Survival Outcomes Following Pregnancy or Assisted Reproductive Technologies After Breast Cancer: A Population-Based Study

Abstract

Background:

We sought to determine the impact of pregnancy or assisted reproductive technologies (ART) on breast-cancer-specific survival among breast cancer survivors.

Methods:

We performed a cohort study using a novel data linkage from the California Cancer Registry, the California birth cohort, and the Society for Assisted Reproductive Technology Clinic Outcome Reporting System data sets. We performed risk-set matching in women with stage I-III breast cancer diagnosed between 2000 and 2012. For each pregnant woman, we matched at the time of pregnancy comparable women who were not pregnant at that point but were otherwise similar based on observed characteristics. After matching, Cox proportional hazards models were used to estimate hazard ratios (HRs) and 95% CIs for the association of pregnancy with breast-cancer-specific survival. We repeated these analyses for women who received ART.

Results:

Among 30,021 women with breast cancer; 553 had a pregnancy and 189 attempted at least 1 cycle of ART. In Cox proportional hazards modeling, the pregnancy group had a higher 5-year disease-specific survival rate; 95.6% in the pregnancy group and 90.6% in the non-pregnant group (HR, 0.43; 95%CI, 0.24–0.77). In women with hormone receptor–positive cancer, we found similar results (HR, 0.43; 95%CI, 0.2–0.91). In the ART analysis, there was no difference in survival between groups; the 5-year disease-specific survival rate was 96.9% in the ART group and 94.1% in the non-ART group (HR, 0.44; 95%CI, 0.17–1.13).

Conclusion:

Pregnancy and ART are not associated with worse survival in women with breast cancer.

Lay summary:

We sought to determine the impact of pregnancy or assisted reproductive technologies (ART) among breast cancer survivors. We performed a study of 30,021 women by linking available data from California and the Society for Assisted Reproductive Technology Clinic Outcome Reporting System. For each pregnant woman, we matched at the time of pregnancy comparable women who were not pregnant at that point but were otherwise similar based on observed characteristics. We repeated these analyses for women who received ART. We found that pregnancy and ART were not associated with worse survival.

Precis:

Pregnancy or assisted reproductive technologies does not hamper disease-specific survival among women with breast cancer, including those with positive hormone-receptor status.

Introduction

Reproductive-age cancer survivors make difficult decisions about parenthood and reproductive health while navigating the physical and psychological effects of cancer and its therapy. Involuntary iatrogenic childlessness profoundly affects reproductive-age survivors, and the American Society of Clinical Oncology, American Society of Reproductive Medicine, European Society of Human Reproduction and Embryology, and European Society for Medical Oncology empower oncologists to discuss reproductive goals before initiating cancer treatment.^1–4 As literature about the positive role that family-building has on quality of life and the ability to cope with cancer evolves and more women delay pregnancy,^5–9 concerns regarding fertility have become increasingly relevant. However, uncertainty regarding the oncologic safety of pregnancy and assisted reproductive technologies (ART) among breast cancer survivors may contribute to inconsistent recommendations from clinicians and low desire to conceive in some cancer survivors.^10,11

While some have theorized that pregnancy- or ART-related hormone exposure may stimulate dormant micrometastases and beget a clinical recurrence,¹² multiple observational studies have suggested that women who conceive after breast cancer treatment have equivalent, and often better, cancer outcomes than those who do not.^13–21 Similar findings have been described among women who received ART.²² However, some researchers^20,23,24 have questioned whether these data were influenced by immortal time bias (a period during follow-up in which, by design, the study outcome cannot occur or is excluded from the analysis due to an incorrect definition of the start of follow-up) or the “healthy mother effect,”²⁵ a type of selection bias that underestimates the risk of pregnancy or ART. Healthy, cancer-free patients who become pregnant likely have better oncologic outcomes than those who are not healthy enough to conceive; thus, the groups are not comparable in terms of outcomes. Moreover, because it is challenging to properly balance observed covariates among such disparate groups, traditional statistical methods, including propensity score matching^15–18 and meta-analyses that pool estimates from these studies,^{19–21,24,26} may fail to adjust for this bias.

