Risk factors for Epstein Barr virus-associated cancers: a systematic review, critical appraisal, and mapping of the epidemiological evidence

Background Epstein Barr Virus (EBV) infects 90%-95% of all adults globally and causes ~ 1% of all cancers. Differing proportions of Burkitt’s lymphoma (BL), gastric carcinoma (GC), Hodgkin’s lymphoma (HL) and nasopharyngeal carcinoma (NPC) are associated with EBV. We sought to systematically review the global epidemiological evidence for risk factors that (in addition to EBV) contribute to the development of the EBV-associated forms of these cancers, assess the quality of the evidence, and compare and contrast the cancers. Methods MEDLINE, Embase and Web of Science were searched for studies of risk factors for EBV-associated BL, GC, HL and NPC without language or temporal restrictions. Studies were excluded if there was no cancer-free comparator group or where analyses of risk factors were inadequately documented. After screening and reference list searching, data were extracted into standardised spreadsheets and quality assessed. Due to heterogeneity, a narrative synthesis was undertaken. Results 9916 hits were retrieved. 271 papers were retained: two BL, 24 HL, one GC and 244 NPC. The majority of studies were from China, North America and Western Europe. Risk factors were categorised as dietary, environmental/non-dietary, human genetic, and infection and clinical. Anti-EBV antibody load was associated with EBV-associated GC and BL. Although the evidence could be inconsistent, HLA-A alleles, smoking, infectious mononucleosis and potentially other infections were risk factors for EBV-associated HL. Rancid dairy products; anti-EBV antibody and EBV DNA load; history of chronic ear, nose and/or throat conditions; herbal medicine use; family history; and human genetics were risk factors for NPC. Fresh fruit and vegetable and tea consumption may be protective against NPC. Conclusions Many epidemiological studies of risk factors in addition to EBV for the EBV-associated forms of BL, GC, HL and NPC have been undertaken, but there is a dearth of evidence for GC and BL. Available evidence is of variable quality. The aetiology of EBV-associated cancers likely results from a complex intersection of genetic, clinical, environmental and dietary factors, which is difficult to assess with observational studies. Large, carefully designed, studies need to be strategically undertaken to harmonise and clarify the evidence. Registration PROSPERO CRD42017059806.


VIEWPOINTS PAPERS
Epstein Barr Virus (EBV) is a herpesvirus estimated to infect 90%-95% of all humans [1]. Infection is lifelong and in low-income settings occurs in childhood; in higher income settings it may also occur in childhood but is often delayed until late adolescence or adulthood. In a minority of people infection is associated with the development of cancer [2]. Indeed, Epstein, Achong and Barr' s work describing the presence of EBV in a cultured Burkitt' s lymphoma (BL) cell line was the scientific foundation implicating the first known human tumour virus [3]. EBV is now associated with 1% of global cancers, which are mostly lymphomas and carcinomas; approximately 140 000 people die of EBV-associated cancers each year [4,5].
Despite decades of intensive research, the aetiology of EBV-linked cancers remains unclear. As only a fraction of EBV-infected individuals develop cancer, other factors must also be influential. Several reviews have been undertaken of the biological and epidemiological evidence for the association between EBV and different cancers [5][6][7][8][9][10][11][12][13][14][15][16], but none have systematically collated the epidemiological evidence for additional risk factors beyond EBV infection, or assessed the quality of the presented evidence. Pragmatically, such an undertaking requires a focus on a small number of contrasting cancer types that have strong evidence of a causal relationship with EBV, are of significant burden and/or have globally disparate distributions, such as BL, Hodgkin' s lymphoma (HL), gastric carcinoma (GC), and nasopharyngeal carcinoma (NPC).
In 2016, there were approximately 461 000 incident cases of non-Hodgkin' s lymphoma (NHL) globally [17]. BL is a highly aggressive and fast-growing NHL classified in three forms: endemic, sporadic and AIDS-associated [5]. The endemic form is thought to have EBV present in more than 95% of tumours; this decreases to 15%-88% for sporadic tumours and 30%-40% for those that are AIDS-associated [5]. Endemic BL is generally a paediatric condition of the jaw found in equatorial Africa and Papua New Guinea which has been strongly linked to malaria [5]. Sporadic BL is associated largely with white populations, across a wider age range and disease of the abdomen [5]. All types of BL are more common in males than females [5].
HL is a lymphoma characterised by the presence of Hodgkin and Reed-Sternberg cells which has various histological appearances. There were an estimated 73 000 incident cases globally in 2016 [17]. 40%-50% are thought to be associated with EBV, but the degree of association is highly population-dependent; EBV positive HL as a percentage of all HL is highest (90%-100%) in lower income countries [18][19][20]. Incidence rates of this form of cancer peak in children and individuals over 70 years of age, are more common in non-white ethnic groups, and in males than females [5].
With 830 000 deaths and 1.2 million cases globally (2016), GC is a major cause of morbidity and mortality [17]. Around 95% of GC is gastric adenocarcinoma; approximately 10% of this is EBV-associated, with variation between populations [5]. Due to the commonality of GC, EBV-associated GC is likely to be the most common EBV-associated malignancy, but is poorly studied [5]. The cancer is associated with being male; for men prevalence decreases with age [21]. 96 000 cases of cancer of the nasopharynx occurred in 2016 [17]. NPC forms the majority of this burden and occurs in the epithelial cells of the nasopharynx. Tumours are classified depending upon the level of differentiation in these cells: type I (squamous), type II (non-keratinising), type III (undifferentiated). EBV is strongly associated with types II and III and more controversially with type I [22]. NPC incidence is geographically extremely variable, with the disease associated with northern Africa and southeast Asia, particularly southern China and East Malaysia [5]. The disease is more common in males; age associations vary between settings.
Given the global importance of EBV-associated BL, HL, GC and NPC, their contrasting features yet common link with EBV, and their complex but unclear aetiology, we undertook a systematic review to map the current global epidemiological evidence base of risk factors for the EBV-associated forms of these tumours in addition to EBV itself. We present an evaluative narrative summary of that evidence, a quality assessment, and an appraisal of the critical gaps in the literature.

