Abstract
Background
Real-world data studies usually consider biases related to measured confounders. We emulate a target trial implementing study design principles of randomized trials to observational studies; controlling biases related to selection, especially immortal time; and measured confounders.
Methods
This comprehensive analysis emulating a randomized clinical trial compared overall survival in patients with HER2-negative metastatic breast cancer (MBC), receiving as first-line treatment, either paclitaxel alone or combined to bevacizumab. We used data from 5538 patients extracted from the Epidemiological Strategy and Medical Economics–MBC cohort to emulate a target trial using advanced statistical adjustment techniques including stabilized inverse-probability weighting and G-computation, dealing with missing data with multiple imputation, and performing a quantitative bias analysis for residual bias due to unmeasured confounders.
Results
Emulation led to 3211 eligible patients, and overall survival estimates achieved with advanced statistical methods favored the combination therapy. Real-world effect sizes were close to that assessed in the existing E2100 randomized clinical trial (hazard ratio = 0.88, P = .16), but the increased sample size allowed to achieve a higher level of precision in real-world estimates (ie, reduced confidence intervals). Quantitative bias analysis confirmed the robustness of the results with respect to potential unmeasured confounding.
Conclusion
Target trial emulation with advanced statistical adjustment techniques is a promising approach to investigate long-term impact of innovative therapies in the French Epidemiological Strategy and Medical Economics–MBC cohort while minimizing biases and provides opportunities for comparative efficacy through the synthetic control arms provided.
Database registration
clinicaltrials.gov Identifier NCT03275311.
Real-world evidence based on real-world data (RWD) increasingly influences health-care decisions (1,2). In oncology, RWD issued from electronic health records provide access to large datasets of actively treated cancer patients (3-6). Beyond descriptive characteristics related to the use of treatments in routine clinical practice, RWD analyses may open promising opportunities to overcome challenges in exploring the impact of cancer treatments for which designs such as randomized controlled trials (RCTs) are not feasible, as in rare disease situations, or powered for intermediate outcomes and not for more definitive endpoints or long-term outcomes. RWD may provide enhanced assessment of treatment effectiveness from larger populations over an extended period of time, more representative than the highly selected groups of patients that RCTs include.
Calibration of real-world evidence studies using robust methodology is one of the prerequisites to support causal conclusions in real-world efficacy of cancer treatment strategies. Indeed, identifying and eliminating sources of biases such as selection, measured confounding, and residual confounding biases are major challenges to increase the validity of comparative real-world efficacy analyses. To this end, Hernán and Robins (7) theorized the methodological framework of target trial emulation, mainly to minimize selection biases (7-11). This approach consists of applying the RCT methodological and design principles to observational data. Explicitly emulating a target trial emerges therefore as a useful approach to provide robust real-world efficacy estimates by preventing several biases including immortal time bias (8,9). Moreover, emulation of a target trial can be successfully combined with appropriate statistical adjustment techniques to control for confounding biases and thereby reach causal inference capability (9,12,13).
This study aims to assess whether the efficacy of cancer treatment strategies investigated in RCTs can be approximated through the combination of a target trial emulation and advanced confounding adjustment techniques. The reliability of the approach was tested in the French observational Epidemiological Strategy and Medical Economics (ESME) metastatic breast cancer (MBC) cohort (NCT03275311); we applied target trial emulation principles to the existing trial E2100 (NCT00028990) aiming at comparing paclitaxel plus bevacizumab (PBVZ) combination to paclitaxel alone as first-line treatment in patients with HER2-negative MBC.
Methods
ESME-MBC cohort
The retrospective cohort ESME-MBC from UNICANCER, the federation of the French comprehensive cancer centers (FCCCs), gathers routinely collected data from all patients aged older than 18 years who initiated MBC treatment in 1 of 18 FCCCs. Patient demographics included age and sex and were collected from electronic medical records. More details can be found in previously published descriptions of the cohort (5,6,14) and in Supplementary Methods (available online).
Ethics approval and consent to participate
This study used data from MBC patients registered into the cohort between January 1, 2008, and December 31, 2020. ESME research program received approval from an independent ethics committee (Comité de Protection des Personnes Sud-Est II-2015-79). The ESME-MBC was authorized by the database French data protection authority in 2013 (registration ID 1704113; authorization N_DE-2013-117; NCT03275311). No formal dedicated informed consent was required, but all patients were informed about the use of their electronically recorded data and can access, rectify, limit, or require withdrawal of their data on a dedicated web platform at any time. In compliance with the applicable European regulations, a complementary authorization was obtained on October 14, 2019, regarding the ESME research data warehouse.
Emulating the E2100 RCT following the target trial emulation framework
The open-label randomized phase III trial E2100 (NCT00028990) enrolled 722 patients from December 2001 to May 2004 and compared the efficacy of PBVZ (a monoclonal antibody against vascular endothelial growth factor) as initial treatment for patients with HER2-negative MBC. The primary endpoint was progression-free survival (PFS), defined as the time from random assignment to disease progression or death from any cause. Overall survival (OS) was a secondary endpoint (15). Combination of PBVZ statistically prolonged progression-free survival (PFS) as compared with paclitaxel alone (median PFS = 11.8 vs 5.9 months; hazard ratio [HR] = 0.60; P<.001). However, no statistically significant effect on OS was demonstrated (median OS = 26.7 vs 25.2 months; HR = 0.88; P = .16).
The emulation process of the E2100 trial consists of outlining 7 key elements of the target trial protocol (details are given in Supplementary Table 1, available online) and then applying them to the ESME data. These key elements included the following.
Eligibility criteria
We first selected ESME-MBC patients who initiated paclitaxel with or without (w/o) bevacizumab as first-line treatment and applied eligibility criteria according to the data availability; major criteria were applied, whereas minor secondary criteria, such as laboratory data, were not, because they were not captured in the database.
Treatment assignment
According to the guidelines of Hernán and Robins (7), a clinically relevant “grace period” for treatment initiation was defined around the MBC diagnosis, and patients were eligible if their treatment strategy was initiated within 1 month before and up to 4 months after confirmation of the diagnosis of MBC.
Follow-up period
Time zero was defined as the initial time at which both eligibility criteria and treatment initiation were met and indicates either treatment initiation or MBC diagnosis. Each patient was followed until the first event, specified as the date of death, last follow-up, or 60 months after baseline to ensure comparability with E2100.
