figure b

Introduction

Prostate cancer contributes substantially to morbidity and mortality among men worldwide, and no modifiable risk factors have been firmly established [1]. A reduced risk of prostate cancer has been suggested in association with treatment with glucagon-like peptide-1 receptor agonists (GLP-1RAs) [2,3,4,5], which are indicated for type 2 diabetes and obesity [6,7,8]. In post hoc analyses of the randomised clinical controlled LEADER (Liraglutide Effect and Action in Diabetes: Evaluation of Cardiovascular Outcome Results) trial, patients assigned to GLP-1RAs (liraglutide) had a reduced risk of prostate cancer (HR 0.54, 95% CI 0.34, 0.88) compared with patients assigned to placebo in addition to standard care. The trial included 9340 patients above 50 years with type 2 diabetes and a high risk of CVD who were followed for maximum 5 years [2]. However, large-scale population-based studies evaluating this association are few [3, 5].

The incretin hormone GLP-1 has multiple metabolic effects, most prominently reduction of blood glucose and body weight [7, 9]. Moreover, several cardiovascular trials have demonstrated a reduced risk of adverse cardiovascular events among type 2 diabetes patients with CVD or high cardiovascular risk who were treated with GLP-1RAs [10]. Some preclinical studies have suggested that GLP-1RAs inhibit proliferation and increase apoptosis of human prostate cancer cells [11, 12]. Moreover, GLP-1RAs have been shown to decrease markers of systemic inflammation [13, 14], which may be implicated in the pathogenesis of prostate cancer [15].

Prompted by the preclinical and clinical findings, we conducted a nationwide cohort study in Denmark examining the potential preventive properties of GLP-1RAs against prostate cancer in a real-world setting.

Methods

Study design

Our study was designed as a nationwide register-based cohort study using an active-comparator, new-user design. We used basal insulin as an active comparator to GLP-1RAs, as both GLP-1RAs and basal insulin are recommended as second-line therapies for type 2 diabetes by Danish national guidelines [16]. Moreover, insulin does not appear to be associated with prostate cancer risk [4].

Data sources

We retrieved data from Danish nationwide registries including the Danish Cancer Registry [17], the Danish National Prescription Registry [18], the Danish National Patient Registry [19], the Danish Civil Registration System [20], the Danish Pathology Register [21] and socioeconomic registries at Statistics Denmark [22, 23]. Unambiguous linkage of registry data was secured via the personal identification number assigned to all Danish residents [20]. Details of the registries and codes used for cancer diagnoses, drug exposure and covariates are provided in electronic supplementary material (ESM) Methods and ESM Tables 1–3.

Study population

We included all men in Denmark aged ≥50 years with a first-time prescription for GLP-1RAs or basal insulin (naive to both drugs prior to initiation) during 2007–2019. Additional inclusion criteria were residence in Denmark throughout the previous 10 years and no history of cancer (except non-melanoma skin cancer). Time of entry (index) was the date of the first filled prescription for GLP-1RAs or basal insulin during 2007–2019. GLP-1RAs were introduced in Denmark in 2007, and five specific GLP-1RA agents were marketed during the study period, i.e. exenatide, liraglutide, lixisenatide, dulaglutide and semaglutide [24]. Men with a prescription of liraglutide for obesity (Saxenda, Novo Nordisk, Denmark) as the primary GLP-1RA were excluded from the study population. We also excluded men with missing information for the selected covariates (<4%) or with no exposure time, i.e. first prescription for study drug on last day of follow-up (31 December 2019) or date of death (<0.05%). A flow chart illustrating selection of the study cohort is shown in Fig. 1.

Fig. 1
figure 1

Flow chart for the cohort selection of initiators of GLP-1RA and basal insulin treatment. Incomplete confounder information was due to missing information on either education, region of residence or income (<4% of the total population). Men excluded due to no exposure time filled their first prescription for a study drug on the last day of follow-up or date of death (<0.05% of the total population). DK, Denmark, excl., excluding

Full size image

Prostate cancer outcomes

The study outcome was a first diagnosis of prostate cancer. Prostate cancer outcomes were further classified according to clinical stage (localised or non-localised), histology (restriction to verified adenocarcinoma) and tumour aggressiveness (Gleason score <8 or ≥8) at diagnosis (definitions and codes in ESM Table 3).

