Vol.:(0123456789)
1 3
Cancer Causes & Control (2020) 31:431–449
https://doi.org/10.1007/s10552-020-01291-3
REVIEW ARTICLE
Systematic review andmeta‑analysis oftheassociations
betweenbody mass index, prostate cancer, advanced prostate cancer,
andprostate‑specic antigen
SeanHarrison
1,2
· KateTilling
1,2
· EmmaL.Turner
1
· RichardM.Martin
1,3
· RosieLennon
4
· J.AtheneLane
1,3
·
JennyL.Donovan
1,5
· FreddieC.Hamdy
6
· DavidE.Neal
6,7
· J.L.H.RuudBosch
8
· HayleyE.Jones
1
Received: 23 August 2019 / Accepted: 27 February 2020 / Published online: 11 March 2020
© The Author(s) 2020
Abstract
Purpose The relationship between body mass index (BMI) and prostate cancer remains unclear. However, there is an inverse
association between BMI and prostate-specific antigen (PSA), used for prostate cancer screening. We conducted this review
to estimate the associations between BMI and (1) prostate cancer, (2) advanced prostate cancer, and (3) PSA.
Methods We searched PubMed and Embase for studies until 02 October 2017 and obtained individual participant data
from four studies. In total, 78 studies were identified for the association between BMI and prostate cancer, 21 for BMI and
advanced prostate cancer, and 35 for BMI and PSA. We performed random-effects meta-analysis of linear associations of
log-PSA and prostate cancer with BMI and, to examine potential non-linearity, of associations between categories of BMI
and each outcome.
Results In the meta-analyses with continuous BMI, a 5kg/m
2
increase in BMI was associated with a percentage change
in PSA of −5.88% (95% CI −6.87 to −4.87). Using BMI categories, compared to normal weight men the PSA levels of
overweight men were 3.43% lower (95% CI −5.57 to −1.23), and obese men were 12.9% lower (95% CI −15.2 to −10.7).
Prostate cancer and advanced prostate cancer analyses showed little or no evidence associations.
Conclusion There is little or no evidence of an association between BMI and risk of prostate cancer or advanced prostate
cancer, and strong evidence of an inverse and non-linear association between BMI and PSA. The association between BMI
and prostate cancer is likely biased if missed diagnoses are not considered.
Keywords Prostate cancer· Prostate-specific antigen· Body mass index· Screening· Meta-analysis· Systematic review
Electronic supplementary material The online version of this
article (https ://doi.org/10.1007/s1055 2-020-01291 -3) contains
supplementary material, which is available to authorized users.
* Sean Harrison
sean.harrison@bristol.ac.uk
1
Department ofPopulation Health Sciences, Bristol Medical
School, University ofBristol, Bristol, England
2
Medical Research Council Integrative Epidemiology Unit,
University ofBristol, Bristol, England
3
National Institute forHealth Research Bristol Biomedical
Research Centre, University Hospitals Bristol NHS
Foundation Trust andUniversity ofBristol, Bristol, England
4
Department ofEnvironment andGeography, University
ofYork, York, England
5
National Institute forHealth Research Collaboration
forLeadership inApplied Health Research andCare West,
University Hospitals Bristol NHS Trust, Bristol, England
6
Nuffield Department ofSurgical Sciences, University
ofOxford, Oxford, England
7
Department ofOncology, Addenbrooke’s Hospital,
University ofCambridge, Cambridge, England
8
Department ofUrology, University Medical Centre Utrecht,
Utrecht, TheNetherlands
432 Cancer Causes & Control (2020) 31:431–449
1 3
Background
Prostate cancer is the second commonest male cancer
worldwide, [1] and the most commonly diagnosed cancer
in men in the UK, with an estimated 47,151 diagnoses in
2015 [2]. Generally, most prostate cancers are slow grow-
ing, but can metastasize to the bones, lungs, and brain.
Worldwide, there were an estimated 307,000 deaths from
prostate cancer in 2012 [1], and in the UK, around 11,600
men died from prostate cancer in 2016 [2].
Body mass index (BMI) has been associated with
many cancers [3], but its association with prostate can-
cer is unclear. Previous meta-analyses and reviews have
suggested that BMI is not associated with prostate cancer
[4, 5], positively associated with prostate cancer [6, 7],
inversely associated with localized prostate cancer [8], and
positively associated with advanced [8], aggressive [9],
high-grade, and fatal prostate cancers [4]. These meta-
analyses were either limited to cohort studies [4, 5, 7, 8] or
in need of updating [6, 7]. Additionally, no meta-analysis
assessed potential non-linear associations between BMI
and risk of prostate cancer or advanced prostate cancer.
