Estimates of Overall Survival in Patients With Cancer Receiving Different Treatment Regimens

Key Points Question Does the Surveillance, Epidemiology, and End Results–Medicare linked database contain sufficient information to estimate the effectiveness of adding a drug to an existing treatment regimen for elderly individuals with cancer who are systematically underrepresented in randomized trials? Findings After explicitly emulating target trials, findings from this comparative effectiveness study using Surveillance, Epidemiology, and End Results–Medicare data were not meaningfully different from those in elderly subgroup analyses reported from randomized trials. Naive observational estimates, however, were not compatible with those from previous trials. Meaning Analyses using Surveillance, Epidemiology, and End Results–Medicare data may be informative for some research questions examining the comparative effectiveness of oncological therapies for elderly individuals.

− Resection margins and peritoneal washings negative for malignant cells

Treatment Strategies
A. Initiate any dose of fluorouracil as first line treatment up to 3 months after post-surgery hospital discharge.

B. Do not initiate any chemotherapy within 3 months
A. 30 doses of fluorouracil (370mg/m2 intravenously), given either as six 5-day courses with 4 weeks between the start of the courses or as 30 once-Participants are randomized to a strategy by phone call to a central office. A "minimized" randomization procedure was used, ensuring balance with respect to age-group, site of cancer, stage, portal-vein infusion, preoperative radiotherapy, planned postoperative radiotherapy, and chemotherapy schedule (weekly versus not). Treatments were balanced within participating centers.

Follow-up Period
Time zero of follow-up is the first time an individual meets all eligibility criteria (when the person is assigned to one of the treatment strategies), here assumed to be the date of postsurgery discharge from the hospital.
Follow-up ends at the earliest of death, loss to follow-up (loss of enrollment in Medicare Parts A or B; enrollment in an HMO), or administrative end of follow-up (December 31, 2013 or 60 months after time zero) Follow-up begins at randomization.
Follow-up ends at the earliest of death, loss to follow-up, or administrative end of follow-up (January 2005 or 10 years after time zero).

Outcome
All-cause mortality. Death certified by a physician, reported to Medicare and confirmed by the National Death Index within 5 years of time zero.
All-cause mortality within 10 years of time zero.

Causal contrasts of interest
Intention-to-treat effect: effect of being assigned to the strategies at baseline, regardless of whether individuals adhere to them during follow-up Per-protocol effect: effect of adhering to the strategies (as defined in the protocol) during follow-up Intention-to-treat effect only.

Analysis Plan
Intention-to-treat effect estimated via comparison of 5-year risk of all-cause mortality among individuals assigned to each treatment strategy from a pooled logistic regression model adjusted for baseline covariates.

Assignment Procedures
Participants are randomized to either treatment strategy at baseline, and are aware of the strategy they are assigned to.
Patients are randomized to either treatment strategy at baseline, stratified by center, performance status (ECOG 0 versus 1-2), and stage (locally advanced versus metastatic). Patients and physicians are blinded to treatment assignment.

Follow-up Period
Time zero of follow-up is the first time an individual meets all eligibility criteria (when the person is assigned to one of the treatment strategies). Follow-up ends at the earliest of death, loss to follow-up (loss of enrollment in Medicare Parts A, B, or D; enrollment in an HMO), or administrative end of follow-up (December 31, 2013 or 18 months after time zero) Follow-up begins at randomization. Follow-up ends at the earliest of death, loss to follow-up, or administrative end of follow-up (September 2004 or 24 months after time zero).

Outcome
All-cause mortality. Death certified by a physician, reported to Medicare and confirmed by the National Death Index within 18 months of time zero.
All-cause mortality within 24 months of baseline.

Causal contrasts of interest
Intention-to-treat effect: effect of being assigned to the strategies at baseline, regardless of whether individuals adhere to them during follow-up Intention-to-treat effect only.
Per-protocol effect: effect of adhering to the strategies (as defined in the protocol) during follow-up Analysis Plan Intention-to-treat effect estimated via comparison of 18-month risk of all-cause mortality among individuals assigned to each treatment strategy from a pooled logistic regression model adjusted for baseline covariates.
Per-protocol effect estimates are calculated from an inverse probability weighted pooled logistic regression model, adjusted for baseline and postbaseline covariates: number of emergency department visits, Charlson Comorbidity Index, cholangitis, and pneumonia (each defined using claims from the previous week).
Intention-to-treat effect estimated via comparison of 24-month risk of all-cause mortality among individuals assigned to each treatment strategy using the Kaplan-Meier method.  E: erlotinib trial emulation; F: fluorouracil trial emulation; B: both emulations b "Other chemotherapy" excludes revenue center, erlotinib and gemcitabine codes for the erlotinib trial emulation; fluorouracil codes for the fluorouracil trial emulation c Comorbidity identification required one claim only. All available positions were used.

eAppendix 2. Details of Statistical Analysis
The statistical analysis to estimate the per-protocol effect had three steps: cloning (to avoid immortal time bias), censoring at deviation from protocol (to ensure adherence), and inverse probability weighting (to adjust for selection bias). For simplicity, we will primarily discuss this process in the context of the fluorouracil trial emulation.

