Body Mass Index (BMI), like most human phenotypes, is substantially heritable. However, BMI is not normally distributed; the skew appears to be structural, and increases as a function of age. Moreover, twin correlations for... [ view full abstract ]
Body Mass Index (BMI), like most human phenotypes, is substantially heritable. However, BMI is not normally distributed; the skew appears to be structural, and increases as a function of age. Moreover, twin correlations for BMI frequently violate the assumptions of the most common variety of the classical twin model, with the MZ twin correlation greater than twice the DZ correlation. This study aimed to decompose twin correlations for BMI using more general skew-t distributions. Same sex MZ and DZ twin pairs (N = 7,086) from the community-based Washington State Twin Registry were included. We used latent profile analysis (LPA) to decompose twin correlations for BMI into multiple mixture distributions. LPA was performed using the default normal mixture distribution and the skew-t mixture distribution. Similar analyses were performed for height as a comparison. Our analyses were then replicated in an independent dataset. A two-class solution under the skew-t mixture distribution fit the BMI distribution for both genders. The first class consists of a relatively normally distributed, highly heritable BMI with a mean in the normal range. The second class is a positively skewed BMI in the overweight and obese range, with lower twin correlations, especially in DZ. In contrast, height is normally distributed, highly heritable, and is well-fit by a single latent class. The finding was then replicated in an independent dataset consisting of 13,553 (5,965 MZ, 7,588 DZ) same-sex male twin pairs from the National Academy of Sciences-National Research Council Twin Registry. Results in the replication dataset were highly similar.
Our findings suggest that two distinct processes underlie the skew of the BMI distribution, only one of which fits the conventional model of heritable variance resulting from independent assortment of a large number of loci. The second latent profile, highly skewed with a mean in the overweight to obese range, lower MZ correlations and DZ correlations close to zero, is consistent with reciprocal phenotype-environment processes in which small initial differences within pairs are magnified by non-random exposure to subsequent environments. The contrast between height and weight is in accord with subjective psychological experience: both are under obvious genetic influence, but BMI is also subject to behavioral control, and thus to complex person-environment interactions, whereas height is not. Dual process models of this kind have potential relevance to a wide variety of psychopathological phenotypes that plausibly consist of a heritable, normally distributed trait in the normal range, and a positively skewed pathological upper tail. In traits of this kind, the psychopathology is located in one distribution but not the other. The goal of treatment programs for BMI or drinking is not to produce teetotalers or people who are as thin as possible, but rather to prevent individuals from drifting into the skewed upper tail.
