European Journal of Clinical Nutrition (2014) 68, 409–410 & 2014 Macmillan Publishers Limited All rights reserved 0954-3007/14 www.nature.com/ejcn
LETTER TO THE EDITOR
Re: Coffee consumption and risk of prostate cancer: an up-to-date meta-analysis European Journal of Clinical Nutrition (2014) 68, 409–410; doi:10.1038/ejcn.2013.292; published online 22 January 2014
In their recent meta-analysis, Zhong et al.1 reported a 17% reduction in prostate cancer (PCa) risk when comparing highest versus non/lowest coffee consumption categories among 12 cohort studies1 (relative risk (RR) ¼ 0.83 (95% confidence interval (CI) ¼ 0.72–0.96)). Furthermore, they observed, from a doseresponse meta-analysis, a 7% decreased risk of PCa for every two cups per day increase in coffee consumption (RR ¼ 0.93 (95% CI ¼ 0.88–0.99)). The authors also graphically presented that result in Figure 3b1. In the present letter, we would like to discuss two issues of that meta-analysis that we hope might draw attention to two fundamental steps of a meta-analysis: collection of the data and presentation of the results. First, Zhong et al.1 used in their meta-analysis crude RRs extracted from our cohort study2 instead of multivariable-adjusted RRs, as incorrectly reported in Table 1.1 For example, they reported a RR of 0.52 (95% CI ¼ 0.41–0.65) and of 0.55 (95% CI ¼ 0.36–0.86) for localized and advanced PCa respectively, comparing men who drank six or more cups per day with nondrinkers. In our study the risk of PCa increased exponentially with age and older men tended to drink less coffee. Therefore, one might expect some differences in the association between coffee consumption and PCa risk when further adjusting for age as well as other possible confounders. The choice of calculating crude RRs from the numbers of cases and person-years reported in Table 22 of our study might have depended on the authors’ need of changing the reference category from ‘1–3 cups per day’, as originally reported, to nondrinkers. A viable solution for changing the reference category of adjusted RRs using published data only is by calculating the effective number of cases and person-years for each category of coffee consumption (also known as pseudocounts) using the methods proposed by Greenland and Longnecker3 and by Hamling et al.4 Practically, the pseudoTable 1. Reanalysis of the association between coffee consumption and risk of localized prostate cancer in the study by Discacciati et al.2 Category of coffee consumption (cups per day) None o1 1–3 4–5 X6
Number of cases (Hamling)a
Person-years (Hamling)b
RR (95% CI)c
RR (95% CI)d
114.26 985.97 985.97 495.30 164.38
20 082.48 195 822.50 195 822.50 105 774.61 40 306.23
1.00 (ref ) 0.885 (0.729–1.074) 0.885 (0.729–1.074) 0.823 (0.671–1.009) 0.717 (0.564–0.910)
1.00 (ref ) 0.889 (0.710–1.113) 0.888 (0.732–1.076) 0.823 (0.666–1.016) 0.723 (0.565–0.925)
Abbreviations: CI, confidence interval; RR, relative risk. aEffective number of cases estimated using the Hamling method. bEffective number of personyears estimated using the Hamling method. cRRs and 95% CIs calculated using the number of cases and person-years estimated with the Hamling method. For example, for X6 cups per day versus nondrinkers: (164.38/ 40 306.23)/(114.26/20 082.48) ¼ 0.717. dRelative risks and 95% confidence intervals obtained reanalyzing the original data.
