Randomised controlled trials (RCTs) and preventive efficacy

We aimed to identify all RCTs so far carried out, extending the standard search by using unconventional methods such as citation chasing and scanning of conference proceedings ‐ methods that have proved effective in yielding additional high‐quality evidence for systematic reviews and meta‐analyses for other topics (Greenhalgh 2005; Vinceti 2017c). Using this approach, we identified a total of 10 RCTs that investigated monoselenium supplements for prevention of non‐melanoma skin cancer, prostate cancer, any cancer, and other site‐specific cancers. Overall, clear and consistent evidence indicates that selenium supplementation did not reduce subsequent cancer incidence, whether this endpoint was considered a primary or secondary outcome. Most of these trials raised concerns about possible harmful effects of selenium supplements, including increased incidence of non‐melanoma skin cancer in the Nutritional Prevention of Cancer Trial (NPCT), dermatological effects in the Selenium and Vitamin E Cancer Prevention Trial (SELECT), and type 2 diabetes in all RCTs, although with generally limited and statistically imprecise risk ratios (RRs).

Of the three liver cancer prevention trials, one reported a strongly reduced risk of liver cancer for male carriers of the hepatitis B surface antigen (HBs‐Ag) taking inorganic selenium supplements (sodium selenite) for three years, and the other two studies reported little effect of organic selenium supplements (selenium yeast) for the same cancer site (Li 2000; Yu 1991; Yu 1997). Owing to several methodological concerns related to randomisation and completeness of outcome data, we judged the risk of bias as unclear for all three of these RCTs. Therefore, we could not conclude that we found strong support for selenium supplements as agents for prevention of liver cancer. Unfortunately, the other trials did not include liver cancer among their secondary outcomes, with the exception of ECOG 5597 (Karp 2013). In this RCT, investigators reported new cases of liver, gallbladder, and bile duct cancer only among selenium‐treated participants; however, trial authors observed a total of only six cases, making risk estimates highly statistically unstable. In addition, the population included in this trial, which comprised patients with a history of resected non‐small‐cell lung cancer, was rather different from the general population.

The NPCT (NPCT 2002) reported strongly decreased risk for all cancers (‐22%), and for oesophageal (‐59%), colorectal (‐52%), lung (‐28%), and prostate (‐46%) cancers, showing lesser decreases compared with the ad interim report (Clark 1996, in: NPCT 2002), but still indicative of a strong cancer preventive effect. In addition, when participants were categorised into tertiles according to baseline serum selenium, evidence suggested an inverse relationship between selenium status and effects of supplementation for all cancers and for prostate cancer in the lower two tertiles, and no effect in the upper tertile. However, interpretation of these results is difficult because in 2003, the trial authors acknowledged the occurrence of a detection bias, namely, a considerably higher rate of prostate biopsy in the placebo group, whose cause was not specified. It is unclear whether this detection bias applied only to prostate cancer or applied more generally to other outcomes (as would be the case if the bias was due to unblinding, for example). This major detection bias forced us to downgrade the reliability of this study. Data show an increase in the incidence of its primary outcome ‐ non‐melanoma skin cancer ‐ in selenium‐supplemented participants, as well as in the incidence of five other cancer types, including melanoma, bladder cancer, breast cancer, head and neck cancer, and lymphoma and leukaemia. Trial authors stated: "These results, although non‐significant and based on small case numbers, may indicate potential increased risk with selenium supplementation"; these authors also relied on previous observational studies to provide some support for these positive associations (Duffield‐Lillico 2002, in: NPCT 2002).

The turning point of research on selenium and cancer was the SELECT trial (SELECT 2009), a large, well‐conducted prostate cancer prevention trial carried out in the male general population of North America not at high risk of prostate cancer (≤ 4 ng/mL in serum prostate‐specific antigen (PSA) and digital rectal examination not suspicious for cancer). This trial, widely considered a milestone in cancer prevention and research, found no difference in prostate cancer incidence for selenium–supplemented participants as compared with placebo participants after a median follow‐up of 5.5 years (hazard ratio (HR) 1.04, 95% confidence interval (CI) 0.90 to 1.18), and no effect of selenium on risk of overall cancer or on risk of other cancers (as well as cardiovascular disease). Median selenium at baseline (135 µg/L in serum in the selenium arm vs 137.6 µg/L in the placebo arm) was higher than in the NPCT (average plasma selenium 114 µg/L). The intervention used in this trial was different from that used in the NPCT (selenomethionine in SELECT, and selenised yeast in the former), although this is unlikely to have been responsible for observed differences (Waters 2013); in both cases, the intervention comprised organic selenium species (Block 2004).

In a small study of organ transplant recipients (Dreno 2007), an unexpected increase in non‐melanoma skin cancer incidence emerged; this was a matter of concern in the light of results of the NPCT. In the Polish trial Lubinski 2011, which included 1135 women with high genetic susceptibility to breast cancer due to BRCA1 mutations, evidence was more consistent with increased risk of both all cancers and primary breast cancer than with decreased risk, although with statistically unstable HRs (1.4, 95% CI 0.9 to 2.0; and 1.3, 95% CI 0.7 to 2.5, respectively). In this trial, the intervention consisted of administration of 250 µg/d of inorganic tetravalent selenium (selenite).