VIEWPOINTS PAPERS
America) were searched for studies of risk factors for a) EBV-associated HL and BL and b) EBV-associated NPC and GC in addition to EBV in June to August 2017 (Appendix S1 in the Online Supplementary Document). Search strategies were compiled for MEDLINE and then adapted for the other databases. Reference lists of included papers and review articles were also searched ('snowballing'). This review was registered on PROSPERO as CRD42017059806.

Study selection
Inclusion criteria: • Epidemiological studies (observational and interventional) of risk factors for EBV-associated HL, BL, NPC and/or GC with extractable data and a cancer-free comparator population.
• No date restrictions.
• No population restrictions.
• No language restrictions.
• Studies of risk factors for EBV infection, not EBV-associated cancers.
• Studies where the tumours were not proven to be EBV positive, unless for NPC, where EBV is thought to be associated with 95% of tumours [23].
• Comparator population had cancer or pre-cancerous lesions (including both EBV-negative HL/BL/ GC/NPC and other forms of cancer).
• Gene/protein expression or genetic (human or EBV) studies of samples taken from cancer tissues, due to the potential for post-tumour mutations.
• Genetics studies where samples were taken from different bodily sites for the cancer and comparator groups.
• EBV genetic and anti-EBV antibody expression studies where samples were taken cross-sectionally from cancer patients, due to the potential for tumour-induced expression, or where antibody load was not documented.
Hits were screened at the title, abstract and full text stages by three reviewers per cancer, with at least a 10% overlap. Disagreements were resolved by an independent reviewer. Data extraction was undertaken by the same three reviewers, with at least 10% overlap. Discrepancies were resolved by consensus.

Data extraction
Data were extracted into a pre-designed spreadsheet which included data on the study design, population and risk factors assessed for EBV-associated cancers.

Quality assessment
The quality of included studies was assessed using a checklist adapted from Downs and Black [24], as per the guidance issued by Deeks et al. [25]. Quality assessment was made for the main risk factor(s) of interest; if multiple risk factors were documented the most conservative score was taken. The minimum sample size required to detect a relative increase in cancer of 50% from a statistically conservative baseline of 50% among the unexposed was calculated at different powers (with a significance level of 5%) using the Fleiss method within Epi Info. Different thresholds were set for cohort/cross-sectional studies and (unmatched) case-control studies. Conservatively, a ratio of one to one between exposure strata was assumed, as well as only two strata. This criteria was scored from 0 (<70% power) to 5 (>99% power).
We pragmatically defined a minimal confounder set as age, sex and ethnicity. Three reviewers undertook the assessment, with at least 10% overlap. Discrepancies were resolved by consensus.

