Predictors of mammographic density: insights gained from a novel regression analysis of a twin study

Abstract

Understanding which factors influence mammographically dense and nondense areas is important because percent mammographic density adjusted for age is a strong, continuously distributed risk factor for breast cancer, especially when adjusted for weight or body mass index. Using computer-assisted methods, we measured mammographically dense areas for 571 monozygotic and 380 dizygotic Australian and North American twin pairs ages 40 to 70 years. We used a novel regression modeling approach in which each twin's measure of dense and nondense area was regressed against one or both of the twin's and co-twin's covariates. The nature of changes to regression estimates with the inclusion of the twin and/or co-twin's covariates can be evaluated for consistency with causal and/or other models. By causal, we mean that if it were possible to vary a covariate experimentally then the expected value of the outcome measure would change. After adjusting for the individual's weight, the co-twin associations with weight were attenuated, consistent with a causal effect of weight on mammographic measures, which in absolute log cm(2)/kg was thrice stronger for nondense than dense area. After adjusting for weight, later age at menarche, and greater height were associated with greater dense and lesser nondense areas in a manner inconsistent with causality. The associations of dense and nondense areas with parity are consistent with a causal effect and/or within-person confounding. The associations between mammographic density measures and height are consistent with shared early life environmental factors that predispose to both height and percent mammographic density and possibly breast cancer risk.

Figure 1

A path diagram showing the measured variables represented by squares (the outcome variables, Y_i, and the corresponding covariates, X_i, for each twin i = 1,2 in a pair) and the potential causes represented by circles (trait-specific causes common to or shared by twins, S_Y and S_X, and individual-specific between-trait confounders, I₁ and I₂). Causes that are specific to individuals and traits have been deleted to simplify the figure, as they do not contribute to the argument espoused in this paper. A single-headed arrow between a cause and a measured variable indicates the presumed direction of causation, while a double-headed arrow indicates that the causes are correlated. The regression coefficients representing the associations on Y₁ of X₁ (β) and X₂ (β′) are shown as dotted lines.