Outcome
The outcome of interest was OS, defined as the time from time zero to death from any cause within 60 months from baseline. Date of death was certified by a physician and reported to ESME-MBC database or directly extracted from the online French National Institute of Statistics and Economic Studies database.
OS was favored over intermediate outcomes such as PFS, which cannot be considered in HER2-negative MBC as an approved surrogate for OS, and because, as opposed to disease progression measured with different accuracy in RCT and RWD settings, date of death is an undisputed endpoint.
Causal effect
The association of interest was the RWD analogue of the intention-to-treat hazard ratio, representing the average treatment effect of initiating PBVZ vs initiating paclitaxel alone during the grace period regardless of subsequent adherence.
Identification and selection of confounders
The classification of the confounders was supported by a literature review and the availability of data of sufficient quality. We modeled our hypotheses with causal directed acyclic graph (DAG) using the DAGitty browser-based environment (16-19). This method provides reasonable arguments to differentiate confounders from other variables that could introduce biases if controlled in the analysis (colliders and mediators). Details of selected confounders and the DAG are given in Supplementary Methods and Supplementary Figure 1 (available online).
Statistical analysis
To control for differences in baseline clinical characteristics across the treatment arms, the following strategy was used. First, patients’ characteristics identified as confounders were included in a propensity-score model that captured the probability of receiving bevacizumab. Common propensity-score overlap and covariate balance between treatment arms before and after weighting were assessed using standardized mean differences (SMD < 0.1 implied negligible balance). Second, the real-world efficacy of PBVZ compared with paclitaxel alone was gradually estimated with hazard ratio using 1) univariable (crude effect) and multivariable Cox regressions (conditional effect), 2) G-computation (GC) (marginal effect), and 3) weighted Cox regressions with stabilized inverse-probability weighting (SIPTW) and propensity score–overlap weighting (PSOW) (marginal effects) (20-23).
Missing baseline confounders were imputed using multiple imputation (24-26). Imputed datasets (m = 30) were generated using the Multiple Imputation with Factorial Analysis of Mixed Data (MIFAMD) model from the missMDA R package (27). More details are given in Supplementary Methods (available online).
A sensitivity analysis explored unmeasured confounding through quantitative bias analysis (QBA) for unmeasured confounding (28,29). To this end, we applied the array approach (28) (Supplementary Methods, available online). In addition, the E-value was calculated as described in VanderWeele and Ding (30) (Supplementary Methods, available online). Finally, the influence of the emulation process was explored by conducting an analysis on a “naive” population obtained by selecting patients on only few major clinical criteria. More details are given in Supplementary Methods (available online). All analyses were performed in R statistical software version 4.2.0.
Results
Study population
Among 30 459 patients enrolled in the ESME-MBC cohort, the naive analysis considered 5538 patients, including 2940 patients who initiated PBVZ and 2598 patients who initiated paclitaxel alone (Figure 1; Supplementary Table 1, available online). As part of the emulation process, only 3211 patients met additional eligibility criteria. These 3211 patients constituted the so-called emulated population of which 1740 initiated PBVZ and 1471 initiated paclitaxel alone.
Figure 1.
Flowchart. ESME = Epidemiological Strategy and Medical Economics; MBC = metastatic breast cancer; ECOG PS = Eastern Cooperative Oncology Group Performance Status; PBVZ = paclitaxel plus bevacizumab; P/BVZ = paclitaxel combined or not with bevacizumab; CT = Chemotherapy; RT = radiotherapy; HT = hormonal therapy.
In the naive population, major patients’ characteristics showed imbalance between treatment arms (Table 1). Compared with paclitaxel alone–treated patients, PBVZ-treated patients were younger (mean age [SD] = 54.7 [11.5] years vs 59.1 [13.4] years), more likely to have visceral metastases (n = 1823 [70.2%] vs 1814 [61.7%] patients), and more heavily pretreated in (neo)adjuvant setting ((neo)adjuvant hormonal therapy: n = 1302 [50.2%] vs 1208 [41.2%] patients; (neo)adjuvant chemotherapy: n = 1868 [72%] vs 1426 [48.6%] patients). By contrast, PBVZ-treated patients were less likely to have de novo disease (n = 517 [19.9%] vs 1116 [38.0%] patients) and were less pretreated in the first-line of metastatic setting before paclitaxel w/o bevacizumab initiation (n = 724 [27.9%] vs 1268 [43.1%] patients).
Table 1.
Unadjusted patient characteristics by treatment arm in the naive and emulated population
| Characteristics | Naive population |
Emulated population |
||||
|---|---|---|---|---|---|---|
| Paclitaxel arm (n = 2940) n (%) |
PBVZ arm (n = 2598) n (%) |
uSMD (range)a | Paclitaxel arm (n = 1471) n (%) |
PBVZ arm (n = 1740) n (%) |