We therefore sought to perform an updated review of the
literature, including more studies, and additionally exam-
ining non-linear associations.
BMI has also been inversely associated with prostate-
specific antigen (PSA) [10], although no previous meta-
analysis of this relationship exists. The presence of such an
association could bias observed relationships between BMI
and prostate cancer as PSA testing often plays a key role in
diagnosis. More specifically, a negative association between
BMI and PSA could lead to a spurious negative association
or mask a positive association between BMI and localized
prostate cancer, as obese men, with lower PSA values, would
be less likely to be offered a biopsy as the result of a PSA
test. A negative association between BMI and PSA could
also induce a spurious positive association between BMI
and advanced prostate cancer, as obese men may be diag-
nosed later, due to their lower PSA levels. In addition, if the
association between BMI and prostate cancer (or advanced
prostate cancer) is non-linear, then studies with different
distributions of BMI will give rise to different estimates
of the BMI-prostate cancer association. There may also be
an association between BMI and prostate cancer screening
behavior (including uptake of PSA testing), though stud-
ies have shown conflicting results. In the USA, men with
high BMI values were more likely to receive PSA tests [11],
whereas in the UK men with both very low and high BMI
values were less likely to receive a PSA test [12]. This fur-
ther complicates the relationship between BMI and prostate
cancer diagnosis (though not BMI and PSA values), and this
review does not aim to assess this association.
We systematically reviewed the literature for all relevant
studies and performed meta-analyses. We also examined
these relationships using individual participant data (IPD)
from prostate cancer studies. In analyzing the IPD studies,
we aimed to account for incomplete and PSA-dependent
diagnosis by imputing prostate cancer status for all men who
did not receive a biopsy, and in doing, avoid potential bias
resulting from an association between BMI and PSA.
Our objectives were to i) precisely quantify the (assumed
linear) associations between BMI and prostate cancer,
advanced prostate cancer, and PSA; ii) update previ-
ous meta-analyses using all relevant evidence, including
case–control studies; and iii) explore potential non-linear-
ity in these associations. Our overall aim was to understand
whether BMI is a risk factor for prostate cancer, and to iden-
tify whether failure to account for the role of PSA in many
prostate cancer diagnoses is likely to lead to biased estimates
of the association between BMI and prostate cancer.
Methods
Eligibility criteria
We performed a systematic review in which we included
original articles published in peer reviewed journals that
measured an association between BMI and total prostate
cancer incidence and/or advanced prostate cancer; and stud-
ies that measured an association between BMI and PSA,
including supplements and meeting abstracts; human rand-
omized controlled trials (RCTs), case–control, cohort, cross-
sectional, and non-randomized experimental studies. If the
abstract did not specifically mention BMI but mentioned
height or weight, we acquired the full text to determine if
BMI was calculable from data included in the publication.
We excluded reviews, books, commentaries, letters, and
animal and cell-line studies; studies examining pre-malig-
nant disease if there was no mention of prostate cancer or
PSA; studies where BMI was measured after diagnosis of
prostate cancer, as this increases the likelihood of reverse
causality; and studies that we considered to be at critical risk
of bias (see ‘Risk of Bias Assessment’ below).
We determined the effect estimate to be for advanced
prostate cancer if the individual studies labeled the effect as
“advanced” or “aggressive,” or if the effect was for locally
advanced, extra-prostatic, nodular or metastatic prostate
cancer. Advanced prostate cancer represents clinically
meaningful cancer, with lower survival rates than non-
advanced cancers. High-grade prostate cancer on its own
was not considered equivalent to advanced prostate cancer
and was not extracted, as the definition of “high-grade” has
been inconsistent over time, incorporating Gleason scores
433Cancer Causes & Control (2020) 31:431–449
1 3
(the definition of which has changed over time [13]), tumor,
node, metastases [TNM] scores, and PSA levels.
Data sources
We searched Medline and Embase databases up to 02 Octo-
ber 2017 for studies in humans associating BMI with either
prostate cancer or PSA. The search query was as follows
(each term as a text word search): (BMI or body mass index
or obese or obesity or body weight or body size or adipos-
ity) AND (prostate cancer or prostate neoplasm or PSA or
prostate-specific antigen) NOT psoriatic arthritis. Psoriatic
arthritis was excluded as its initialism is also PSA. We also
reviewed the reference lists of previous meta-analyses for
further studies for inclusion [6, 8, 14]. Duplicate studies
were removed prior to download using the Ovid deduplica-
tion tool.