More recently, results of three well‐conducted phase 3 trials in participants at higher risk for prostate cancer than the general male population indicated that 200 µg/d of selenium (as selenomethionine in one study ‐ Marshall 2011 ‐ and as selenised yeast in the other two ‐ Algotar 2013; Karp 2013) did not decrease subsequent cancer incidence compared with placebo. The baseline selenium status of populations included in these RCTs was comparable with that in SELECT for Southwest Oncology Group (SWOG) S9917 (135 to 138 µg/L in the two arms) (Marshall 2011), slightly lower in the Negative Biopsy Trial (NBT) (126.1 µg/L) (Algotar 2013), and unfortunately unspecified for Eastern Cooperative Oncology Group (ECOG) 5597 (Karp 2013). Results of these high‐quality RCTs, all characterised by low risk of bias and two of which were discontinued before their planned end for futility, were consistent and showed no beneficial effect of selenium treatment on cancer risk.

Although not eligible for our meta‐analyses because their outcome was non‐malignant neoplasms rather than cancer, two recently published RCTs on colorectal adenoma risk in participants receiving selenium are worth noting. One of these trials was embedded in SELECT (Lance 2017), and the other, the SELCEL trial (an intervention study investigating the effect of selenium supplementation or celecoxib for prevention of colorectal adenoma recurrence), allocated 1374 men and women who had undergone removal of colorectal adenomas to either 200 µg/d selenium as selenised yeast, or placebo (Thompson 2016). Both RCTs did not find a beneficial effect of selenium for prevention of colorectal adenoma.

The RCTs carried out on selenium have generated clear evidence of adverse effects associated with selenium exposure, showing both the health effects related to overexposure and the amount at which these effects become evident, thus providing much more reliable evidence than that generated by environmental studies such as Vinceti 2017a for use in risk assessments of the safe upper limit of selenium exposure in humans. The trial that provided the most evidence about selenium‐associated adverse effects was SELECT. These effects include an excess risk of dermatitis and alopecia, non‐melanoma skin cancer, high‐grade prostate cancer, and type 2 diabetes. The excess risk of dermatological effects was anticipated as a potential side effect based on previous knowledge of health consequences of human overexposure to this element (Vinceti 2001), although such effects had been predicted to occur at higher amounts of selenium exposure than those experienced by SELECT supplemented participants, thus calling for reassessment of the upper limit of selenium exposure. The increased incidence of non‐melanoma skin cancer in NPCT and of advanced prostate cancer in SELECT was extremely disappointing, as they were the primary endpoints in these studies, and the expectation was that they would be reduced. The excess risk of diabetes in selenium‐supplemented NPCT participants, which was also an unanticipated finding, was mostly limited to participants in the two highest tertiles of baseline plasma selenium (> 105.2 µg/L), raising concern about the safety of selenium amounts that thus far had been considered entirely safe (i.e. on the order of 200 µg/d) (Stranges 2007). Therefore, subsequent RCTs added this endpoint to monitored adverse effects that contributed to interruption of the SELECT trial, together with the null effect on cancer mortality and adverse effects of vitamin E on prostate cancer risk (Lippman 2009, in: SELECT 2009). So far, all RCTs that included diabetes among trial endpoints, including trials investigating risk of colorectal adenoma, have shown an increased incidence of type 2 diabetes among selenium‐allocated participants, with RRs ranging from 1.08 to 1.71, although most estimates were statistically imprecise (Vinceti 2017b). In addition, in SELECT, a slight decrease in excess risk of diabetes in the intervention arm followed completion of selenium supplementation, further suggesting a causal relation between selenium administration and the disease (Lippman 2009 and Klein 2011, in: SELECT 2009). Currently, an excess risk of type 2 diabetes appears to be one of the adverse effects of selenium of greatest concern, and its plausibility is supported by the results of observational human studies (cohort, case‐control, and cross‐sectional), as well as by some biological plausibility (Galan‐Chilet 2017; Su 2016; Thompson 2016; Vinceti 2015; Vinceti 2017b; Zhou 2013). These side effects, in addition to the null results of RCTs, particularly of those of the highest quality, make implementation of new trials very unlikely owing to ethical concerns.

Observational studies and aetiological association

From our meta‐analyses of 16 prospective observational studies on overall cancer risk, we found lower cancer risk associated with highest selenium exposure compared with lowest exposure. Risk of cancer was 28% (95% CI 7% to 45%) lower in the highest category of selenium exposure than in the lowest, and risk of death from cancer was 24% (95% CI 3% to 41%) lower. Subgroup analyses by sex yielded increased evidence of this inverse association between selenium exposure and cancer risk in men compared with women.