counts can be obtained for example using an Excel spreadsheet or a SAS macro (both available online)4 or the R package dosresmeta (available on CRAN).5 The Greenland and Longnecker and the Hamling methods are based on different sets of assumptions that are unlikely to be perfectly met in practice,6 but in this particular case they gave similar results that, most importantly, also agreed with those obtained by reanalyzing the original data. For example, using the Hamling method to change the reference group to nondrinkers yielded multivariable-adjusted RRs of 0.72 (95% CI ¼ 0.56–0.91) and 0.91 (95% CI ¼ 0.59–1.38) for localized and advanced PCa, respectively (see Table 1). A pooled RR of 0.90 (95% CI ¼ 0.84–0.97) was observed when reanalyzing the data coming from the cohort studies using a random-effects model. This estimate suggests a weaker inverse association between coffee consumption and PCa risk as compared with what observed by Zhong et al.1 (RR ¼ 0.83 (95% CI ¼ 0.72–0.96)). The meta-analysis on prostate cancer mortality suffers from the same issue: in our study, the crude RR for high coffee drinkers (X6 cups per day) compared with nondrinkers (RR ¼ 0.36 (95% CI ¼ 0.20–0.64)) was very different from the multivariable-adjusted RR calculated, for example, using the Hamling method from the published results (RR ¼ 0.71 (95% CI ¼ 0.40–1.25)). Again, the pooled RR form a random-effects model suggested a weaker association between coffee consumption and fatal PCa (RR ¼ 0.71 (95% CI ¼ 0.54–0.94)) as compared with what observed by Zhong et al.1 (RR ¼ 0.61 (95% CI ¼ 0.42–0.90)). A 29% reduction in PCa mortality when comparing the highest with the lowest coffee consumption category (mean range ¼ 8 cups per day) is consistent with a RR of 0.89 for every three cups per day increase in coffee consumption that we observed in a recent dose–response meta-analysis7 (exp(ln(0.89)/3*8) ¼ 0.73). Secondly, even ignoring the aforementioned issue, we observed an inconsistency between the RR from the dose–response metaanalysis of cohort studies reported in the Results section (RR ¼ 0.93 (95% CI ¼ 0.88–0.99) for every two cups per day increase in coffee consumption) and its graphical representation (Figure 3b).1 In particular, the linear trend shown in Figure 3b is more consistent with a 3% (rather than a 7%) decreased risk of PCa for every two cups per day increase in coffee consumption. To give an idea on the magnitude of this discrepancy, we can for example calculate the RR for a man who drank six cups per day versus a nondrinker using the RR presented in the Results section (exp(ln(0.93)/2*6) ¼ 0.80) and the RR derived from Figure 3b (exp(ln(0.97)/2*6) ¼ 0.91). The former RR corresponds to a 20% decreased risk, whereas the latter corresponds to a more modest 9% decreased risk. The issues we have discussed in this letter do not substantially alter the interpretation of the final results of the meta-analysis by Zhong et al.1 However, given the importance of meta-analyses in quantitatively summarizing and describing the body of existing epidemiological evidence, and also given their intrinsic limitations,8 it is of paramount importance that the data collection and the presentation of the results are carefully carried out. CONFLICT OF INTEREST The authors declare no conflict of interest.
Letter to the Editor
410 A Discacciati1,2 and N Orsini1,2 Unit of Nutritional Epidemiology, Institute of Environmental Medicine, Karolinska Institutet, Stockholm, Sweden and 2 Unit of Biostatistics, Institute of Environmental Medicine, Karolinska Institutet, Stockholm, Sweden E-mail: [email protected] 1
REFERENCES 1 Zhong S, Chen W, Yu X, Chen Z, Hu Q, Zhao J. Coffee consumption and risk of prostate cancer: an up-to-date meta-analysis. Eur J Clin Nutr 2013; 68: 330–337. 2 Discacciati A, Orsini N, Andersson SO, Andre´n O, Johansson JE, Mantzoros CS et al. Coffee consumption and risk of localized, advanced and fatal prostate cancer: a population-based prospective study. Ann Oncol 2013; 24: 1912–1918.
European Journal of Clinical Nutrition (2014) 409 – 410
3 Greenland S, Longnecker MP. Methods for trend estimation from summarized dose-response data, with applications to meta-analysis. Am J Epidemiol 1992; 135: 1301–1309. 4 Hamling J, Lee P, Weitkunat R, Ambu¨hl M. Facilitating meta-analyses by deriving relative effect and precision estimates for alternative comparisons from a set of estimates presented by exposure level or disease category. Stat Med 2008; 27: 954–970. 5 Crippa A, Orsini N. Dosresmeta: performing multivariate dose-response meta-analysis. Available at: http://CRAN.R-project.org/package=dosresmeta. 6 Orsini N, Li R, Wolk A, Khudyakov P, Spiegelman D. Meta-analysis for linear and nonlinear dose-response relations: examples, an evaluation of approximations, and software. Am J Epidemiol 2012; 175: 66–73. 7 Discacciati A, Orsini N, Wolk A. Coffee consumption and risk of nonaggressive, aggressive and fatal prostate cancer--a dose-response meta-analysis. Ann Oncol 2013; e-pub ahead of print 24 November 2013; doi:10.1093/annonc/mdt420. 8 Greenland S. Can meta-analysis be salvaged? Am J Epidemiol 1994; 140: 783–787.