More sophisticated matching techniques, however, are better equipped to address immortal time bias and covariate imbalances. Using risk-set matching methods, we sought to provide a more precise real-world estimation of the oncologic outcomes among women with a history of breast cancer who subsequently gave birth or used ART. We used a novel linkage of data from the California Cancer Registry, the California Office of Statewide Health Planning and Development (OSHPD) birth cohort, and the Society for Assisted Reproductive Technology Clinic Outcome Reporting System (SART CORS).

Methods

Data sources

We performed a population-based cohort study using linked data from the California Cancer Registry, the California OSHPD birth cohort, and SART CORS. Institutional Review Boards from The University of Texas MD Anderson Cancer Center, OSHPD, the California Cancer Registry, and SART CORS approved the study. The linked database included California Cancer Registry data²⁷ from January 2000 through December 2015 linked to OSHPD data files for patients treated from January 2000 through December 2012, the most recent year for which OSPHD data are available (linkage details are found in eAppendix 1). The California OSHPD data include maternal antepartum and postpartum hospital records for the 9 months before delivery and 1 year after delivery. The California Cancer Registry data were also linked to SART CORS data for patients treated for cancer between January 2004 and December 2015 (ART treatment details before 2004 were not available). The SART CORS database contained data from more than 90% of all clinics performing ART in the United States; data are submitted by individual clinics and verified by the practice director of each clinic and validated annually.^28,29

Study population

We identified all women aged 18 to 45 years who were diagnosed with American Joint Committee on Cancer pathologic stage I-III breast cancer between January 1, 2000, and December 31, 2012 (N=30,021). Each eligible woman with cancer was screened for her deliveries (“pregnancy group”) and initiation of an ART cycle using fresh autologous oocytes (“ART group”) in the linked datasets respectively.

In the pregnancy group, we determined date of conception to establish a timeline that indicted when a pregnancy occurred relative to the breast cancer diagnosis. To calculate date of conception, we subtracted the estimate of gestational age from the date of delivery. As treatment completion is not a defined variable in the California Cancer Registry, we defined cases as those who conceived at least 12 months after diagnosis based on the rationale that this duration would account for patients who received adjuvant chemotherapy after surgery (date of delivery timeline details in eAppendix 2). We excluded patients whose births were missing a delivery date (N=2325) as we could not establish when pregnancy occurred relative to cancer diagnosis. In order to cleanly establish a group of patients who had the exposure (“pregnancy”) after their breast cancer diagnosis, we excluded all patients who conceived only prior to breast cancer diagnosis (and not after) or who may have been pregnant at the time of breast cancer diagnosis or treatment by excluding all who conceived before or within 12 months of their cancer diagnosis (N=4673). To avoid including inconsistent data, we excluded deliveries that had implausible combinations of living status and gestational age (<22 weeks or >45 weeks; N=328) as is a common practice when utilizing this dataset and whose date of birth in the California Cancer Registry was inconsistent with the date of birth in the linked OSHPD file (N=48). In the ART group, patients who underwent autologous ART cycles (defined as the start date of the first cycle of controlled ovarian hyperstimulation with intention to retrieve oocytes) initiated after cancer diagnosis were included. ART included fresh embryo transfers and cycles in which all oocytes and/or embryos were cryopreserved without a fresh embryo transfer. Frozen embryo transfer cycles were also excluded because natural cycle preparation protocols (in which embryo transfer is synchronized with ovulation) may not involve any exogenous hormone exposure. The data for each patient were summarized into a single record that included data from the initial ART treatment.

Matching and survival analysis

We assessed pregnancy and receipt of ART as exposures of interest in separate analyses. A key principle in the design of observational studies is to approximate a controlled randomized study that could have been conducted under ideal circumstances.^30,31 Thus, we used risk-set matching to balance the observed covariates between exposed and control patients while respecting the temporal structure of the data.³² At each time point when a patient was exposed (to pregnancy or ART), a risk set was formed by selecting 1 control patient who was diagnosed with breast cancer at the same date (±3 months) as the exposed patient and was alive and unexposed at the same time point. From among the pool of eligible controls for each risk set, one control was matched to each exposed patient to balance observed covariates that were identified a priori as potential confounders. This avoided using future information to categorize patients as exposed or unexposed, thereby avoiding immortal time bias. A consequence of this approach is that the women included in a risk set as controls who were not pregnant or exposed to ART at the time of matching could be exposed at a later follow-up time. Such patients were categorized only as controls throughout the analyses and are better described as “not yet exposed,” rather than “unexposed” (Figure 1).