The inverse association between overall cancer risk and baseline selenium levels was mainly attributable to lower risks of gastrointestinal, lung, and bladder cancer, and for men also prostate cancer. No association was seen between selenium and risk of breast cancer in women. However, when the amount of baseline exposure was taken into consideration, no clear and consistent trend between baseline selenium exposure and risk emerged for any of the major outcomes investigated in observational studies. Lack of lower risk of cancer in the highest versus the lowest selenium category among participants with the lowest baseline exposure levels compared with those with intermediate or high levels, for overall cancer, lung cancer, and prostate cancer, argues against a causal association between selenium exposure and cancer risk. This is supported by lack of a relation between differences in the highest and lowest categories of selenium exposure and the corresponding RR, further suggesting that larger differences in exposure are not associated with large and consistent decreases in RR. Finally, further uncertainty of the evidence generated by observational studies arises from the inconsistent and sometimes sharply conflicting results on the same cancer type that emerged from different studies.

We saw little evidence of any effect of modification on the relation of selenium and cancer by geographical area of residence. It should however be noted that most of the observational cohort studies that we examined were conducted in Europe and in the USA, and none were conducted in Africa or South America. This regional distribution seems to reflect the under representation of non‐Western and resource‐poor countries in epidemiological research (Pearce 2004). Differential regional representation in epidemiological studies is of special interest for this review, as selenium levels in humans around the world vary significantly. Even if selenium levels measured in included cohorts reflect a broad range of naturally occurring selenium exposure, investigators have reported some of the lowest and highest levels of selenium exposure in populations from South America (Jaffé 1992), Africa (Hurst 2013b), China (Li 2012), and India (Chawla 2016) ‐ regions not investigated by any of the reviewed observational studies, with the exception of three Chinese trials. Concerning sex‐related effects, our meta‐analysis of longitudinal studies revealed an inverse association between RR of cancer and selenium status in some cases in men but not in women for the same cancer type. Unfortunately, although more than half of reviewed studies included mixed‐sex populations, most did not report sex‐disaggregated results. In available sex‐specific results, men are over represented ‐ a fact that may potentially hamper assessment of the relation between selenium exposure and cancer risk in women. Theoretically, factors such as variations in body composition between men and women, including lean body mass versus fat composition, or differences in metabolism or in nutritional requirements (e.g. higher antioxidant requirements, particularly for the urological system) between the two sexes might be associated with differential effects of selenium for prevention of cancer.

Concerning the indicator used to assess selenium exposure and its relation with cancer risk, we observed generally null associations when evaluating selenium status through assessment of dietary intake, although some inverse associations at specific cancer sites emerged when we used biomarkers such as blood or toenail selenium levels. We extensively reviewed in the previous version of this review the characteristics and limitations of indicators of selenium exposure, with particular reference to dietary assessment methods and biomarkers, and inconsistencies across studies assessing the validity of different indicators (Ashton 2009; Fairweather‐Tait 2011; Jablonska 2015a; Vinceti 2014). In particular, a large body of literature concerns the limitations of dietary assessment methods, mainly linked to large variations of selenium content in single food types, and the limitations of biomarkers of exposure. Concerning the latter, a major source of exposure misclassification consists of the different behaviours of inorganic and organic selenium species, whose tendency to be retained in the body and to accumulate in specific body tissues greatly varies, although this does not necessarily correlate with their biological activity (Behne 1996; Behne 2010; Kim 2001; Michalke 2017; Panter 1996; Slavik 2008; Solovyev 2013; Steen 2008; Tiwary 2006; Vinceti 2013c). Investigators have frequently proposed that selenoprotein activity may be an indicator of selenium status and may be tested in association with cancer risk (Vinceti 2017b), but this relation has been questioned because different sources of oxidative stress, paradoxically including pro‐oxidant selenium species themselves, may upregulate selenoprotein activity (Jablonska 2015a). Furthermore, intake of heavy metals and other dietary factors such as vitamins, metalloids, and amino acids (e.g. methionine) may modify the health effects of selenium, or the relations between selenium exposure and biomarkers (Jablonska 2015a; Vinceti 2000), owing to metabolic interactions or changes in tissue‐specific deposition and retention of selenium (Behne 1996; Zeng 2005; Zwolak 2012).

Overall, available evidence indicates the potential for exposure misclassification in observational studies on selenium, as well as the pitfalls associated with an approach based on assessment of total selenium content in peripheral biomarkers, suggesting that in some instances, measurements of nutritional intake might provide better exposure estimates than are provided by biomarkers, particularly in the light of relative exposure to inorganic and organic species of the element. In general, observational cohort studies on selenium and cancer are expected to have been characterised by random exposure misclassification, thus shifting RRs towards the unity and reducing the ability to detect real associations. However, some exposure misclassification may have been non‐random, such as that induced by smoking, which although it is a source of selenium exposure also induces lower body selenium levels, possibly owing to an effect of cadmium in increasing selenium excretion (Vinceti 2000). In such cases, exposure misclassification based on biomarkers (serum/plasma selenium levels) may have substantially biased risk estimates and may have been associated with some degree of confounding due to the well‐known effect of smoking on cancer risk, which could not have been adequately captured and controlled for. Inadequate control for smoking has been suggested to be a major confounder inducing spurious associations between low selenium levels and enhanced cancer risk in observational studies (Beane Freeman 2015).