Data synthesis
Due to between-study heterogeneity, the findings of the included publications were narratively synthesised by theme.

Ethics approval
As a systematic review of previously published data, this work did not require ethical approval.

Search results and included studies
After de-duplication, 5373 hits were retrieved by the lymphomas search and 2823 by the carcinomas search (Figure 1). This was reduced to 229 and 260 hits, respectively, after title and abstract screening. 271 papers were extracted, including those identified through snowballing: two on EBV-associated BL, 24 on EBV-associated HL, one on EBV-associated GC and 244 on NPC (Table S1 in the Online Supplementary Document). All BL, HL and GC publications used case-control designs (although one of the HL papers also contained a reconstructed cohort) and 227/244 (93.0%) of those for NPC. Papers analysed populations ranging in size between 11 and hundreds of thousands (longitudinal studies). The majority of NPC analyses were undertaken in China, GC and HL papers in Europe and the Americas, and BL papers in Africa (Figure 2).

Quality assessment of included studies
In the overall quality assessment, 32/271 (11.8%) of publications were very under-powered, with a score of zero or one, ie, less than 80% power (  EBV-associated nasopharyngeal carcinoma and gastric carcinoma. EBV -Epstein Barr virus.

VIEWPOINTS PAPERS
Across the publications, none were thought to be at high risk for observer bias for the outcome, but the potential for recall bias was higher (85/271, 31.4%), with exceptions being studies of genetic risk factors.

Risk factors for Burkitt's lymphoma
Two publications met our inclusion criteria for EBV-associated BL. Both were undertaken on the same population in Uganda in the 1970s. They examined infection and clinical factors, specifically anti-EBV antibody load prior to cancer diagnosis (anti-viral capsid antigen [VCA], anti-early antigen [EA] and anti-EBV nuclear antigen [EBNA]) [27,28]. Only levels of anti-VCA were found to be associated with the development of BL. Both had small sample sizes and largely utilised descriptive statistics.
Of the papers partially or totally examining diet as a risk factor for HL (Table 1), one looked at dietary patterns [43], one alcohol [38], and one dietary fats [44]. None of the papers found any statistically associated risk factors. Two of these publications used substantially overlapping study populations [43,44]. Due to the retrospective collection of dietary information, recall bias may have been an issue.
Within the infection and clinical factors publications, three from Europe provided an indication that personal infectious mononucleosis (IM) may be a risk factor for HL ( Table 2) [32,40,42]. Although the two publications from the Americas had a high degree of uncertainty (confidence intervals (CI) crossed the null), the reported effect estimates were somewhat suggestive of risk for either personal or familial IM [47,49]. Importantly, all exposure data was self-reported.

VIEWPOINTS PAPERS
Three publications examined the impact of childhood infections. Using a combined variable, Glaser et al. found that measles, mumps or rubella may be protective against HL diagnosed in younger patients aged 19-44 years [47]. When these factors were considered individually CIs crossed the null, but the effect estimates tended in the same direction. In older adolescents and young adults, Alexander et al. demonstrated that two or more infections (measles, mumps, chicken pox, pertussis) were protective without examining each infection individually [40]. Within Linabery et al. (children and adolescents), the direction of effect was protective for mumps, neutral for measles and in favour of risk for rubella (all with CIs that crossed the null), but Strep or sore throat/Scarlet fever/Tonsillitis as a combined variable was associated with greater HL risk and likely contributed strongly to overall findings for infections [49]. Again, all exposure data was self-reported. Another publication documented a potential association with cytomegalovirus (CMV) serostatus, but minimal information was available as it was a conference abstract [51].
Two papers examined the impact of autoimmune and allergic conditions on the risk of HL [33,49]. Although each found specific risk factors among the autoimmune conditions (eg, personal rheumatoid arthritis, parental autoimmune conditions), the evidence was not consistent. Neither scored higher for quality than the other.