Figure 1:

Diagram showing an example of risk-set matching for patient A; both patients B and C are both at risk and eligible matches.

Patient B was diagnosed with breast cancer at about the same date (± 3 months) as patient A; she was alive at the time of matching (patient A’s date of delivery); and she never became pregnant. Similarly, patient C was diagnosed with breast cancer around the same date (± 3 months) as patient A and was alive at the time of matching (patient A’s date of delivery). Since patient C achieved her first pregnancy after the matching date, she could also serve as a control. If patient C was matched before achieving pregnancy, she was excluded from the pregnancy group.

Abbreviations: ART, assisted reproductive technologies.

We used optimal risk matching based on mixed integer programming to balance the following covariates across groups³³: age (continuous variable), year of diagnosis; stage (I,II,III); grade (1,2,3,unknown); estrogen, progesterone, and ERBB2 (HER2) receptor status (positive,negative,unknown); chemotherapy, radiotherapy, surgery (lumpectomy,mastectomy,none); race and ethnicity based on medical record abstractions supplemented with surname matching for Hispanic ethnicity [White,Black,Hispanic,Asian or Pacific Islander, and other (including American-Indian, unknown race)]; median household income of the zip code of residence (quintiles)^34,35; insurance at diagnosis (uninsured,insured,Medicaid,no specifics,unknown); marital status (married,single,separated,divorced,widowed,other); Charlson comorbidity score (0,1,≥2). Matching was conducted using the designmatch package (version 0.3.1) for R.³⁶

Statistical Analyses

The primary outcome of interest was time from exposure to death from breast cancer. We focused on breast-cancer specific deaths to reduce bias from the differential risk of non-cancer mortality among those that achieved pregnancy and those that did not. For control patients, the date of exposure was the date of delivery or ART exposure of the risk set-matched exposed patient. We compared 5-year disease-specific survival rates between exposed patients and matched controls using the Kaplan-Meier method and the log-rank test and used Cox proportional hazards models to estimate hazard ratios (HRs) and 95% confidence intervals (CI) for the association of pregnancy or ART with time to death. Additionally, we repeated the analysis to estimate overall survival. We then tested for the statistical significance of the interaction of pregnancy by time from cancer diagnosis to pregnancy or matching for the unexposed group to assess whether the magnitude or direction of association between pregnancy and breast cancer mortality hazard changed with the interval between cancer diagnosis and pregnancy. We also repeated analyses stratified by pregnancy (or match date) ≤2 years after cancer diagnosis and those who had a pregnancy >2 years after diagnosis. We repeated the pregnancy analysis restricting to women with hormone receptor (estrogen or progesterone receptor)-positive, women with stage I, and in women with triple negative breast cancers.

As a stability analysis, we repeated analyses with a 1:5 matching ratio, after excluding the 25 pregnant patients who received ART. All statistical tests were 2-sided. P values of <0.05 and 95% CI not inclusive of the null (1.0) were considered statistically significant. These analyses were implemented in SAS Enterprise Guide version 7.11 (SAS Institute Inc.)

As previously highlighted, a key idea in the design of observational studies is to approximate or emulate the randomized experiment that would have been performed under controlled circumstances.^31,37 A critical feature of our data is its temporal structure. Respecting this structure, our matching approach seeks to emulate a sequentially randomized controlled trial, where patients are randomly assigned to the treatment (pregnancy or ART) or control (expectant management) group at different time points. For each patient treated at a given time point, the matching approach finds untreated patients with very similar observed characteristics until that time point, forming a risk-set. Under the assumption of “no unmeasured confounders”^38,39 the matched cohort approximates a randomized trial, in which each risk-set represents a comparable group of subjects randomized on the same day. Consequently, we compare cancer-specific survival between treated and untreated patients using proportional hazard models, which are conventional in randomized trials with time-to-event outcomes.⁴⁰ We consider the day of risk-set assignment as each subjects origin time, just as an analogous trial would treat the day of randomization.⁴¹ Among the strengths of this approach are its prospective structure, transparent covariate adjustments, the reduced reliance of the estimates on parametric modeling, and the intuitive interpretation or hazard ratios estimates by analogy to a hypothetical randomized trial.

