Testosterone Therapy in Relation to Prostate Cancer in a U.S. Commercial Insurance Claims Database

In recent decades, testosterone therapy (TT) has dramatically increased in the United States (1), although trends have recently plateaued and slightly decreased (2). TT is approved by the FDA for men with hypogonadism—confirmed morning serum testosterone of <300 ng/dL—due to disorders of the testicles, pituitary gland, or brain. There are few specific signs or symptoms of hypogonadism (incomplete/delayed sexual development, body hair loss, very small testes) and the majority of TT is prescribed for nonspecific signs or symptoms (e.g., fatigue, reduced muscle bulk, increased body fat) that are age related (3). Although testosterone trials have provided some evidence that TT in men older than 65 years may aid some of the maladies associated with hypogonadism (sexual function, physical function, mood, and depressive symptoms; ref. 4), no trial was designed to evaluate risk of prostate cancer, and most observational studies of this relation have been underpowered.

In 1941, Huggins and Hodges showed that prostate cancer was androgen dependent (5), and this led to the development of androgen deprivation therapy as well as a plethora of basic, animal, and epidemiologic studies of sex steroid hormones and prostate cancer. These studies have clearly demonstrated the importance of the androgen pathway in prostate cancer progression, and the largest epidemiologic study to date of prediagnostic circulating hormones has recently found that men with very low endogenous testosterone may have a reduced risk of developing prostate cancer (6). Whether exogeneous testosterone alters risk of prostate cancer is largely unknown. Exogeneous supplementation of a single metabolite within a complex biochemical pathway with a multitude of phenotypic effects deserves careful scientific study that cannot be substituted with studies focused on endogenous androgens (7, 8). Therefore, we conducted a study using a large U.S. commercial insurance research database to assess whether TT was associated with risk of prostate cancer.

The HealthCore Integrated Research Database (HIRD) is a repository of administrative claims beginning in 2006 with linked medical, pharmacy, and eligibility data for approximately 59 million researchable covered lives (at the time of this study) with a median continuous membership of 3 years enrolled in 14 commercial health plans across the United States (9, 10). From the HIRD, we selected men ages 30 years or greater with medical and pharmacy coverage who were newly dispensed a TT prescription (see Supplementary Table S1 for codes) during January 1, 2007, to July 31, 2015. We selected two control groups to which we compared the TT-exposed group: (i) unexposed men and (ii) men dispensed one or more phosphodiesterase type 5 inhibitor (PDE5i) prescriptions. Our study design included the PDE5i group as to provide a comparison with a group of men who were willing, able, and motivated to seek medical care when symptomatic and fill prescriptions. In addition, PDE5i medications are not associated with risk of prostate cancer (11). The unexposed comparison group was matched with a target ratio of 10:1 using date of birth (±183 days), outpatient physician visit (±60 days of TT subject's dispensed prescription date), and U.S. region of residence. For the PDE5i comparison group, we selected all men newly dispensed a PDE5i during the study period. Men of comparison groups who were dispensed TT during follow-up were right censored and entered, at that time, into the TT-exposed group. Once exposed to TT, men remained in the TT group until event date or right censoring. Male sex was determined by both self-report and a lack of ICD-9 codes indicating transgenderism or GID.

Index dates were date of first dispensing of TT for exposed subjects; date of the matched outpatient physician visit for unexposed comparison group subjects; and date of first dispensing of PDE5i for the PDE5i comparison group subjects. We required a minimum 12-month continuous enrollment prior to index date. Men were excluded if, in the preindex period, there was evidence suggesting they may have had prevalent prostate cancer. Thus, we excluding men who had any code indicating prostate cancer, prostate cancer–specific treatment, elevated prostate-specific antigen (PSA; or lab test showing >4 ng/mL), prostate ultrasound guidance, prostate biopsy, medications containing estrogen, transgenderism or GID (any time point), prostatectomy, congenital absence of prostate, TT dispensing, or PDE5i dispensing or erectile dysfunction (TT vs. unexposed comparison only; see Supplementary Tables S1 and S2 for codes). Matching of unexposed men occurred after exclusions had been applied to maximize algorithm efficiency. All subjects were followed from index date until the earliest of end of study (July 31, 2015), health plan disenrollment, estrogen use, or prostate cancer (outcome).

TT exposure durations were calculated as days supplied plus a bridge rule to account for nonadherence and differences in dispensing and use. Based on exploratory analysis of missing days' supply (for oral, topical, patch, implant) and recommended usage (for injection), the bridge rule for all TT formulations was 30 days, except for mail-order TT formulations with a bridge rule of 90 days.