Infection:
Childhood infections (not IM) Two or more infections (measles, mumps, chicken pox, pertussis, rubella) associated with reduced risk [40] Increased risk associated with Strep or sore throat/scarlet fever/tonsillitis, infections in siblings.
Other personal infections also considered [49] Measles, mumps, rubella (as a single variable, but not individually) -potentially protective [47] Measles, mumps, rubella in older adult patients and chicken pox across all age groups -no association found [47] IM (personal) Personal IM associated with increased risk (across all ages and in younger adults alone; association not seen in older adults alone) [32] Personal IM associated with increased risk [40] Personal IM associated with increased risk [42] Personal IM -no association found [47,49]

Autoimmune diseases and allergies
Rheumatoid arthritis associated with increased risk. Other autoimmune and allergic conditions also examined [33] Autoimmune conditions in parents associated with increased risk. Personal autoimmune or allergic conditions, autoimmune conditions and allergies in siblings, allergies in parents -no association [49] BMI, weight, height No association (BMI) [39] Higher weight and BMI associated with protection [45] No association [48] Physical activity Participating in (strenuous) physical activity as an adult associated with protection [48]

Family history of cancer
Earlier age of cancer onset among family members may be associated with an increased risk. Different types of tumour examined [50] BMI -body mass index, CMV -cytomegalovirus, IM -infectious mononucleosis

VIEWPOINTS PAPERS
A second publication by Linabery et al., using an overlapping population to their infection, autoimmune and allergy study [49], examined the impact of family cancer history, but found no clear associations [50].
Body mass index (BMI), weight, height and measures of physical activity were examined as risk factors by three publications [39,45,48]. Physical activity as an adult was associated with protection from HL. Two papers documented no BMI association and one found that higher BMI/weight was protective. The direction of effect was not associated with the quality of the evidence.

HLA and associated genes:
HLA-A A*01 associated with increased risk, A*02 with decreased risk [36] A*01:01 associated with increased risk [41] HLA-B B*08:01 associated with increased risk [41] B*07 and B*08 no association [29] HLA region Locus D6S265 allele 126 and locus D6S510 allele 284 heterozygotes and homozygotes (both HLA class I) associated with HL risk in a classic association analysis. (Other, weaker, associations also found.) Association lost when haplotype sharing statistic analysed [35]. Later narrowed down as above [36].

KIR
No clear association [52] HL -Hodgkin' s lymphoma, HLA -human leukocyte antigen, KIR -killer-cell immunoglobulin-like receptors, SNP -single nucleotide polymorphism, TNF -tumour necrosis factor Among the HLA papers, the evidence was balanced in favour of a HLA-A association, particularly an increased risk with the HLA-A*01 serotype group. Three publications represented a gradual narrowing down to A*01 [35][36][37]. For B alleles, the evidence was contradictory. Various other alleles and loci were also associated.
Of the publications examining environmental and non-diet lifestyle factors, three looked at smoking as a risk factor for HL (Table 4) [31,38,46]. All three provided evidence supportive of an association between current smoking and HL (and two of ever having smoked and HL), including from Willet et al., which had previously found no association with alcohol. [38] More detailed breakdowns of smoking status were less clear, however. There was a risk of recall bias in all of the analyses and only one adjusted for all of our a priori confounders, but analysis choices were deemed appropriate throughout.

Risk factors for gastric carcinoma
One publication met our inclusion criteria for EBV-associated GC. In American men of Japanese ancestry, Levine et al. examined infection and clinical factors for GC, specifically anti-EBV antibody load prior to cancer being diagnosed (IgG and IgA anti-VCA, IgG anti-EA and IgG anti-EBNA) [53]. IgG anti-VCA was specifically found to be associated with EBV-positive GC vs non-cancer controls. Power was very low and the minimal confounder set was not adjusted for. Among the dietary risk factors papers, there were a series of common factors that were assessed ( Table  5). The first of these was alcohol (often in combination with smoking), where generally the evidence was not in favour of involvement [56,[58][59][60][61][62][63][64][65][66][67][68][69][70][71][72][73].
Two papers were in agreement that preserved plums were a risk factor for NPC [84,93]. Broadly, fresh fruit and vegetables were deemed protective, as was the consumption of tea in various forms, including herbal.
The infection and clinical papers examined risk factors that fell into eight main categories-EBV serology/ DNA load/genetics, other infections, medical history, medication, oral hygiene, T-cells, and family history of NPC or cancer ( Table 6).
In terms of medical history, the general trend of evidence was that a history of (chronic) ear nose and/or throat conditions was associated with NPC risk, as were herbal medicines. Nasal balms, drops, ointments, oils or sprays were generally not associated (five of seven papers).