Covariates

Covariates were established at index, and included age; calendar year; education (basic or high school, vocational, higher education); income; region of residence; comorbid conditions, including individual covariates of ischaemic heart disease, heart failure, cerebrovascular disease, chronic obstructive pulmonary disease, moderate to severe kidney disease or moderate to severe liver disease; concomitant drug use (one or more filled prescriptions before index), including individual covariates of metformin, sulfonylureas, combinations of non-metformin oral glucose-lowering drugs, α-glucosidase inhibitors, thiazolidinediones, dipeptidyl peptidase-4 inhibitors, sodium–glucose co-transporter 2 inhibitors, other blood glucose-lowering drugs, statins, aspirin, non-aspirin non-steroidal anti-inflammatory drugs, antihypertensives, other cardiovascular drugs, 5α-reductase inhibitors or psychotropic drugs; and proxies for diabetes severity, i.e. duration of metformin use prior to index, number of different preparations (0, 1, 2, 3, ≥4) and duration of use of non-metformin glucose-lowering drugs prior to index. Information on ethnicity was not available.

Statistical analyses

Patients were followed from initiation of the study drugs (after 1 January 2007) until first diagnosis of prostate cancer or other cancer (except non-melanoma skin cancer), emigration, death or 31 December 2019. Patients were considered exposed from initiation (i.e. first prescription) of the study drugs throughout follow-up, analogous to an intention-to-treat (ITT) analysis in a clinical trial. Exposure to the study drugs was additionally assessed continuously during follow-up (~per-protocol trial analysis), with censoring at the time of discontinuation or shift to the comparator drug. The time of discontinuation was set to 180 days after filling a prescription for a study drug, if no additional prescriptions for the drug had been filled within these 180 days. The 180-day period was chosen to account for irregular drug use. We further calculated cumulative duration of use of the two study drugs by summing the duration of individual treatment periods.

We used Cox proportional hazards regression with time since treatment initiation as the time scale for calculations of HRs with 95% CIs for prostate cancer. We applied inverse probability of treatment weighting based on propensity scores. Propensity scores were estimated using the covariate balancing propensity score method, and weights were chosen such that we estimated the average treatment effect, i.e. comparing the effect if all study participants used GLP-1RAs vs the effect if all participants used basal insulin. We used multivariable logistic regression analysis, including the above covariates, to estimate the predicted probabilities, i.e. propensity scores, of being prescribed GLP-1RAs vs basal insulin. The weights from the covariate balancing propensity score and the robust variance (sandwich) estimator were used to account for inflation of the sample size due to the weights. The proportional hazards assumption was evaluated using scaled Schoenfeld residuals. Additionally, we constructed inverse probability-weighted cumulative incidence curves for the patients commencing treatment with GLP-1RAs or basal insulin.

We estimated HRs for prostate cancer with use of GLP-1RAs compared with use of basal insulin in separate ‘ITT’ and ‘per-protocol’ analyses and according to cumulative duration of use. We further assessed potential effect modification according to clinical stage and Gleason score at prostate cancer diagnosis, i.e. each specific subgroup of prostate cancer outcome in turn constituted the event of interest, and the remaining prostate cancer cases were accordingly censored at date of diagnosis. We also repeated the main analyses stratified according to age (<65 or ≥65 years and <70 or ≥70 years) and history of CVD at index (i.e. ischaemic heart disease, heart failure or cerebrovascular disease; codes in ESM Table 2). Potential heterogeneity of effects was evaluated using a Wald test.

We performed six secondary or sensitivity analyses. First, we repeated the analyses with restriction of follow-up time to a maximum of 5 years to allow comparison of results with the LEADER trial. Second, we performed analyses accounting for competing risk from other cancer diagnoses or death, using the method proposed by Fine and Gray [25,26,27]. Third, we restricted the study outcome to histologically verified prostate cancer to increase the specificity of the case definition. Fourth, we repeated the main analyses excluding patients with one or more prescriptions for a glucose-lowering drug in 1995 to increase the accuracy of the estimate of duration of non-metformin glucose-lowering drug use prior to index. Fifth, we performed the main analysis restricting the study population to men with one or more prescriptions for an oral glucose-lowering drug prior to index to avoid including patients with type 1 diabetes mellitus. Finally, we estimated HRs for non-melanoma skin cancer as a negative control outcome, assuming no causal relationship between use of GLP-1RAs or basal insulin and non-melanoma skin cancer.

All analyses were performed using R statistical software version 4.1.0 (2021-05-18) [28]. In Denmark, ethical approval is not required for studies that are entirely register-based.