In addition to exposure misclassification, and probably more important than this, a major issue affecting observational studies is unmeasured confounding (Vinceti 2016a). This potential bias is a matter of greater concern than exposure misclassification because it may have systematically biased RRs in one direction, particularly for some cancer types. Moreover, detection (and control) of this bias is extremely difficult and nearly impossible, given the hundreds of nutritional and non‐nutritional lifestyle variables that may be associated with both variations in selenium intake and cancer risk. Among these factors are smoking (Beane Freeman 2015; Vinceti 2013b), socioeconomic status ‐ which appears to be positively associated with socioeconomic position in both men and women (Gundacker 2006; Niskar 2003) ‐and most likely hundreds of nutritional and toxicological factors that may vary in the diet, together with selenium intake. An approach that would reduce the risk of unmeasured confounding in observational studies might include investigation of dietary patterns rather than single nutrients, but these investigations seem not to have made adjustments for diet quality. Finally, it should be noted that most studies did not take into account the role of genetic factors (related to selenoproteins or otherwise) in the relation between selenium exposure and cancer risk, although some studies have suggested the importance of such relations (Jablonska 2016; Meplan 2014); the true relevance of genetic factors has not yet been well defined. Some studies examining selenoprotein‐related single‐nucleotide polymorphisms have suggested a role for genetic variants among genes coding for selenoproteins in modifying cancer risk, or in determining the relation between selenium exposure and subsequent cancer risk, although results have not been consistent (Geybels 2013; Meplan 2012; Penney 2010; Penney 2013; Slattery 2012; Takata 2011).

With awareness of the fundamental limitations of observational studies, even of those of longitudinal design, which may avoid selection bias or reverse causality, investigators designed and carried out in the 1990s and the 2000s several experimental studies as RCTs investigating the effect of selenium supplements on cancer risk. The evidence base from these intervention studies has become so large and complete as to allow a comprehensive evaluation of cancer risk associated with selenium supplementation for some specific cancer types. It is interesting to note that major interest in the cancer preventive activity of selenium originated not just from observational studies (mainly of ecological and cohort design) (Vinceti 2013b), but from a randomised trial ‐ the ad interim analysis of the NPCT, which was published in 1996 and attracted great interest from both the scientific community and the general public because of the apparently large beneficial effect that it reported (Clark 1996, in: NPCT 2002). Null results of the most recent low‐bias RCTs ‐ Algotar 2013; Marshall 2011; SELECT 2009 ‐ also do not suggest a major or strong role of genetic factors in modifying selenium and cancer relations, given their generally null or troubling results. An exception can be seen in recent data from SELECT, which suggest that a genetic variant of the NKX3.1 androgen‐regulated prostate tumour suppressor protein may modify, or increase, the risk of prostate cancer associated with selenium supplementation (Martinez 2014).

From a methodological perspective, we acknowledge that comparison of risks between highest and lowest exposure categories in observational studies, as performed in the present meta‐analysis, is most suitable for identifying an effect when a consistent decrease or increase is seen across absolute exposure levels. Other associations (e.g. threshold effects, U‐shaped relations) may have been missed by this method of meta‐analysis, or their true effect might have been diminished.

RCTs and preventive efficacy

This review investigated a diverse range of cancers, substantially extending the analysis compared with that performed in previous reviews. However, cancer is not a uniform condition, and malignant neoplasms show great differences in tumour biology. Only non‐melanoma skin cancer, liver cancer, and prostate cancer have been investigated as primary outcomes in the included prevention trials, and, regarding these main outcomes, specific characteristics of study populations may limit the generalisability of results. Participants in included RCTs on skin and liver cancer belonged to populations at high risk for the outcome under investigation, and participants in high‐quality prostate cancer trials were at average risk (Karp 2013; SELECT 2009), or at high risk (Algotar 2013; Marshall 2011), for this disease. Most participants in the NPCT were older and white, predominantly male inhabitants of the United States, and the most recent trials were limited to the USA male population.

Average baseline selenium exposure in the NPCT was less than that characterising subsequent trials carried out in the United States, although it was more similar to that seen in some European populations. Although the NPCT suggested that selenium supplementation was beneficial only at the lowest range of baseline selenium exposure, the most recent studies, carried out in populations generally characterised by higher average selenium exposure, did not confirm such an interaction. The NPCT also found an indication of strong effect modification for sex, as demonstrated, for example, by the HR for all cancers associated with selenium supplementation ‐ 0.67 (95% CI 0.50 to 0.89) in men and 1.20 (95% CI 0.66 to 2.20) in women (NPCT 2002).

Participants in the SELECT study on prostate cancer prevention were apparently healthy men over 50 years of age from the general population of North America (SELECT 2009). A large sample size and inclusion of non‐white participants from different socioeconomic backgrounds support the generalisability of study findings to other adequately nourished populations.

