound plus the width of the second-highest category. Richardson and Loomis (21) used simulation studies to examine the bias in dose-response estimates that can result from exposure categorization, which they found to be greater when the exposure distribution was nonnormal or when categories were defined by quintiles rather than deciles of the exposure distribution. These issues may imply that authors should publish direct estimates of the dose-response association as well as results based on exposure categories.
One of the most important reasons that a dose-response odds, risk, or hazard ratio could not be estimated was that the range of the exposure, or for dose-response data the unit of increase, was not reported (82 percent of nonusable results). For example, an odds ratio comparing seven or fewer versus more than seven glasses of beer weekly cannot be used because mean consumption in one of the groups cannot be estimated. We hope that this paper will help alert authors and editors to the improved reporting of epidemiologic associations needed to facilitate the conduct of dose-response meta-analyses. Guidelines on how to report observational studies, the STROBE statement (22), explicitly recommend reporting this information and may also be expected to improve matters in this regard. It would be desirable for journals publishing observational studies to amend their editorial guidelines in light of the STROBE statement to encourage adequate reporting.
Meta-analyses based on "high versus low" category comparisons are commonly used to summarize results of observational studies (e.g., Etminan et al. (23), Gao et al. (24)). However, results of such meta-analyses can be difficult to interpret because the exposure ranges compared will typically vary substantially between studies and will often be unknown. The results presented here suggest that dose-response meta-analyses are an attractive alternative. A similar proportion of results are usable in meta-analyses, and standardization of the scale of the exposure means that the pooled estimate has a direct biologic interpretation. Dose-response meta-analyses will be misleading when associations are clearly nonlinear (e.g., the association between alcohol consumption and risk of coronary heart disease). Therefore, linearity assumption should be checked for each study, before dose-response meta-analyses are conducted.
Although more recently published results were more likely to be usable in dose-response meta-analyses, 28 percent of results published since 2001 were not. Contacting authors might alleviate this problem but involves substantial extra work by reviewers. Results of observational studies should be presented with a description of the unit and range of exposure and numbers in each category, and odds, risk, or hazard ratios should be given with confidence intervals. Table 11 presents the information required for data to be usable for dose-response meta-analysis for each type of result. Addressing problems such as the omission of ranges or numbers of individuals would require at most minor increases in the space required for tables of results. Furthermore, the increasing use of Web appendices and online publication means that space constraints should not be seen as an obstacle to complete reporting. Since, currently, a substantial proportion of results are likely to be excluded from meta-analyses of observational research, reviews of such research should clearly state the numbers of results that were and were not included in meta-analyses, describe results that were not included, and discuss the likely direction of biases resulting from omission of incomplete reporting.
|
|
ACKNOWLEDGMENTS |
|---|
This work was supported by a grant from the World Cancer Research Fund, as one of a number of systematic reviews of the association of diet, nutrition, physical activity, and anthropometry with different types of cancer that were conducted according to a common protocol. The statistical methods described here were developed to facilitate statistical analysis of data arising from these reviews.
C. B. was supported by a Leverhulme Fellowship.
The funder had no involvement in the decision to submit this work for publication.
Conflict of interest: none declared.
|
References |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
- Stroup DF, Berlin JA, Morton SC, et al. Meta-analysis of observational studies in epidemiology: a proposal for reporting. Meta-analysis Of Observational Studies in Epidemiology (MOOSE) group. JAMA (2000) 283:2008–12.
[Abstract/Free Full Text] - Egger M, Schneider M, Davey Smith G. Meta-analysis. Spurious precision? Meta-analysis of observational studies. BMJ (1998) 316:140–4.
[Free Full Text] - Bradford Hill A. The environment and disease: association or causation? Proc R Soc Med (1965) 58:295–300.[Web of Science][Medline]
- Berlin JA, Longnecker MP, Greenland S. Meta-analysis of epidemiologic dose-response data. Epidemiology (1993) 4:218–28.[Web of Science][Medline]
- Weed DL. Meta-analysis under the microscope. J Natl Cancer Inst (1997) 89:904–5.
[Free Full Text] - Weed DL. Interpreting epidemiological evidence: how meta-analysis and causal inference methods are related. Int J Epidemiol (2000) 29:387–90.
[Abstract/Free Full Text] - WCRF. World Cancer Research Fund/American Institute for Cancer Research. Food, nutrition, physical activity, and the prevention of cancer: a global perspective. (2007) Washington, DC: AICR. (http://www.dietandcancerreport.org).
- Chêne G, Thompson SG. Methods for summarizing the risk associations of quantitative variables in epidemiologic studies in a consistent form. Am J Epidemiol (1996) 144:610–21.
[Abstract/Free Full Text] - Greenland S, Longnecker MP. Methods for trend estimation from summarized dose-response data, with applications to meta-analysis. Am J Epidemiol (1992) 135:1301–9.
[Abstract/Free Full Text] - Key TJ, Silcocks PB, Davey GK, et al. A case-control study of diet and prostate cancer. Br J Cancer (1997) 76:678–87.[Web of Science][Medline]
- Michaud DS, Spiegelman D, Clinton SK, et al. Fruit and vegetable intake and incidence of bladder cancer in a male prospective cohort. J Natl Cancer Inst (1999) 91:605–13.
[Abstract/Free Full Text] - Kelley GA, Kelley KS, Tran ZV. Retrieval of missing data for meta-analysis: a practical example. Int J Technol Assess Health Care (2004) 20:296–9.[Web of Science][Medline]
- Walker AM. Reporting the result of epidemiologic studies. Am J Public Health (1986) 76:556–8.