Section B.1. Cloning and censoring
The cloning process involved duplicating the original data for each individual and assigning each clone or replicate (two per individual) to either strategy A or B, as visualized in Supplemental Figure 1. We created a new variable to indicate which treatment strategy the replicate was assigned to -this variable is "treated" in the model summaries in Appendices C, D, E, and F.
Replicates were then censored when they deviated from the protocol of the treatment strategy we had assigned them to follow. Each individual's treatment strategy was completely determined by the end of the grace period, so at most only one replicate from each individual still contributes person-time to the analysis by the end of the grace period.
Supplemental figure 1 illustrates how four types of individuals would be treated in this setting, which we describe in detail here.
Subject 1 initiates fluorouracil (green circle) during the grace period, and then dies or is censored after the grace period ends (black circle). Their complete person-time contributes to the Strategy A clone. However, only their person-time before initiating fluorouracil contributes to the Strategy B clone, resulting in "artificial" censoring (red circle) at the time of fluorouracil initiation.
Subject 2 does not initiate fluorouracil during the grace period, and then dies or is censored after the grace period ends. Their complete person-time contributes to the Strategy B clone. However, only their person-time during the grace period contributes to the Strategy A clone, as they have not initiated fluorouracil by the end of the grace period. The Strategy A clone here is "artificially" censored at the end of the grace period.
Subject 3 initiates fluorouracil after the end of the grace period, and then dies or is censored. Like Subject 2, only their person-time during the grace period contributes to the Strategy A clone, as they have not initiated fluorouracil. Their Strategy B clone only includes the person-time contributed before they initiate fluorouracil -they are "artificially" censored at that time.
Subject 4 dies or is censored during the grace period. In addition to the censoring for administrative or insurance reasons, this censoring includes initiating other chemotherapy for Strategy B. For Strategy A, it also includes initiating other chemotherapy before initiating fluorouracil. Individuals like Subject 4 contribute their complete person-time to both clones.
Note: For simplicity, we consider censoring due to losing insurance to happen at random, so we do not account for it in our analysis (for example, by using time-varying inverse probability of censoring weights).

Section B.2. Weighting process
For each target trial emulation, we estimated subject-specific time-varying stabilized inverse-probability (IP) weights, which create a pseudopopulation where time-varying prognostic factors are independent of future treatment. To introduce the IP weights, we first have to introduce a bit of notation. Ak is an indicator for use of fluorouracil (or erlotinib) at time k (1: ever initiated, 0: never initiated), L0 is the vector of baseline prognostic factors, and Lk is the vector of time-varying prognostic factors at time k. The overbar denotes the history of a variable since start of follow-up. The stabilized IP weights can then be written as: where ( | | ̅ −1 , 0 ) = { Pr( = 1 | ̅ −1 , 0 ), = 1 1 − Pr ( = 1 | ̅ −1 , 0 ), = 0.

Fluorouracil in stage II colorectal cancer
To estimate the probabilities in the numerator and denominator, we fit two separate pooled logistic regression model for initiation of fluorouracil in the original, unexpanded study population (n = 9,549). Each model also included a function of time ( ) as restricted cubic splines with knots pre-selected at 3, 16, 30, 44, and 57 months.
The numerator model included baseline covariates: year of diagnosis, sex, race, marital status at diagnosis, region of the US, metropolitan county, median household income in census tract, % households under poverty line in census tract, time between diagnosis and surgery, prolonged hospitalization after surgery, preoperative radiotherapy, cancer type, tumor grade, and comorbidities (anemia, abdominal distention, abnormal weight loss, asthenia, change in bowel movements, constipation, diarrhea, irritable bowel syndrome, # of emergency department visits, colonoscopy, and abdominal or pelvic CT scan). The denominator model included the baseline covariates as well as the most recent measurement of the following time-varying covariates: anemia, abdominal distention, abnormal weight loss, asthenia, change in bowel movements, constipation, diarrhea, irritable bowel syndrome, # of emergency department visits, colonoscopy, and abdominal or pelvic CT scan. During months in which a covariate measurement was not available, we carried forward the most recently recorded measurement.

Erlotinib in metastatic pancreatic cancer
To estimate the models in the numerator and denominator, we fit a pooled logistic regression model for initiation of erlotinib in the original, unexpanded study population (n = 940). Each model also included a function of time ( ) as linear and quadratic terms.
The numerator model included baseline covariates: tumor stage, age at diagnosis, and in the year before diagnosis, number of emergency department visits, Charlson Comorbidity index, performance status, cholangitis, and pneumonia. The denominator model included the baseline covariates as well as the most recent measurement of the following timevarying covariates: number of emergency department visits, Charlson Comorbidity index, cholangitis, and pneumonia. During weeks in which a covariate measurement was not available, we carried forward the most recently recorded measurement.
To calculate a single summary (average) hazard ratio as reported in trials, we use the predicted values from the weighted model to simulate the trajectory of each original individual under complete follow-up (10 simulations per individuals were used to reduce simulation uncertainty), as previously described (Toh et al., 2010). That is, we used the estimated probability of death for a random Bernoulli flip to determine if an individual was alive at a given time. The first instance of death was deemed to be end of follow-up. We then fit an unadjusted Cox proportional hazards model in the simulated data, using the predicted time of end of follow-up as the outcome, and treatment assignment (as determined by the end of the grace period) as the sole predictor. The exponentiated coefficient from this model can be interpreted as the average hazard ratio comparing, say, fluorouracil initiators to non-initiators.
95% confidence intervals were generated using a nonparametric bootstrap with 500 resamples. The estimated weights were then truncated at the 99 th percentile.

eAppendix 3. Models Used in the Emulation of the Fluorouracil Target Trial Section C.1. Model coefficients for hazard ratio estimates
Note: In all reported models, t represents the linear term for time, and t*, t**, and t*** represent the estimates for the 1 st , 2 nd , and 3 rd spline basis terms (knots prespecified at 3, 16, 30, 44, and 57 months