Selenium supplements generally contain organic or inorganic species of selenium, or a mixture of both (e.g. in the form of selenised yeast). Different species of selenium may exhibit different effects on human health and more specifically on proteomic endpoints, as also suggested by human controlled randomised trials though with inconsistent results (Ravn‐Haren 2008; Richie 2014). High‐quality RCTs using selenised yeast supplements, almost entirely comprising organic selenium forms (Block 2004; Waters 2013), found no effect of supplementation on the main study outcome and an indication of a harmful effect (i.e. an excess diabetes risk) (Vinceti 2017b). The SELECT trial used supplements of L‐selenomethionine, which is the major component of selenised yeast, and also found no preventive efficacy. The only two RCTs investigating sodium selenite supplements found a protective effect against liver cancer, and null or adverse effects on breast cancer risk, but we considered these trials to have unclear risk of bias. It is unclear how applicable these results are in other settings and in populations with a different nutritional status. Interpretation of the results of clinical trials using selenium supplements should consider the different chemical forms of selenium, as well as their potentially different health effects when used as supplements (Vinceti 2013c; Weekley 2013). Most studies used organic selenium as selenised yeast (Algotar 2013; NPCT 2002), or as selenomethionine (Marshall 2011; SELECT 2009). However, the chemical form used is unlikely to explain the differences in results between NPCT and the other trials (Waters 2013). With reference to this issue, of interest are the results of a 'natural experiment' that occurred in Northern Italy, wherein a small population unintentionally consumed for several years drinking water with an unusually high content of selenium in its inorganic hexavalent form ‐ selenate (Vinceti 2000). Follow‐up of that population revealed increased risk of neurodegenerative disease ‐ a not entirely unexpected finding owing to the potential neurotoxicity of inorganic selenium (Vinceti 2014a), along with a slightly increased risk of cancer, mainly due to excess risk of oropharyngeal cancer, melanoma, kidney cancer, and lymphoid malignancies (Vinceti 2016b).

An important issue is the possibility that participants with low baseline selenium status may experience an inverse association between selenium exposure and cancer risk, as suggested by some trial authors (Lu 2016; Rayman 2009). This has been suggested to explain the different results of SELECT and the NPCT, and could also hypothetically explain, at least in part, the different relations found in experimental as compared with observational studies. NPCT found a strong beneficial effect of selenium supplementation among participants at the lowest tertiles of baseline selenium levels; however, the risk of cancer changed abruptly from an apparently protective effect in the two lower tertiles (HR 0.51 and 0.70) to an excess risk in the highest tertile of plasma selenium (HR 1.20, 95% CI 0.77 to 1.86). This occurred despite a difference of only 16.4 µg/L between lowest and highest tertiles, corresponding to a change in dietary selenium intake as low as around 10 µg. This would imply that such a small a change in selenium dietary intake would change a strongly protective effect of the element on cancer risk into a possibly detrimental effect ‐ an implausible scenario given the wide range of selenium intake (from about 20 to several hundred micrograms) characterising Western populations. Moreover, the intermediate tertile of baseline plasma selenium in the NPCT (105.6 to 122.0 µg/L) appeared to be associated not only with reduced overall cancer risk but also with an excess risk of squamous cell skin carcinoma (HR 1.49, 95% CI 1.05 to 2.12) and overall non‐melanoma skin cancer (NPCT 2002), as well as diabetes (RR 1.36, 95% CI 0.60 to 3.09), whose risk also considerably increased at the highest tertile of baseline selenium (Stranges 2007). Overall, this occurrence of both adverse and beneficial effects is unlikely if the selenium supplementation was serving to remedy a selenium deficiency. In addition, the strongest effect of selenium on overall cancer risk at lower levels of baseline selenium status was due to a considerable decrease in prostate cancer, but this finding was subject to detection bias because of a decreased biopsy rate in selenium‐supplemented participants, particularly in those with lowest baseline selenium status, as recognised by investigators of the NPCT (NPCT 2002).