uSMD (range)a | |
| Age, years | ||||||
| Mean (SD) | 59.1 (13.4) | 54.7 (11.5) | −0.35 (−0.37 to −0.33) | 60.9 (13.1) | 54.8 (11.5) | −0.49 (−0.51 to −0.46) |
| Younger than 50 | 740 (25.2) | 862 (33.2) | 0.17 (0.15-0.19) | 310 (21.1) | 570 (32.8) | 0.26 (0.24-0.29) |
| 50-70 | 1548 (52.7) | 1527 (58.8) | 0.12 (0.10-0.14) | 784 (53.3) | 1034 (59.4) | 0.11 (0.08-0.15) |
| Older than 70 | 652 (22.2) | 209 (8.0) | −0.39 (−0.42 to −0.37) | 377 (25.6) | 136 (7.8) | −0.48 (−0.52 to −0.45) |
| Menopause, yes | 2054 (69.9) | 1621 (62.4) | −0.15 (−0.17 to −0.13) | 1090 (74.1) | 1087 (62.5) | −0.24 (−0.26 to −0.22) |
| Performance status | ||||||
| 0 | 874 (50.1) | 720 (54.2) | 385 (44.2) | 521 (56.4) | ||
| 1 | 871 (49.9) | 608 (45.8) | −0.01 (−0.05 to 0.02) | 486 (55.8) | 402 (43.6) | −0.13 (−0.18 to −0.08) |
| Missing | 1195 | 1270 | 600 | 817 | ||
| Estrogen receptor status | ||||||
| Negative | 1060 (36.8) | 1023 (39.8) | 622 (43) | 815 (47.2) | ||
| Positive | 1820 (63.2) | 1547 (60.2) | −0.06 (−0.07 to −0.05) | 823 (57) | 911 (52.8) | −0.08 (−0.09 to −0.06) |
| Missing | 60 | 28 | 26 | 14 | ||
| Progesterone status | ||||||
| Negative | 1682 (59.7) | 1519 (61.1) | 889 (63.2) | 1076 (64) | ||
| Positive | 1134 (40.3) | 968 (38.9) | −0.03 (−0.05 to −0.00) | 517 (36.8) | 606 (36) | −0.01 (−0.04 to 0.01) |
| Missing | 124 | 111 | 65 | 58 | ||
| Hormone receptor status | ||||||
| Negative | 1028 (35.7) | 985 (38.3) | 606 (41.9) | 786 (45.5) | ||
| Positive | 1852 (64.3) | 1585 (61.7) | −0.05 (−0.06 to −0.04) | 839 (58.1) | 940 (54.5) | −0.07 (−0.09 to −0.05) |
| Missing | 60 | 28 | 26 | 14 | ||
| HER2 status | ||||||
| Negative | 2801 (100) | 2506 (100) | NA | 1398 (100) | 1684 (100) | NA |
| Missing | 139 | 92 | 73 | 56 | ||
| M+ disease-free interval, monthsb | ||||||
| Mean (SD) | 60.4 (81.5) | 58.1 (66.7) | −0.03 (−0.04 to −0.01) | 71.1 (86.3) | 58.1 (68.8) | −0.17 (−0.18 to −0.15) |
| <6, de novo | 1116 (38) | 517 (19.9) | −0.41 (−0.43 to −0.40) | 458 (31.1) | 397 (22.8) | −0.18 (−0.21 to −0.17) |
| 6-24 | 247 (8.4) | 477 (18.4) | 0.30 (0.28-0.32) | 143 (9.7) | 336 (19.3) | 0.27 (0.25-0.29) |
| >24 | 1577 (53.6) | 1604 (61.7) | 0.17 (0.15-0.18) | 870 (59.1) | 1007 (57.9) | −0.03 (−0.05 to 0.01) |
| Metastatic sites | ||||||
| Bone-only metastasis | 433 (14.7) | 375 (14.4) | −0.00 (−0.01 to 0.02) | 167 (11.4) | 208 (12) | 0.02 (0.00-0.04) |
| Nonvisceral metastases, skin, and lymph nodes | 693 (23.6) | 400 (15.4) | −0.22 (−0.24 to −0.21) | 251 (17.1) | 287 (16.5) | −0.02 (− 0.04 to −0.00) |
| Visceral metastases, excluding brain | 1814 (61.7) | 1823 (70.2) | 0.19 (0.18- 0.20) | 1053 (71.6) | 1245 (71.6) | 0.01 (−0.01 to 0.03) |
| Year of M+ diagnosis | ||||||
| 2008-2015 | 1385 (47.1) | 2290 (88.1) | 781 (53.1) | 1500 (86.2) | ||
| 2016-2020 | 1555 (52.9) | 308 (11.9) | −1.00 (−1.02 to −0.98) | 690 (46.9) | 240 (13.8) | −0.79 (−0.81 to −0.77) |
| Center activity level, by No. of patients enrolled in ESME database | ||||||
| <1500 | 736 (25) | 840 (32.3) | 0.17 (0.14-0.19) | 350 (23.8) | 565 (32.5) | 0.20 (0.17-0.22) |
| [1500-2000] | 962 (32.7) | 923 (35.5) | 0.05 (0.03-0.07) | 522 (35.5) | 641 (36.8) | 0.02 (0.00-0.04) |
| ≥2000 | 1242 (42.2) | 835 (32.1) | −0.21 (−0.23 to −0.19) | 599 (40.7) | 534 (30.7) | −0.21 (−0.23 to −0.18) |
| History of other cancer, yes | 164 (5.9) | 111 (4.4) | −0.06 (−0.08 to −0.04) | 100 (7.2) | 80 (4.7) | −0.10 (−0.12 to −0.07) |
| Missing | 149 | 66 | 73 | 45 | ||
| Family history of breast and /or ovarian cancer, yes | 545 (22.2) | 449 (19.2) | −0.06 (−0.08 to −0.05) | 261 (21.5) | 300 (19.2) | −0.05 (−0.08 to −0.02) |
| Missing | 485 | 263 | 257 | 178 | ||
| (Neo)adjuvant chemotherapy, IV, yes | 1426 (48.6) | 1868 (72.0) | 0.49 (0.47-0.51) | 807 (55) | 1203 (69.2) | 0.28 (0.26-0.31) |
| Missing | 8 | 2 | 5 | 1 | ||
| (Neo)adjuvant hormonal therapy, yes | 1208 (41.2) | 1302 (50.2) | 0.18 (0.16-0.20) | 579 (39.5) | 700 (40.3) | 0.00 (−0.02 to 0.02) |
| Missing | 8 | 2 | 5 | 1 | ||
| (Neo)adjuvant anthracyclines, yes | 1316 (44.8) | 1797 (69.2) | 0.50 (0.49-0.53) | 739 (50.2) | 1151 (66.1) | 0.31 (0.29-0.34) |
| (Neo)adjuvant taxanes, yes | 1022 (34.8) | 1374 (52.9) | 0.37 (0.36-0.39) | 557 (37.9) | 875 (50.3) | 0.24 (0.22-0.27) |
| Other drug(s) in neoadjuvant setting, yes | 1165 (39.6) | 1396 (53.7) | 0.28 (0.27-0.30) | 642 (43.6) | 893 (51.3) | 0.14 (0.13-0.16) |
| Prior m1L drugs before paclitaxel w/o bevacizumab, yesc | 1268 (43.1) | 724 (27.9) | −0.34 (−0.35 to −0.32) | 83 (5.6) | 113 (6.5) | 0.03 (0.01-0.05) |
Unadjusted average standardized mean difference (SMD) computed after multiple imputation; SMD < 0.1 implies negligible imbalance. No. = Number of patients; PBVZ = Paclitaxel plus bevacizumab; M+ = metastatic; ESME = Epidemiological Strategy and Medical Economics; IV = intravenous; m1L = first-line of metastatic treatment; w/o = with or without; NA = not applicable.
Metastasis-free interval is the period between the date of primary tumor diagnosis and the date of metastatic diagnosis.
Prior metastatic drugs for the naive population could be chemotherapy or hormonal therapy regimens, whereas for the emulated population, prior drugs were only hormonal therapy regimens (because no chemotherapy is a mandatory eligibility criterion).