We used the “E-value” to evaluate the robustness of derived estimates to potential unmeasured confounding.⁴² The E-value uses estimates from the study data to quantify how strong an unmeasured confounder must be to fully explain an observed treatment-outcome relationship. We used the E-value to calculate the magnitude of the association an unmeasured confounder with both the exposure (pregnancy or ART) and survival necessary to fully explain the derived estimate (eAppendix 3). We also calculated the E-value required to explain the confidence limit closest to the null and to explain an inverse relationship with survival.

Results

Among 30,021 women aged 18 to 45 years diagnosed with stage I-III breast cancer from 2000 through 2012, 553 had a pregnancy at least 12 months after cancer diagnosis. eAppendix 4 shows the comparison between the patients included and those excluded from the pregnancy analysis. In terms of ART exposure, 189 initiated at least 1 cycle of ART (Figure 2). Tables 1 and 2 summarize the demographic and clinical characteristics of the pregnancy and ART cohorts and their respective unexposed comparison groups before and after risk-set matching (details of all covariates are in eAppendices 5 and 6). The covariates were well balanced in the matched cohort of 520 women who were pregnant (33 patients matched as controls prior to pregnancy were not considered exposed) and their respective controls, with all standardized differences <0.2 (Table 1). The median time to pregnancy was 32.6 [interquartile range (IQR) 21.8–49.2] months and 32.9 (IQR 21.8–49.9) months to exposure for controls. Similarly, the 186 patients who received ART and their matched controls were well balanced (3 patients matched prior to receiving ART were not considered exposed) (Table 2). The median time to ART was 1.9 [interquartile range (IQR) 1.09–3.4] months and 2.7 (IQR 0.3–5.3) months to exposure for controls.

Figure 2:

Cohort selection process.

Abbreviations: CCR, California Cancer Registry; DOB, date of birth; OSHPD, California Office of Statewide Health Planning and Development.

Table 1.

Pregnancy and nonpregnant cohorts before and after risk-set matching (selected covariates).

Covariates	Pregnancy* (N=553)a	Nonpregnant (N=28,976)	Standardized difference	Pregnancy at matching (N=520)	Nonpregnant at matching (N=520)	Standardized difference
Age, median (range), y	32 (29–35)	41 (38–44)	1.89	32 (29–35)	33 (28–37)	0.07
Race and ethnicity
White	269 (48.6)	14,678 (50.7)	0.04	253 (48.7)	278 (53.5)	0.10
Black	34 (6.1)	2232 (7.7)	0.06	34 (6.5)	25 (4.8)	0.08
Hispanic	149 (26.9)	6973 (24.1)	0.07	136 (26.2)	131 (25.2)	0.02
Asian/Pacific Islander	99 (17.9)	4837 (16.7)	0.03	95 (18.3)	84 (16.2)	0.06
Insurance (Insured)	413 (74.7)	21,760 (75.1)	0.01	387 (74.4)	413 (79.4)	0.12
Marital status (Married)	352 (63.7)	18,657 (64.4)	0.02	335 (64.4)	343 (66.0)	0.03
Cancer stageb
I	221 (40.0)	10,323 (35.6)	0.09	207 (39.8)	195 (37.5)	0.05
II	282 (51.0)	13,779 (47.6)	0.07	265 (51.0)	278 (53.5)	0.05
III	50 (9.0)	4874 (16.8)	0.23	48 (9.2)	47 (9.0)	0.01
Estrogen receptor +	259 (46.8)	19,739 (68.1)	0.44	240 (46.2)	273 (52.5)	0.13
Progesterone receptor +	232 (42.0)	17,467 (60.3)	0.37	213 (41.0)	241 (46.3)	0.11
ERBB2 +	131 (23.7)	5568 (19.2)	0.11	123 (23.7)	110 (21.2)	0.06
Triple Negative (Yes)	128 (23.2)	3794 (13.1)	0.26	124 (23.9)	127 (24.4)	0.01
Chemotherapy (Yes)	382 (69.1)	19,835 (68.5)	0.01	361 (69.4)	399 (76.7)	0.17
Radiotherapy (Yes)	272 (49.2)	14,640 (50.5)	0.03	255 (49.0)	275 (52.9)	0.08
Surgery type
Lumpectomy	301 (54.4)	13,372 (46.1)	0.17	286 (55.0)	255 (49.0)	0.12
Mastectomy	247 (44.7)	15,012 (51.8)	0.14	229 (44.0)	264 (50.8)	0.14
Charlson score =0	532 (96.2)	26,605 (91.8)	0.19	501 (96.3)	504 (96.9)	0.03

^aData represent number of patients (%) unless otherwise indicated.

