Body mass index and incidence of nonaggressive and aggressive prostate cancer: a dose-response meta-analysis of cohort studies

The relationship between body mass index (BMI) and incidence of prostate cancer is still inconclusive. We performed a dose-response meta-analysis of eligible cohort studies to evaluate potential association of BMI with prostate cancer risk by subtype of prostate cancer (nonaggressive and aggressive). A comprehensive literature search was performed in PubMed and Web of Science databases through March 22, 2017. Linear and non-linear dose-response meta-analyses were carried out to evaluate the effects of BMI on incidence of prostate cancer. A total of 21 cohort or nested case-control studies (17 for nonaggressive and 21 for aggressive prostate cancer) were included in this meta-analysis. For nonaggressive prostate cancer, the pooled relative risk (RR) per 5 kg/m2 increment of BMI with 95% confidence interval (CI) was 0.96 (95% CI 0.92–1.00). Sensitivity analysis indicated that this result was not robust and steady. For aggressive prostate cancer, a significant linear direct relationship with BMI (RR, 1.07; 95% CI 1.03–1.12) for every 5 kg/m2 increase was observed. Statistically significant heterogeneity was detected for nonaggressive prostate cancer (P = 0.020, I2 = 46.1%) but not for aggressive prostate cancer (P = 0.174, I2 = 22.4%). In conclusion, BMI level may be positively associated with aggressive prostate cancer risk. Further large prospective cohort studies are warranted to confirm the findings from our study.


INTRODUCTION
Prostate cancer has become the most common malignancy in males in several developed countries and the second most common one worldwide after lung cancer [1]. The precise etiology of prostate cancer is still virtually unknown and the only well-established risk factors are those that are inherited and uncontrollable, including age, race, and family history of prostate cancer [2]. Emerging evidence indicates that environmental factors may also play an important role in the carcinogenesis and progression of prostate cancer. A high incidence of prostate cancer in the USA and European countries suggests that prostate cancer may be related to the ''Western'' lifestyle pattern [3].
To date, a large number of well-designed prospective cohort studies have been performed to evaluate the potential relationship between body mass index (BMI) and prostate cancer risk with positive, negative, or null results. Several studies indicated that body adiposity may have a dual effect on two subtypes of prostate cancer (localized/ non-aggressive and advanced/aggressive cancer) [4,5]. Therefore, a recent meta-analysis published in 2012 reviewed all eligible cohort studies investigating this topic and assessed the potential association separately by tumor characteristics. As a result, a high BMI appeared to increase the risk of advanced prostate cancer while reducing the risk of localized disease. Thus, these results support the hypothesis of a dual effect of BMI on prostate cancer carcinogenesis [6]. Since then, more high-quality
In this study, we performed an updated randomeffects dose-response meta-analysis of all available cohort studies up to now in order to comprehensively evaluate the association between BMI and incidence of prostate cancer separately by tumor characteristics.

Galbraith plot analysis
Galbraith plot was used to detect the studies that contributed to the heterogeneity. As a result, two studies by Kelly et al. and Rundle et al. [7,11] led to the heterogeneity among those contributing for nonaggressive prostate cancer ( Figure 3A). After removing these two studies, the heterogeneity became small (P = 0.285, I 2 = 15.0%) and the direction of the combined RR did not change (RR = 0.94, 95% CI 0.91-0.97) ( Figure 3B).

Sensitivity analysis
In the sensitivity analysis, we removed each included study at a time and repeated the meta-analysis. The combined RRs for non-aggressive prostate cancer were not robust and steady ( Figure 4A). In contrast, the pooled risk estimates for aggressive prostate cancer were not influenced greatly by any individual study ( Figure 4B).

Publication bias
Potential publication bias was detected for nonaggressive prostate cancer ( Figure 5A and 5B, Egger's test: P = 0.003; Begg's test: P = 0.044). The application of trim and fill analysis did not virtually change the pooled risk estimate for non-aggressive prostate cancer ( Figure 5C, RR = 0.95, 95% CI 0.90-0.99). No significant evidence of publication bias was observed for aggressive   Figure 5D, P = 0.339) or Begg's test ( Figure 5E, P = 0.608). Figure 6 shows the results of non-linear dose-response meta-analysis for aggressive prostate cancer. There was no evidence of a nonlinear relationship between BMI level and aggressive prostate cancer (P = 0.181 for nonlinearity).

