Figure 1 depicts the usual directed acyclic graph of MR adapted to the present study. Increasing individual proportions of Native American ancestry have been linked to decreasing BC risk, but this association may be attributed to some degree to unknown confounders and non-random measurement error in known confounders (e.g. socio-economic status), motivating the use of AIMs as instrumental variables to assess the unconfounded effect of the proportion of Native American ancestry on BC risk.

Typical directed acyclic graph of Mendelian randomization adapted to the present study

Figure 2 represents the analyses conducted from AIM pre-selection to the two-sample MR analyses. The left part of the flowchart shows some intermediate results based on Colombian women, while the right part refers to the analyses in Mexican women. The ancestry reference panels in the primary analyses included 107 Iberians from Spain (IBS) and 108 Yorubans from Ibadan, Nigeria (YRI) from the 1000 Genomes Project, complemented with 29 Native Americans from Colombia (nine Inga, five Arhuaco, four Embera, three Kogi and three Waunana described by Reich et al. and five Piapoco individuals from the Human Genome Diversity Project [HGDP]) or 177 Native Americans from Mexico (49 Maya, 43 Zapotec, 30 Pima, 25 Tepehuano, 17 Mixe, five Mixtec, one Yaqui and one Purepecha described by Reich et al. and four Maya and one Pima from the HGDP) [29, 30]. Relying on the informativeness for assignment measure, we preselected 36,403 AIMs for Colombian women and 32,835 AIMs for Mexican women. Among the preselected markers, 623 were robustly associated (p < 5 × 10–8) with the proportion of Native American ancestry among Colombian female controls; the corresponding number in Mexican female controls was 6118. Downstream phenome-wide association analyses (PheWAS, using a p-value threshold of 5 × 10–8) and linkage disequilibrium (LD) pruning (r2 > 0.01) resulted in 121 preliminary IVs for the subsequent Colombian analyses and 150 IVs for the subsequent Mexican analyses. We then complemented age-adjusted summary statistics on the association between the IVs and the proportion of Native American ancestry with age-adjusted summary statistics on the association between the IVs and BC risk based on BC patients and controls to perform radial MR and exclude potentially outlying instruments (p < 0.1). Finally, 98 and 134 IVs were utilised to investigate the confounder-free effect of the proportions of Colombian and Mexican Native American ancestry on BC risk.

Flowchart representing the statistical analyses for Colombian and Mexican study participants from the preselection of markers of Native American ancestry to the two-sample Mendelian randomization (MR) analysis

Figure 3 represents the first versus the second, and the first versus the third genetic principal components, along with the reference panels of Native American, European and African ancestry. The first three principal components explained a genetic variance of 4.07%, 1.34% and 0.18% in Colombian women and 5.89%, 2.00% and 0.26% for Mexican women, respectively. The first principal component distinguished between African and non-African ancestry in both admixed Colombian women (panel A) and admixed Mexican women (panel B). The second principal component distinguished between European and Native American ancestry. The third principal component separated three subtypes of Native American ancestry in Mexican women: the group of Native American reference individuals most similar to the study population (M1) included 115 individuals (49 Maya, 43 Zapotec, 17 Mixe, five Mixtec and one Purepecha).

Scatter plots of first versus second, and first versus third genetic principal components (PC) of study participants (crosses: BC patients, circles: population-based controls) and reference panels of African, European and Native American ancestry (represented by triangles; M1 includes Maya, Mixe, Mixtec, Purepecha and Zapotec, M2 includes Tepehuano and Yaqui, and M3 includes Pima individuals); panel A: Colombian study, panel B: Mexican study

Table 1 shows the results for Colombian women, while the corresponding results based on Mexican women are summarised in Table 2. The 98 IVs utilised for two-sample MR in Colombian women explained a cumulative variance of 7.1% in the proportion of Native American ancestry (F-statistic = 55.57). The p-value of Cochran’s Q statistic (p > 0.99) revealed no instrument heterogeneity as an indicator of pleiotropy, the MR-Egger intercept (p = 0.60) was consistent with no horizontal pleiotropy, and departing instruments were not evident in scatter and funnel plots (Fig. 4, panels A and B), prompting the use of inverse variance-weighted (IVW) odds ratios (ORs) for the primary analyses. The cumulative variance explained by the 134 IVs used for the Mexican analyses amounted to 38.5% (F-statistic = 412.29), which is an atypically high proportion in Mendelian randomization studies. After inspection of the p-value of Cochran’s Q statistic (p > 0.99), the MR-Egger intercept (p = 0.66) and the scatter and funnel plots (Fig. 4, panels C and D), we also decided to use the IVW OR for the primary Mendelian randomization analyses in Mexican women.