Data extraction
One author (SH) screened the titles and abstracts of all
papers for inclusion and retrieved full texts for all studies
that met the inclusion criteria. Full texts were also sought
if no abstract was available or if the abstract did not include
sufficient information to decide on inclusion. We also sought
full texts for conference abstracts, if a corresponding full text
was not found in the original search. If no full text could be
found, and the abstract provided insufficient information for
inclusion, the study was excluded. We excluded one pub-
lished paper where we could not locate a full text [15].
One author (SH) screened all full texts for inclusion, and
one of three independent reviewers (KT, ET, HJ) reviewed
the first 60 full texts to check for consistency. We resolved
any inconsistency with discussion to clarify screening crite-
ria. A random subset of the remaining studies [30 full texts]
was also reviewed by the independent reviewers to check for
drift from inclusion/exclusion criteria.
Both SH and RL independently extracted all relevant data
from included studies, with disagreements resolved by dis-
cussion. The first ten extractions were also performed by
HEJ, KT, and ELT to check for consistency.
We categorized prostate cancer studies as “before” if BMI
was measured on average at least two years before diagnosis
(prospective studies), and “same time” if BMI was measured
on average less than two years before diagnosis. In general,
“before” studies were cohort studies and “same time” stud-
ies were case–control studies. We considered the “before”
studies to be at lower risk of reverse causation.
We extracted data that were (or could be transformed to)
an odds ratio (OR) or hazard ratio (HR) quantifying the con-
tinuous association between BMI and total and advanced
prostate cancer risk, and a regression coefficient for the
association between BMI and log-PSA. Log-PSA was used
as an outcome rather than PSA as we assumed a multiplica-
tive association between BMI and PSA, which fits with the
theory that haemodilution is responsible for any observed
association [16]. Studies reported associations in a variety
of ways; a detailed list of the statistical conversions used to
estimate the ORs, HRs, and regression coefficients and their
standard errors (SEs) is in Supplementary appendix 1.
We estimated linear associations, taking BMI as a con-
tinuous exposure variable, and assessing the possibility
of non-linear associations by coding BMI as a categori-
cal exposure. Specifically, we estimated linear associa-
tions between BMI and the log odds of prostate cancer or
advanced prostate cancer, and between BMI and log trans-
formed PSA. For simplicity, we refer to linear associations
as “continuous” throughout. The following BMI categories
were used: normal weight (BMI < 25kg/m
2
), overweight
(25kg/m
2
≤ BMI < 30kg/m
2
), and obese (BMI ≥ 30kg/m
2
).
We refer to these as “categorical” associations throughout.
When several papers reported on the same study, for con-
tinuous associations we prioritized papers that presented
continuous effect estimates (e.g., HR or OR per 1kg/m
2
increase in BMI) over papers presenting categorical effect
estimates (e.g., HR or OR for overweight and obese groups
versus normal weight), and these were prioritized over mean
differences. For categorical associations, we extracted esti-
mates from papers presenting categorical associations only.
If duplicate studies presented the same effect estimate types
in multiple papers, the paper with the largest number of
participants was used in the meta-analysis. If both adjusted
(e.g., for potential confounders such as age, ethnicity, etc.)
and unadjusted results were given in the same paper, the
most-adjusted model was used in the meta-analysis.
If the data were insufficient to estimate a regression coef-
ficient, OR or HR and SE, we extracted a p value, the num-
ber of participants and direction of association from the most
relevant analysis for use in an albatross plot [17].
Risk ofbias assessment
SH and RL assessed the risk of bias in each study inde-
pendently using an assessment tool created for a previous
meta-analysis [18], with disagreements resolved by discus-
sion. This tool uses the categories of assessment from a draft
of the ROBINS-I tool [19], and questions from the CASP
case–control and cohort questionnaires [20, 21], see Sup-
plementary appendix 2.
We assessed risk of bias in six categories: confounding,
selection of participants, missing data, outcome measure-
ment, exposure measurement, and results’ reporting. We
assigned overall and category-specific risks of bias: either
low, moderate, high, critical, or unclear (if there was insuffi-
cient information to assign a risk). We based the overall risk
of bias on a subjective combination of the category-specific
434 Cancer Causes & Control (2020) 31:431–449
1 3
risk of biases, looking at the maximum risk of bias that
could have been introduced into the study by each category.
The overall risk of bias was not low in any study, as all
studies were observational and thus potentially subject to
unmeasured confounding.