Imbalance between arms was reduced after the emulation process (eligibility criteria implementation) to achieve a suitable balance for some of the confounders, especially regarding visceral metastases (average unadjusted SMD < 0.1), (neo)adjuvant hormonal therapy (average uSMD < 0.1), and prior metastatic drugs administration before paclitaxel w/o bevacizumab initiation (average uSMD < 0.1). However, imbalance remained regarding other confounders, for example, in metastasis-free interval (de novo, average SMD = 0.18), (neo)adjuvant chemotherapy (average uSMD = 0.28), and year of diagnosis (average uSMD = -0.79). For other parameters, imbalance increased, such as in age (average uSMD = -0.49), menopause (average uSMD = -0.24), and performance status (average uSMD = -0.13) (Table1).
The propensity score distributions showed acceptable overlaps in both the naive and emulated population (Figure 2, A). However, in the emulated population, the shape of the distribution in the control arm was right shifted, indicating a reduction in imbalance between arms.
Figure 2.
Common support of the probability of receiving bevacizumab in the naive and emulated populations (A) and covariate balance before and after weighting in the emulated population (B). B) Only the confounders that were the most imbalanced at baseline among the selected ones are presented. Absolute Standardized Mean Differences are averaged and presented with the range across all m = 30 imputed datasets. Center activity level is defined according to the number of patients enrolled by the center in the ESME cohort. no.: number; ESME = Epidemiological Strategy and Medical Economics; M+ = metastatic; ECOG PS = Eastern Cooperative Oncology Group Performance Status; CT = chemotherapy; IV = intravenous; HT = hormonal therapy; m1L = first-line metastatic treatment; w/o = with or without; SIPTW = stabilize inverse probability of treatment weighting.
Baseline covariates measurement before and after weighting in the emulated population are shown in Figure 2, B. As expected, after weighting by SIPTW, balance was achieved for all covariates except for those aged older than 70 years (average SIPTW SMD = 0.11). More details are given in Supplementary Results (available online).
Real-world treatment efficacy
In the naive population, the median OS was similar between PBVZ-treated and paclitaxel alone–treated patients (unadjusted median = 27.8 months, 95% confidence interval [CI] = 27.5 to 28.0, vs unadjusted median = 27.3 months, 95% CI = 26.9 to 27.6 months, respectively; HR = 1.02, 95% CI = 0.95 to 1.10) (Figure 3, A). After emulation, the median survival time was statistically longer for PBVZ-treated patients than for paclitaxel alone–treated patients in both unadjusted analyses (PBVZ = 28.1 months, 95% CI = 27.9 to 28.3, months vs Paclitaxel alone = 23.6 months, 95% CI = 23.4 to 24.1, months; HR = 0.89, 95% CI = 0.82 to 0.98) and SIPTW analyses (PBVZ = 26.5 months, 95% CI = 26.1 to 27.2, months vs Paclitaxel alone: 20.8 months, 95% CI = 20.5 to 21.1, months; HR = 0.84, 95% CI = 0.77 to 0.92) (Figure 3, B).
Figure 3.
Unadjusted overall survival in the naive population (A) and unadjusted and SIPTW-weighted overall survival in the emulated population (B) according to the treatment arm. Numbers at risk are averaged sample sizes across all m = 30 imputed datasets. Number at risk corresponding to the dashed distributions are SIPTW. *The proportionality hazards assumption is not satisfied in the naive population; the crude hazard ratio is given for exploratory purposes only. OS = overall survival; mo = months; CI = confidence interval; HR = hazard ratio; SIPTW = stabilized inverse probability of treatment weighting; mths = months.
To explore the impact of statistical methods used to control confounding bias in the emulated population, concurrent analyses were performed using the following methods: multivariable-adjusted Cox model, SIPTW, propensity score–overlap weighting (PSOW), and G-computation (GC). The various methods showed consistent results (Figure 4).
Figure 4.
Summary of hazard ratios for overall survival. *The proportionality hazards assumption is not satisfied in the naive population; the crude hazard ratio is given for exploratory purposes only. **N indicates sample size for efficacy analyses in the E2100 trial or corresponds to the number of individuals with analyzable data in the present study. Pseudo sample sizes after weighting and averaged across the m = 30 imputed datasets are shown for SIPTW and PSOW methods. CI = confidence interval; RCT = randomized control trial; pop = population; SIPTW = stabilized inverse probability of treatment weighting; PSOW = propensity score-overlap weighting.
QBA of unmeasured confounding
QBA used the array approach to investigate robustness of our findings under varying assumptions relative to a hypothetical unmeasured confounder. Based on a GC-adjusted hazard ratio of 0.81 (95% CI = 0.74 to 0.88; apparent relative risk = 0.86, 95% CI = 0.81 to 0.92), we estimated a fully adjusted exposure relative risk for a hypothetical unmeasured binary confounder representing an unconfounded association between the treatment and the mortality (Supplementary Figure 2, available online). Results showed that to reverse our findings, an unmeasured confounder would require simultaneously fulfilling the following 2 conditions: 1) to be highly correlated with mortality and 2) to be highly imbalanced between the treatment arms with a higher prevalence among paclitaxel alone–treated patients than in PBVZ-treated patients. Hypertension was investigated as a key potential unmeasured confounder (Supplementary Figure 2, available online). Results about the E-value are given in Supplementary Results (available online).
Discussion
To our knowledge, this is the first analysis combining a target trial emulation with recent statistical adjustment to estimate the real-world efficacy of bevacizumab added to paclitaxel as first-line treatment in HER2-negative MBC in the real-world clinical setting. This work based on the existing E2100 RCT (15) used data from a large observational cohort (14) and was designed to minimize selection bias by design and confounding bias by statistical adjustment in combining the target trial emulation with robust statistical methods.
Explicit emulation of a target trial prevents selection bias by design (8,9). The emulation process achieved a suitable balance or a reduced imbalance between treatment arms regarding some confounders. Moreover, we observed that emulation affected preferentially patients from the control paclitaxel-alone arm than patients from the experimental PBVZ arm. Many patients from the paclitaxel-alone arm in the naive population would never have been included in RCT because of their characteristics and so would not be eligible to receive bevacizumab. The inclusion of such patients in a standard comparative analysis from observational data leads to biased estimates because the probability of being treated for those patients is actually close to zero, a phenomenon called positivity violation in the causal inference framework. Indeed, the theoretical positivity violation is a consequence of a conceptual issue in the design of observational studies (31). The strength of the emulation process is to avoid this pitfall by excluding patients with contraindications to a study treatment. However, we cannot ignore that imbalance may remain and even sometimes be greater in the emulated cohort than in the naive population regarding some confounders, because the emulation framework mainly addresses the selection bias and not directly the confounding bias at this step of the process.