& 2014 Macmillan Publishers Limited
LETTER TO THE EDITOR
Re: Coffee consumption and risk of prostate cancer: an up-to-date meta-analysis European Journal of Clinical Nutrition (2014) 68, 409–410; doi:10.1038/ejcn.2013.292; published online 22 January 2014
In their recent meta-analysis, Zhong et al.1 reported a 17% reduction in prostate cancer (PCa) risk when comparing highest versus non/lowest coffee consumption categories among 12 cohort studies1 (relative risk (RR) ¼ 0.83 (95% confidence interval (CI) ¼ 0.72–0.96)). Furthermore, they observed, from a doseresponse meta-analysis, a 7% decreased risk of PCa for every two cups per day increase in coffee consumption (RR ¼ 0.93 (95% CI ¼ 0.88–0.99)). The authors also graphically presented that result in Figure 3b1. In the present letter, we would like to discuss two issues of that meta-analysis that we hope might draw attention to two fundamental steps of a meta-analysis: collection of the data and presentation of the results. First, Zhong et al.1 used in their meta-analysis crude RRs extracted from our cohort study2 instead of multivariable-adjusted RRs, as incorrectly reported in Table 1.1 For example, they reported a RR of 0.52 (95% CI ¼ 0.41–0.65) and of 0.55 (95% CI ¼ 0.36–0.86) for localized and advanced PCa respectively, comparing men who drank six or more cups per day with nondrinkers. In our study the risk of PCa increased exponentially with age and older men tended to drink less coffee. Therefore, one might expect some differences in the association between coffee consumption and PCa risk when further adjusting for age as well as other possible confounders. The choice of calculating crude RRs from the numbers of cases and person-years reported in Table 22 of our study might have depended on the authors’ need of changing the reference category from ‘1–3 cups per day’, as originally reported, to nondrinkers. A viable solution for changing the reference category of adjusted RRs using published data only is by calculating the effective number of cases and person-years for each category of coffee consumption (also known as pseudocounts) using the methods proposed by Greenland and Longnecker3 and by Hamling et al.4 Practically, the pseudoTable 1. Reanalysis of the association between coffee consumption and risk of localized prostate cancer in the study by Discacciati et al.2 Category of coffee consumption (cups per day) None o1 1–3 4–5 X6
Number of cases (Hamling)a
Person-years (Hamling)b
RR (95% CI)c
RR (95% CI)d
114.26 985.97 985.97 495.30 164.38
20 082.48 195 822.50 195 822.50 105 774.61 40 306.23
1.00 (ref ) 0.885 (0.729–1.074) 0.885 (0.729–1.074) 0.823 (0.671–1.009) 0.717 (0.564–0.910)
1.00 (ref ) 0.889 (0.710–1.113) 0.888 (0.732–1.076) 0.823 (0.666–1.016) 0.723 (0.565–0.925)
Abbreviations: CI, confidence interval; RR, relative risk. aEffective number of cases estimated using the Hamling method. bEffective number of personyears estimated using the Hamling method. cRRs and 95% CIs calculated using the number of cases and person-years estimated with the Hamling method. For example, for X6 cups per day versus nondrinkers: (164.38/ 40 306.23)/(114.26/20 082.48) ¼ 0.717. dRelative risks and 95% confidence intervals obtained reanalyzing the original data.