[Free Full Text] - Squires BP, Elmslie TJ. Cohort studies: what editors want from authors and peer reviewers. CMAJ (1990) 143:179–80.[Medline]
- Guidelines for documentation of epidemiologic studies. Epidemiology Work Group of the Interagency Regulatory Liaison Group. Am J Epidemiol (1981) 114:609–13.
[Free Full Text] - Tooth L, Ware R, Bain C, et al. Quality of reporting of observational longitudinal research. Am J Epidemiol (2005) 161:280–8.
[Abstract/Free Full Text] - Pocock SJ, Collier TJ, Dandreo KJ, et al. Issues in the reporting of epidemiological studies: a survey of recent practice. BMJ (2004) 329:883–7.
[Abstract/Free Full Text] - Chan A, Hrobjartsson A, Haahr MT, et al. Empirical evidence for selective reporting of outcomes in randomized trials. Comparison of protocols to published articles. JAMA (2004) 291:2457–65.
[Abstract/Free Full Text] - Riley RD, Abrams KR, Sutton AJ, et al. Reporting of prognostic markers: current problems and development of guidelines for evidence-based practice in the future. Br J Cancer (2003) 88:1191–8.[CrossRef][Web of Science][Medline]
- Il'yasova D, Hertz-Picciotto I, Peters U, et al. Choice of exposure scores for categorical regression in meta-analysis: a case study of a common problem. Cancer Causes Control (2005) 16:383–8.[CrossRef][Web of Science][Medline]
- Richardson DB, Loomis D. The impact of exposure classification for grouped analyses of cohort data. Occup Environ Med (2004) 61:930–5.
[Abstract/Free Full Text] - von Elm E, Altman DG, Egger M, et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. Lancet (2007) 370:1453–7.[CrossRef][Web of Science][Medline]
- Etminan M, FitzGerald JM, Gleave M, et al. Intake of selenium in the prevention of prostate cancer: a systematic review and meta-analysis. Cancer Causes Control (2005) 16:1125–31.[CrossRef][Web of Science][Medline]
- Gao X, LaValley MP, Tuckler KL. Prospective studies of dairy product and calcium intakes and prostate cancer risk: a meta-analysis. J Natl Cancer Inst (2005) 97:1768–77.
[Abstract/Free Full Text] - Davies J, Dickerson J. Nutrient content of food portions. (1991) Cambridge, United Kingdom: Royal Society of Chemistry.
- Crawley H. Food portion sizes. (1988) London, United Kingdom: HMSO.
- Key TJ, Allen N, Appleby P, et al. Fruits and vegetables and prostate cancer: no association among 1104 cases in a prospective study of 130544 men in the European Prospective Investigation into Cancer and Nutrition (EPIC). Int J Cancer (2004) 109:119–24.[CrossRef][Web of Science][Medline]
- Hayes RB, Ziegler RG, Gridley G, et al. Dietary factors and risks for prostate cancer among blacks and whites in the United States. Cancer Epidemiol Biomarkers Prev (1999) 8:25–34.
[Abstract/Free Full Text] - Cerhan JR, Torner JC, Lynch CF, et al. Association of smoking, body mass, and physical activity with risk of prostate cancer in the Iowa 65+ Rural Health Study (United States). Cancer Causes Control (1997) 8:229–38.[CrossRef][Web of Science][Medline]
- Norrish AE, Ferguson LR, Knize MG, et al. Heterocyclic amine content of cooked meat and risk of prostate cancer. J Natl Cancer Inst (1999) 91:2038–44.
[Abstract/Free Full Text] - Hsing AW, McLaughlin JK, Schuman LM, et al. Diet, tobacco use, and fatal prostate cancer: results from the Lutheran Brotherhood Cohort Study. (See comment). Cancer Res (1990) 50:6836–40.
[Abstract/Free Full Text] - Schuurman AG, Goldbohm RA, Dorant E, et al. Vegetable and fruit consumption and prostate cancer risk: a cohort study in The Netherlands. Cancer Epidemiol Biomarkers Prev (1998) 7:673–80.[Abstract]
- Alavanja MC, Samanic C, Dosemeci M, et al. Use of agricultural pesticides and prostate cancer risk in the Agricultural Health Study cohort. Am J Epidemiol (2003) 157:800–14.
[Abstract/Free Full Text] - Platz EA, De Marzo AM, Erlinger TP, et al. No association between pre-diagnostic plasma C-reactive protein concentration and subsequent prostate cancer. Prostate (2004) 59:393–400.[CrossRef][Web of Science][Medline]
- Chan JM, Pietinen P, Virtanen M, et al. Diet and prostate cancer risk in a cohort of smokers, with a specific focus on calcium and phosphorus (Finland). Cancer Causes Control (2000) 11:859–67.[CrossRef][Web of Science][Medline]
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  B. E. G. Rothberg, M. B. Bracken, and D. L. Rimm Tissue Biomarkers for Prognosis in Cutaneous Melanoma: A Systematic Review and Meta-analysis J Natl Cancer Inst, April 1, 2009; 101(7): 452 - 474. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 M. J.J. Van Hemelrijck, D. S. Michaud, G. N. Connolly, and Z. Kabir Secondhand Smoking, 4-Aminobiphenyl, and Bladder Cancer: Two Meta-analyses Cancer Epidemiol. Biomarkers Prev., April 1, 2009; 18(4): 1312 - 1320. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||