Smoking
Having ever smoked and being a current smoker associated with increased risk [38] Having ever smoked and current smoking associated [31] Current smoking associated [46] Years smoked, pack years and years since stopped smoking -no association [38] Age at initiation of smoking, duration, intensity, cumulative exposure, time since cessation -not associated [31] Having ever smoked, intensity, duration, age at initiation, years since cessation and childhood exposure -no association [46] Childhood environmental factors Numbers of younger siblings associated with protection [32] Being an older sibling potentially protective among younger adult patients [47] Number of older siblings -no association [32] Childhood household size no association [47] Bedroom sharing associated with reduced risk among younger adult patients [47] Persons per room, attendance at kindergarten, mother' s age at birth, personal and parental education levels -no association [32] Number of playmates -no association [47] VIEWPOINTS PAPERS Potential risk factor summary of results Preserved plums risk factor [84,93] Preserved vegetables generally risk factor (although salted vegetables and picked Chinese cabbage protective) [93] Preserved vegetables no association [86,88] Preserved fruit no association [79] Pickled vegetables (and fungus on pickles) risk factor [S7] Mouldy bean curd no association [84] Vegetables, beans and fruit Chung choi not consistently associated, could be risk factor [83] Fresh fruits in childhood protective [82] Fresh fruit associated protective [61] Fresh fruit and vegetables protective [79] Fresh green vegetables protective, others no association [84] Fruit and vegetables protective [68] [S6] Leafy vegetables protective [85] Dark vegetables and fresh fruit protective [S5] Grapes protective [91,93] Non-preserved fruits and vegetables generally protective [93] Cooked vegetables and citrus fruits not consistently associated [75] Fresh fruit and vegetables (including green and leafy) no association [74,86] Servings per week of fruit and vegetables no association [88] Carrots no association [64,86,91]

Medical history
Allergic rhinitis associated risk factor [S15] Acute and/or chronic rhinosinusitis risk factor [S9,S16] Paranasal sinusitis risk factor [S17] Sinusitis potential risk factor [92] (Chronic) ear, nose (and throat) conditions risk factor [60,62,73] [S18] (Chronic) ear and nose diseases -inconclusive [83] Ear, nose and throat conditions -no association [64] Hayfever, tonsillectomy, heart disease, diabetes, cold sores, canker sores no association [ Twenty publications analysed whether a family history of NPC was associated with personal risk, with the balance of evidence on the side of this being a risk factor (15 publications), in line with the genetic evidence presented in Table 7. Three other publications contained effect estimates that trended in the same direction. Among the 15, only one adjusted for genetic factors in the analysis; four others adjusted for other shared factors eg, diet. Five of the seven analyses examining whether family history of cancer as a whole was associated with personal NPC also had positive results; the likelihood and strength of the associations could be linked to the setting in which this work was undertaken, ie, how large a proportion of all cancer cases are due to NPC, although the studies were geographically confined within Asia.

General dust and smoke
General smoke exposure risk factor [S1] General dust exposure no association [S1] Broad analyses of the impact of dust, fumes and smoke in occupational settings were inconclusive across included papers, however, within the different analyses of dust, wood dust was found to be associated in three analyses, but not in a fourth. Other occupational wood exposure was not found to be associated, but use of wood fuels in the home was found to be a risk factor within two publications. Poor ventilation in the workplace was not found to be associated with NPC risk. Within the home, the balance of evidence suggested that poor ventilation could be a risk factor. The association with socioeconomic status and education was inconsistent.