Incident prostate cancer was defined as diagnosis of prostate cancer (ICD-9: 185.xx) within 4 weeks following a prostate biopsy (see Supplementary Table S3 for codes) using biopsy date as date of diagnosis. To assess the validity of this definition, study-eligible men who had ever lived in Georgia were submitted to the Georgia Comprehensive Cancer Registry (GCCR) for probabilistic linkage (9). We also assessed metastatic prostate cancer, using the Dolan algorithm (12). This study was approved by the New England IRB and the Georgia Department of Public Health IRB. HIRD data are accessible under contract with HealthCore Inc.

We first calculated unadjusted matched/crude prostate cancer incidence rates and 95% confidence intervals (CI) for each of the exposure groups. We then used logistic regression to compute propensity scores (13) and estimated standardized differences to assess covariate balance (14). Propensity score models included age, region, index date, preindex time, and the following preindex period covariates: Deyo–Charlson comorbidity index, obesity, benign prostatic hyperplasia, family history of prostate cancer, inflammatory diseases of the prostate, other prostate disorders, urinary symptoms, osteoarthritis, biological treatment, antineoplastics treatment, anti-TNF treatment, alpha-reductase treatments, HIV therapy, presence of any oncologist visit, prostate cancer screening, annual examinations, ER utilization during the past year, and inpatient hospitalization utilization during the past year. Inverse probability of treatment weights was used in Poisson regression models to estimate adjusted prostate cancer incidence rates, incidence rate ratios (IRR), and their respective 95% CIs using doubly robust estimation (15).

We conducted subgroup analyses by preindex prostate cancer screening, hypogonadism, and/or benign prostatic hyperplasia. We assessed associations by amount of preindex enrollment time, calendar year of index date, and follow-up time since index date.

Among the TT group, we estimated IRRs by time on TT, number of TT refills, route of administration, and change in circulating testosterone concentration during TT.

Using the cancer registry data from GCCR as a gold standard, we estimated the positive predictive value (PPV) and sensitivity of the prostate cancer case definition in each group, and corrected effect estimates for outcome misclassification (16).

There were 76,159 men in the TT group who were matched with 721,326 unexposed men (Table 1; Supplementary Table S4). For the PDE5i comparison, there were 113,041 TT men available for comparison with 147,620 PDE5i users. Median TT exposure time was 65 days. The TT versus unexposed groups were closely matched on age and region (Table 1). Propensity score weighting achieved good comparability (d < |0.20|) for both comparison groups (Table 1). Fatigue, hypogonadism, and psychosexual dysfunction were not included in the PS due to convergence issues. We conducted sensitivity analyses in which we additionally controlled for these factors. Our prostate cancer case definition had high sensitivity (91.2%; 95% CI, 87.7%–94.0%) and PPV (81.7%; 95% CI, 77.9%–85.0%) that did not differ substantially by the analytic group (9).

We identified 335 prostate cancers in 178,704 person-years in the TT group and 4,133 cases in 1.6 million person-years in the unexposed group (Table 2). Unadjusted age- and region-matched prostate cancer incidence rates were 187.5 per 100,000 person-years in the TT group and 245.5 in the unexposed group, resulting in an IRR of 0.76 (95% CI, 0.68–0.85; Fig. 1). This estimate was similar to the propensity score–adjusted IRR of 0.77 (95% CI, 0.68–0.86). An unmeasured confounder would have to have an IRR of 1.92 with both TT and prostate cancer to explain away this observed association (17, 18). In analyses further adjusted for fatigue, hypogonadism, and psychosexual dysfunction, results were slightly attenuated with an adjusted IRR of 0.86 (95% CI, 0.74–1.00), whereas results were similar when further adjusting for outcome misclassification (adjusted IRR, 0.75; 95% CI, 0.67–0.84). Inverse associations between TT and prostate cancer were also observed when stratified by prostate cancer screening, hypogonadism, or benign prostatic hyperplasia in the preindex period (Table 2). Effect estimates did not differ appreciably by duration of preindex time, although incidence rates did decrease in both TT and unexposed groups that appeared to be related to decreasing prostate cancer incidence by calendar year. Effect estimates attenuated with time from index date, from stronger estimates of 0.51 (<6 months), to 0.72 (6–12 months), to 1.03 (12–24 months). The last period analyzed, of 24 months and greater post index, had an IRR of 0.74.