^bCancer stage was determined using the 8th edition of the AJCC staging system.

^*For simplicity we excluded some categories, and the number of patients for each category might not add to 553. All the percentages are calculated using a denominator of 553. Appendix 4 includes all categories.

Table 2.

ART and non-ART cohorts before and after risk-set matching (selected covariates).

Covariates	ART* (N=189)a	Non-ART (N=36,279)	Standardized difference	ART (N=186)	Non-ART (N=186)	Standardized difference
Age, median (range), y	35 (32–38)	41 (37–44)	1.16	35 (31–38)	36 (31–40)	0.10
Race and ethnicity
White	124 (65.6)	17,874 (49.3)	0.34	121 (65.1)	116 (62.4)	0.06
Black	b	2698 (7.4)	0.19	b	b	<0.001
Hispanic	19 (10.1)	9121 (25.1)	0.40	19 (10.2)	22 (11.8)	0.05
Asian/Pacific Islander	40 (21.2)	6256 (17.2)	0.10	40 (21.5)	42 (22.6)	0.03
Insurance (Insured)	161 (85.2)	27,352 (75.4)	0.25	159 (85.5)	160 (86.0)	0.02
Marital status (Married)	102 (54.0)	23,384 (64.5)	0.21	99 (53.2)	112 (60.2)	0.14
Cancer stagec
I	70 (37.0)	12,893 (35.5)	0.03	69 (37.1)	67 (36.0)	0.02
II	93 (49.2)	17,207 (47.4)	0.04	91 (48.9)	92 (49.5)	0.01
III	26 (13.8)	6179 (17.0)	0.09	26 (14.0)	27 (14.5)	0.02
Estrogen receptor +	137 (72.5)	25,089 (69.2)	0.07	136 (73.1)	135 (72.6)	0.01
Progesterone receptor +	123 (65.1)	22,308 (61.5%)	0.07	123 (66.1)	124 (66.7)	0.01
ERBB2 +	33 (17.5)	6568 (18.1)	0.02	31 (16.7)	26 (14.0)	0.07
Triple Negative (Yes)	27 (14.3)	4516 (12.5)	0.05	26 (14)	32 (17.2)	0.09
Chemotherapy (Yes)	137 (72.5)	24,771 (68.3)	0.09	134 (72.0)	140 (75.3)	0.07
Radiotherapy (Yes)	95 (50.3%)	17,566 (48.4)	0.04	95 (51.1)	84 (45.2)	0.12
Surgery
Lumpectomy	80 (42.3)	15,979 (44.0)	0.04	80 (43.0)	64 (34.4)	0.18
Mastectomy	107 (56.6)	19,223 (53.0)	0.07	104 (55.9)	117 (62.9)	0.14
Charlson score = 0	176 (93.1)	33,124 (91.3)	0.07	173 (93.0)	180 (96.8)	0.17

^aData represent number of patients (%) unless otherwise indicated.

^bNo cell containing a value of 1 to 10 can be reported directly.

^cCancer stage was determined using the 8th edition of the AJCC staging system. ART, assisted reproductive technologies.

^*For simplicity we excluded some categories, and the number of patients for each category might not add to 189. All the percentages are calculated using a denominator of 189. Appendix 5 includes all categories.