We determined that a study had a critical risk of bias if
i) age was not accounted for in either the design or analy-
sis of the study and, for BMI-prostate cancer case–control
studies, if there was more than a 3-year difference in the
mean or median ages of cases and controls, because age is
strongly associated with BMI [22], prostate cancer risk [23],
and PSA [23]; or ii) if the design of the study was such that
participation was conditional upon PSA levels, both for the
association between BMI and PSA (as this would involve
conditioning on the outcome) and the association between
BMI and prostate cancer (as this would involve conditioning
on a collider) [24].
Studies with a critical risk of bias were excluded prior to
analysis and were not considered further.
In the studies found in the systematic review, it was gen-
erally unclear whether men considered as not having prostate
cancer had received biopsies. Usually, the controls were “not
known to have prostate cancer,” rather than “known not to
have prostate cancer.” Therefore, screening could have intro-
duced bias in the association between BMI and prostate can-
cer. Although we did not consider this a critical risk of bias,
we sought to investigate and quantify this bias using large
studies where biopsy status was known, and IPD available.
Individual participant data studies
Studies that offered prostate biopsies if the participants’
PSA were above threshold values (screening studies) were
excluded from our systematic review for having a critical
risk of bias. However, we noted that some of the largest
potentially relevant studies for our research questions were
screening studies, and that bias due to screening could
potentially be accounted for using imputation of prostate
cancer status if IPD were available. This would then allow
these studies to be included in the meta-analyses.
We approached four prospective studies looking at pros-
tate cancer to obtain IPD: Krimpen [25], Prostate Cancer
Prevention Trial (PCPT) [26], Prostate, Lung, Colorectal,
and Ovarian cancer screening trial (PLCO) [27] and Pros-
tate Testing for cancer and treatment trial (ProtecT) [28].
These studies were chosen because they were large studies of
prostate cancer with known PSA screening protocols, or the
biopsy status of all participants was known. Key to inform-
ing the imputation model was PCPT, which offered biopsies
to all participants regardless of PSA level. This information
allowed us to predict prostate cancer status for men with a
PSA level below the threshold for biopsy in the other three
studies using multiple imputation. However, PCPT only
included men with a PSA less than 3.0ng/ml, biasing both
the BMI-PSA and BMI-prostate cancer analyses, and as such
was excluded from the meta-analyses due to the critical risk
of bias from conditioning on a collider or outcome. Imputa-
tion is valid if the missing data (prostate cancer status) is
missing at random given other variables in the imputation
model, so imputing prostate cancer is not biased even though
PCPT is restricted to men with a PSA less than 3.0ng/ml, as
PSA is in the imputation model [29].
For each IPD study, we requested data measured at
baseline on BMI and PSA, as well as age, family history
of prostate cancer and ethnicity. We also requested data on
prostate cancer status (including tumor, node, metastases
[TNM], and Gleason scores). For each man who was not
biopsied, we imputed prostate cancer status by the end of
the study in which he participated using multiple imputa-
tion. We included baseline age, BMI, log-PSA, family his-
tory of prostate cancer, and study as explanatory variables
to predict prostate cancer status using multiple imputation.
BMI, log-PSA, and family history of prostate cancer were
also imputed if missing.
We checked the validity of the imputation model by
checking whether the predicted incidence of prostate cancer
among men without prostate biopsies was credible, given
results from autopsy studies [30]. Additionally, we visually
inspected a plot of estimated prostate cancer risk against
PSA for the imputed studies, to see whether the predicted
risk of prostate cancer at low PSA levels for each study was
plausible (see Supplementary Appendix3.4).
In each of the three included IPD studies, we estimated
associations between BMI and (1) prostate cancer, (2)
advanced prostate cancer, and (3) PSA. We restricted the
analyses to men with white ethnicity (due to low numbers
of non-white men and therefore difficulties in imputation),
and adjusted the analyses for age, family history of prostate
cancer (for prostate cancer analyses), and prostate cancer
status (for the PSA analyses). Full details of the IPD studies,
the imputation method, and statistical analyses are available
in Supplementary Appendix3.
Combining data
Meta‑analysis
We combined estimates from studies identified through the
systematic review and the IPD studies using random-effects
and fixed-effect meta-analyses. We performed separate meta-
analyses of continuous and categorical associations for each
outcome (prostate cancer, advanced prostate cancer, and
PSA). All meta-analysis results are presented in forest plots.
