This paper examines the relationships between body mass index (BMI) and body composition in different population groups where low BMIs might be expected to occur and assesses the extent to which BMIs are influenced by size and shape. The relationship between BMI and fat as a percentage of body weight is approximately linear although theoretically a curvilinear relationship is to be expected. However, by allowing for a variable composition of weight differences, an approximately linear theoretical relationship is obtained. There are few direct data (e.g. from densitometry, hydrometry etc.) on body composition in the groups in question and to examine the relationships indirect data were used. The regression coefficients of fat-free mass on BMI for 285 samples of Africans, people of Asian origin, Indo-Mediterraneans and Pacific people were not significantly different in the various groups of each sex; % fat on BMI was similarly related in four groups of women. Intercept terms were all significantly different. Using the sitting height-to-stature ratio (SH/S) as an index of body shape in 158 groups, the regression coefficient of BMI on SH/S was 0.90 kg/m2 per 0.01 SH/S. Mean SH/S lies between 0.50 and 0.55 in most populations so that shape could affect BMI by 5 kg/m2 and influence markedly the interpretation of BMI. However, allowing for SH/S is not straightforward as there is as much variation within as between groups. In conclusion, low BMI approximates to low weight, fat mass and fat-free mass. There are differences in the relationships of BMI to body composition but over the range 20-25 kg/m2 these may not be important in epidemiological studies. To interpret BMI in terms of body composition in more detail it is necessary to take into account sex, age, shape and ethnicity.