We conducted a population-based cohort study in Denmark, linking individual-level data from nationwide registries, clinical databases, and biobanks. Data linkage was accomplished with the civil personal registration number, a unique identifier assigned to all residents of Denmark at birth or immigration [16].

Denmark has a tax-supported healthcare system, ensuring free-of-charge care provided to all residents by general practitioners in the community and by hospitals [16]. Patients with invasive breast cancer are registered in the Danish Breast Cancer Group (DBCG) clinical database, which compiles clinical and demographic characteristics at the time of breast cancer diagnosis including tumor stage, grade, human epidermal growth factor receptor 2 (HER2), estrogen receptor (ER) status, guideline-based recommended treatments (including chemotherapy), received treatments, age, and menopausal status [17, 18]. In Denmark, all diagnostic formalin-fixed paraffin-embedded (FFPE) tissue biopsies are routinely archived in perpetuity at local pathology departments, and logged in the Danish National Pathology Registry [19].

In addition to tax-supported healthcare, the Danish government offers subsistence and unemployment payments, student grants, and a range of other social and health-related benefits to residents in need. Such transfer payments have been registered on a weekly basis in the Danish Register for Evaluation of Marginalization (DREAM) since 1991 [20, 21].

Information on comorbidities, cohabitation, and level of education were obtained from the Danish National Patient Registry [22], Statistics Denmark [23], and the Danish Population’s Education Registry [24], respectively.

The ProBe CaRe cohort includes 5959 premenopausal women diagnosed with stage I–III breast cancer across Denmark from 2002 to 2011 and registered in the DBCG [25, 26]. Only women with ER-positive tumors who received tamoxifen (ER +/TAM +) and women with ER-negative tumors who did not receive tamoxifen (ER −/TAM −) were eligible for cohort entry. In 2007, docetaxel was designated as guideline chemotherapy (in combination with cyclophosphamide and, for most, epirubicin [27, 28]). Thus, we restricted the study population to women diagnosed from 2007 onwards for whom chemotherapy was recommended [28] (Supplementary file1, Supplemental Fig. 1). We included women aged 18–55 years at diagnosis to ensure women were of working age throughout the study period. Finally, we restricted our sample to women with available FFPE tumor blocks who resided in Denmark during the year preceding diagnosis.

The study was approved by the Danish Data Protection Agency (AU 2016-051-000001, #808), the Regional Ethics Committee (record no. 1-10-72-4-18), and the DBCG (DBCG-2018-01-04).

We investigated 21 candidate SNPs in 16 genes coding for enzymes involved in taxane metabolism (CYP1B1, CYP3A, CYP3A4, CYP3A5, and GSTP1), taxane transport (ABCB1, ABCC2, ABCG2, SLCO1B1, and SLCO1B3), DNA repair (ERCC1 and ERCC2), and taxane-related toxicity or neural repair (EPHA4, EPHA5, EPHA6, and FGD4) [14, 15, 29]. Archived breast tumor tissue blocks were collected for eligible ProBe CaRe participants. DNA was extracted and genotyped for the selected SNPs using commercially available TaqMan assays on a StepOne Plus real-time instrument (Applied Biosystems, Thermo Fischer Scientific, Foster City, CA, USA), as previously described [14, 15, 29]. RefSeq numbers (rs) and assay details of the SNPs are provided in Supplementary file1, Supplementary Table 1.

We categorized women based on number of variant alleles as non-carriers (carrying no variant allele on either locus) or carriers (carrying the variant on one locus [heterozygotes] or on both loci [homozygotes]) for each gene.

We used the DREAM database, which contains weekly records of all transfer payments, to access information on receipt of social benefits. We collected information from 1 year before to 5 years after the breast cancer diagnosis date (i.e., date of breast cancer surgery). Given the diagnostic period of our study population (2007–2011), we examined data on social benefits from 2006–2016.

Specifically for sick leave, only ‘long-term’ leave is captured by DREAM [21, 30]. Conversely, ‘short-term’ sick leave is not registered. ‘Short-term’ is defined as the initial period in which the employer must cover the cost of sick leave compensation, ranging from 14 to 30 days during our study period [31]. Absence lasting longer than this period is defined as ‘long-term’. After the initial period, the employer is reimbursed by the municipality for the sick leave compensation, which is registered in DREAM [21, 30]. In addition, the initial period is also registered, and DREAM thus reflects the entire length of ‘long-term’ sick leave (from the first day of absence).

For each week, we categorized women into one of three social benefit categories: receiving health-related benefits, receiving labor market-related benefits, and self-supporting or receiving student grants (referred to as ‘self-supporting’; Table 1). We categorized women who received benefits due to maternity leave according to their social benefit category during the week immediately before the start of maternity leave. If women had retired, emigrated, or died they were placed into a fourth censored category. Full details of DREAM codes used for categorizing cohort members are provided in Supplementary file1, Supplementary Table 2.

Because CYP family genes are also involved in tamoxifen metabolism [14], associations between SNPs and social benefit receipt could potentially be affected by ER-status and receipt of tamoxifen. We therefore used the DBCG database to ascertain data on ER-status and receipt of tamoxifen. Other covariates considered were age at diagnosis, additional clinical and treatment characteristics (including tumor stage, grade, HER2, type of surgery, and radiotherapy [17, 18]), comorbidities categorized using the Charlson Comorbidity Index (CCI) Score [33], cohabitation status (cohabiting; living alone), and the highest attained educational level at diagnosis (short [~ 9–11 years of schooling]; intermediate [~ 12–13 years of schooling]; long [~ 14–20 years of schooling]).

Participant characteristics were described with frequencies and proportions.

We visually described receipt of social benefits by plotting the proportion of women in each of the social benefit categories in weekly intervals.

We estimated the association of SNPs with the number of weeks spent in each of the three social benefit categories (i.e., receiving health-related benefits, receiving labor market-related benefits, and self-supporting) and the associated rate ratio (RR), 95% confidence intervals (CI) and p-values using negative binominal regression models with a robust variance estimator. Differences in log counts between variant carriers and non-carriers can be interpreted as log RR when follow-up is constant. To account for differences in social benefit receipt by time period with respect to breast cancer diagnosis, we stratified all analyses by time (− 12–0, 1–6, 7–12, 13–24, and 25–60 months from the diagnosis date). We used unadjusted models because variant alleles are allocated randomly at conception and thus causally unrelated to covariates. We used volcano plots (− log(p) against RR) to visually summarize the results. To represent the uncertainty of the RR, we plotted the ratio between the upper and lower limit of the 95% CI (UL/LL) against the RR; a smaller ratio represents a narrower CI.

In exploratory analyses, we ran the same models separately for heterozygotes and homozygotes compared with non-carriers. Because of the low minor allele frequency of some SNPs and resulting low statistical power, we only report results for SNPs with at least 100 homozygous variant women. We also repeated the main analyses in women in the ER +/TAM + group.

All analyses were conducted using SAS software version 9.4 (SAS Institute Inc., Cary, NC, USA).