Sex differences in the association between diabetes and cancer: a systematic review and meta-analysis of 121 cohorts including 20 million individuals and one million events

1Cancer is the second leading causes of death in the world. In 2015, there were 17.5 million incident cancer cases and 8.7 million cancer deaths globally, and it is estimated that one in four women and one in three men develop cancer during their lifetime. The incidence of cancer is expected to increase in the next decades, emphasising the importance of efficient prevention and treatment of cancer worldwide.

2The prevalence of diabetes has also grown rapidly. In 2015, one in 11 adults (415 million) were reported to have diabetes, five million deaths were attributed to diabetes, and 12% of global health expenditure was spent on diabetes and its complications. Diabetes has been associated with the risk of all-site and some site-specific cancers in several systematic reviews and meta-analyses. However, only a minority of these associations are based on robust supporting evidence without question of significant bias. To date, there has been no systematic overview of the evidence available on sex differences in the association between diabetes and cancer. We have previously published compelling evidence that women with diabetes are at an increased risk of stroke, coronary heart disease and dementia compared with their male peers. We now question whether this is also true for cancer. In this study, we conducted the most comprehensive systematic review and meta-analysis, to date, to estimate the relative effect of diabetes on the risk of cancer in women compared with men.

3A systematic search was performed in PubMed (https://www.ncbi.nlm.nih.gov/pubmed/) on 23 December 2016 using a combined text word and medical subject heading search strategy (electronic supplementary material [ESM] Table 1). The reference lists of identified reports were also checked for other potentially relevant studies.

4Observational cohort studies in general populations were included if they had provided relative risks (RRs), or equivalents, for the association between diabetes and cancer in both women and men. Studies were excluded if they had not adjusted at least for age or did not provide information about the variability around the point estimate, or if they only had data for one sex. In case of duplicate reports from the same study, the study providing the longest follow-up or the highest number of cases was included. We also used individual participant data from the Asia Pacific Cohort Studies Collaboration (APCSC), treated as two separate combinations of data from cohorts in Asia and cohorts from Australia or New Zealand, as in our previous work. One author (TO) did the search and extracted the data. Uncertainties regarding the inclusion or exclusion of articles and data extraction were discussed by all authors and resolved by mutual consent. The meta-analysis was done in accordance with the PRISMA criteria.

5The primary endpoint was all-site cancer events (incident or, if this was all that was presented, mortal only). The secondary endpoints were all-site cancer incidence (i.e. omitting studies that only reported mortality), all-site cancer mortality and, for those cancers that could present in both sexes, site-specific cancer events, site-specific incidence and site-specific mortality. In sensitivity analysis we also compared all-site cancer incidence and mortality when restricting to the studies that reported both.

6The primary metrics were the pooled adjusted RRs and the women-to-men ratios of RRs (RRRs) for individuals with diabetes vs those without diabetes. For each study, we extracted the sex-specific RRs and 95% CIs for individuals with diabetes vs those without diabetes, from which we estimated the RRRs and 95% CIs. To include the largest set of individuals and cancer endpoints, studies that reported either age-adjusted or multiple-adjusted (maximum-available-adjusted, i.e. the maximum set of adjustments available for each study) results were included in our primary analyses. In pooling multiple-adjusted results, the set of adjustments made were allowed to vary by study, but had to include at least one other risk factor for cancer, in addition to age. We obtained pooled estimates of sex-specific RRs across studies using random-effects meta-analyses applied on the loge scale. Individual studies were weighted according to the inverse variance of loge RRs. The same method was used to pool the RRRs.

7The I2 statistic was used to estimate the percentage of variability across studies due to between-study heterogeneity and the Q test was used to assess whether there was a significant lack of homogeneity. The possibility of publication bias was explored using funnel plots and Egger’s and Begg’s tests. Random-effects meta-regression analyses were used to test for differences between pre-assigned subgroups: study region (Asia or Non-Asia), year of baseline study (pre-1985 or 1986 onwards, and also examined as a continuous variable), ascertainment of diabetes (self-reported only or others), type of diabetes (type 1 or type 2, where studies which did not differentiate type were classified as type 2), level of adjustment (age-adjusted or multiple-adjusted), and study quality (the Newcastle–Ottawa Scale [ESM Table 2], ≥7 or <7 points, and also examined as a continuous variable). Post hoc, we also considered absolute risk difference, examined as a categorical and continuous variable) (ESM Table 3). A p value of below 0.05 was considered to be statistically significant in analyses for the primary analyses, i.e. all-site cancer. As many statistical tests were envisaged, a p value of below 0.01 was taken to denote significance for site-specific cancers. All analyses were performed using Stata software (release 13; StataCorp, College Station, TX, USA).

8Of the 6371 articles identified through the systematic search, 371 articles qualified for full-text evaluation, and 107 articles provided summary data on the association between diabetes and the risk of cancer for both sexes. In addition, 36 cohorts with individual participant data from the APCSC were included (Fig. 1).

9Characteristics of the studies that reported the association between diabetes and all-site cancer incidence or mortality are shown in Table 1 and ESM Table 4. Data on all-site cancer were available from 47 studies, involving 121 cohorts, 19,239,302 individuals (not counting one study that did not state the total number of participants), and 1,082,592 events (not counting one study that did not state the total number of cancer events).

