Use of an emulated trial to investigate the association between use of nitrogen-based bisphosphonates and risk of epithelial ovarian cancer

Abstract Background Epithelial ovarian cancer (EOC) is the eighth most common cancer in women, with poor survival outcomes. Observational evidence suggests that nitrogen-based bisphosphonate (NBB) use may be associated with reduced risk of EOC, particularly the endometrioid and serous histotypes; however, confounding by indication is a concern. An alternative approach to investigate the chemo-preventive potential of NBBs is to emulate a target trial by identifying all women who initiate use of NBBs and investigate the risk of EOC for continued users compared with discontinued users. Methods Using population-based linked data, we identified all Australian women aged over 50 years who first used NBBs over 2004–12. We used the year after first use to define treatment for each woman as either continued or discontinued use. We emulated randomization using stabilized inverse probability weights to balance the treatment groups using covariates including age, comorbidities and socioeconomic status. We followed women from treatment assignment until EOC diagnosis, death or 31 December 2013. We assessed the risk of EOC (overall and by histotype) using flexible parametric time-to-event models allowing for time-varying effects, and produced time-varying coefficients. Results Of the 313 383 women in the study, 472 were diagnosed with EOC during follow-up (261 serous EOC), with an average age at diagnosis of 72 years. Continued use of NBBs was associated with reduced risk of EOC overall (HR = 0.87, 95% CI: 0.69, 1.10), and serous EOC (HR = 0.71, 95% CI: 0.53, 0.96), compared with discontinued treatment, with estimates remaining constant over the 9-year follow-up. Conclusions Results from our emulated trial suggest that in women who initiated NBB treatment, those who continued use had 13% and 29% lower hazards of being diagnosed with EOC overall and serous EOC, respectively, compared with women who discontinued use.


Covariates
We used postcode at Medicare enrolment to define state of residence, area-level socioeconomic status using the Socio-Economic Indexes for Areas (SEIFA) Index of Relative Socio-Economic Disadvantage, 1 and remoteness category using the Accessibility/Remoteness Index of Australia 2 (ARIA).We classified participants into SEIFA quintiles from the most disadvantaged (1st quintile) to the least (5th quintile) using the distribution for all women aged 18 years or older enrolled for Medicare in 2002.Remoteness categories included major cities, inner regional, outer regional, remote/very remote.We used the earliest available SEIFA (2001)   and ARIA (2006) indices; however, if the relevant score was missing for a postcode, the 2006 or 2011 score was used.

Pharmaceutical Benefits Scheme
The Pharmaceutical Benefits Scheme (PBS) provides timely, reliable and affordable access to necessary medicines for Australians. 3Under the PBS, the government subsidises the cost of medicine for most medical conditions.All Australian residents registered for Medicare are entitled to a subsidy under the PBS, however Australians eligible for a concession also receive a subsidy for low-cost medicines.Australians who hold a Pensioner concession card, Commonwealth Seniors Health Card, Health Care Card, or Department of Veterans' Affairs card are eligible for a concession.Co-payment (out-of-pocket cost) for concessional beneficiaries for each PBS medicine ranged from $3.60 to $5.90 in 2002 to 2013, with the remaining cost of the medicine paid for by the Australian Government.The co-payment for general beneficiaries, however, ranged from $22.40 to $36.10 during this time. 3Prior to 2012, only dispensed medicines that attracted an Australian government subsidy were recorded on the PBS records.
Our study includes records from the PBS, which includes all dispensed medicines that attracted an Australian government subsidy until 2012, and all dispensed PBS medicines after 2012.The cost of many medicines is above the co-payment amount for both general and concessional beneficiaries, and therefore the PBS records include most prescribed medicines from 2002 to 2013.All nitrogen-based bisphosphonates were above co-payment throughout our study period and therefore would be captured by PBS records for all women in our study.
We used PBS data to identify comorbidities at baseline using Rx-Risk comorbidity categories, 4 which has been mapped to Australian PBS item codes. 5We were able to use the PBS data to ascertain most comorbidities included in the Rx-Risk score calculation.Some low-cost medicines that fall under the government co-payment threshold may be under ascertained prior to 2012 for example, over-the-counter pain medications.However, over 80% of the women in the cohort were concessional beneficiaries during the study period, therefore medicines for most women will be captured in our analysis.

Sensitivity analyses
We performed several sensitivity analyses to explore the potential for bias in our main analyses.