The median follow-up time in the matched pregnancy analysis was 111 months. We observed 23 breast cancer deaths in the pregnancy group and 48 in the non-pregnant group during the follow-up period. The Kaplan-Meier curve is shown in Figure 3. The 5-year breast-cancer-specific survival rate was 95.6% in the pregnancy group and 90.6% in the non-pregnant group (P=0.002). In Cox proportional hazards models, women who had a pregnancy had a lower risk of 5-year breast cancer-related mortality (hazard ratio [HR]=0.43; 95%CI=0.24–0.77). The overall survival was similar (HR=0.47; 95%CI=0.27–0.79). The association was similar for women who achieved a pregnancy within 1–2 years after diagnosis (HR=0.83; 95%CI=0.36–1.93) or beyond 2 years from diagnosis (HR=0.32; 95%CI=0.16–0.66; P for interaction=0.09). The association between pregnancy and risk of breast cancer death was similar among women with stage I (HR=0.31; 95%CI=0.06–1.55), women with hormone receptor-positive breast cancer (HR=0.43; 95%CI=0.2–0.91), and those with triple negative breast cancer (HR=0.3; 95%CI=0.08–1.11). We did not find differences after repeating the analyses using a 1:5 matching ratio (HR=0.47; 95%CI=0.29–0.75; eAppendix 7), after excluding women who received ART after cancer (HR=0.47; 95%CI=0.27–0.79).

Figure 3.

Kaplan-Meier survival curves for the risk set-matched groups in the pregnancy and non-pregnant cohorts.

The median follow-up time for the ART analysis was 76 months. We observed 5 breast cancer deaths in the ART group and 10 in the non-ART group during the follow-up period. The Kaplan-Meier curve is shown in Figure 4. The 5-year breast cancer-specific survival rate was 96.9% in the ART group and 94.1% in the non-ART group (P=0.20); risk of death from breast cancer did not differ statistically between the groups (HR=0.44; 95%CI=0.17–1.13). The overall survival was similar (HR=0.44; 95%CI=0.14–1.44). Results were similar for women with shorter or longer time (<2y vs. ≥2y) from breast cancer diagnosis to pregnancy, women with stage I, those with hormone receptor-positive cancers, women with triple negative breast cancer, after repeating the analyses using a 1:5 matching ratio, and after removing women who received ART and then became pregnant (eAppendix 7).

Figure 4.

Kaplan-Meier survival curves for the risk set-matched groups for the assisted reproductive technologies (ART) and non-ART cohorts.

We calculated E-values to assess the sensitivity of our findings to unmeasured confounding (eAppendix 8). For the pregnancy group, the observed HR of 0.47, conditional on measured covariates, could be explained by an unmeasured confounder that was associated with both pregnancy and breast cancer survival by a risk ratio (RR) of 3.68. To move the CI to include the null, an unmeasured confounder associated with both variables with an RR of at least 1.85 was required. Among women who received ART, the CI included 1.0. The observed HR of 0.44 would move to include 1.0 with an unmeasured confounder that was associated with both the exposure and survival by an RR of 3.97. An unmeasured confounder would have to be associated with both pregnancy and survival by an RR of 7.98 (CI, 4.5) to shift the HR to 2.0 and would have to be associated with both ART and survival by an RR of 8.56 (CI, 2.12) to shift the HR to 2.0.

Discussion

Reproductive-age cancer survivors rate fertility as one of the most important determinants of their quality of life after treatment.⁴³ By combining cancer registry, birth, and ART data from California, the current study provides a comprehensive outlook on the oncologic outcomes among a large cohort of women with history of breast cancer who had a pregnancy or used ART. This novel and unique linkage method is robust in both its sample size and granularity, allowing for well-designed epidemiologic assessments of complex questions. Our findings suggest that pregnancy or ART after breast cancer does not lead to worse oncological outcomes. Moreover, our findings were robust to various analytical strategies.

Estrogen and progesterone levels are elevated during pregnancy and ART, and there is a concern that these hormonal changes might increase risk of recurrence in women with a history of breast cancer.⁴⁴ In the absence of randomized trials, addressing these important clinical questions requires evidence from well-designed observational studies. Prior studies have found that pregnancy is associated with longer survival among women with a history of breast cancer,^{13–15,17,18,45,46} including women with estrogen receptor-positive breast cancer^24,25 and BRCA mutations.¹³ However, many studies do not account for “immortal time” between diagnosis pregnancy, producing results that are biased towards in favor of pregnancy.^42,43 Even when immortal time is accounted for, because healthy, recurrence-free women are more likely to conceive, “healthy mother bias” may contribute to superior oncologic outcomes.⁴⁸