counts can be obtained for example using an Excel spreadsheet or a SAS macro (both available online)4 or the R package dosresmeta (available on CRAN).5 The Greenland and Longnecker and the Hamling methods are based on different sets of assumptions that are unlikely to be perfectly met in practice,6 but in this particular case they gave similar results that, most importantly, also agreed with those obtained by reanalyzing the original data. For example, using the Hamling method to change the reference group to nondrinkers yielded multivariable-adjusted RRs of 0.72 (95% CI ¼ 0.56–0.91) and 0.91 (95% CI ¼ 0.59–1.38) for localized and advanced PCa, respectively (see Table 1). A pooled RR of 0.90 (95% CI ¼ 0.84–0.97) was observed when reanalyzing the data coming from the cohort studies using a random-effects model. This estimate suggests a weaker inverse association between coffee consumption and PCa risk as compared with what observed by Zhong et al.1 (RR ¼ 0.83 (95% CI ¼ 0.72–0.96)). The meta-analysis on prostate cancer mortality suffers from the same issue: in our study, the crude RR for high coffee drinkers (X6 cups per day) compared with nondrinkers (RR ¼ 0.36 (95% CI ¼ 0.20–0.64)) was very different from the multivariable-adjusted RR calculated, for example, using the Hamling method from the published results (RR ¼ 0.71 (95% CI ¼ 0.40–1.25)). Again, the pooled RR form a random-effects model suggested a weaker association between coffee consumption and fatal PCa (RR ¼ 0.71 (95% CI ¼ 0.54–0.94)) as compared with what observed by Zhong et al.1 (RR ¼ 0.61 (95% CI ¼ 0.42–0.90)). A 29% reduction in PCa mortality when comparing the highest with the lowest coffee consumption category (mean range ¼ 8 cups per day) is consistent with a RR of 0.89 for every three cups per day increase in coffee consumption that we observed in a recent dose–response meta-analysis7 (exp(ln(0.89)/3*8) ¼ 0.73). Secondly, even ignoring the aforementioned issue, we observed an inconsistency between the RR from the dose–response metaanalysis of cohort studies reported in the Results section (RR ¼ 0.93 (95% CI ¼ 0.88–0.99) for every two cups per day increase in coffee consumption) and its graphical representation (Figure 3b).1 In particular, the linear trend shown in Figure 3b is more consistent with a 3% (rather than a 7%) decreased risk of PCa for every two cups per day increase in coffee consumption. To give an idea on the magnitude of this discrepancy, we can for example calculate the RR for a man who drank six cups per day versus a nondrinker using the RR presented in the Results section (exp(ln(0.93)/2*6) ¼ 0.80) and the RR derived from Figure 3b (exp(ln(0.97)/2*6) ¼ 0.91). The former RR corresponds to a 20% decreased risk, whereas the latter corresponds to a more modest 9% decreased risk. The issues we have discussed in this letter do not substantially alter the interpretation of the final results of the meta-analysis by Zhong et al.1 However, given the importance of meta-analyses in quantitatively summarizing and describing the body of existing epidemiological evidence, and also given their intrinsic limitations,8 it is of paramount importance that the data collection and the presentation of the results are carefully carried out. CONFLICT OF INTEREST The authors declare no conflict of interest.
Letter to the Editor
410 A Discacciati1,2 and N Orsini1,2 Unit of Nutritional Epidemiology, Institute of Environmental Medicine, Karolinska Institutet, Stockholm, Sweden and 2 Unit of Biostatistics, Institute of Environmental Medicine, Karolinska Institutet, Stockholm, Sweden E-mail: [email protected] 1
REFERENCES 1 Zhong S, Chen W, Yu X, Chen Z, Hu Q, Zhao J. Coffee consumption and risk of prostate cancer: an up-to-date meta-analysis. Eur J Clin Nutr 2013; 68: 330–337. 2 Discacciati A, Orsini N, Andersson SO, Andre´n O, Johansson JE, Mantzoros CS et al. Coffee consumption and risk of localized, advanced and fatal prostate cancer: a population-based prospective study. Ann Oncol 2013; 24: 1912–1918.
European Journal of Clinical Nutrition (2014) 409 – 410
3 Greenland S, Longnecker MP. Methods for trend estimation from summarized dose-response data, with applications to meta-analysis. Am J Epidemiol 1992; 135: 1301–1309. 4 Hamling J, Lee P, Weitkunat R, Ambu¨hl M. Facilitating meta-analyses by deriving relative effect and precision estimates for alternative comparisons from a set of estimates presented by exposure level or disease category. Stat Med 2008; 27: 954–970. 5 Crippa A, Orsini N. Dosresmeta: performing multivariate dose-response meta-analysis. Available at: http://CRAN.R-project.org/package=dosresmeta. 6 Orsini N, Li R, Wolk A, Khudyakov P, Spiegelman D. Meta-analysis for linear and nonlinear dose-response relations: examples, an evaluation of approximations, and software. Am J Epidemiol 2012; 175: 66–73. 7 Discacciati A, Orsini N, Wolk A. Coffee consumption and risk of nonaggressive, aggressive and fatal prostate cancer--a dose-response meta-analysis. Ann Oncol 2013; e-pub ahead of print 24 November 2013; doi:10.1093/annonc/mdt420. 8 Greenland S. Can meta-analysis be salvaged? Am J Epidemiol 1994; 140: 783–787.
& 2014 Macmillan Publishers Limited