Per protocol analysis
We performed per-protocol sensitivity analyses to assess the effect of non-adherence to the treatment allocation.The effect that NBB use may have on the risk of epithelial ovarian cancer (EOC) is likely to have a latency period, we designed our per protocol analyses to allow for a 12month lag-period from non-adherence to censoring.For women in the continued use group, non-adherence was defined as 12 months of no dispensed prescriptions for NBBs.For women in the discontinued use, non-adherence occurred when a NBB prescription was dispensed during follow-up.We censored follow-up time 12 months after the date of non-adherence.Supplementary Figure S1 shows the treatment allocation and censoring times for the per protocol analysis.
After censoring, we replicated the models used for the intention-to-treat analyses, first in an unweighted Cox model adjusted for age (model S1.1) and then in a fitted flexible parametric model allowing for time-varying effects using inverse probability (IP) weights for treatment assignment (model S1.2).We then used IP of censoring weights (IPCW) in pooled logistic regression models 6 with one-year time intervals.Our IPCW models included pre-baseline variables: age, MHT use, and four comorbidities that had the greatest variation between treatment groups (ischaemic heart disease: angina, gastro-oesophageal reflux disease, steroidresponsive diseases and pain), and time-varying covariates: MHT use (during follow-up year) and another cancer diagnosis.We first used pooled logistic regression in an unweighted model adjusting for age (model 1.3) and then applying the IPCW in a doubly robust logistic regression model adjusting for age and the pre-baseline variables (model S1.4).

Other intention-to-treat (ITT) sensitivity analyses
We performed additional sensitivity analyses using the intention-to-treat survival models.These are described below.a) Exclusion of zoledronic acid (model A2.1)In this analysis, women who used zoledronic acid in the first 12 months after baseline were excluded, therefore, only women who used alendronate and risedronate were included.In our main ITT analysis, women with a single zoledronic acid injection during the first 12 months of use were categorised in the continued use group.By excluding these women in this sensitivity analysis, we could assess the extent any observed association might have been influenced by including women with a single injection of zoledronic acid in the continued use group.

b) Starting follow-up at two years after baseline (model S2.2)
We performed a sensitivity analysis conditional upon having no EOC diagnosis during the first year of follow-up.This was to assess whether there were any changes in our models when women diagnosed with EOC in the first year were excluded.In this analysis, follow-up commenced 12 months after treatment assignment, i.e. two years after the first NBB use.We explored the potential for bias due to unmeasured confounding for the association between use early use patterns of NBBs and future cancer diagnosis using a negative control cancer outcome.The purpose of this analysis was to investigate whether the same effect would be found for a cancer outcome that is not expected to be associated with NBB use.NBB use has been associated with other cancers, such as breast and other gynaecological cancers. 7erefore, we selected pancreatic cancer as the negative control outcome, as prior studies have not found an association between NBB use and risk of pancreatic cancer. 7) Alternative treatment definition (model S2. 6)   In our final sensitivity analysis, we used a different definition of continued NBB use.For this analysis, we excluded continuing users with 168 or less defined daily doses from the analysis.
The definition of discontinued users did not change.We used both ITT and per protocol analyses for this model.

Per protocol analyses (models S1.1-S1.4)
In the per protocol sensitivity analyses, 20% of women in the discontinued use group recommenced use of NBBs during follow-up, while 40% of women in the continued use group stopped using NBBs for 12 or more months.Supplementary Table S8 shows the results for each of the per protocol sensitivity analyses.There were 379 women diagnosed with EOC overall and 210 with serous EOC after excluding those diagnosed after censoring.The results of the flexible parametric model with IP weighting for treatment allocation (model 1.2) showed slightly stronger associations with both reduced risk of EOC overall (HR=0.81,95%CI:0.64,1.03)and serous EOC (HR=0.67,95%CI:0.49,0.91)compared to our main ITT analyses.The inverse association, however, strengthened over the follow-up period, showing associations from four years onwards (three years for serous EOC) (Supplementary Table S8, Supplementary Figures S4-S7).These results show that women who continued to use NBBs during follow-up had a lower risk EOC overall after 4 years and lower risk of serous EOC after 3 years, compared to those who discontinued use and did not recommence use.

ITT sensitivity analyses a) Exclusion of zoledronic acid Exclusion of zoledronic acid (model S2.1)
For this sensitivity analysis, we excluded 4,702 women whose initial NBB use was zoledronic acid and 1729 women who used zoledronic acid within 12 months of baseline.Results, as presented in Supplementary Table S9 and Supplementary Figures S8-S11, did not differ appreciably from our main ITT analysis.

b) Conditional upon having no EOC diagnosis during the first year of follow-up (model S2.2)
In the sensitivity analysis conditional upon having no EOC diagnosis during the first year of follow-up, 64 and 22 EOC diagnoses were excluded from the continued and discontinued treatment groups, respectively.Adjusted results were slightly weaker for EOC overall (HR=0.91,95%CI:0.70,1.18),but similar for serous EOC (HR=0.73,95%CI:0.53,1.00)suggesting that the first year of follow-up was not strongly influencing our overall results (Supplementary Table S10, Supplementary Figures S12-S15 (Supplementary Table S9, Supplementary Figures S16-S19).