DISCUSSION
BL, HL, GC and NPC are established as EBV-associated cancers. We present the first systematic review of epidemiological studies of risk factors in addition to EBV for the EBV-associated forms of these cancers, with an associated quality assessment. The 271 included publications provide a rich overview of our knowledge on the causes of EBV-associated BL, HL, GC and NPC. We highlight smoking, IM, and the HLA genetic region as risk factors for HL and being an older sibling as potentially protective. Rancid dairy products, salted fish, anti-EBV antibody and EBV DNA load, history of chronic ear, nose and/or throat conditions, herbal medicine use, family history, and genetics are NPC risk factors, as well as potentially smoking. In addition, fresh fruit and vegetables, slow cooked soup, and tea consumption are potentially protective against NPC. Anti-EBV antibody load was found to be associated with both BL and GC. We thus demonstrate a wider range of risk factors for NPC than the other EBV-associated cancers, which could either represent a more complex aetiology for NPC, or simply the limited number of publications for the other diseases.
In addition to the extensive array of studies and thus risk factors documented within this review, we note that our fundamental understanding of EBV and its associated cancers is underpinned by many important VIEWPOINTS PAPERS laboratory and epidemiological studies not captured by our specific and rigorous inclusion criteria [98]. Critically, we note the detailed evidence accumulated over decades linking endemic malaria to BL, which has been derived largely from settings where the EBV association with BL is robust [99]. Less strongly, HLA type, plants in the Euphorbiaceae family, and sociodemographic factors have been suggested to be associated with BL [5,100,101]. For GC, inference from studies where EBV status is unknown or negative is more problematic, as the EBV-associated form of the disease represents a distinct subtype with markedly different genomic, immunologic and pathological features [102], many of which have direct therapeutic relevance [103].
The major strength of this work is its systematic approach to examining published risk factors for EBV-associated cancers, including a quality assessment and mapping of the available literature. The global and temporal scope of our review allows a wide-reaching consolidation of the literature to date, as well as an identification of our knowledge gaps and targeting of future studies. As to its limitations, many of the NPC papers (particularly genetics studies) were only captured during the snowballing process, likely due to our decision to follow the STROBE guidelines for indexing when choosing our search terms. We excluded studies of genetic factors where samples were taken from cancer tissues, due to concerns about malignancy-induced genetic changes. This may, however, have excluded some genetics studies of loci that are unlikely to have been mutated in a cancerous cell. It should be noted that, where we document that a publication provides no evidence of an association, this does not mean that the converse is true; some studies were simply under-powered, thus confidence intervals were wide and estimates statistically uncertain. Our sample size calculations did not consider the implications of matching. We only documented analyses answering the declared aims or hypotheses of a publication; while post hoc exploratory data analyses can provide valuable insights, they require formal validation by subsequent studies and can sometimes arise due to data dredging [104].
Regarding the methodological quality of the included publications, we note several common features.
Testing of multiple factors was common, increasing the likelihood of chance findings. Slightly over a third of publications used inappropriate statistical methods and only a quarter adjusted for our pragmatically minimal confounder set of age, ethnicity and sex. We acknowledge, however, that studies from certain settings would have recruited from ethnically homogenous populations, removing the need to control for ethnicity. Particular care should be taken interpreting the findings of dietary and environmental factor studies that do not adjust for upstream socioeconomic determinants of these factors and generally of recall bias. A few publications only presented results for factors positively associated with cancer.
As this review was specifically one of EBV-associated forms of our chosen cancers, we were limited by the number of studies that did not report EBV status. For GC, this reflects the relatively recent documentation of GC' s EBV-association and the fact that histological samples are not tested for the virus as standard. Without such testing, future meta-analyses will be substantially hampered. The close association between BL and EBV in endemic areas means that we could have included all publications from these regions regardless of EBV tumour status, however we opted for a consistent approach per tumour.
The major implications of our findings for public health are as follows. First, some of the critical risk and protective factors documented are modifiable eg, dietary elements, such as the consumption of rancid dairy products. These present opportunities for governmental interventions. Second, risk profiles can be built from these data for use as screening tools in areas of high cancer incidence. Third, our thorough documentation of the literature to date provides a signpost for future studies a) examining promising, but not fully proven, risk or protective factors in a broader span of geographical locations and b) to ensure that causal networks can be thoroughly mapped and thus confounders appropriately adjusted for [105]. Finally, it is importantly to reflect on the fact that the EBV-associated forms of certain cancers, eg, GC-differ from the non-EBV-associated forms. Thus reviews such as ours, where EBV-associated forms are considered separately, are critical to order to direct control efforts.

CONCLUSIONS
We document 271 epidemiological publications on risk factors in addition to EBV for the EBV-associated forms of BL, GC, HL and NPC; the majority focussed on NPC. The quality of the available evidence was variable. The aetiology of EBV-associated cancers likely results from a complex intersection of genetic, clinical and dietary factors, which are difficult to pull apart through observational studies. A more strategic approach to building the evidence base should be undertaken with large, well-designed studies, in order to harmonise and clarify the evidence, particularly for GC and BL.