In the present study, we address these challenges by implementing an optimal risk-set matching strategy. Our analytic framework seeks to recapitulate a randomized control trial in observational data by identifying risk sets that achieve optimal covariate balance, while respecting the temporal structure of the data. In so doing, we avoid bias from introducing immortal time, since patients enter risk sets when they become pregnant (or a similar time since diagnosis for controls). Furthermore, by allowing controls to become pregnant subsequent to their inclusion in a risk set, we avoid matching based on future information,¹⁴ a practice that could introduce bias analogous to the well-known fallacy of comparing survival distributions of responder and non-responders in cancer trials.⁴¹ Additionally, by addressing the time-varying nature of exposures, we were better equipped to assess the potential time-dependent effects of either pregnancy or ART. Another important methodologic strength of our analysis is the implementation of an optimal matching algorithm that allowed better covariate balance (directly, for each of the covariates via a mean balance constraint, and optimally, for all of the covariates in aggregate via a Mahalanobis distance), while prioritizing characteristics most strongly linked with survival.^13–17 For example, because older age is associated with a lower probability of getting pregnant and worse survival, failure to achieve a tight match by age will result in residual confounding. Yet frequently cited studies^13,15,17 that used traditional matching approaches failed to achieve close age matching. Furthermore, our novel data linkage allowed us to include a larger number of covariates as control variables in our analyses compared to prior studies.

Similar to prior investigations,^15,16 we found that patients who achieved a pregnancy had better breast-cancer-specific survival than those who did not. This better survival was not evident among women who achieved a pregnancy within 1–2 years after diagnosis or received ART, but the sample size in these groups was smaller and CI were wide. Although we adjusted for numerous patient and tumor characteristics, we could not adjust for potential unmeasured confounders, such as better overall well-being and healthier lifestyles.^49,50 We used a sensitivity parameter to quantify the magnitude of unobserved confounding that would be needed to change the observed effects of either pregnancy or ART on breast-cancer-specific survival. We found that an unmeasured confounder that was associated with both breast cancer and survival by an RR of only 1.85 each could shift the CI to include the null (therefore, the estimate of better survival with pregnancy would not be statistically significant). More importantly, we found that substantial confounding—above and beyond measured covariates—would be required for pregnancy or ART to be associated with an increased risk of death from breast cancer. Moreover, the consistency of the oncologic outcomes in the pregnancy and ART cohorts provides further reassurance that the relatively short hormonal exposure associated with pregnancy and ART is unlikely to be associated with worse long-term survival among breast cancer survivors.

The results of this study should be interpreted in light of some limitations, including lack of random allocation and absence of granular clinical information. Disease-specific survival requires all causes of death to be classified as either a death attributable to cancer or not attributable to cancer, misclassification of cause of death is a potential risk. Additionally, disease-free survival was not available. The database also lacks detailed data regarding adjuvant treatment, in particular start and duration of endocrine therapy. There are populations that these data do not address, namely patients who had first and second trimester terminations, or early miscarriages. Early pregnancies, particularly those that result in a first trimester loss, are not consistently reported in administrative data, as many patients may not have established obstetric care or not have been seen in a hospital at the time of the miscarriage. Moreover, our data only include records for the 9 months before delivery and 1 year after delivery. If a miscarriage occurred close to the time of a new conception, it may have been captured by discharge data, but in most cases, would not have been captured in our dataset. Given that miscarriages and terminations are very likely to be under-reported in this dataset we do not think this is an adequate database to establish an early pregnancy as an “exposure”. While this is certainly a limitation of the dataset as these outcomes may have a profound psychological impact that may affect survival outcomes, the hormonal exposure is considerably smaller in these cases. Furthermore, we were not able to account for history of breastfeeding or duration of breastfeeding as a modifier. Finally, since 80% of the patients received only one ART cycle, we were not able to assess the risks of receiving multiple ART cycles.

Well-designed observational studies may well produce the highest quality evidence available to evaluate the safety of pregnancy or ART after breast cancer. This rigorous analysis using risk-set matching, which is less susceptible to immortal time and selection bias than other methods, offers a more detailed portrait of the role of fertility preservation in the care of reproductive-age women with breast cancer than has been previously reported. Our findings suggest that pregnancy or ART are not associated with worse disease-specific survival among women with breast cancer, including those with positive hormone-receptor status.