Our results showed that combining the emulation process with advanced adjustment methods is appropriate to strengthen the correction of the imbalances between treatment arms. Balance was indeed achieved after weighting by SIPTW for all covariates, except for age for which little imbalances persisted. After emulation and statistical adjustment for observed confounders, PBVZ-treated patients had statiscally longer OS compared with paclitaxel alone–treated patients, for all concurrent adjustment methods. These estimates are consistent with that reported in the E2100 trial with a hazard ratio of 0.88 (95% CI = 0.74 to 1.05). As a reminder, in the early 2000s, 3 pivotal trials (E2100, AVADO, and RIBBON-1) investigating the efficacy of the addition of bevacizumab to a chemotherapy were conducted in first-line treatment of HER2-negative MBC (15,32-35). AVADO and RIBBON-1 were not powered for OS, whereas E2100 had 80% power to detect an OS hazard ratio of 0.71 (35). All failed to demonstrate a statistically significant OS benefit.
Conversely, our results slightly differ from those previously reported by Delaloge et al. (36) using data available at that time in the real-life ESME-MBC cohort. Indeed, the benefit was higher for paclitaxel combined with bevacizumab vs paclitaxel alone after adjustment on propensity score (HR = 0.7, 95% CI = 0.635 to 0.771). In addition to the advanced confounding adjustment methods and the use of multiple imputation to manage missing data in our analyses, we hypothesize that this difference can also be explained by the emulation framework providing a control arm with highly selective criteria, similar to those used in randomized trials, and therefore excluding patients with contraindications to receive bevacizumab. Applying the emulation allows to overcome the limits often formulated in regard to the RWD (37,38). Moreover, the effect sizes estimated in our analyses may better address selection bias owing to design and confounding bias by adjustment, whereas Delaloge et al. (36) used propensity score matching and adjustment on the propensity score—2 methods known to be less robust than weighting (39-41)—and no imputation of missing data was performed.
In the present study, a major issue was to raise awareness on potential misuse of statistical adjustment methods to address confounding bias in a poorly selected population such as the naive cohort. We consequently proposed a step-by-step analysis by performing univariable analysis in the naive population, then in the emulated population, and finally applying confounding adjustment only in the emulated population. We therefore emphasize the control of selection bias first ensuring in the selected population that positivity assumption via target trial emulation is met and then the control of the confounding bias by applying statistical adjustment methods. Our results highlight that the target trial emulation step was the main process correcting the results, revealing a protective effect of PBVZ compared with paclitaxel alone. The different methods used for baseline confounding adjustment had lower impact on the association between treatment and OS, slightly strengthening the protective effect. These results emphasize that the target trial emulation framework is fundamental when investigating a causal treatment effect. By contrast, the choice of the adjustment technique is not decisive if the emulation is carried out correctly and if sufficient data of good quality on confounders are available (9). The target trial approach was previously investigated in oncology by using data from population-based sources like in the study by Petito et al. (42) where the authors emulated 2 existing RCTs from the Surveillance, Epidemiology, and End Results–Medicare database. The strengths and specificities of our work are as follows: first, our analysis is based on a causal diagram for the selection of the confounders; second, we used a large national database dedicated to MBC, which gave us the possibility to include more patients to support both emulation and adjustment processes; and finally, we were able to explore the issue of residual bias.
The QBA analyses showed that only an unmeasured confounder that would be highly imbalanced between the treatment arms and highly correlated with mortality would change the values of our treatment estimates. None of the other unmeasured confounders such as uncontrolled hypertension or a history of deep thrombosis, which could both differ the medical decision to administer bevacizumab and affect OS, fulfilled these conditions. Especially, uncontrolled hypertension (grade II or III) is lowly represented in the overall French female population displaying a similar pattern of age as in our emulated population (43).
Residual bias because of unmeasured confounders should also be investigated in the future with other approaches, such as the negative controls, for example, to support these findings (44,45). This approach was not tested in our analyses because it requires matching the ESME data to the database from the French national health insurance information system to access data outside cancer therapy in ESME patients (46).
The generalizability of our findings would need further target trial emulations in different treatment settings to test and confirm the robustness of using the same approach within real-life ESME-MBC data. Moreover, large real-world databases with long follow-up periods could also be helpful to assess real-world treatment efficacy and provide a promising opportunity to investigate the long-term effects of innovative therapies that cannot be accurately evaluated in RCT, especially in rare subgroups of patients.
This study is based on a nation-wide, real-life cohort ESME-MBC, annually updated, and containing data of high quality and granularity—2 key points to substantiate the feasibility and the relevance of an emulation framework. In addition, it was possible to satisfactorily emulate critical components of the E2100 RCT protocol, including treatment strategies and inclusion and exclusion criteria, with minor exceptions for rare features not exactly emulated owing to differences in data source (data capturing). The positivity assumption was not violated although bevacizumab was removed from the reimbursement list in France for hormone receptor–positive [HR+] MBC (HR+ and HER2-) on August 3, 2016, with a probability greater than zero to prescribe bevacizumab in HR+ MBC after 2016 (data not shown).
If confounding bias related to differences in some available baseline patient characteristics is commonly investigated in RWD studies, selection bias is rarely discussed, especially in oncology. This work, by contrast, included a rigorous effort to minimize by design the selection bias, including the immortality bias in aligning the time zero. In addition, measured confounding bias was conventionally minimized by statistical adjustment, with caution to identify confounders by the elaboration of a DAG. We explored several robust causal inference methods, and we compared the estimates gradually from crude effect to conditional and marginal effects obtained by using the current state-of-the-art statistical methods.
