Re: Pregnancies, Breastfeeding, and Breast Cancer Risk in the International BRCA1/2 Carrier Cohort Study (IBCCS)

Andrieu et al. ( 1 ) recently reported on the association between pregnancies and breast cancer risk in a large cohort of 1601 BRCA1/2 mutation carriers. They showed that later age at first full-term pregnancy was associated with a reduction in the risk of breast cancer for BRCA1 mutation carriers. This issue is important because it may affect the type of family planning counseling that young women receive. The decreasing trend between later age at pregnancy and breast cancer risk reported by Andrieu et al. can either result from a real protective effect of late pregnancies on breast cancer or may only be due to the fact that women who were diagnosed with breast cancer at older ages had more time to become pregnant and have a child before their diagnosis than their younger counterparts. The analysis of time-dependent data, such as age at first pregnancy, may lead to immortal time bias ( 2–4 ) if the dynamic nature of the variable is not taken into account in the analysis. In this case, the cancer-free time to a woman’s first pregnancy is the immortal time, and it must not be credited toward the effect of the first pregnancy on the age at cancer diagnosis. One method to avoid immortal time bias is to use models in which the categories describing age at first pregnancy are modeled as time-dependent variables. Although Andrieu et al. explicitly reported using parity as a time-dependent variable, they did not mention any time-dependent modeling of the other pregnancy-related variables. Thus, we cannot assess whether they used time-dependent methods to produce their results on the association between age at first pregnancy and cancer diagnosis. We used simulations to explicitly show the difference in hazard ratios obtained with non–time-dependent and time-dependent analyses. This methodologic issue is important because immortal time bias affects the results of many clinical cohorts ( 4 , 5 ). We simulated 1000 cohorts of 1000 women for whom the distribution of age at first pregnancy mimicked that of the control subjects (unaffected women) among the BRCA1 mutation carriers in the study by Andrieu et al. We then randomly assigned these women a breast cancer status (yes/no) and an age at diagnosis/censure using a distribution of age at diagnosis/censure with the same mean age and standard deviation as the one in Table 1 of Andrieu et al. Therefore, the way we simulated the data implied no association between age at pregnancy and age at breast cancer diagnosis. Results of simulations of 1000 cohorts of 1000 women carrying a BRCA1 gene mutation * HR = hazard ratio; CI = confidence interval; – = not applicable. Results of simulations of 1000 cohorts of 1000 women carrying a BRCA1 gene mutation * HR = hazard ratio; CI = confidence interval; – = not applicable. We estimated two statistical models of the association between age at pregnancy and age at breast cancer. In both models, the categorization of age at first pregnancy was the same as in Andrieu et al. ( 1 ). In model 1, membership in an age-at-first-pregnancy category was determined at the onset of the study. By contrast, in model 2, we used time-dependent indicators of categories of age at first pregnancy that used a value of 0 for the years during which the woman had not yet had a child and a value of 1 for the year that she became pregnant as well as the subsequent years. Because the categories of age at first pregnancy in model 1 are determined at the onset of the analysis, model 1 wrongly attributes the time elapsed before the pregnancy as the effect of the pregnancy. This creates immortal time bias and suggests a reduction in the risk of breast cancer for pregnancies at later ages. Model 2 provides unbiased estimates because it assumes that any potential effect of pregnancy on age at diagnosis can only start at the date of pregnancy and not at the onset of the study. Although in our simulations there was no association between age at first pregnancy and age at breast cancer diagnosis, model 1 created an artificial trend in the hazard ratio ( Table 1 ) that gives the impression that later age at first pregnancy is associated with a reduced risk of breast cancer. In roughly 30% of the simulated datasets, the age-at-first-pregnancy category of “older than 30 years” was, in fact, statistically significantly associated with age at breast cancer diagnosis. In model 2, however, the estimated hazard ratio showed no association between age at first pregnancy and age at breast cancer diagnosis. Our simulations illustrate how an association between breast cancer risk and maternal age may be influenced by the choice of statistical analysis. If Andrieu et al. only considered parity as a time-dependent variable as was stated in the methods ( 1 ), then the reported effects of age at first pregnancy on age at breast cancer diagnosis must be interpreted with caution. Our results may have clinical implications: if immortal time bias is not taken into account in the interpretation of the results, BRCA1 mutation carriers may be wrongly encouraged to postpone having children to reduce their risk of breast cancer. Andrieu et al. never made such recommendation, but as women, we believe that maternity is such an important issue that we want to warn readers against an overinterpretation of these results, as they might be, at least in part, due to a statistical artifact.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.