Scatter and funnel plots on the association between Native American ancestry and risk of breast cancer (A and B: Colombian study, C and D: Mexican study)

We found evidence of a putatively causal protective effect of Colombian Native American ancestry on BC risk (IVW OR = 0.974 per 1% increase in ancestry proportion, 95% confidence interval [CI] 0.970–0.978, p = 3.1 × 10–40). This estimate can also be interpreted as a 2.6% decrease in BC risk (95% CI 2.2–3%) per each 1% increase in the proportion of Native American ancestry (median 42% and interquartile range 36%-49% in Colombian female controls). MR results based on Mexican women also pointed to a protective effect of Native American ancestry on BC risk (IVW OR 0.988, 95% CI 0.987–0.990, p = 1.4 × 10–44). The median and interquartile range of the proportion of Native American ancestry were 65% and 54%-79% in Mexican control women, respectively.

Table 1 also shows the results of sensitivity analyses for Colombian women. The MR-Egger regression estimate (OR = 0.983, 95% CI 0.950–1.017) was consistent (overlapping 95% CI), and the weighted median estimate (OR = 0.975, 95% CI 0.969–0.980, p = 0.33) was virtually identical to the IVW OR (see also the regression lines on Fig. 4, panel A). To assess the influence on the results of the reference panel of Native American ancestry, we replaced the original 29 Native Americans from Colombia with 57 unrelated Native Americans from the HGDP (21 Maya, 13 Pima, 11 Karitiana, seven Piapoco and five Surui, see Additional file 1: Fig. 1) [31]. The use of Native Americans from the HGDP for calculation of ancestry proportions reduced the explained variance (from 7.1%-6.6%) and increased the p-value (to p = 6.1 × 10–37), but resulted in practically identical IVW OR estimates.

Stratified analyses in Colombian women suggested stronger protective effects of the proportion of Native American ancestry among women with a family history of BC and/or ovarian cancer (4.2% decrease in BC risk per 1% increase in proportion of Native American ancestry, 95% CI 3.6%-4.8%, p = 2.0 × 10–41), and on ER-positive BC (IVW OR = 0.968) versus ER-negative BC (IVW OR = 0.988) or triple-negative BC (IVW OR = 0.993; non-overlapping 95% CIs with ER-positive BC).

The results of the sensitivity analyses based on the Mexican women are shown in Table 2. As in the Colombian analyses, the IVW, MR-Egger and weighted median estimates were consistent (overlapping 95% CIs and very similar regression lines on Fig. 4, panel C). To examine the impact of the large variance explained by the IVs on the results, we reduced the number of IVs to 17, resulting in a similar explained variance as in the Colombian analyses (8.4%), but the IVW estimates barely changed (from OR = 0.988 to OR = 0.990). Neither the use of Native Americans from the HGDP nor the use of a subset of 115 Native Americans from Mexico (49 Maya, 43 Zapotec, 17 Mixe, five Mixtec and one Purepecha) as the reference panels for estimation of Native American ancestry proportions affected the IVW OR estimates.

Stratified analyses in Mexican women confirmed the stronger protective effect of Native American ancestry against familial BC found in Colombian women (2.7% decrease in familial BC risk compared with 1.2% decrease in overall BC risk per 1% increase in the proportion of Native American ancestry). Also in agreement with the Colombian results, the protective effect was stronger against ER-positive BC (IVW OR = 0.989) than against ER-negative BC (IVW OR = 1.004) or triple-negative BC (IVW OR = 1.004; non-overlapping 95% CIs and associated p < 2.2 × 10–16 with ER-positive BC).

Interestingly, most IVs (Colombian: 96%, Mexican: 97%) used as instrumental variables were population-specific, but differences between causal effect estimates based on common and population-specific markers did not reach statistical significance (overlapping 95% confidence intervals). For example, the IVW OR based on the four joint Colombian and Mexican IVs was 0.969 (95% CI 0.949–0.988) compared to 0.974 (95% CI 0.970–0.978) based on the 94 Colombia-specific IVs.