d) Death as a competing risk (model S2.4)
There were 47,619 women who died during follow-up.When we repeated our analysis with death as a competing event (and other EOC histotypes for our analysis with serous EOC as the outcome) our results did not change materially for either EOC overall (HR=0.88,95%CI: 0.71,1.09)or serous EOC (HR=0.72,95%CI: 0.55,0.94).Continued use of NBBs was associated with a small inverse association with death, compared to discontinued users (HR=0.96,95%CI: 0.94,0.98).e) Negative control cancer outcome (model S2.5)Our sensitivity analysis using a negative control cancer outcome showed no association between continued use of NBBs and risk of pancreatic cancer (HR=1.03,95%CI:0.85,1.25),compared to discontinued use (Supplementary Table S12 and Supplementary Figure S20).There were 803 women diagnosed with pancreatic cancer during follow-up.

f) Alternative treatment definition (model S2.6)
For our final sensitivity analysis where we used an alternative definition for continued users (>168 daily doses), 17,279 women were excluded from the continued use group (Supplementary Table S9).For the ITT analysis, the overall results were not materially different from our main ITT analysis, however the association appeared to strengthen up to four years after follow-up commenced before weakening slightly.For the per protocol analysis (Supplementary Table S9 and Supplementary Figures S21-S24), however, the results were not materially different to our main per protocol analysis also using flexible parametric models (model 1.2).Supplementary Table S1

Assignment procedures
Participants will be randomly assigned to either strategy at baseline and will be blinded to their treatment.
To emulate the random assignment of treatment strategies at baseline, inverse probability weighting for propensity for treatment will be applied using pre-baseline characteristics.Follow-up period Follow-up commences 12 months after first NBB use.Women who were diagnosed with cancer or who died prior to this time will be excluded.Follow-up ends at diagnosis of epithelial ovarian cancer, death, or end of follow-up (31 December 2013).
As per target trial.

Outcome
Epithelial ovarian cancer overall and serous histotypes.Cancers registered on the Australia Cancer Database.
As per target trial.

Causal contrasts of interest
Our main analysis will be an intention-totreat analysis: comparative effect of being assigned to the treatment strategies at the start of follow-up, regardless of whether the individuals continue following the strategies during follow-up.
Per-protocol sensitivity analysis: comparative effect of following the strategies specified in the study protocol.
Women who deviate from treatment assignment are censored 1 year after non-adherence.Non-adherence defined as: • For continued NBB use: 12 months of no use (therefore censored 2 years since last use) • For discontinued NBB use: recommencement of NBB use (therefore censored 1 year after recommenced use.
As per target trial.
Analysis plan Intention-to-treat: Analysis using flexible parametric survival model with timevarying effects.We will also apply a time-varying treatment weight (one-year intervals) in a pooled logistic regression model.
Per-Protocol: We will perform the flexible parametric models as per the intention-to-treat model.We will also use inverse probability of censoring weights in a pooled logistic regression model, using one-year time intervals for time-varying effects.
As per target trial.
c) Restricting to women aged 70 years and younger (model S2.3)We performed a sensitivity analysis to only include women who initiated NBB use at 70 years or younger, to investigate whether our results were influenced by the older ages of women who commence use of bisphosphonates.d) Death as a competing risk (model S2.4)We repeated our ITT flexible parametric models with death as a competing event.For serous EOC as an outcome, we additionally included death and other EOC histotypes as competing events.e) Pancreatic cancer as a negative control cancer outcome (model S2.5) ). c) Women aged 70 years and younger (model S2.3)For the sensitivity analysis including only women commencing NBB use at 70 years or younger (n=128,633), estimates were slightly stronger, although confidence intervals were wider.
: Emulated Trial Description NBB: nitrogen-based bisphosphonateSupplementary TableS2: PBS item codes and defined daily doses for each medicine category.Excluding nitrogen-based bisphosphonates per Supplementary TableS1.Supplementary TableS6: Characteristics of women excluded due to death or cancer diagnosis during treatment assignment.
ATC: Anatomical Therapeutic Chemical; GORD: gastro-oesophageal reflux disease; IHD: Ischaemic heart disease; PBS: Pharmaceutical Benefits Scheme.a Includes combination medicines.b IP: inverse probability; NBB: nitrogen-based bisphosphonate; SEIFA: Socio-Economic Indexes for Areas; SMD: standardised mean difference.a A difference of <0.1 is generally considered acceptable.b IP Weighted frequencies presented for n.