In addition, after NPCT, three of the four high‐quality RCTs on selenium supplementation for cancer prevention investigated the possible modifying effect of baseline selenium exposure and found no evidence of a beneficial effect of the intervention even in the lowest baseline exposure category. For instance, in NBT (Algotar 2013), the average baseline plasma selenium level at the lowest tertile of the study population was 101.1 µg/L ‐ much lower than the corresponding level at the middle tertile of NPCT (114.6 µg/L), in which the HR of prostate cancer had been as low as 0.33 (95% CI 0.13 to 0.82). However, in this ‘low’ NBT subgroup, investigators found no evidence of a beneficial effect of selenium supplementation on prostate cancer risk. In the SWOG S9917 trial (Marshall 2011), data show no change in the null effect of selenium in the two lowest categories (quartiles) of selenium intake, whose boundaries were < 106 and 106 to132 µg/L ‐ similar to cut points of the two bottom NPCT tertiles and of the bottom category of NBT. In these two subgroups of the SWOG population with the lowest baseline selenium status, the RR of prostate cancer was 0.82 (95% CI 0.40 to 1.69) and 1.38 (95% CI 0.68 to 2.78), and in the third upper quartile, the RR was 0.98 (95% CI 0.58 to 1.68), suggesting no consistent trend of an inverse relation between antecedent selenium exposure and effects of supplementation (as was also shown by analysis for trend in this study). Investigators in SELECT reported no reduction in cancer risk among selenium‐supplemented participants, although they did not provide specific RRs according to baseline selenium status. Calculation of blood selenium content distribution in SELECT, as well as in the three other RCTs (NPCT, NBT, SWOG), showed substantial overlap of plasma and serum selenium levels between this large trial population and the other study populations (Figure 7). In addition, a more recent case‐cohort study carried out within SELECT assessed the effect of selenium supplementation on prostate cancer risk, taking into consideration baseline selenium exposure, as assessed through toenail selenium levels. The study, which involved 1739 prostate cancer cases and 3117 controls, was unable to find a beneficial effect of selenium supplementation in the lowest categories (quintiles) of baseline toenail selenium (Kristal 2014). Actually, a dose‐response effect in that SELECT population emerged, but it favoured an increased risk of (high‐grade) prostate cancer induced by selenium supplementation among participants belonging to the two upper quintiles of baseline selenium exposure (Kristal 2014). Therefore, it seems reasonable to agree with this SELECT statement: “The analysis of our data using lower cut points for baseline toenail Se categories, in an attempt to replicate findings from the NPCT, also showed no evidence of benefit from supplementation among men with low baseline Se status (data given in Results). Given these findings, we believe it reasonable to conclude that Se supplementation of men at the low range of Se intake common in USA men will not reduce PCa risk” (Kristal 2014).

Figure 7

---

---

Baseline circulating selenium levels in the NPC trial (Duffield‐Lillico 2003b in: NPCT 2002), the NBT (Algotar 2013), SWOG trial (as plasma selenium) (Marshall 2011), and SELECT (as serum selenium) (Lippman 2009, in: SELECT 2009). When median and interquartile values were reported, we estimated mean and standard deviation according to Cochrane guidelines provided in Higgins 2011a.

Overall, results of recent high‐quality RCTs do not support the hypothesis that differing baseline selenium status may explain conflicting results between NPCT and SELECT (Lu 2016; Rayman 2009). Results of the most recent RCTs seem therefore to be applicable to populations with various degrees of background selenium exposure, with the exception of populations characterised by extremely low (< 20 µg) or high selenium intake.

Observational studies and aetiological association

We reviewed data from prospective observational studies in which investigators measured selenium exposure in populations without evidence of cancer, who were then followed up for a specified period of time. We limited our systematic review to cohort studies to avoid or decrease two major sources of bias in observational investigations, particularly in case‐control and cross‐sectional studies (i.e. selection bias and risk of reverse causality). Data continue to show important differences among included studies in terms of selenium exposure assessment, types of outcomes, and study populations, which may affect their interpretation. The small number of studies that examined most of the meta‐analysed types of cancers prevented a thorough investigation of sources of heterogeneity between study results. In particular, we had limited opportunity to explore the influence of specific sources of bias or the methodological quality of epidemiological studies on heterogeneity.

Participants examined in this review update include more than 2,300,000 individuals, predominantly from Europe and North America, and, to a much lesser extent, from Asia and Australia. We were able to identify no prospective observational studies on selenium and cancer risk from Africa or South America. This regional distribution reflects the under representation of non‐Western and resource‐poor countries in epidemiological research (Pearce 2004). Differential regional representation in epidemiological studies is of special interest for this review, as selenium levels in humans around the world vary significantly. Selenium levels measured in the included cohorts reflect a broad range of naturally occurring selenium exposure, as documented by several epidemiological studies worldwide. However, some of the lowest and highest selenium levels in humans have been reported in populations in South America (Jaffé 1992) ‐ a region not investigated by any of the reviewed observational studies.

More than half of the included studies enrolled mixed‐sex populations, but most did not report sex‐disaggregated results. In available sex‐specific results, men are over represented ‐ a fact that could hamper potential assessment of the relation between selenium exposure and cancer risk in women. Despite this sex imbalance, we systematically saw stronger (inverse) associations with cancer risk among men than among women, for whom such associations with antecedent selenium status was nearly absent. This was true for stomach, colorectal, and lung cancer, and, when added to the inverse association for prostate cancer, led to an impact on overall cancer risk that was clearly lacking in women that could be due to potential confounders (such as smoking, occupational exposures, or other dietary factors) or to a real change in the association between selenium exposure and cancer risk in the two sexes.

The range of selenium exposure experienced by members of cohorts investigated in the observational studies was generally lower than that experienced by participants in RCTs, who added supplemental selenium, generally 200 µg/d and in its organic forms, to their usual background intake, which ranged from about 70 to 90 µg/d as organic selenium, although some RCTs provided no estimate (Jablonska 2015a). It is theoretically possible that a preventive effect of selenium against cancer exists only at low (< 30 to 50 µg/d) intake of the element, and that it disappears at higher intakes, when ‘saturation’ or ‘maximisation’ of selenoprotein expression driven by selenium occurs. Investigations have frequently chosen this proteomic endpoint as a reference point for deriving dietary reference values for selenium (Jablonska 2015a; Vinceti 2017a). Selenium exposure in the range of around 50 to 200 µg of daily selenium intake has not been tested by intervention studies, which have used larger amounts of supplemental selenium, and is unlikely to be tested in RCTs in the future, given the termination of past trials for futility or safety concerns. This possibility must be considered, but within the context of the two fundamental limitations of observational studies ‐ exposure misclassification and unmeasured confounding, which limit the reliability of the evidence they generate and its applicability in terms of cancer prevention.