In addition, even after controlling for measured confounders by design and analysis, we cannot exclude the presence of residual confounding. Systematic differences in terms of patient selection and data collection could still exist between RCT and RWD settings and between RWD treatment arms after controlling by design and adjustment methods. Moreover, some confounders may not have been addressed because of data not collected or unknown. Because laboratory parameters (blood cell count, liver function tests, and renal function) were not collected in the ESME-MBC cohort, the related inclusion and exclusion criteria could not be implemented in the emulation process. However, we assume that the clinicians checked the absence of contraindications before treatment prescription and did not prescribe otherwise. For instance, this could introduce a bias in the paclitaxel alone–treated arm, which could therefore be enriched in patients with altered renal function. In the same way, the E2100 RCT took place in the early 2000s, whereas the ESME cohort included patients only since 2008, so temporality could not be emulated. Finally, observational studies often included some missing data, carrying specific biases. To address this limitation, this study used the multiple imputation approach assuming a missing at random hypothesis for which verification from the data cannot be achieved, even if the multitude of variables available in the ESME database made this hypothesis plausible.
The real-world effect of paclitaxel plus bevacizumab in HER2- MBC patients from the French ESME-MBC cohort was consistent with that observed in the initial E2100 RCT when an appropriate causal inference approach was implemented. These findings suggest that the explicit emulation of a target trial combined with robust statistical adjustment techniques is a powerful approach to assess treatment effectiveness in large observational database. This approach offers large opportunities to investigate long-term impact of treatments in the ESME-MBC cohort, especially in rare subgroups of patients, and to address challenges to demonstrate the comparative effectiveness of innovative therapies including real-world synthetic control arms.
Supplementary Material
Acknowledgements
The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication. Data collection, analysis, and publication are managed entirely by UNICANCER independently of the industrial consortium.
We thank UNICANCER network and the 18 French Comprehensive Cancer Centres (FCCCs) for providing the data and each ESME local coordinator for managing the project at the local level and the ESME Scientific Group and Strategic Committee for their ongoing support. The 18 participating FCCCs included I. Curie, Paris/Saint-Cloud; G. Roussy, Villejuif; I. Cancérologie de l’Ouest, Angers/Nantes; C. F. Baclesse, Caen; ICM Montpellier; C. L. Bérard, Lyon; C. G.-F. Leclerc, Dijon; C. H. Becquerel, Rouen; I. C. Regaud, Toulouse; C. A. Lacassagne, Nice; Institut de Cancérologie de Lorraine, Nancy; C. E. Marquis, Rennes; I. Paoli-Calmettes, Marseille; C. J. Perrin, Clermont Ferrand; I. Bergonié, Bordeaux; C. P. Strauss, Strasbourg; I. J. Godinot, Reims; and C. O. Lambret, Lille.
We also thank Béchir Ben Hadj Yahia, MD, PhD; Rémy Choquet, PhD; and Cyril Esnault, MSc; for kind consideration to the study; all members of the thesis supervisory committee: Michel Cucherat, MD, PhD; Marie-Laure Delignette-Muller, PhD; Roch Giorgi, MD, PhD; and Raphaël Porcher, PhD, for advice and relevant methodological expertise, and Sophie Darnis, PhD, for editorial support.
Contributor Information
Alison Antoine, Clinical Research and Biostatistics Department, Centre Léon Bérard, Lyon, France; UMR CNRS 5558 LBBE, Claude Bernard Lyon 1 University, Villeurbanne, France.
David Pérol, Clinical Research and Biostatistics Department, Centre Léon Bérard, Lyon, France.
Mathieu Robain, Data Direction, UNICANCER, Paris, France.
Suzette Delaloge, Department of Cancer Medicine, Gustave Roussy, Villejuif, France.
Christine Lasset, UMR CNRS 5558 LBBE, Claude Bernard Lyon 1 University, Villeurbanne, France; Prevention & Public Health Department, Centre Léon Bérard, Lyon, France.
Youenn Drouet, UMR CNRS 5558 LBBE, Claude Bernard Lyon 1 University, Villeurbanne, France; Prevention & Public Health Department, Centre Léon Bérard, Lyon, France.
Data availability
The dataset analyzed in the current study is issued from the Epidemiological Strategy and Medical Economics (ESME) Metastatic Breast Cancer (MBC) Data Platform. The database of the ESME program, including the database of the MBC cohort, are currently not accessible. In accordance with the regulatory frame applicable to the ESME Data Platform, data are only available via a secure cloud computing service with limited access to authorized persons; access authorizations to the secure cloud computing service are available from the corresponding author upon reasonable request. Each demand will be examined on a case-by-case basis by the UNICANCER ESME scientific committee. Moreover, subset of data (2008-2014) is made available on the Health Data Hub (https://www.health-data-hub.fr/).
Author contributions
Alison Antoine, MSc (Conceptualization; Formal analysis; Funding acquisition; Methodology; Writing – original draft; Writing – review & editing), David PEROL, MD (Conceptualization; Funding acquisition; Supervision; Writing – original draft; Writing – review & editing), Mathieu Robain, MD, PhD (Conceptualization; Funding acquisition; Writing – review & editing), Suzette Delaloge, MD (Writing – review & editing), Christine Lasset, MD, PhD (Funding acquisition; Supervision; Writing – original draft; Writing – review & editing), and Youenn Drouet, PhD (Conceptualization; Formal analysis; Methodology; Supervision; Writing – original draft; Writing – review & editing).
Funding
This study was supported by the National French Research and Technology Association (ANRT) and Roche (France) via CIFRE doctoral fellowship no. 2020/1054. The ESME-MBC cohort receives financial support from an industrial consortium (Roche, Pfizer, AstraZeneca, MSD, Eisai, and Daiichi Sankyo).
Conflicts of interests
AA reports a grant from the National French Research and Technology Association (ANRT) and Roche (France) via CIFRE doctoral fellowship no. 2020/1054. DP reports personal fees and travel funding from Roche; consulting and personal fees from Takeda; personal fees from AstraZeneca, Bayer, Boehringer-Ingelheim, Bristol-Myers Squibb, Daiichi-Sankyo, Eli-Lilly, Ipsen, Novartis, Merck Sharp and Dohme, Janssen, and Pfizer, outside the submitted work. MR reports consulting fees from Roche and Pfizer, outside the submitted work. SD reports grants and non-financial support from Pfizer and AstraZeneca; grants from Novartis, Roche Genentech, Lilly, Orion, Amgen, Sanofi, Genomic Health, Servier, MSD, BMS, Pierre Fabre, Exact Sciences, Besins, European Commission, French government, Fondation ARC, Taiho, and Elsan, outside the submitted work. All remaining authors have nothing to disclose.
References
- 1. Office of Medical Policy, US Food and Drug Administration. Real-world evidence. Published 2022. https://www.fda.gov/science-research/science-and-research-special-topics/real-world-evidence. Accessed January 17, 2023.