10The maximum-available-adjusted pooled sex-specific RR estimates for combined fatal and non-fatal cancer associated with diabetes were 1.27 (95% CI 1.21, 1.32, p < 0.001) for women and 1.19 (1.13, 1.25, p < 0.001) for men (Fig. 2). The pooled women-to-men RRR was 1.06 (1.03, 1.09, p < 0.001, Fig. 3). The I2 statistic for heterogeneity between studies was 66.7%, with no evidence of publication bias (Egger’s test p = 0.13, Begg’s test p = 0.16, ESM Fig. 1). The corresponding RRR was 1.06 (1.02, 1.11, p = 0.005) for type 1 diabetes and 1.06 (1.03, 1.09, p < 0.001) for type 2 diabetes, without evidence of significant heterogeneity by type of diabetes (p for interaction = 0.88, Fig. 4). Exclusion of 22 studies that provided only age-adjusted results had no appreciable effect on the pooled RR estimates (multiple-adjusted pooled RR in women 1.25 [1.17, 1.34], p < 0.001, RR in men 1.20 [1.11, 1.29], p < 0.001, RRR 1.06 [1.03, 1.10], p < 0.001, I2 = 48.9%) (ESM Figs 2 and 3).

11The pooled RRR did not vary substantially by study region (p = 0.45), year of baseline study (p = 0.54 for categorical analysis, p = 0.18 for continuous analysis), ascertainment of diabetes (p = 0.72), level of adjustment (p = 0.70), quality of study (p = 0.09 for categorical analysis) or absolute risk difference between men and women (p = 0.82 for categorical analysis, p = 0.99 for continuous analysis), with the exception of continuous analysis for quality of study, p = 0.01) (Fig. 4 and ESM Fig. 4).

12Secondary analyses of incidence (fatal or not) and mortality alone for all-site cancer are described in the ESM. The pooled women-to-men RRR for incidence was 1.10 (1.07, 1.13, p < 0.001) (ESM Fig. 5) and for mortality was 1.03 (0.99, 1.06, p = 0.16) (ESM Fig. 6).

13In sensitivity analysis using only those studies which provided the RRs for both incidence and mortality, the pooled maximum-available-adjusted RRR was 1.12 (1.06, 1.17, p < 0.001) for all-site cancer incidence, and 1.10 (1.00, 1.21, p = 0.04) for all-site cancer mortality (ESM Fig. 7).

14Data on site-specific cancer were available for 50 sites (50 sites for incidence and 29 sites for mortality) (https://www.georgeinstitute.org/sites/default/files/esm-table.pdf). Diabetes was associated with an increased risk of cancer in 43 sites in women and 42 sites in men, with a statistically significant increase (p < 0.01) in risk for those with diabetes in 20 sites in women and 18 sites in men (ESM Fig. 8). The pooled maximum-available-adjusted RRR was statistically significantly higher in women than men for kidney (1.11 [99% CI 1.04, 1.18], p < 0.001), oral (1.13 [1.00, 1.28], p = 0.009), stomach cancer (1.14 [1.07, 1.22], p < 0.001) and leukaemia (1.15 [1.02, 1.28], p = 0.002), whereas it was statistically significantly lower for liver cancer (0.88 [0.79, 0.99], p = 0.005) (Fig. 5). Separate results for incidence and mortality by site of cancer are described in the ESM (ESM Figs 5, 6, 9–24).

15This systematic review, with meta-analysis, of 121 cohorts including more than 19 million individuals and over one million all-site cancer events, demonstrated that diabetes was associated with a 6% higher excess risk of all-site cancer in women than men. Diabetes was associated with several site-specific cancers and conferred a significantly greater excess risk in women than men for oral, stomach and kidney cancer and for leukaemia, but a lower excess risk for liver cancer. The findings were broadly consistent for incident and fatal cancers and across a wide range of prespecified subgroups.

16Our findings are in agreement with a previous meta-analysis, which found that the risk of all-site cancer incidence and mortality was significantly increased in both sexes. However, this previous meta-analysis was about a tenth of the size of the current study, and included single-sex studies, and therefore was not able to reliably quantify sex differences as they could have been explained by differences in methods, confounders adjusted for, and the background risks between studies of women and men alone.

17As we found some evidence to suggest that the women-to-men RRRs tended to be smaller in studies of lower quality (Fig. 4 and ESM Fig. 4), our results may underestimate any true sex difference. A significant degree of heterogeneity was also observed between studies conducted in and outside Asia with regards to all-site cancer mortality (ESM Fig. 19). However, we did not find heterogeneity between regions for our primary outcome, nor for the other secondary outcomes (all-site cancer incidence), and thus we speculate that this may be a chance finding consequent to the high number of statistical tests conducted.

18Although we found a slightly higher women-to-men RRR for cancer incidence than cancer mortality, the finding may be explained by chance differences between the included studies, as almost identical pooled RRR estimates were obtained in the sensitivity analysis restricted to five studies which provided the sex-specific RRs for both incidence and mortality from the same study.