Studies presenting HRs and ORs were analyzed and pre-
sented separately. For studies presenting ORs, “same time”
and “before” studies were meta-analyzed in subgroups,
435Cancer Causes & Control (2020) 31:431–449
1 3
and labeled as such in forest plots. Studies presenting HRs
were all classed as “before” studies, and labeled simply
“HR.” The results are presented as the HR or OR for pros-
tate cancer or advanced prostate cancer and percentage
change in PSA for a 5kg/m
2
increase in BMI. Heterogene-
ity was tested for and quantified using the Cochran’s Q and
I
2
statistics [31, 32].
In meta-analyses of categorical associations, studies
from the systematic review were included if they presented
ORs or HRs for overweight and/or obese men relative to
normal weight men (for the outcomes of prostate cancer
and advanced prostate cancer) or means and SDs of PSA
or log-PSA for each of these BMI categories (for the out-
come of PSA). ORs and HRs that were presented for other
categories of BMI were not used (such as morbidly obese,
BMI ≥ 35kg/m
2
), though we combined the mean and SD
of PSA for different categories with neighboring catego-
ries when sufficient information was available.
Meta‑regression
Meta-regression [33] was used to determine if the effect
estimates from individual studies included in the meta-
analyses varied by study-level factors. For all meta-regres-
sions, we considered ethnicity (non-white versus white in
each study, defined as > 80% white participants or from a
country with a majority white population), mid-year of
recruitment, mean BMI in the study, and the overall risk of
bias (high versus medium). For the associations between
BMI and prostate cancer and advanced prostate cancer, we
also considered the mean age at diagnosis, and study mean
time between BMI measurement and diagnosis.
Funnel plots
Funnel plots [34] were drawn to assess for small study
effects in each analysis [35].
Albatross plots
As not all studies reported enough information to be
included in the meta-analyses, we also present albatross
plots containing results from studies with and without suf-
ficient information to be included in the meta-analyses [17].
These are plots of the p value of an association against the
number of participants and can be used to assess heteroge-
neity between studies and assess the rough magnitude of an
association using limited information. By indicating which
studies had insufficient data to contribute to meta-analysis on
the albatross plots, we determined whether inclusion of the
remaining studies would have altered the overall interpreta-
tion of the evidence.
Results
In total, 9,127 papers were found that had keywords for
BMI and prostate cancer or PSA. After title and abstract
screening, 725 papers remained (see Fig.1, PRISMA flow
diagram). After full text screening, risk of bias assessment,
and removal of papers reporting the same studies, 78 stud-
ies examined the association between BMI and prostate
cancer [67 with data for meta-analysis], 21 studies exam-
ined the association between BMI and advanced prostate
cancer [18 with data for meta-analysis], and 35 studies
examined the association between BMI and PSA [20 with
data for meta-analysis, one of which only had data for
categorical associations].
A summary of all results is given in Table1.
BMI andprostate cancer
Continuous BMI
Of the 78 studies examining the association between BMI
and prostate cancer [25, 27, 28, 36–110], 11 (14%) could
not be included in the meta-analysis due to insufficient
data but were included in the albatross plot [100–110].
All studies are detailed in Supplementary Table1, with
the results of the risk of bias assessment in Supplementary
Table2. All studies in the meta-analysis adjusted for age
in either the study design or analysis, while 23 studies
(34%) adjusted for smoking status, 22 (33%) for ethnicity,
20 (30%) for family history of prostate cancer, 13 (19%)
for education, 10 (15%) for area, 10 (15%) for diabetes,
10 (15%) for physical activity, 9 (13%) for alcohol, 6 (9%)
for diet, and 6 (9%) for income. No other variable (of 24
other variables) was adjusted for in more than four studies.
In total, 9,513,326 men from 67 studies were included
in the HR and OR meta-analyses, (9,351,795 in 30 HR
studies, 161,531,383 in 37 OR studies); of these, 201,311
(2.1%) men had prostate cancer (157,990 cases [1.7%] in
HR studies, 41,863 [25.9%] in OR studies). The random-
effects meta-analyses (Figs.2 and 3) estimated the average
HR and OR for prostate cancer for a 5kg/m
2
increase in
BMI to be 1.01 (95% CI 0.99–1.04, p = 0.29) and 0.99
(95% CI 0.96–1.02, p = 0.64), respectively. There was
strong evidence for heterogeneity in effect estimates
across studies for the studies reporting an HR (p < 0.001,
I
2
= 79.9%), and studies reporting an OR (p < 0.001,
I
2
= 65.8%). Pooled estimates from fixed-effect meta-
analyses were essentially the same.