A few lines of evidence suggest that even at low levels of selenium exposure, it is unlikely that such an inverse association with cancer risk exists. First, inconsistencies in the results found in our meta‐analysis for most cancer sites and lack of a dose‐response relation between cancer risk and selenium at varying levels of background selenium exposure, or of a difference between highest and lowest exposure categories, argue against a real relation between selenium and cancer risk. Limited differences between highest and lowest categories of selenium intake, often amounting to a difference of only 20 to 30 µg per day, compared with large variations in selenium intake worldwide (from 10 to 15 µg in low‐selenium areas up to several hundred µg in seleniferous areas), also argue against a true relation. Finally, as previously described, some recent high‐quality RCTs investigated the effect of baseline selenium status on cancer risk associated with selenium supplementation and found no beneficial effect of selenium supplementation, even among participants with the lowest amounts of baseline exposure. Overall, these findings do not support an association between higher selenium status and lower cancer risk independently from factors such as sex, baseline selenium exposure, and cancer type. One additional observational cohort study, which could not be meta‐analysed in this review because it was released in PubMed in July 2017, appears to confirm these conclusions (Sandsveden 2017).

RCTs and preventive efficacy

SELECT (SELECT 2009), SWOG S9917 (Marshall 2011), NBT (Algotar 2013), and ECOG 5597 (Karp 2013) were the only trials considered to have low risk of bias with adequate sequence generation, allocation concealment, blinding, and reporting of findings, and the consistency of their findings for prostate cancer, as well as for other cancer types for the two trials investigating them (SELECT and ECOG 5597), added to the statistical power of the major trial (SELECT), making their overall results highly reliable and suitable for yielding useful evidence to assess the relation between selenium supplementation and cancer prevention. These trials are also of major importance because (with one exception) they have provided information about baseline selenium exposure and its possible modifying role and about the effect of selenium supplementation on subsequent cancer incidence. Another important feature of these trials has been their ability to address the issue of selenium overexposure and related adverse effects owing to a systematic surveillance system for adverse effects, as well as their ability to extend the monitoring programme to additional effects, if suggested by new analyses targeting previously unplanned secondary endpoints, as was the case for diabetes (Stranges 2007). This is particularly relevant because all of these trials were planned under the hypothesis, later found to be erroneous but at that time endorsed by regulatory agencies, that the supplemental selenium dose administered to intervention arms (200 µg/d in almost all RCTs) was entirely safe and was well below the upper safe limit of the element, even with consideration of background selenium exposure.

These trials may continue to yield important results. Secondary analysis of additional endpoints, or based on genetic and non‐genetic biomarkers of exposure to selenium and other factors, is still possible. For example, major contributions were yielded by SELECT in 2017, concerning outcomes such as prevention of colorectal adenoma and of Alzheimer’s disease by selenium supplementation, in both cases with null results (Kryscio 2017; Lance 2017).

We assessed the certainty of evidence from high‐quality RCTs using the GRADE approach (http://gdt.guidelinedevelopment.org/app/handbook/handbook.html#h.svwngs6pm0f2) and reported the results of this assessment in the 'Summary of findings' table (summary of findings Table for the main comparison). From preliminary assignment to a high level of certainty due to the experimental study design, we did not identify reasons to downgrade trial quality according to standard GRADE guidelines for risk of all cancers, for risk of cancer mortality, or for risk of colorectal, lung, bladder, or prostate cancer. In contrast, meta‐analysis for breast cancer risk yielded a statistically imprecise result mainly reflecting the small number of cases, and meta‐analysis for non‐melanoma skin cancer showed high statistical heterogeneity across studies. When addressing factors possibly increasing the certainty of evidence assessment, we considered as non‐applicable the GRADE item “All plausible confounding would reduce the demonstrated effect or increase the effect if no effect was observed”, neither could we evaluate possible dose‐response gradients because unfortunately they were not tested in these RCTs. We therefore rated the certainty of evidence as ‘high’ if it indicates no effect of selenium supplementation on all cancers overall, on cancer mortality, nor on colorectal, lung, bladder, and prostate cancer, and we considered certainty of the evidence as ‘moderate’ if it indicates no effect on non‐melanoma skin cancer and breast cancer, with downgrades due to heterogeneity and imprecision, respectively. However, stating that the evidence supporting no effect of selenium on cancer prevention at these sites is of moderate rather than high certainty does not mean that the only alternative hypothesis is necessarily that selenium decreases risk of cancer at these sites. Actually, the overall results of high‐quality RCTs, when available, suggest a slight to moderate although statistically imprecise increase in the risk of some of these specific cancers following selenium supplementation.