Results

We identified 697 prostate cancer cases among 14,206 initiators of GLP-1RAs (with a mean follow-up of 4.4 years) and 21,756 initiators of basal insulin (with a mean follow-up of 4.9 years) (Fig. 1). Before inverse probability weighting, we observed some differences in characteristics between treatment groups. Compared with users of basal insulin, GLP-1RA users were younger at index, had less comorbidity, a higher prevalence of use of oral glucose-lowering drugs prior to index and slightly higher education and income levels (Table 1). After inverse probability weighting, baseline characteristics were well balanced between the GLP-1RA and basal insulin users, e.g. in both study groups, the mean age was 65 years, 16% had higher education and 28% had a history of ischaemic heart disease (Table 1).

Table 1 Baseline characteristics before and after inverse probability weighting
Full size table

The HR for prostate cancer among GLP-1RA users vs basal insulin users was 0.91 (95% CI 0.73, 1.14) in the ‘ITT’ analysis and 0.80 (95% CI 0.64, 1.01) in the ‘per-protocol’ analysis (Table 2). As illustrated in Fig. 2, the inverse probability-weighted cumulative incidence curves were somewhat overlapping in the ‘ITT’ analysis, but separated in the ‘per-protocol’ analysis. No consistent trends in HRs were seen with cumulative duration of GLP-1RA use (ESM Table 4). In stratified analyses, we observed no major variation in HRs according to clinical stage, but we found reduced HRs for high, but not low, tumour aggressiveness (Table 2). The latter pattern was observed in the ‘per-protocol’ analysis, with an HR for prostate cancer with high tumour aggressiveness of 0.63 (95% CI 0.42, 0.92).

Table 2 Associations of GLP-1RA and basal insulin use with prostate cancer risk overall and according to clinical stage and Gleason score
Full size table
Fig. 2
figure 2

Inverse probability-weighted cumulative incidence curves for prostate cancer among initiators of GLP-1RA and basal insulin treatment. (a) ‘ITT’ analysis; (b) ‘per-protocol’ analysis

Full size image

In analyses stratified according to age at baseline, we observed reduced HRs for prostate cancer with GLP-1RA vs basal insulin use among older men (≥65 years, ≥70 years) in both the ‘ITT’ and ‘per-protocol’ analyses; this was most pronounced among men aged ≥70 years at index (‘ITT’ HR 0.56; 95% CI 0.38, 0.82; ‘per-protocol’ HR 0.47; 95% CI, 0.30, 0.74). No inverse association was observed among younger age groups (Table 3). Stratification according to history of CVD prior to index revealed reduced HRs for prostate cancer with GLP-1RA vs basal insulin use among men with CVD in both the ‘ITT’ analysis (HR 0.75; 95% CI 0.53, 1.06) and the ‘per-protocol’ analysis (HR 0.60; 95% CI 0.39, 0.91), whereas the HRs were close to unity among men without CVD (Table 3).

Table 3 Associations of GLP-1RA and basal insulin use with prostate cancer risk according to baseline age and CVD
Full size table

The association between GLP-1RA use and prostate cancer risk remained largely unchanged across secondary and sensitivity analyses (ESM Tables 5 and 6). Finally, use of non-melanoma skin cancer as a negative control outcome yielded generally neutral HRs with GLP-1RA use across the study analyses (ESM Table 6).

Discussion

In our nationwide cohort study, use of GLP-1RAs was inversely associated with prostate cancer risk compared with use of basal insulin in the ‘per-protocol’ analysis. Strong inverse associations were seen among older men and in patients with CVD in both ‘ITT’ and ‘per-protocol’ analyses. Risk estimates were robust across sensitivity analyses.

In keeping with our findings, a recent cohort study from the USA based on the Explorys Electronic Health Record Database reported that use of GLP-1RAs was associated with reduced odds of prostate cancer (OR 0.85; 95% CI 0.73, 0.98) compared with use of metformin [3]. These results were validated in the Food and Drug Administration Adverse Event Reporting System database (OR 0.72; 95% CI 0.50, 1.01) [3]. Moreover, a cohort study in the UK based on the Clinical Practice Research Datalink reported a reduced risk of prostate cancer among GLP-1RA users compared with users of sulfonylureas (HR 0.65; 95% CI 0.43, 0.99) [5]. In our study, we also observed a moderately reduced risk estimate for prostate cancer in the ‘per-protocol’ analysis, whereas the results of the ‘ITT’ analysis were closer to unity. The US study reported a tendency towards decreasing risks for prostate cancer with increasing duration of GLP-1RA use [3]. In our study, we found no consistent trends in prostate cancer risk according to duration of GLP-1RA use, whereas the Clinical Practice Research Datalink study reported a reduced risk of prostate cancer with 2 years of GLP-1RA use, but no clear trend with increasing duration of use [5].