- 2. Bakker E, Plueschke K, Jonker CJ, Kurz X, Starokozhko V, Mol PGM. Contribution of real-world evidence in European medicines agency’s regulatory decision making. Clin Pharmacol Ther. 2023;113(1):135-151. doi: 10.1002/cpt.2766. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3. Wilkinson S, Gupta A, Scheuer N, et al. Assessment of alectinib vs ceritinib in ALK-positive non–small cell lung cancer in phase 2 trials and in real-world data. JAMA Netw Open. 2021;4(10):e2126306. doi: 10.1001/jamanetworkopen.2021.26306. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. Lux MP, Nabieva N, Hartkopf AD, et al. Therapy landscape in patients with metastatic HER2-positive breast cancer: data from the PRAEGNANT Real-World Breast Cancer Registry. Cancers. 2018;11(1):10. doi: 10.3390/cancers11010010. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Deluche E, Antoine A, Bachelot T, et al. Contemporary outcomes of metastatic breast cancer among 22,000 women from the multicentre ESME cohort 2008-2016. Eur J Cancer. 2020;129:60-70. doi: 10.1016/j.ejca.2020.01.016. [DOI] [PubMed] [Google Scholar]
- 6. Grinda T, Antoine A, Jacot W, et al. Evolution of overall survival and receipt of new therapies by subtype among 20 446 metastatic breast cancer patients in the 2008-2017 ESME cohort. ESMO Open. 2021;6(3):100114. doi: 10.1016/j.esmoop.2021.100114. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Hernán MA, Robins JM. Using big data to emulate a target trial when a randomized trial is not available. Am J Epidemiol. 2016;183(8):758-764. doi: 10.1093/aje/kwv254. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Hernán MA, Sauer BC, Hernández-Díaz S, Platt R, Shrier I. Specifying a target trial prevents immortal time bias and other self-inflicted injuries in observational analyses. J Clin Epidemiol. 2016;79:70-75. doi: 10.1016/j.jclinepi.2016.04.014. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Hernán MA. Methods of public health research—strengthening causal inference from observational data. N Engl J Med. 2021;385(15):1345-1348. doi: 10.1056/NEJMp2113319. [DOI] [PubMed] [Google Scholar]
- 10. García-Albéniz X, Hsu J, Hernán MA. The value of explicitly emulating a target trial when using real world evidence: an application to colorectal cancer screening. Eur J Epidemiol. 2017;32(6):495-500. doi: 10.1007/s10654-017-0287-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Matthews AA, Danaei G, Islam N, Kurth T. Target trial emulation: applying principles of randomised trials to observational studies. BMJ. 2022;378:e071108. doi: 10.1136/bmj-2022-071108. [DOI] [PubMed] [Google Scholar]
- 12. Fu EL, van Diepen M, Xu Y, et al. Pharmacoepidemiology for nephrologists (part 2): potential biases and how to overcome them. Clin Kidney J. 2021;14(5):1317-1326. doi: 10.1093/ckj/sfaa242. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Desai RJ, Franklin JM. Alternative approaches for confounding adjustment in observational studies using weighting based on the propensity score: a primer for practitioners. BMJ. 2019;367:l5657. doi: 10.1136/bmj.l5657. [DOI] [PubMed] [Google Scholar]
- 14. Pérol D, Robain M, Arveux P, et al. The ongoing French metastatic breast cancer (MBC) cohort: the example-based methodology of the Epidemiological Strategy and Medical Economics (ESME). BMJ Open. 2019;9(2):e023568. doi: 10.1136/bmjopen-2018-023568. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Miller K, Wang M, Gralow J, et al. Paclitaxel plus bevacizumab versus paclitaxel alone for metastatic breast cancer. N Engl J Med. 2007;357(26):2666-2676. doi: 10.1056/NEJMoa072113. [DOI] [PubMed] [Google Scholar]
- 16. Lipsky AM, Greenland S. Causal directed acyclic graphs. JAMA. 2022;327(11):1083. doi: 10.1001/jama.2022.1816. [DOI] [PubMed] [Google Scholar]
- 17. VanderWeele TJ. Principles of confounder selection. Eur J Epidemiol. 2019;34(3):211-219. doi: 10.1007/s10654-019-00494-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Hernán MA, Hernández-Díaz S, Robins JM. A structural approach to selection bias. Epidemiology. 2004;15(5):615-625. doi: 10.1097/01.ede.0000135174.63482.43. [DOI] [PubMed] [Google Scholar]
- 19. Textor J, van der Zander B, Gilthorpe MS, Liśkiewicz M, Ellison GT. Robust causal inference using directed acyclic graphs: the R package ‘dagitty’. Int J Epidemiol. 2016;45(6):1887-1894. doi: 10.1093/ije/dyw341. [DOI] [PubMed] [Google Scholar]
- 20. Chatton A, Le Borgne F, Leyrat C, et al. G-computation, propensity score-based methods, and targeted maximum likelihood estimator for causal inference with different covariates sets: a comparative simulation study. Sci Rep. 2020;10(1):9219. doi: 10.1038/s41598-020-65917-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Xu S, Ross C, Raebel MA, Shetterly S, Blanchette C, Smith D. Use of stabilized inverse propensity scores as weights to directly estimate relative risk and its confidence intervals. Value Health. 2010;13(2):273-277. doi: 10.1111/j.1524-4733.2009.00671.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Chesnaye NC, Stel VS, Tripepi G, et al. An introduction to inverse probability of treatment weighting in observational research. Clin Kidney J. 2022;15(1):14–20. doi: 10.1093/ckj/sfab158. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Mlcoch T, Hrnciarova T, Tuzil J, Zadak J, Marian M, Dolezal T. Propensity score weighting using overlap weights: a new method applied to regorafenib clinical data and a cost-effectiveness analysis. Value Health. 2019;22(12):1370-1377. doi: 10.1016/j.jval.2019.06.010. [DOI] [PubMed] [Google Scholar]
- 24. White IR, Royston P. Imputing missing covariate values for the Cox model. Stat Med. 2009;28(15):1982-1998. doi: 10.1002/sim.3618. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Rubin DB. Multiple Imputation for Nonresponse in Surveys. New York: Wiley; 1987. [Google Scholar]
- 26. Barnard J, Rubin DB. Small-sample degrees of freedom with multiple imputation. Biometrika. 1999;86(4):948-955. [Google Scholar]