Similar results were observed when comparing TT with PDE5i users (Table 3). The crude IRR was 0.71 (95% CI, 0.64–0.76; Fig. 2) and the propensity score–adjusted IRR was 0.85 (95% CI, 0.79–0.91). Results were similar to the overall analyses, when further adjusting for unbalanced factors not included in the propensity score (adjusted IRR, 0.87; 95% CI, 0.80–0.94), and when adjusting for outcome misclassification (adjusted IRR, 0.91; 95% CI, 0.85–0.98). Subgroup analyses using the PDE5i comparison group were mostly similar to those using the unexposed comparison group, including attenuation to the null with increased time since index date with sequential 6-month period IRRs of 0.53, 0.83, 0.88, and 0.96. One difference in the PDE5i analyses, to that of the unexposed group analyses, was an attenuated association with increased preindex time, with an adjusted IRR of 1.06 (95% CI, 0.85–1.33) among individuals with at least 5 years.

Among men who received TT, duration of therapy, number of prescription fills, and route of administration were not associated with prostate cancer (Table 4). Within the TT group, men who experienced more extreme changes in circulating testosterone levels (increased or decreased) had inverse associations with prostate cancer, relative to men closer to the average change.

There were 17 prostate cancer cases with metastatic disease at diagnoses during 178,704 person-years in the TT group and 195 cases during 1.7 million person-years in the unexposed group, providing an adjusted IRR of 0.77 (95% CI, 0.46–1.29). For the PDE5i comparison, the TT group had 28 metastatic prostate cancer cases diagnosed during 267,795 person-years and the PDE5i group had 50 cases during 370,507 person-years, providing an adjusted IRR of 1.07 (95% CI, 0.77–1.49).

In men who received TT for whom we had testosterone laboratory data, circulating testosterone concentrations increased from a preindex median of 237 to 351 ng/dL (Supplementary Table S5). This increase did not vary by age but did vary by preindex circulating testosterone concentration, with greater increases observed for men with lower preindex concentrations. Circulating PSA concentrations, in the subset of men we had such data for, were not largely different between TT and comparison groups (Supplementary Tables S6 and S7). PSA concentrations in TT men increased by a median change of 0.1 ng/mL when comparing preindex with postindex periods, whereas the median changes were zero in the comparison groups.

In this study of a large healthcare claims database, men who received TT had a lower rate of prostate cancer compared with unexposed men or men receiving a PDE5i. The inverse association between TT and risk of prostate cancer was observed for a majority of subgroup analyses—including stratifications by prostate cancer screening, hypogonadism, and benign prostatic hyperplasia in the preindex period—yet the association between TT and reduced prostate cancer risk generally attenuated with increased time following initial exposure.

The majority of prior studies to have assessed TT in relation to prostate cancer have had small populations with imprecise effect estimates (19–25). There have been only two previous studies that have had large numbers of topical TT users and that have had sufficient prostate cancer cases to provide precise estimates of association. The first was a study of cases from the National Prostate Cancer Register of Sweden compared with matched controls (26). This study included 284 prostate cancer cases who had previously received a TT prescription and found no association between TT and prostate cancer (OR, 1.03; 95% CI, 0.90–1.17). The second study was a retrospective cohort of 147,593 U.S. male veterans who had one or more laboratory test-based flags for hypogonadism (27). Within this cohort, 58,617 received TT, and a total of 1,439 prostate cancers were diagnosed. The adjusted HR for the association between TT and prostate cancer was 0.90 (95% CI, 0.81–1.01). Associations in these two studies did not vary by route of TT administration, time between therapy and risk period, or duration of treatment.

The Swedish study (26), discussed above, also observed that TT was positively associated with more favorable-risk prostate cancer (OR, 1.35; 95% CI, 1.16–1.56) and negatively associated with aggressive cancer (OR, 0.50; 95% CI, 0.37–0.67). The positive association with favorable-risk prostate cancer was already apparent and strongest within the first year of TT, leading the authors to suggest detection bias. The negative association with aggressive cancer only became apparent after the first year of TT, which the authors speculated may be the result of hypogonadal genesis of poorly differentiated prostate cancer that is reversible with short-term TT. However, the study of U.S. veterans found no association between TT use and aggressive prostate cancer (HR, 0.89; 95% CI, 0.70–1.13). Although we could not assess prostate cancer stage and grade in this study, our findings of inverse associations between TT and the outcomes of metastatic prostate cancer and overall prostate cancer incidence contrast with the Swedish study and are more in line with the U.S. veterans study.