19With regard to cancer at specific sites, previous meta-analyses have yielded inconsistent results of increased (stomach, lung, kidney), similar (oesophagus, colorectum, pancreas, bladder, thyroid) or decreased (liver) excess risk of cancer associated with diabetes in women compared with men. However, unlike our methods, these analyses included single-sex studies as well as studies among both women and men.

20There are several possible explanations for the excess risk of cancer conferred by diabetes in women than men. One possible mechanism is poor glycaemic control in women with diabetes compared with men with diabetes. Hyperglycaemia may have carcinogenic effects by causing DNA damage, which could result from increased oxidative stress due to hyperglycaemia or from hyperglycaemia itself. Historically, women were likely to be undertreated or receive less intensive care compared with men. Further, a recent study showed that adherence to glucose-lowering medication was lower in women than men. As such, the carcinogenic effects of hyperglycaemia may be enhanced in women and subsequently lead to an increased cancer risk compared with men. Alternatively, cumulative exposure to insulin resistance and subsequent hyperinsulinaemia may be longer in women compared with men. The average duration of impaired glucose tolerance or impaired fasting glucose has been found to be more than 2 years longer in women than men, suggesting that women may have more exposure to, often untreated, hyperinsulinaemia in the prediabetic state. Hyperinsulinaemia promotes cancer cell proliferation by stimulating the insulin receptor directly and insulin-like growth factor-1 indirectly. Another factor that may, to some extent, explain the smaller RR for incidence of all-site cancer in men compared with women is the apparent protective effect of diabetes on prostate cancer in men with diabetes. Sex-specific cancers or site-specific cancers in which diabetes conferred greater or lower excess risk in women than men may also account for the association, although the degree of contribution cannot be determined from our analyses. In addition to sex difference for all-site cancer, we found also that diabetes conferred a significantly greater RR in women than men for oral, stomach and kidney cancer and for leukaemia, but a lower RR for liver cancer. The underlying mechanisms for sex differences in each specific association are not clear. However, unmeasured confounding factors specific to each site, such as Helicobacter pylori infection for stomach cancer and hepatitis virus infection for liver cancer, might be involved. However, the literature around mechanisms underpinning the sex differences in site-specific cancers is scant and further studies are required to confirm and clarify these sex differences in site-specific associations. Finally, the studies in our analyses were not adjusted for female-specific factors including pregnancy, menopausal status and use of hormone replacement therapy that have also been associated with diabetes and cancer.

21We quantified sex differences based on RRs rather than risk differences. This might introduce a statistical artefact, in which the generally higher absolute risk for cancer in men, and the same risk difference subsequent to diabetes in each sex, would translate to a greater relative risk in women than men. However, this would require that risks of cancers associated with diabetes are additive rather than multiplicative, which is not generally considered to be the case in epidemiology. Indeed, RRs are much more commonly reported than risk differences in both epidemiological studies and clinical trials. Also, unlike risk differences, RRs are typically fairly stable across populations with different background risks, which make them suitable for summarisation of effects in meta-analyses. Furthermore, our previous meta-analyses on risk factors for cardiovascular diseases demonstrated that detection of a female disadvantage in RRs is not inevitable when men have higher absolute risk. We thus believe that the use of RRs in the present analyses is both practical and justifiable.

22The strengths of this meta-analysis are its size and the inclusion of studies on the sex-specific effects of diabetes on all-site cancer and 50 site-specific cancers, which enabled us to conduct the most comprehensive analyses to date on the sex-specific effects of diabetes on cancer risk. To limit the risk of bias, we only included cohort studies that were conducted in men and women and had adjusted for at least age. Limitations of this study are inherent to the use of published data and the heterogeneity between studies in ascertainment of diabetes, study design and duration, endpoint definition and degree of adjustment for confounders. Nevertheless, a range of subgroup analysis provided broadly consistent results. However, as we compared women and men from within the same study, any effect of differences in methods between studies is likely to have affected women and men similarly. We therefore assume that the sex comparisons reported in this analysis are still valid. Second, the lack of data on duration of diabetes and the degree of glycaemic control precluded more detailed analyses on the effect of diabetes on the risk of cancer. Third, as this meta-analysis largely used published data, endpoint definition varied across the studies. Fourth, in analysis of all-site cancer, the women-to-men RRRs depend not only on the strengths of the RRRs of site-specific cancers (as illustrated by Fig. 5), but also on the relative incidence of site-specific cancers, which varies considerably between populations. This is likely to be a key factor in the high between-study heterogeneity we show in Fig. 3. Finally, studies generally did not adjust for obstetric and gynaecological history and unmeasured confounding is likely in the current estimates. However, confounding is likely to have been non-differentially distributed between women and men from the same study and we therefore assume that it had only a negligible effect on the reported associations.

23In conclusion, diabetes is a risk factor for all-site cancer in both sexes, with a stronger effect in women than men. Sex differences varied across the location of the cancer, heightening the importance of a sex-specific approach to quantification of the role of diabetes in cancer research, prevention and treatment. Further studies are needed to clarify the mechanisms underlying the sex differences in the diabetes–cancer association.