Our results are in line with the suggested preventive effect of GLP-1RAs against prostate cancer reported in the LEADER trial population (HR 0.54; 95% CI 0.34, 0.88) [2]. This trial included patients with a mean age of 64 years and prior CVD or cardiovascular risk factors, who were followed for a maximum of 5 years. In our study, we observed moderately reduced HR estimates in the main ‘per-protocol’ analysis, as well as in the analysis restricted to maximum 5 years of follow-up (HR 0.75; 95% CI 0.58, 0.97). However, substantial HR reductions were seen among men with patient characteristics similar to the LEADER population, i.e. older men and patients with CVD. Further studies are warranted to evaluate potential effect modification of GLP-1RA use on prostate cancer risk according to age and CVD.

In contrast to previous studies [2, 3, 5], we had information on the clinical characteristics of prostate cancer for the majority of the prostate cancer patients. Although we did not observe a variation in the associations between GLP-1RA use and prostate cancer risk according to clinical stage, our finding of an inverse association for prostate cancer with high tumour aggressiveness in the ‘per-protocol’ analysis may indicate some relation to stage of disease.

The main strengths of our study included the use of nationwide high-quality data from continuously updated registries, including prescription data, details of medical conditions, socio-demographic variables, and detailed information on prostate and other cancer diagnoses. Cancer diagnoses recorded in the Cancer Registry [17] are validated by the Pathology Register [21], further enhancing validity. Moreover, analyses restricted to histologically verified prostate adenocarcinomas showed similar results. Use of the Prescription Registry [18] yielded continuous, complete and detailed information on use of the study drugs and other drug use over a period of more than 20 years.

Our study had limitations. Low adherence to the prescribed study drugs may have biased our risk estimates towards the null. This limitation may particularly have influenced the ‘ITT’ analyses, whereas non-compliance was of less concern in the ‘per-protocol’ analyses in which patients were censored at treatment cessation or switch. As liraglutide constituted the vast majority of GLP-1RAs dispensed in Denmark during the study period [24], we were unable to examine associations for various types of GLP-1RAs.

Our choice of an active-comparator design was intended to reduce confounding by indication, severity of diabetes mellitus and unmeasured clinical characteristics, as basal insulin is used in similar clinical settings and stages of type 2 diabetes as GLP-1RAs [16]. However, if use of the comparator drug per se influenced the risk of prostate cancer, bias might have been introduced. In a recent meta-analysis of observational studies, overall use of insulin was not associated with prostate cancer risk [4]. The meta-analysis included a Danish study [29] showing an increased incidence of prostate cancer during the first two years following initiation of insulin. However, a similar pattern was seen for patients with diabetes not using insulin therapy and likely reflects increased surveillance for prostate cancer among individuals initiating insulin or non-insulin glucose-lowering therapy. Regarding choice of comparison group, it is noteworthy that inverse associations between GLP-1RA use and prostate cancer risk have been found in studies using different comparator drugs, i.e. metformin [3] and sulfonylureas [5] as mentioned above.

To substantially reduce confounding, we used inverse probability weighting, including a broad range of patient characteristics and proxies for diabetes severity. Although we did not have information on HbA1c levels or lifestyle factors, such as BMI and smoking, neither of these factors appear to be major risk factors for prostate cancer [30,31,32]. Lack of information on HbA1c, BMI and smoking is thus unlikely to have had a major influence on our results even if the factors were unbalanced between the two treatment groups. We also lacked information on ethnicity, an established risk factor for prostate cancer; however, confounding by ethnicity is unlikely due to the homogeneity of the Danish population.

Although we used an active-comparator, new-user design, and accounted for a wide range of characteristics, the observational design of our study implies that unmeasured and residual confounding are possible. However, the analyses using non-melanoma skin cancer as a negative control outcome yielded neutral results, indicating that our findings for the association between GLP-1RA use and prostate cancer risk were not induced by a general systematic bias related to the use of GLP-1RAs.

Conclusions

In this nationwide register-based study, we observed an inverse association between GLP-1RA use and prostate cancer risk compared with use of basal insulin in the ‘per-protocol’ analysis. Strong inverse associations were seen in both the ‘per-protocol’ and ‘ITT’ analyses among older men and patients with CVD. Our findings warrant further investigation of the potential preventive effect of GLP-1RAs against prostate cancer risk.