- 27. Josse J, Husson F. missMDA: a package for handling missing values in multivariate data analysis. J Stat Soft. 2016;70(1):31. doi: 10.18637/jss.v070.i01 [DOI] [Google Scholar]
- 28. Schneeweiss S. Sensitivity analysis and external adjustment for unmeasured confounders in epidemiologic database studies of therapeutics. Pharmacoepidemiol Drug Saf. 2006;15(5):291-303. doi: 10.1002/pds.1200 [DOI] [PubMed] [Google Scholar]
- 29. Sammon CJ, Leahy TP, Gsteiger S, Ramagopalan S. Real-world evidence and nonrandomized data in health technology assessment: using existing methods to address unmeasured confounding?. J Comp Eff Res. 2020;9(14):969-972. doi: 10.2217/cer-2020-0112. [DOI] [PubMed] [Google Scholar]
- 30. VanderWeele TJ, Ding P. Sensitivity analysis in observational research: introducing the e-value. Ann Intern Med. 2017;167(4):268-274. doi: 10.7326/M16-2607. [DOI] [PubMed] [Google Scholar]
- 31. Petersen ML, Porter KE, Gruber S, Wang Y, van der Laan MJ. Diagnosing and responding to violations in the positivity assumption. Stat Methods Med Res. 2012;21(1):31-54. doi: 10.1177/0962280210386207. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32. Miles DW, Chan A, Dirix LY, et al. Phase III study of bevacizumab plus docetaxel compared with placebo plus docetaxel for the first-line treatment of human epidermal growth factor receptor 2–negative metastatic breast cancer. J Clin Oncol. 2010;28(20):3239-3247. doi: 10.1200/JClinOncol.2008.21.6457. [DOI] [PubMed] [Google Scholar]
- 33. Robert NJ, Diéras V, Glaspy J, et al. RIBBON-1: randomized, double-blind, placebo-controlled, phase III trial of chemotherapy with or without bevacizumab for first-line treatment of human epidermal growth factor receptor 2–negative, locally recurrent or metastatic breast cancer. J Clin Oncol. 2011;29(10):1252-1260. doi: 10.1200/JClinOncol.2010.28.0982. [DOI] [PubMed] [Google Scholar]
- 34. Rossari JR, Metzger-Filho O, Paesmans M, et al. Bevacizumab and breast cancer: a meta-analysis of first-line phase III studies and a critical reappraisal of available evidence. J Oncol. 2012;2012:e417673. doi: 10.1155/2012/417673. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35. Miles DW, Diéras V, Cortés J, Duenne A-A, Yi J, O’Shaughnessy J. First-line bevacizumab in combination with chemotherapy for HER2-negative metastatic breast cancer: pooled and subgroup analyses of data from 2447 patients. Ann Oncol. 2013;24(11):2773-2780. doi: 10.1093/annonc/mdt276. [DOI] [PubMed] [Google Scholar]
- 36. Delaloge S, Pérol D, Courtinard C, et al. Paclitaxel plus bevacizumab or paclitaxel as first-line treatment for HER2-negative metastatic breast cancer in a multicenter national observational study. Ann Oncol. 2016;27(9):1725-1732. doi: 10.1093/annonc/mdw260. [DOI] [PubMed] [Google Scholar]
- 37. Buyse M, Vansteelandt S. The potential and perils of observational studies. Ann Oncol. 2017;28(1):182. doi: 10.1093/annonc/mdw539. [DOI] [PubMed] [Google Scholar]
- 38. Delaloge S, Pérol D, Robain M, Cailliot C. Reply to ‘The potential and perils of observational studies’ by M. Buyse et al. Ann Oncol. 2017;28(2):436-437. doi: 10.1093/annonc/mdw572. [DOI] [PubMed] [Google Scholar]
- 39. Austin PC. The performance of different propensity score methods for estimating marginal hazard ratios. Stat Med. 2013;32(16):2837-2849. doi: 10.1002/sim.5705. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40. Austin PC, Grootendorst P, Normand SLT, Anderson GM. Conditioning on the propensity score can result in biased estimation of common measures of treatment effect: a Monte Carlo study. Stat Med. 2007;26(4):754-768. doi: 10.1002/sim.2618. [DOI] [PubMed] [Google Scholar]
- 41. Austin PC. The relative ability of different propensity score methods to balance measured covariates between treated and untreated subjects in observational studies. Med Decis Making. 2009;29(6):661-677. doi: 10.1177/0272989X09341755. [DOI] [PubMed] [Google Scholar]
- 42. Petito LC, García-Albéniz X, Logan RW, et al. Estimates of overall survival in patients with cancer receiving different treatment regimens: emulating hypothetical target trials in the Surveillance, Epidemiology, and End Results (SEER)-Medicare linked database. JAMA Netw Open. 2020;3(3):e200452. doi: 10.1001/jamanetworkopen.2020.0452. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 43. Vallée A, Gabet A, Grave C, Sorbets E, Blacher J, Olié V. Patterns of hypertension management in France in 2015: the ESTEBAN survey. J Clin Hypertens (Greenwich). 2020;22(4):663-672. doi: 10.1111/jch.13834. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44. Arnold BF, Ercumen A, Benjamin-Chung J, Colford JMJ. Brief report: negative controls to detect selection bias and measurement bias in epidemiologic studies. Epidemiology. 2016;27(5):637-641. doi: 10.1097/EDE.0000000000000504. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45. Lipsitch M, Tchetgen Tchetgen E, Cohen T. Negative controls: a tool for detecting confounding and bias in observational studies. Epidemiology. 2010;21(3):383-388. doi: 10.1097/EDE.0b013e3181d61eeb. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46. Scailteux LM, Droitcourt C, Balusson F, et al. French administrative health care database (SNDS): the value of its enrichment. Therapies. 2019;74(2):215-223. doi: 10.1016/j.therap.2018.09.072. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The dataset analyzed in the current study is issued from the Epidemiological Strategy and Medical Economics (ESME) Metastatic Breast Cancer (MBC) Data Platform. The database of the ESME program, including the database of the MBC cohort, are currently not accessible. In accordance with the regulatory frame applicable to the ESME Data Platform, data are only available via a secure cloud computing service with limited access to authorized persons; access authorizations to the secure cloud computing service are available from the corresponding author upon reasonable request. Each demand will be examined on a case-by-case basis by the UNICANCER ESME scientific committee. Moreover, subset of data (2008-2014) is made available on the Health Data Hub (https://www.health-data-hub.fr/).