DISCUSSION
This updated dose-response meta-analysis summarized all available cohort studies that explored the relationship between BMI level and incidence of prostate cancer separately by tumor characteristics. As a result, high BMI may be related with an increased risk of aggressive prostate cancer. On the other hand, a borderline inverse relationship between high BMI and nonaggressive prostate cancer was observed, but sensitivity analysis indicated that this result was not robust and steady.
A previous meta-analysis also examined two prostate cancer subtypes (localized and advanced) separately based on 12 and 13 cohort studies, respectively. They observed an inverse linear relationship for localized prostate cancer (RR = 0.94, 95% CI 0.91-0.97 for every 5 kg/m 2 increase) but a linear positive relationship for advanced prostate cancer (RR = 1.09, 95% CI 1.02-1.16 for every 5 kg/m 2 increase) [6]. In contrast, although we also found a linear dose-response relationship for aggressive prostate cancer, the evidence that supported the inverse relationship between BMI and nonaggressive prostate cancer was weak.
Several potential mechanisms might mediate the positive association between BMI and risk of aggressive prostate cancer. Obesity leads to high circulating concentrations of insulin, leptin and insulin-like growth factor-I (IGF-I) and low levels of adiponectin, which have been described to promote prostate cancer growth and progression, thus increasing the risk of advanced prostate cancer [26]. In addition, obese men might have a lower concentration of serum testosterone [27]. Although testosterone could stimulate the growth and development of prostate cancer, testosterone could also help maintain the normal differentiated state of the prostate [28]. Hence reduced testosterone may be associated with a higher risk of less differentiated and more aggressive prostate cancer [29].
This meta-analysis has several limitations that should be acknowledged. Firstly, exposure assessment was based on information collected at the baseline. Participants may have changed their BMI over the long follow-up period, which may have led to some bias in risk estimation. Secondly, a meta-analysis is not able to avoid the problems of confounders that could be inherent in the included studies, which may result in either an overestimation or an underestimation of an effect measure. Thirdly, the criteria used to definite nonaggressive and aggressive prostate cancer varied between different cohorts and involved  two studies led to the heterogeneity. (B) Pooled risk estimate with its 95% CI for the association between BMI and non-aggressive prostate cancer risk after removing the studies that led to heterogeneity.
Oncotarget 97589 www.impactjournals.com/oncotarget Gleason score, World Health Organization grading system, TNM (tumor-node-metastasis), and so on, which may lead to a misclassification bias. Fourthly, obvious heterogeneity was observed for nonaggressive prostate cancer, which would throw some doubts on the reliability of the pooled estimates. Fifthly, as lack of individual participant data, we are not able to determine the independent effect of individual variables on study outcomes. Sixthly, like all meta-analyses, this study also has the limitation of being a retrospective analysis. Lastly, potential publication bias was detected by Begg's and Egger's tests, which may be due to the fact that small studies with null findings tend not to be published.
This present meta-analysis also has some strengths. Linear and non-linear dose-response analyses were adopted to explore the potential relationship between BMI level and prostate cancer risk. Sensitivity analyses and Galbraith plot analysis were performed to assess the robustness and stability of the pooled risk estimates. All included studies were cohort studies and came from several countries. The risk estimates were extracted from the models adjusting for most established risk factors in each study.
In conclusion, this systematic review and doseresponse meta-analysis indicates that BMI level is associated with aggressive prostate cancer risk. Further

Literature search
Two investigators (BX and XX) independently searched PubMed and Web of Science databases from their inception to March 22, 2017 with the following keywords: ("body mass index" OR "BMI" OR "body size" OR "overweight" OR "obesity" OR "adiposity") AND ("prostate cancer" OR "prostate neoplasm"). No language or date restrictions were applied. The reference lists of retrieved studies and related reviews were also checked for additional eligible articles.

Selection criteria
Studies included in this meta-analysis had to meet all of the following criteria: i) cohort or nested casecontrol studies; ii) the exposure was BMI; iii) the outcome of interest was incidence rate of prostate cancer; iv) risk estimates with 95% CIs were provided.

Data extraction and quality assessment
Data were extracted independently with a standardized data collection form by two reviewers (BX and XX). The following information were recorded: first author's surname, publication date, country, cohort name, sample size, the number of cases, follow-up duration, adjusted variables, BMI exposure levels, and corresponding risk estimates with 95% CIs. Quality assessment of each included study was performed with the Newcastle-Ottawa scale (NOS) (http://www.ohri. ca/programs/clinical_epidemiology/oxford.asp). Any discrepancies were resolved by discussing with a third reviewer.

Statistical methods
RR with its corresponding 95% CI was used to evaluate the relationship between the BMI and incidence of prostate cancer. A random-effects model was adopted to account for both within-and between-study heterogeneity. For dose-response meta-analysis, we assigned the reported median or mean BMI level of each category to the corresponding RR. If medians or means were not available, the midpoint of the lower and upper bounds of that category was adopted. When the highest category was open-ended, the width of the category was assumed to be the same as the closest adjacent category [6]. The method proposed by Greenland and Longnecker [30] was used to estimate the dose-response trend for each study. These dose-response trends were then pooled with a randomeffects meta-analysis. Finally, a potential non-linear doseresponse relationship was also explored by using restricted cubic regression splines with three knots at the 25th, 50th, and 75th percentiles of the distribution [31].
Publication bias was evaluated by Begg's test [32], Egger's test [33] and a trim-fill analysis. Heterogeneity was assessed by Cochran's Q (significance level was set to P < 0.10) and I 2 [34]. Galbraith plot was used to identify Oncotarget 97591 www.impactjournals.com/oncotarget the studies that contributed to the heterogeneity. Sensitivity analyses were performed by removing one study at a time and recalculating the pooled risk estimate for the remaining studies. Statistical analyses were completed with Stata version 10 (StataCorp, College Station, TX).