There have been two prior studies of intravenous TT in older men of the SEER-Medicare database. The first study found evidence for an inverse association with high-grade prostate cancer (OR, 0.84; 95% CI, 0.67–1.05) and no association with the high-risk disease proxy on receipt of primary androgen deprivation therapy (OR, 0.97; 95% CI, 0.74–1.30; ref. 28). The second study found no increased risk of higher grade or higher stage prostate cancer (29). In fact, this latter SEER-Medicare study, which assessed TT exposure retrospectively among a cohort of confirmed prostate cancer cases, found that men who were TT exposed were more likely to be diagnosed with moderately differentiated than less-differentiated prostate cancer, and were more likely to be diagnosed with clinical stage T3 over T4 prostate cancer, each relative to prostate cancer cases who had not previously received TT. Combined with our findings herein, the results of these four large observational studies support a hypothesis of no increased risk of prostate cancer among men who receive TT and acknowledge the possibility that an inverse association may exist (26).

An inverse association between TT and prostate cancer could be attributable to residual confounding. For example, very low endogenous-free testosterone levels could decrease the risk of prostate cancer (6) and increase the likelihood of receiving TT; although it is important to note that, among those diagnosed with hypogonadism, approximately 80% of TT prescriptions are based on tests of testosterone and approximately 15% are based on tests of free testosterone (27). Another example is diabetes has been inversely associated with prostate cancer (30) and correlates with hypogonadism, increasing the likelihood of TT (31). Residual confounding would have to be strong, given our estimate that an unmeasured confounder would have to have an IRR of 1.92 with both TT and with prostate cancer to account for the TT–prostate cancer association. However, residual confounding is supported by attenuation to the null with increased preindex time in the PDE5i comparison, which increases the likelihood of ascertaining confounding factors. Contraindication bias could also explain the observed association, whereby factors and symptoms perceived by the physician to be related with a higher risk of prostate cancer (e.g., family history, borderline PSA, urogenital symptoms) may reduce the likelihood of the physician prescribing TT (32), which would have the effect of causing a decreased prostate cancer incidence rate in the TT group and an increased rate in the unexposed group. Finally, the inverse association between TT and prostate cancer could be the result of a biological effect, such as causing an increase in proapoptotic signaling during early prostate carcinogenesis (33). Any biological effect would have to be able to explain the immediacy of the observed association and the general attenuation of effect over the 2-year period following initial exposure. It is true that the estimate with the unexposed comparison was similar in the last time period (adjusted IRR_{≥24 months}, 0.74; 95% CI, 0.62–0.88) as the overall association, but the equivalent estimate with the PDE5i comparison group was null (adjusted IRR_{≥24 months}, 0.96; 95% CI, 0.87–1.07), a pattern that mirrors the results of another recent study (25). Finally, there was no evidence that length of time of TT exposure altered the rate of prostate cancer within the TT group.

Strengths of our study include the use of a large database that enabled assessment of a younger population (unlike SEER-Medicare), and use of a validated prostate cancer definition with high sensitivity (91.2%) and PPV (81.7%) that were similar between comparison groups, thus mitigating outcome misclassification bias (9). Limitations include lack of cancer stage and grade, lack of an ability to offer a precise estimate of TT in relation to metastatic or aggressive prostate cancer, reliance on claims/payer data to infer clinical variables, availability of PSA and testosterone concentrations for only a subset of subjects, data are not informative about long latency between exposure and outcome, and limited ability or inability to assess differences by age at initial TT exposure, race, specific forms/regimens of TT, or hypogonadal subtype.

This study provides evidence that men who receive TT do not have a higher rate of prostate cancer than unexposed men or men receiving PDE5i. The inverse association between TT and prostate cancer could be the result of residual confounding, contraindication bias, or undefined biological effect.

D.C. Beachler is an associate director at Anthem. L.E. Parlett is a research manager at Anthem. P.T. Cochetti has ownership interest (including patents) in Anthem. S. Lanes is Principal Epidemiologist at HealthCore. No potential conflicts of interest were disclosed by the other authors.

Conception and design: M.B. Cook, D.C. Beachler, P.T. Cochetti, W.D. Finkle, S. Lanes

Development of methodology: D.C. Beachler, L.E. Parlett, W.D. Finkle, S. Lanes

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): L.E. Parlett

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): M.B. Cook, D.C. Beachler, L.E. Parlett, P.T. Cochetti, W.D. Finkle, S. Lanes

Writing, review, and/or revision of the manuscript: M.B. Cook, D.C. Beachler, L.E. Parlett, P.T. Cochetti, W.D. Finkle, S. Lanes, R.N. Hoover

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): M.B. Cook, D.C. Beachler, P.T. Cochetti, S. Lanes

Study supervision: M.B. Cook, S. Lanes

The authors wish to thank Bola Ekezue, PhD, Nandini Selvam, PhD, Gayathri Sridhar, PhD, and Anna Wallace, PhD, for their help and support in conducting this study. This study was financially supported by the Intramural Program of the NCI, NIH, Department of Health and Human Services (Intramural Funding Project Number: Z01 CP010180).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.