Concerning the RCTs that we downgraded in our appraisal of risk of bias, we considered the quality of reporting to be an issue in the three trials on liver cancer prevention, thus leading to their classification as having unknown risk of bias. Several papers reported the individual trials, in some cases discrepantly, and essential questions regarding sequence generation, allocation concealment, handling of dropouts and withdrawals, and detection of outcomes remain unanswered. This might be due to inadequate reporting but might also hint at flaws in trial design and implementation. We were uncertain about whether the only trial that reported positive results for selenium supplements in liver cancer prevention randomly assigned participants individually. Cluster randomisation of participants who lived in the same area/village, which may have been the procedure used in this investigation, might have introduced additional bias to the study results (e.g. as the result of different environmental factors contributing to liver cancer development or detection) and might have led to an overestimation of the protective efficacy of selenium. Duplication of results of trials based on a rigorous study design would be necessary to assess the effects of sodium selenite on liver cancer incidence. With regard to the NPCT (NPCT 2002) and the trial of Dreno 2007, indications of serious detection bias for the USA study and of unclear methodological details (such as blinding) for the French investigation led us to consider these experimental studies to be at unclear risk of bias, as discussed in greater detail elsewhere in this review. As far as the trial on breast cancer is concerned (Lubinski 2011), our downgrade of evidence certainty was based on incomplete information provided in the only report that we could retrieve (an abstract), although we acknowledge the relevance of that trial ‐ the only trial specifically targeting breast cancer and a genetically specific population ‐ and the fact that complete reporting of trial procedures may lead to reassessment of trial quality and its upgrade.

Observational studies and aetiological association

The 70 observational studies were heterogeneous, not only in methodology, but also in the quality and level of detail of reporting and in their potential biases. We assessed our confidence in the evidence from these studies using the GRADE approach and reported our findings in summary of findings Table 2; we reported judgements only for those outcomes evaluated in the 'Summary of findings' table for RCTs with low risk of bias.

RCTs and preventive efficacy and observational studies and aetiological association

The literature search included major international databases in the English and German languages, and we applied a broad search strategy supplemented by handsearching for references. We assume that we identified all randomised controlled studies and prospective observational studies relevant to our review questions. As we did not search databases in other languages (e.g. Chinese, Russian), we cannot rule out that we might have missed smaller studies that were not published in international journals. However, we consider it unlikely that we could have missed major sources of evidence through our approach. We also might have missed observational studies whose results on selenium exposure and cancer were reported in the body of a paper but were not mentioned in the paper's title or abstract, even if the paper is indexed in the searched databases. However, our systematic use of backward and forward citation chasing and our search for relevant abstracts in conference proceedings or related material should have substantially decreased the risk of missing literature that could have been relevant for our assessment.

When needed because of lack of complete or appropriate participant data (e.g. when cohorts including cancer and non‐cancer participants were mixed in data analysis), we contacted study investigators to ask for data missing from their studies. We also did this when we did not have enough data from published reports to adequately appraise study risk of bias. Sometimes we were unable to obtain answers to questions that we had regarding methods or outcomes, but frequently investigators kindly gave us the information we needed. We were sometimes unable to obtain answers, particularly for earlier epidemiological studies from which primary investigators may have relocated or died, or we found that data were not available in a current electronic format. Similarly, we could not make contact with primary investigators of Chinese RCTs.

We based our risk of bias assessment on information included in the original publications, unless the trial authors that we contacted gave us additional details. This means that in some instances, we may have overestimated the true risk of bias of studies that did not adequately describe their design in the original publications, such as Lubinski 2011.

Another concern, especially with epidemiological studies, is publication bias. Cohort and nested case‐control studies often are not exclusively designed to test for a specific exposure‐outcome association but enable investigators to investigate a range of questions. It is conceivable that unfavourable results were less likely to be published, although we could not find evidence supporting such a hypothesis. Our analysis of this issue through use of a funnel plot gave some support to publication bias for prostate cancer.

We systematically meta‐analysed RCTs even when only two studies were available (Karp 2013; NPCT 2002). Finally, we carried out two meta‐analyses of intervention studies ‐ one on all studies, and another on RCTs assessed through a standard appraisal tool as being at low risk of bias ‐ and we emphasised results of the latter, as derived from high‐quality experimental studies. For observational studies, we decided a priori to conduct meta‐analyses only when five or more studies were available for a study outcome, thus excluding from meta‐analysis the few endpoints for which up to four studies were available (Table 1). Our primary intention was to facilitate the investigation of heterogeneity between studies included in meta‐analyses, to avoid producing more precise, but still unexplainably biased, results. However, our emphasis was clearly given to experimental studies because this trial design is widely recognised as the only one that may provide convincing evidence of an association between a factor and disease risk, or more generally biological endpoints, and this may be particularly true in nutritional epidemiology (Vinceti 2016a).

Finally, the authors of this review, as already noted in the previous version of the review, came from different disciplines and have different areas of focus (e.g. epidemiology, biostatistics, clinical medicine, nutrition). We continue to consider such variety of expertise to be a strength of this review, and we made use of it by applying multiple checking procedures during the entire review process whenever possible.